CN107067396A - A kind of nuclear magnetic resonance image processing unit and method based on self-encoding encoder - Google Patents
A kind of nuclear magnetic resonance image processing unit and method based on self-encoding encoder Download PDFInfo
- Publication number
- CN107067396A CN107067396A CN201710283992.7A CN201710283992A CN107067396A CN 107067396 A CN107067396 A CN 107067396A CN 201710283992 A CN201710283992 A CN 201710283992A CN 107067396 A CN107067396 A CN 107067396A
- Authority
- CN
- China
- Prior art keywords
- mentioned
- artificial intelligence
- encoding encoder
- layer
- hippocampus
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/0002—Inspection of images, e.g. flaw detection
- G06T7/0012—Biomedical image inspection
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
- G06F18/243—Classification techniques relating to the number of classes
- G06F18/2431—Multiple classes
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/0002—Inspection of images, e.g. flaw detection
- G06T7/0012—Biomedical image inspection
- G06T7/0014—Biomedical image inspection using an image reference approach
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/10—Image acquisition modality
- G06T2207/10072—Tomographic images
- G06T2207/10088—Magnetic resonance imaging [MRI]
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/20—Special algorithmic details
- G06T2207/20081—Training; Learning
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/20—Special algorithmic details
- G06T2207/20084—Artificial neural networks [ANN]
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/30—Subject of image; Context of image processing
- G06T2207/30004—Biomedical image processing
- G06T2207/30016—Brain
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- Quality & Reliability (AREA)
- Radiology & Medical Imaging (AREA)
- Nuclear Medicine, Radiotherapy & Molecular Imaging (AREA)
- Medical Informatics (AREA)
- General Health & Medical Sciences (AREA)
- Health & Medical Sciences (AREA)
- Bioinformatics & Computational Biology (AREA)
- General Engineering & Computer Science (AREA)
- Evolutionary Computation (AREA)
- Evolutionary Biology (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Artificial Intelligence (AREA)
- Life Sciences & Earth Sciences (AREA)
- Magnetic Resonance Imaging Apparatus (AREA)
Abstract
The present invention relates to a kind of nuclear magnetic resonance image processing unit and method based on self-encoding encoder.It is related to the hippocampus head to hippocampus, CA1, CA2, CA3, CA4/ dentate fascias, the method that fimbria of hippocampus and hippocampus tail carry out Computer assisted identification, realized using the two-dimentional self-encoding encoder of the deep learning of artificial intelligence to Elderly people group, three kinds of situations for being clinically diagnosed as Amnestic mild cognitive impairment and being clinically diagnosed as alzheimer disease carry out the sorting technique of three classification, utilize Elderly people group, clinically be diagnosed as Amnestic mild cognitive impairment and be clinically diagnosed as the nuclear magnetic resonance of alzheimer disease image data determine artificial intelligence deep learning training set, checking collection and test set, the three-dimensional data of nuclear magnetic resonance can be converted into 2-D data using two dimension slicing is chosen.
Description
Technical field
The present invention relates to a kind of nuclear magnetic resonance image processing unit and method based on self-encoding encoder, especially, it is related to base
A Erzi seas are carried out to the substructure of hippocampus in the Magnetic resonance imaging using head of the two-dimentional self-encoding encoder of deep learning
The method of the Computer assisted identification of the differentiation between any two of silent disease, Amnestic mild cognitive impairment and Elderly people group.
Background technology
Alzheimer disease is a kind of nerve degenerative diseases.Existing research work has confirmed, in the nuclear-magnetism of head
In the image file of resonance image-forming, although there is individual difference in the image data of the Magnetic resonance imaging of each subject, in life
In the sense that thing statistics, there is core in three classifications of alzheimer disease, Amnestic mild cognitive impairment and Elderly people group
Difference between three classifications of the image data of magnetic resonance imaging.Utilize the side of the deep learning of the machine learning of artificial intelligence
Method is using the image data of the Magnetic resonance imaging of head to alzheimer disease, Amnestic mild cognitive impairment and Elderly people
Group carries out Computer assisted identification can provide advisory opinion for clinician, improve the operating efficiency of clinician.
Three-dimensional image number of the favourable manually intelligent deep learning to the Magnetic resonance imaging of head in the prior art
According to the document for carrying out alzheimer disease, Amnestic mild cognitive impairment and Elderly people group progress Computer assisted identification, see
The arXiv of Preprint website:1607.00556v1.But, three-dimensional voxel of the document only to the Magnetic resonance imaging of head
Data are handled, and the technical scheme used is Three dimensional convolution self-encoding encoder.Due to needing to the head of each subject
All three-dimensional voxel datas of Magnetic resonance imaging are handled, and the technical scheme needs huge operand, even if making
With high-performance computer or high-performance computer cluster, the technical scheme also is difficult to even not use clinically.In addition,
The image data of the Magnetic resonance imaging of the head of clinical plain scan is 20 view data to 40 two-dimentional sections,
Above-mentioned arXiv:Method in 1607.00556v1 document is not used to the nuclear magnetic resonance of the head of the plain scan of clinic
The image data of imaging.
The content of the invention
The present invention is proposed aiming at problems of the prior art.
The present invention proposes a kind of nuclear magnetic resonance image processing unit based on self-encoding encoder, and the device includes, for depositing
The view data for storing up the Magnetic resonance imaging of the program of image preprocessing, the program of the deep learning of artificial intelligence and head is deposited
Storage media, the meter that the view data of the Magnetic resonance imaging of head is handled and analyzed using the deep learning of artificial intelligence
Machine host or computer cluster are calculated, program, the program of the deep learning of artificial intelligence, various heads for display image pretreatment
The display device of the running and operation result of the view data of the Magnetic resonance imaging of cranium and various programs,
It is characterized in that:Said apparatus can realize that Elderly people group, forgetting type to any given quantity are slightly recognized
Know and damage any given with the view data of the Magnetic resonance imaging of the three-dimensional head of the subject of alzheimer disease
The two dimension slicing of quantity realizes the supervised learning of the self-encoding encoder of the deep learning of artificial intelligence, according to above-mentioned to any given
The Elderly people group of quantity, the nuclear-magnetism of the three-dimensional head of the subject of Amnestic mild cognitive impairment and alzheimer disease
The two dimension slicing of the view data of resonance image-forming is carried out obtained by the supervised learning of the self-encoding encoder of the deep learning of artificial intelligence
Deep learning neural network model, utilize the nuclear-magnetism of the three-dimensional head of the subject to be identified of any given quantity
The two dimension slicing of any given quantity of the view data of resonance image-forming carries out normal old to above-mentioned subject to be identified
The Division identifications of three classification of year group, Amnestic mild cognitive impairment and alzheimer disease.
The present invention proposes a kind of nuclear magnetic resonance image processing method based on self-encoding encoder, the device that the above method is used
Including the Magnetic resonance imaging of the program, the program of the deep learning of artificial intelligence and the head that are pre-processed for storage image
The storage medium of view data, using the deep learning of artificial intelligence to the view data of the Magnetic resonance imaging of head at
Reason and the main frame or computer cluster of analysis, the program pre-processed for display image, the deep learning of artificial intelligence
Program, the display dress of the running of the view data of the Magnetic resonance imaging of various heads and various programs and operation result
Put,
It is characterized in that:The above method can realize that Elderly people group, forgetting type to any given quantity are slightly recognized
Know and damage any given with the view data of the Magnetic resonance imaging of the three-dimensional head of the subject of alzheimer disease
The two dimension slicing of quantity realizes the supervised learning of the self-encoding encoder of the deep learning of artificial intelligence, according to above-mentioned to any given
The Elderly people group of quantity, the nuclear-magnetism of the three-dimensional head of the subject of Amnestic mild cognitive impairment and alzheimer disease
The two dimension slicing of the view data of resonance image-forming is carried out obtained by the supervised learning of the self-encoding encoder of the deep learning of artificial intelligence
Deep learning neural network model, utilize the nuclear-magnetism of the three-dimensional head of the subject to be identified of any given quantity
The two dimension slicing of any given quantity of the view data of resonance image-forming carries out normal old to above-mentioned subject to be identified
The Division identifications of three classification of year group, Amnestic mild cognitive impairment and alzheimer disease.
The present invention proposes a kind of method using Magnetic resonance imaging and Alzheimer disease progress neuroimaging is examined
Disconnected method, can be held the initial data of the image of the three-dimensional voxel of the Magnetic resonance imaging of the full brain of multiple subjects
The storage changed long, the initial data to the three-dimensional voxel of the Magnetic resonance imaging of the full brain of subject carries out denoising and signal intensity
The pretreatment such as normalization, the image progress to the three-dimensional voxel of the Magnetic resonance imaging of the full brain of subject is semi-automatic or full-automatic
Image segmentation, carry out skull stripping, obtain the three-dimensional of the left hippocampus of each subject and the Magnetic resonance imaging of right hippocampus
The image of voxel, using typical or equalization hippocampus free hand drawing collection or many atlas as reference picture or template to tested
The image of the three-dimensional voxel of the left hippocampus of person and the Magnetic resonance imaging of right hippocampus carries out semi- or fully automated image registration,
Automatic image segmentation is carried out to the hippocampus head of hippocampus, fimbria of hippocampus, CA1, CA2, CA3, CA4/ dentate fascia and hippocampus tail,
It is characterized in that:By the view data of the Magnetic resonance imaging of the subject of the Elderly people group of specified quantity, refer to
The view data of the Magnetic resonance imaging of the subject for being clinically diagnosed as Amnestic mild cognitive impairment of fixed number amount
Done with the view data of the Magnetic resonance imaging of the subject for being clinically diagnosed as alzheimer disease of specified quantity
Into the training set of the supervised learning of the two-dimentional self-encoding encoder of the deep learning of artificial intelligence, by the Elderly people group of specified quantity
The view data of the Magnetic resonance imaging of subject, specified quantity are clinically diagnosed as Amnestic mild cognitive impairment
Subject Magnetic resonance imaging view data and specified quantity be clinically diagnosed as alzheimer disease
The view data of the Magnetic resonance imaging of subject makes the supervised learning of the two-dimentional self-encoding encoder of the deep learning of artificial intelligence
Checking collection, by the view data of the Magnetic resonance imaging of the subject of the Elderly people group of specified quantity, specified quantity
The view data and specified quantity of the Magnetic resonance imaging of subject through being clinically diagnosed as Amnestic mild cognitive impairment
The view data of Magnetic resonance imaging of the subject for being clinically diagnosed as alzheimer disease make artificial intelligence
Deep learning two-dimentional self-encoding encoder supervised learning test set, record the ID or sequence number of each subject, structure
The two-dimentional self-encoding encoder of the deep learning of artificial intelligence is built, base is realized using the two-dimentional self-encoding encoder of the deep learning of artificial intelligence
In full brain and/or the Computer assisted identification and/or the depth using artificial intelligence of the deep learning of the morphological feature of hippocampus
The two-dimentional self-encoding encoder of degree study realizes that the area of computer aided of the deep learning of the textural characteristics based on full brain and/or hippocampus is known
Not,
The morphological feature of hippocampus head, fimbria of hippocampus, CA1, CA2, CA3, CA4/ dentate fascia and hippocampus tail based on hippocampus
Assisting in identifying for deep learning of artificial intelligence refer to:From the hippocampus head of each subject, fimbria of hippocampus, CA1, CA2, CA3,
Chosen and coronal-plane, sagittal plane and horizontal plane in the image of the three-dimensional voxel of the Magnetic resonance imaging of CA4/ dentate fascias and hippocampus tail
Parallel specified quantity and the size using voxel that specifies Spacing is the two dimension slicings of thickness, and the size of above-mentioned voxel is big
It is small arbitrarily to be set according to the quality of the image of Magnetic resonance imaging, to above-mentioned training set, to each from normal
The label of all upper Elderly people groups of above-mentioned two dimension slicing mark of the subject of old age group, to each from clinically
It is diagnosed as on all above-mentioned two dimension slicing marks of the subject of Amnestic mild cognitive impairment and is clinically diagnosed as forgetting
The label of type mild cognitive impairment, to each from all upper of the subject for being clinically diagnosed as alzheimer disease
The label that alzheimer disease is clinically diagnosed as on two dimension slicing mark is stated, above-mentioned each two dimension slicing is divided into greatly
The mutual nonoverlapping multiple regions of small identical, by the region of the reflection based on the numerical value of the intensity of the gray scale in each region
Image information feature as the input of the two-dimentional self-encoding encoder of the deep learning of artificial intelligence, utilize above-mentioned artificial intelligence
The two-dimentional self-encoding encoder of deep learning the morphological feature of the above-mentioned two dimension slicing of all above-mentioned subjects is carried out
Identification, according to the identification to above-mentioned morphological feature obtain be directed to above-mentioned two dimension slicing to Elderly people group, clinically
Three kinds of situations for being diagnosed as Amnestic mild cognitive impairment and being clinically diagnosed as alzheimer disease carry out point of three classification
Class method, using it is above-mentioned for above-mentioned two dimension slicing to Elderly people group, be clinically diagnosed as forgetting type and slightly recognize
Know infringement and be clinically diagnosed as the sorting techniques of the classification of three kinds of situations progress three of alzheimer disease to above-mentioned checking
Collection is verified, the structure and relevant parameter of above-mentioned two-dimentional self-encoding encoder are adjusted in the verification, until to above-mentioned three classification
Sorting technique reaches the rate of accuracy reached of the differentiation between any two of three kinds of above-mentioned situations to more than 85%, in above-mentioned three classification
Sorting technique reaches that the rate of accuracy reached of the differentiation between any two of three kinds of above-mentioned situations, will be above-mentioned in the case of more than 85%
The sorting technique of three classification, which is used to provide the image of clinical nuclear magnetic resonance, refers to diagnostic comments;
The artificial intelligence of textural characteristics based on hippocampus head, fimbria of hippocampus, CA1, CA2, CA3, CA4/ dentate fascia and hippocampus tail
Assisting in identifying for deep learning refer to:From hippocampus head, fimbria of hippocampus, CA1, CA2, CA3, CA4/ dentate fascia of each subject
It is specified with coronal-plane, sagittal plane or plane-parallel with being chosen in the image of the three-dimensional voxel of the Magnetic resonance imaging of hippocampus tail
The two dimension slicing group of quantity, each two dimension slicing group there is given quantity with coronal-plane, sagittal plane or plane-parallel
The size using (thickness is) voxel of continuous adjacent is the two dimension slicing of thickness, and the size of above-mentioned voxel can root
Arbitrarily set according to the quality of the image of Magnetic resonance imaging, two layers adjacent of the two dimension slicing being parallel to each other refers to two mutually
Any one voxel of any one layer of two dimension slicing in addition to fringe region in parallel two dimension slicing is mutual with the two
One voxel of another layer of two dimension slicing of parallel two dimension slicing is adjacent, that is, adjacent two layers being parallel to each other
Two dimension slicing refers to do not have other voxel between this two layers of two dimension slicing, in the centre of each above-mentioned two dimension slicing group
All voxels of 3 to 9 voxels for not including edge of one layer of two dimension slicing carry out the extraction of textural characteristics, above-mentioned
The extracting method of textural characteristics be the change of the numerical value of the intensity level of the gray scale according to adjacent voxel in a different direction come
Realize, by one layer of two dimension slicing of the centre of each above-mentioned two dimension slicing group in addition to 3 to 6 voxels at edge
All voxels textural characteristics as two-dimentional self-encoding encoder input, using artificial intelligence deep learning it is two-dimentional self-editing
The textural characteristics of the above-mentioned two dimension slicing group of all above-mentioned subjects are identified code device, according to above-mentioned textural characteristics
Identification obtain for above-mentioned two dimension slicing hinder to Elderly people group, be clinically diagnosed as Amnestic mild cognitive impairment
Three kinds of situations for being clinically diagnosed as alzheimer disease carry out the sorting technique of three classification, using above-mentioned for above-mentioned
Two dimension slicing group to Elderly people group, be clinically diagnosed as Amnestic mild cognitive impairment and be clinically diagnosed as Ah
The sorting techniques for the classification of three kinds of situations progress three that Wurz sea is write from memory sick are verified to above-mentioned checking collection, in the verification in adjustment
The structure and relevant parameter of two-dimentional self-encoding encoder are stated, until the sorting technique to above-mentioned three classification reaches three kinds of above-mentioned situations
Differentiation between any two rate of accuracy reached to more than 85%, reach three kinds of above-mentioned situations in the sorting techniques of above-mentioned three classification
Differentiation between any two rate of accuracy reached in the case of more than 85%, above-mentioned three sorting techniques classified are used for clinic
Nuclear magnetic resonance image provide refer to diagnostic comments.
The artificial intelligence of textural characteristics based on hippocampus head, fimbria of hippocampus, CA1, CA2, CA3, CA4/ dentate fascia and hippocampus tail
Assisting in identifying for deep learning refer to:From hippocampus head, fimbria of hippocampus, CA1, CA2, CA3, CA4/ dentate fascia of each subject
It is specified with coronal-plane, sagittal plane or plane-parallel with being chosen in the image of the three-dimensional voxel of the Magnetic resonance imaging of hippocampus tail
The two dimension slicing group of quantity, each two dimension slicing group there is given quantity with coronal-plane, sagittal plane or plane-parallel
The size using voxel of continuous adjacent is the two dimension slicing of thickness, and the size of above-mentioned voxel can be common according to nuclear-magnetism
The quality of image of imaging of shaking arbitrarily is set, two layers adjacent of the two dimension slicing being parallel to each other refer to two be parallel to each other two
Tie up any one voxel of any one layer of two dimension slicing in addition to fringe region in section is parallel to each other with the two two
It is adjacent to tie up a voxel of another layer of two dimension slicing of section, that is, two layers adjacent of the two dimension slicing being parallel to each other
Refer to there is no other voxel between this two layers of two dimension slicing, to one layer of the centre of each above-mentioned two dimension slicing group two
All voxels of 3 to 9 voxels for not including edge of dimension section carry out the extraction of textural characteristics, above-mentioned textural characteristics
Extracting method be that the numerical value change in a different direction of intensity level of gray scale according to adjacent voxel is realized, will
One layer of two dimension slicing of the centre of each above-mentioned two dimension slicing group it is all in addition to 3 to 6 voxels at edge
The textural characteristics of voxel as two-dimentional self-encoding encoder input, using artificial intelligence deep learning two-dimentional self-encoding encoder to upper
The textural characteristics for stating two dimension slicing group are identified, and are obtained being directed to above-mentioned two dimension slicing according to the identification to above-mentioned textural characteristics
Resistance to Elderly people group, be clinically diagnosed as Amnestic mild cognitive impairment and be clinically diagnosed as alzheimer disease
Three kinds of situations carry out the sorting techniques of three classification, using it is above-mentioned for above-mentioned two dimension slicing group to Elderly people group,
Three kinds of situations for being clinically diagnosed as Amnestic mild cognitive impairment and being clinically diagnosed as alzheimer disease carry out three
The sorting technique of classification verifies to above-mentioned checking collection, adjust in the verification the structure of above-mentioned two-dimentional self-encoding encoder with it is relevant
Parameter, until to above-mentioned three classification sorting techniques reach three kinds of above-mentioned situations differentiation between any two rate of accuracy reached
To more than 85%, the rate of accuracy reached of the differentiation between any two of three kinds of above-mentioned situations is reached in the sorting technique of above-mentioned three classification
In the case of more than 85%, the image offer reference that above-mentioned three sorting techniques classified are used for clinical nuclear magnetic resonance is examined
Disconnected opinion.
The present invention can only realize the area of computer aided of the deep learning of the above-mentioned artificial intelligence based on morphological feature
Identification.
The present invention can only realize the area of computer aided of the deep learning of the above-mentioned artificial intelligence based on texture feature
Identification.
The present invention can realize the computer aided manufacturing of the deep learning of the above-mentioned artificial intelligence based on morphological feature simultaneously
The Computer assisted identification of identification and the above-mentioned deep learning based on texture feature is helped, by two kinds of above-mentioned diagnostic methods
Advisory opinion of the result of Computer assisted identification collectively as diagnosis.
The structure of the selecting method of the two-dimentional view data of the nuclear magnetic resonance of the head of the present invention and two-dimentional self-encoding encoder
Design method and training method be all that selection can be carried out according to the performance of computer, can use common computer and
High-performance computer or high-performance computer cluster.
The present invention can be carried out to the two-dimentional view data of the nuclear magnetic resonance of the head of the plain scan clinically used
Processing.
Brief description of the drawings
Fig. 1 is the schematic diagram of the structure of the two-dimentional self-encoding encoder of the present invention.
Fig. 2 is the schematic diagram of the two dimension slicing of the coronal-plane of the full brain of the nuclear magnetic resonance of the head of the present invention.
Fig. 3 is the schematic diagram of the two dimension slicing of the sagittal plane of the full brain of the nuclear magnetic resonance of the head of the present invention.
Fig. 4 is the schematic diagram of the two dimension slicing of the horizontal plane of the full brain of the nuclear magnetic resonance of the head of the present invention.
Fig. 5 is the schematic diagram of the two dimension slicing of the coronal-plane of the hippocampus of the nuclear magnetic resonance of the head of the present invention.
Fig. 6 is the schematic diagram of the two dimension slicing of the sagittal plane of the hippocampus of the nuclear magnetic resonance of the head of the present invention.
Fig. 7 is the schematic diagram of the two dimension slicing of the horizontal plane of the hippocampus of the nuclear magnetic resonance of the head of the present invention.
Fig. 8 is the schematic diagram of the division methods of the two dimension slicing of the view data of the nuclear magnetic resonance of the present invention.
Embodiment
Embodiment 1
Embodiment 1 utilizes alzheimer disease neuroimaging plan (Alzheimer ' the s Disease in the U.S.
Neuroimaging Initiative, abbreviation ADNI) data realize the present invention the Magnetic resonance imaging to head image
The hippocampus heads of data, fimbria of hippocampus, that the view data of CA1, CA2, CA3, CA4/ dentate fascia and hippocampus tail carries out area of computer aided is auxiliary
Help the technical scheme of identification.Downloaded from the official website of the alzheimer disease neuroimaging plan in the U.S. by clinical definite
The shadow of the nuclear magnetic resonance of the head of alzheimer disease (AD), Amnestic mild cognitive impairment (MCI) and Elderly people group (NC)
As data,
The data of the alzheimer disease neuroimaging plan in the U.S., 3013 scanning results (scans), 321 are tested
Person (subject), can divide training set, checking collection and test set
The situation of training set
Verify the situation of collection
The situation of test set
The nuclear magnetic resonance image number of above-mentioned head is obtained from the website of the alzheimer disease neuroimaging plan in the U.S.
According to full brain three-dimensional data.The pretreatment of view data is first carried out to above-mentioned three-dimensional data, eddy current is mainly carried out
Calibration (Eddy Current Correction), remove skull (Skull Stripping), utilize existing image segmentation skill
Art carries out automatic image segmentation to the hippocampus head of hippocampus, fimbria of hippocampus, CA1, CA2, CA3, CA4/ dentate fascia and hippocampus tail.Will
The hippocampus head of hippocampus, fimbria of hippocampus, the 3-D view of CA1, CA2, CA3, CA4/ dentate fascia and hippocampus tail are converted to two dimensional image,
91 two dimension slicings of the size that thickness is voxel, the chi of above-mentioned voxel are got on the direction with plane-parallel
Very little size can arbitrarily be set according to the quality of the image of Magnetic resonance imaging, then according to the size of two dimension slicing,
The less section of area of hippocampus head, fimbria of hippocampus, CA1, CA2, CA3, CA4/ dentate fascia and the hippocampus tail of hippocampus is filtered out, is stayed
Lower 62 two dimension slicings, zoom to 96 × 96 size, that is, be divided into 96 × 96 mutual nonoverlapping regions.Then enter
Row normalizing operation, makes above-mentioned two dimension slicing be transformed into conventional coordinates.Data enhancement operations are carried out to image, are such as rotated,
Height displacement, scaling, width displacement etc..
The pretreatment of the image of the nuclear magnetic resonance of the alzheimer disease neuroimaging plan (ADNI) in the U.S. and two dimension are cut
Piece obtain can by the various Medical Imagings of nibabel or other increasing income or business software realize, can also use
The programming languages such as matlab are according to image denoising, the principle programming realization of image segmentation and image registration.
To above-mentioned training set, all above-mentioned two dimension slicing marks to each subject from Elderly people group
The label of upper Elderly people group, to each from all of the subject for being clinically diagnosed as Amnestic mild cognitive impairment
Above-mentioned two dimension slicing mark on be clinically diagnosed as the label of Amnestic mild cognitive impairment, derived to each in clinic
On be diagnosed as on all above-mentioned two dimension slicings marks of the subject of alzheimer disease and be clinically diagnosed as A Erzi seas
The label for disease of writing from memory,
The depth of embodiment 1 is realized using the Computational frame of python language and the keras deep learning increased income
The programming of the software section of habit, hardware is used as using common personal computer or high-performance computer.
Scikit is the storehouse increased income of a machine learning, and Scikit OneHotEncoder functions can realize only heat
Coding, one-hot coding is also referred to as an efficient coding, utilizes the OneHotEncoder function pairs in the storehouse of scikit machine learning
Above-mentioned label realizes one-hot coding.The function of one-hot coding is to carry out one-hot coding by the label to subject, will be tested
The value of the label of person is converted into the vector of three-dimensional, and every one-dimensional value of the three-dimensional vector is 0 or 1.For example, (1,0,
0) Elderly people group is represented, (0,1,0) represents Amnestic mild cognitive impairment, (0,0,1) represents alzheimer disease.For by
The label of examination person, if it has 3 possible values (alzheimer disease, Amnestic mild cognitive impairment, Elderly people group),
So after one-hot coding, the value of label is converted to 3 binary features.Also, this 3 binary feature mutual exclusions, if
It is alzheimer disease, is not just Amnestic mild cognitive impairment, nor Elderly people group;If forgetting type mild cognitive
Infringement, is not just alzheimer disease, nor Elderly people group;It is not just alzheimer disease if Elderly people group,
Nor Amnestic mild cognitive impairment.By realizing one-hot coding to label, it is easy to the deep learning in two-dimentional self-encoding encoder
Training of the middle realization to neutral net.
Above-mentioned each two dimension slicing is divided into size identical nonoverlapping 96 × 96=9216 region mutually, often
The numerical value of the intensity of the gray scale of the image in individual region as each element of 96 × 96 matrix numerical value, by above-mentioned 96 × 96
The element of matrix be used as the input of the two-dimentional self-encoding encoder of above-mentioned artificial intelligence, the two-dimentional own coding of above-mentioned artificial intelligence
The first layer of device is the convolutional layer with 128 core, and the size of core is 3 × 3, and activation primitive is relu, above-mentioned artificial intelligence
The second layer of two-dimentional self-encoding encoder be maximum pond layer, the size of core is 2 × 2, the two-dimentional own coding of above-mentioned artificial intelligence
The third layer of device is the convolutional layer for having 64 core, and the size of core is 3 × 3, and activation primitive is relu, the two of above-mentioned artificial intelligence
It is pond layer to tie up the 4th layer of self-encoding encoder, and the size of core is 2 × 2, the 5th of the two-dimentional self-encoding encoder of above-mentioned artificial intelligence the
Layer is the convolutional layer with 32 core, and the size of core is 3 × 3, and activation primitive is relu, above-mentioned artificial intelligence it is two-dimentional self-editing
The layer 6 of code device is pond layer, and the size of core is 2 × 2, each step of above-mentioned two-dimentional self-encoding encoder in the way of convolution
The coding step of realization have it is corresponding realized in the way of deconvolution decoding the step of, pass through convolution and batch normalization
The processing that (batch normalization) is encoded, passes through deconvolution and batch normalization (batch
Normalization the processing) decoded, pond index refers to the size of the core of pond layer, and up-sampling refers to and pond
The opposite operation of effect.The output of above-mentioned layer 6 is launched into one-dimension array, the two dimension of above-mentioned artificial intelligence is connected to certainly
First hidden layer of encoder, above-mentioned first hidden layer of the two-dimentional self-encoding encoder of above-mentioned artificial intelligence has 200 nerves single
Member, activation primitive is Dropout, and the output of above-mentioned first hidden layer is connected into the two-dimentional self-encoding encoder of above-mentioned artificial intelligence
The second hidden layer, above-mentioned second hidden layer of the two-dimentional self-encoding encoder of above-mentioned artificial intelligence has 200 neural units, then will
The output of above-mentioned second hidden layer is connected to the output layer of the two-dimentional self-encoding encoder of above-mentioned artificial intelligence, and above-mentioned is artificial
The activation primitive of the output layer of the two-dimentional self-encoding encoder of intelligence is softmax.Each layer of neuron is to next layer of neuron
Transformation parameter be train come parameter, all parameters trained are the models of the two-dimentional self-encoding encoder trained, will
The parameter for representing the model of this neutral net is stored in local hard disk.
After the training of above-mentioned two-dimentional self-encoding encoder is completed, with the model of the two-dimentional self-encoding encoder trained to testing
Card collection is verified, and test set is tested.
During checking, the model pair of the Computer assisted identification of the above-mentioned two-dimentional self-encoding encoder trained is used
Each two dimension slicing of 62 two dimension slicings obtained according to above-mentioned method of each subject differentiated, each
The differentiation result of two dimension slicing is that the two dimension slicing belongs to alzheimer disease, Amnestic mild cognitive impairment and Elderly people group
Three kinds of situations probability, by this 62 differentiation results be averaged, it is possible to learn the subject belong to alzheimer disease,
Any probability in three kinds of situations of Amnestic mild cognitive impairment and Elderly people group.
That is, using above-mentioned artificial intelligence deep learning two-dimentional self-encoding encoder to all above-mentioned subjects
The morphological feature of above-mentioned two dimension slicing be identified, obtained according to the identification to above-mentioned morphological feature for above-mentioned
Two dimension slicing to Elderly people group, be clinically diagnosed as Amnestic mild cognitive impairment and be clinically diagnosed as A Erzi
Write from memory three kinds of situations of disease of sea carry out the sorting techniques of three classification, using it is above-mentioned for above-mentioned two dimension slicing to Elderly people
Group, three kinds of situations for being clinically diagnosed as Amnestic mild cognitive impairment and being clinically diagnosed as alzheimer disease are carried out
The sorting technique of three classification is verified to above-mentioned checking collection, and the structure of above-mentioned two-dimentional self-encoding encoder is adjusted in the verification and is had
The parameter of pass, until the sorting technique to above-mentioned three classification reaches the accuracy rate of the differentiation between any two of three kinds of above-mentioned situations
More than 85% is reached, the accuracy rate of the differentiation between any two of three kinds of above-mentioned situations is reached in the sorting technique of above-mentioned three classification
Reach in the case of more than 85%, advisory opinion can be provided for clinical diagnosis.
Embodiment 2
Embodiment 2 utilizes alzheimer disease neuroimaging plan (Alzheimer ' the s Disease in the U.S.
Neuroimaging Initiative, abbreviation ADNI) data realize the present invention the Magnetic resonance imaging to head image
The hippocampus head of the hippocampus of data, fimbria of hippocampus, the view data of CA1, CA2, CA3, CA4/ dentate fascia and hippocampus tail are calculated
The technical scheme that machine assists in identifying.Downloaded from the official website of the alzheimer disease neuroimaging plan in the U.S. by facing
The nuclear-magnetism of the head of the alzheimer disease (AD) that bed is made a definite diagnosis, Amnestic mild cognitive impairment (MCI) and Elderly people group (NC)
The image data of resonance,
The data of the alzheimer disease neuroimaging plan in the U.S., 3013 scanning results (scans), 321 are tested
Person (subject), can divide training set, checking collection and test set
The situation of training set
Verify the situation of collection
The situation of test set
The nuclear magnetic resonance image number of above-mentioned head is obtained from the website of the alzheimer disease neuroimaging plan in the U.S.
According to full brain three-dimensional data.The pretreatment of view data is first carried out to above-mentioned three-dimensional data, eddy current is mainly carried out
Calibration (Eddy Current Correction), remove skull (Skull Stripping), utilize existing image segmentation skill
Art carries out automatic image segmentation to the hippocampus head of hippocampus, fimbria of hippocampus, CA1, CA2, CA3, CA4/ dentate fascia and hippocampus tail.Will
The hippocampus head of hippocampus, fimbria of hippocampus, the 3-D view of CA1, CA2, CA3, CA4/ dentate fascia and hippocampus tail are converted to two dimensional image,
91 two dimension slicings of the size that thickness is voxel, the chi of above-mentioned voxel are got on the direction with plane-parallel
Very little size can arbitrarily be set according to the quality of the image of Magnetic resonance imaging, then according to the size of two dimension slicing,
The less section of area of hippocampus head, fimbria of hippocampus, CA1, CA2, CA3, CA4/ dentate fascia and the hippocampus tail of hippocampus is filtered out, is stayed
Lower 62 two dimension slicings, zoom to 96 × 96 size.Then operation is standardized, above-mentioned two dimension slicing is transformed into mark
Conventional coordinates.Data enhancement operations are carried out to image, are such as rotated, height displacement, scaling, width displacement etc..Flat with coronal-plane
91 two dimension slicings are got on capable direction, then according to the size of two dimension slicing, brain area are filtered out less
Section, leaves 62 two dimension slicings, zooms to 96 × 96 size, that is, be divided into 96 × 96 mutual nonoverlapping regions.
Then operation is standardized, above-mentioned two dimension slicing is transformed into conventional coordinates.Data enhancement operations are carried out to image, such as
Rotation, height displacement, scaling, width displacement etc..91 two dimension slicings, Ran Hougen are got on the direction parallel with sagittal plane
According to the size of two dimension slicing, the less section of brain area is filtered out, 62 two dimension slicings is left, zooms to 96 × 96
Size.Then operation is standardized, above-mentioned two dimension slicing is transformed into conventional coordinates.Data enhancing behaviour is carried out to image
Make, such as rotate, height displacement, scaling, width displacement etc..
The pretreatment of the image of the nuclear magnetic resonance of the alzheimer disease neuroimaging plan (ADNI) in the U.S. and two dimension are cut
Piece obtain can by the various Medical Imagings of nibabel or other increasing income or business software realize, can also use
The programming languages such as matlab are according to image denoising, the principle programming realization of image segmentation and image registration.
To above-mentioned training set, all above-mentioned two dimension slicing marks to each subject from Elderly people group
The label of upper Elderly people group, to each from all of the subject for being clinically diagnosed as Amnestic mild cognitive impairment
Above-mentioned two dimension slicing mark on be clinically diagnosed as the label of Amnestic mild cognitive impairment, derived to each in clinic
On be diagnosed as on all above-mentioned two dimension slicings marks of the subject of alzheimer disease and be clinically diagnosed as A Erzi seas
The label for disease of writing from memory,
The depth of embodiment 1 is realized using the Computational frame of python language and the keras deep learning increased income
The programming of the software section of habit, hardware is used as using common personal computer or high-performance computer.
Scikit is the storehouse increased income of a machine learning, and Scikit OneHotEncoder functions can realize only heat
Coding, one-hot coding is also referred to as an efficient coding, utilizes the OneHotEncoder function pairs in the storehouse of scikit machine learning
Above-mentioned label realizes one-hot coding.The function of one-hot coding is to carry out one-hot coding by the label to subject, will be tested
The value of the label of person is converted into the vector of three-dimensional, and every one-dimensional value of the three-dimensional vector is 0 or 1.For example, (1,0,
0) Elderly people group is represented, (0,1,0) represents Amnestic mild cognitive impairment, (0,0,1) represents alzheimer disease.For by
The label of examination person, if it has 3 possible values (alzheimer disease, Amnestic mild cognitive impairment, Elderly people group),
So after one-hot coding, the value of label is converted to 3 binary features.Also, this 3 binary feature mutual exclusions, if
It is alzheimer disease, is not just Amnestic mild cognitive impairment, nor Elderly people group;If forgetting type mild cognitive
Infringement, is not just alzheimer disease, nor Elderly people group;It is not just alzheimer disease if Elderly people group,
Nor Amnestic mild cognitive impairment.By realizing one-hot coding to label, it is easy to the deep learning in two-dimentional self-encoding encoder
Training of the middle realization to neutral net.
Pair above-mentioned three group two dimension slicing parallel with sagittal plane with horizontal plane, coronal-plane is trained respectively, obtains phase
The model of the Computer assisted identification for the two-dimentional self-encoding encoder answered.
Above-mentioned each two dimension slicing is divided into size identical nonoverlapping 96 × 96=9216 region mutually, often
The numerical value of the intensity of the gray scale of the image in individual region as each element of 96 × 96 matrix numerical value, by above-mentioned 96 × 96
The element of matrix be used as the input of the two-dimentional self-encoding encoder of above-mentioned artificial intelligence, the two-dimentional own coding of above-mentioned artificial intelligence
The first layer of device is the convolutional layer with 128 core, and the size of core is 3 × 3, and activation primitive is relu, above-mentioned artificial intelligence
The second layer of two-dimentional self-encoding encoder be maximum pond layer, the size of core is 2 × 2, the two-dimentional own coding of above-mentioned artificial intelligence
The third layer of device is the convolutional layer for having 64 core, and the size of core is 3 × 3, and activation primitive is relu, the two of above-mentioned artificial intelligence
It is pond layer to tie up the 4th layer of self-encoding encoder, and the size of core is 2 × 2, the 5th of the two-dimentional self-encoding encoder of above-mentioned artificial intelligence the
Layer is the convolutional layer with 32 core, and the size of core is 3 × 3, and activation primitive is relu, above-mentioned artificial intelligence it is two-dimentional self-editing
The layer 6 of code device is pond layer, and the size of core is 2 × 2, each step of above-mentioned two-dimentional self-encoding encoder in the way of convolution
The coding step of realization have it is corresponding realized in the way of deconvolution decoding the step of, pass through convolution and batch normalization
The processing that (batch normalization) is encoded, passes through deconvolution and batch normalization (batch
Normalization the processing) decoded, pond index refers to the size of the core of pond layer, and up-sampling refers to and pond
The opposite operation of effect.The output of above-mentioned layer 6 is launched into one-dimension array, the two dimension of above-mentioned artificial intelligence is connected to certainly
First hidden layer of encoder, above-mentioned first hidden layer of the two-dimentional self-encoding encoder of above-mentioned artificial intelligence has 200 nerves single
Member, activation primitive is Dropout, and the output of above-mentioned first hidden layer is connected into the two-dimentional self-encoding encoder of above-mentioned artificial intelligence
The second hidden layer, above-mentioned second hidden layer of the two-dimentional self-encoding encoder of above-mentioned artificial intelligence has 200 neural units, then will
The output of above-mentioned second hidden layer is connected to the output layer of the two-dimentional self-encoding encoder of above-mentioned artificial intelligence, and above-mentioned is artificial
The activation primitive of the output layer of the two-dimentional self-encoding encoder of intelligence is softmax.Each layer of neuron is to next layer of neuron
Transformation parameter be train come parameter, all parameters trained are the models of the two-dimentional self-encoding encoder trained, will
The parameter for representing the model of this neutral net is stored in local hard disk.
After the training of above-mentioned two-dimentional self-encoding encoder is completed, with the model of the two-dimentional self-encoding encoder trained to testing
Card collection is verified, and test set is tested.
During checking, the model pair of the Computer assisted identification of the above-mentioned two-dimentional self-encoding encoder trained is used
62 parallel and parallel with sagittal plane with plane-parallel, with coronal-plane obtained according to above-mentioned method of each subject
Each two dimension slicing of × 3=186 two dimension slicing is differentiated that the differentiation result of each two dimension slicing is that the two dimension is cut
Piece belongs to the probability of three kinds of situations of alzheimer disease, Amnestic mild cognitive impairment and Elderly people group, and this 186 are sentenced
Other result is averaged, it is possible to learn that the subject belongs to alzheimer disease, Amnestic mild cognitive impairment and normal old
Any probability in three kinds of situations that year is organized.
That is, using above-mentioned artificial intelligence deep learning two-dimentional self-encoding encoder to all above-mentioned subjects
The morphological feature of above-mentioned two dimension slicing be identified, obtained according to the identification to above-mentioned morphological feature for above-mentioned
Two dimension slicing to Elderly people group, be clinically diagnosed as Amnestic mild cognitive impairment and be clinically diagnosed as A Erzi
Write from memory three kinds of situations of disease of sea carry out the sorting techniques of three classification, using it is above-mentioned for above-mentioned two dimension slicing to Elderly people
Group, three kinds of situations for being clinically diagnosed as Amnestic mild cognitive impairment and being clinically diagnosed as alzheimer disease are carried out
The sorting technique of three classification is verified to above-mentioned checking collection, and the structure of above-mentioned two-dimentional self-encoding encoder is adjusted in the verification and is had
The parameter of pass, until the sorting technique to above-mentioned three classification reaches the accuracy rate of the differentiation between any two of three kinds of above-mentioned situations
More than 85% is reached, the accuracy rate of the differentiation between any two of three kinds of above-mentioned situations is reached in the sorting technique of above-mentioned three classification
Reach in the case of more than 85%, advisory opinion can be provided for clinical diagnosis.
Embodiment 3
Embodiment 3 utilizes the alzheimer disease neuroimaging plan (Alzheimer ' sDisease in the U.S.
Neuroimaging Initiative, abbreviation ADNI) data realize the present invention the Magnetic resonance imaging to head hippocampus
Head, fimbria of hippocampus, the view data of CA1, CA2, CA3, CA4/ dentate fascia and hippocampus tail carry out the skill that area of computer aided assists in identifying
Art scheme.The alzheimer by clinical definite is downloaded from the official website of the alzheimer disease neuroimaging plan in the U.S.
The image data of the nuclear magnetic resonance of the head of sick (AD), Amnestic mild cognitive impairment (MCI) and Elderly people group (NC),
The data of the alzheimer disease neuroimaging plan in the U.S., 3013 scanning results (scans), 321 are tested
Person (subject), can divide training set, checking collection and test set
The situation of training set
Verify the situation of collection
The situation of test set
The nuclear magnetic resonance image number of above-mentioned head is obtained from the website of the alzheimer disease neuroimaging plan in the U.S.
According to full brain 3-dimensional image data.The pretreatment of view data is first carried out to above-mentioned three-dimensional data, whirlpool is mainly carried out
The calibration (Eddy Current Correction) of electric current, removes skull (Skull Stripping), using it is existing increase income it is soft
The method that part or existing Research Literature are delivered extracts hippocampus from the three-dimensional data of the full brain of the nmr image data of head
The 3-dimensional image data of the part of body, using existing image Segmentation Technology to the hippocampus head of hippocampus, fimbria of hippocampus, CA1, CA2,
CA3, CA4/ dentate fascia and hippocampus tail carry out automatic image segmentation.By the hippocampus head of hippocampus, fimbria of hippocampus, CA1, CA2, CA3,
The 3-D view of CA4/ dentate fascias and hippocampus tail is converted to two dimensional image, and thickness is got on the direction with plane-parallel and is
91 two dimension slicings of the size of voxel, the size of above-mentioned voxel can be according to the image of Magnetic resonance imaging
Quality is arbitrarily set, then according to the size of two dimension slicing, filter out the hippocampus head of hippocampus, fimbria of hippocampus, CA1,
The less section of the area of CA2, CA3, CA4/ dentate fascia and hippocampus tail, leaves 62 two dimension slicings, zoom to 96 × 96 it is big
It is small, that is, to be divided into 96 × 96 mutual nonoverlapping regions.Then operation is standardized, turns above-mentioned two dimension slicing
Change to conventional coordinates.Data enhancement operations are carried out to image, are such as rotated, height displacement, scaling, width displacement etc..
The pretreatment of the image of the nuclear magnetic resonance of the alzheimer disease neuroimaging plan (ADNI) in the U.S. and two dimension are cut
Piece obtain can by the various Medical Imagings of nibabel or other increasing income or business software realize, can also use
The programming languages such as matlab are according to image denoising, the principle programming realization of image segmentation and image registration.
To above-mentioned training set, all above-mentioned two dimension slicing marks to each subject from Elderly people group
The label of upper Elderly people group, to each from all of the subject for being clinically diagnosed as Amnestic mild cognitive impairment
Above-mentioned two dimension slicing mark on be clinically diagnosed as the label of Amnestic mild cognitive impairment, derived to each in clinic
On be diagnosed as on all above-mentioned two dimension slicings marks of the subject of alzheimer disease and be clinically diagnosed as A Erzi seas
The label for disease of writing from memory,
The depth of embodiment 1 is realized using the Computational frame of python language and the keras deep learning increased income
The programming of the software section of habit, hardware is used as using common personal computer or high-performance computer.
Scikit is the storehouse increased income of a machine learning, and Scikit OneHotEncoder functions can realize only heat
Coding, one-hot coding is also referred to as an efficient coding, utilizes the OneHotEncoder function pairs in the storehouse of scikit machine learning
Above-mentioned label realizes one-hot coding.The function of one-hot coding is to carry out one-hot coding by the label to subject, will be tested
The value of the label of person is converted into the vector of three-dimensional, and every one-dimensional value of the three-dimensional vector is 0 or 1.For example, (1,0,
0) Elderly people group is represented, (0,1,0) represents Amnestic mild cognitive impairment, (0,0,1) represents alzheimer disease.For by
The label of examination person, if it has 3 possible values (alzheimer disease, Amnestic mild cognitive impairment, Elderly people group),
So after one-hot coding, the value of label is converted to 3 binary features.Also, this 3 binary feature mutual exclusions, if
It is alzheimer disease, is not just Amnestic mild cognitive impairment, nor Elderly people group;If forgetting type mild cognitive
Infringement, is not just alzheimer disease, nor Elderly people group;It is not just alzheimer disease if Elderly people group,
Nor Amnestic mild cognitive impairment.By realizing one-hot coding to label, it is easy to the deep learning in two-dimentional self-encoding encoder
Training of the middle realization to neutral net.
Above-mentioned each two dimension slicing is divided into size identical nonoverlapping 96 × 96=9216 region mutually, often
The numerical value of the intensity of the gray scale of the image in individual region as each element of 96 × 96 matrix numerical value, by above-mentioned 96 × 96
The element of matrix be used as the input of the two-dimentional self-encoding encoder of above-mentioned artificial intelligence, the two-dimentional own coding of above-mentioned artificial intelligence
The first layer of device is the convolutional layer with 128 core, and the size of core is 3 × 3, and activation primitive is relu, above-mentioned artificial intelligence
The second layer of two-dimentional self-encoding encoder be maximum pond layer, the size of core is 2 × 2, the two-dimentional own coding of above-mentioned artificial intelligence
The third layer of device is the convolutional layer for having 64 core, and the size of core is 3 × 3, and activation primitive is relu, the two of above-mentioned artificial intelligence
It is pond layer to tie up the 4th layer of self-encoding encoder, and the size of core is 2 × 2, the 5th of the two-dimentional self-encoding encoder of above-mentioned artificial intelligence the
Layer is the convolutional layer with 32 core, and the size of core is 3 × 3, and activation primitive is relu, above-mentioned artificial intelligence it is two-dimentional self-editing
The layer 6 of code device is pond layer, and the size of core is 2 × 2, each step of above-mentioned two-dimentional self-encoding encoder in the way of convolution
The coding step of realization have it is corresponding realized in the way of deconvolution decoding the step of, pass through convolution and batch normalization
The processing that (batch normalization) is encoded, passes through deconvolution and batch normalization (batch
Normalization the processing) decoded, pond index refers to the size of the core of pond layer, and up-sampling refers to and pond
The opposite operation of effect.The output of above-mentioned layer 6 is launched into one-dimension array, the two dimension of above-mentioned artificial intelligence is connected to certainly
First hidden layer of encoder, above-mentioned first hidden layer of the two-dimentional self-encoding encoder of above-mentioned artificial intelligence has 200 nerves single
Member, activation primitive is Dropout, and the output of above-mentioned first hidden layer is connected into the two-dimentional self-encoding encoder of above-mentioned artificial intelligence
The second hidden layer, above-mentioned second hidden layer of the two-dimentional self-encoding encoder of above-mentioned artificial intelligence has 200 neural units, then will
The output of above-mentioned second hidden layer is connected to the output layer of the two-dimentional self-encoding encoder of above-mentioned artificial intelligence, and above-mentioned is artificial
The activation primitive of the output layer of the two-dimentional self-encoding encoder of intelligence is softmax.Each layer of neuron is to next layer of neuron
Transformation parameter be train come parameter, all parameters trained are the models of the two-dimentional self-encoding encoder trained, will
The parameter for representing the model of this neutral net is stored in local hard disk.
After the training of above-mentioned two-dimentional self-encoding encoder is completed, with the model of the two-dimentional self-encoding encoder trained to testing
Card collection is verified, and test set is tested.
During checking, the model pair of the Computer assisted identification of the above-mentioned two-dimentional self-encoding encoder trained is used
Each two dimension slicing of 62 two dimension slicings obtained according to above-mentioned method of each subject differentiated, each
The differentiation result of two dimension slicing is that the section belongs to the three of alzheimer disease, Amnestic mild cognitive impairment and Elderly people group
The probability of the situation of kind, this 62 differentiation results are averaged, it is possible to learn that the subject belongs to alzheimer disease, forgetting
Any probability in three kinds of situations of type mild cognitive impairment and Elderly people group.
That is, using above-mentioned artificial intelligence deep learning two-dimentional self-encoding encoder to all above-mentioned subjects
The morphological feature of above-mentioned two dimension slicing be identified, obtained according to the identification to above-mentioned morphological feature for above-mentioned
Two dimension slicing to Elderly people group, be clinically diagnosed as Amnestic mild cognitive impairment and be clinically diagnosed as A Erzi
Write from memory three kinds of situations of disease of sea carry out the sorting techniques of three classification, using it is above-mentioned for above-mentioned two dimension slicing to Elderly people
Group, three kinds of situations for being clinically diagnosed as Amnestic mild cognitive impairment and being clinically diagnosed as alzheimer disease are carried out
The sorting technique of three classification is verified to above-mentioned checking collection, and the structure of above-mentioned two-dimentional self-encoding encoder is adjusted in the verification and is had
The parameter of pass, until the sorting technique to above-mentioned three classification reaches the accuracy rate of the differentiation between any two of three kinds of above-mentioned situations
More than 85% is reached, the accuracy rate of the differentiation between any two of three kinds of above-mentioned situations is reached in the sorting technique of above-mentioned three classification
Reach in the case of more than 85%, advisory opinion can be provided for clinical diagnosis.
Embodiment 4
Embodiment 4 utilizes alzheimer disease neuroimaging plan (Alzheimer ' the s Disease in the U.S.
Neuroimaging Initiative, abbreviation ADNI) data realize the present invention the Magnetic resonance imaging to head hippocampus
The hippocampus head of body, fimbria of hippocampus, the view data of CA1, CA2, CA3, CA4/ dentate fascia and hippocampus tail carry out area of computer aided auxiliary
The technical scheme of identification.From the official website of the alzheimer disease neuroimaging plan in the U.S. download by clinical definite Ah
Wurz sea is silent sick (AD), the image of the nuclear magnetic resonance of the head of Amnestic mild cognitive impairment (MCI) and Elderly people group (NC)
Data,
The data of the alzheimer disease neuroimaging plan in the U.S., 3013 scanning results (scans), 321 are tested
Person (subject), can divide training set, checking collection and test set
The situation of training set
Verify the situation of collection
The situation of test set
The nuclear magnetic resonance image number of above-mentioned head is obtained from the website of the alzheimer disease neuroimaging plan in the U.S.
According to full brain three-dimensional data.The pretreatment of view data is first carried out to above-mentioned three-dimensional data, eddy current is mainly carried out
Calibration (Eddy Current Correction), remove skull (Skull Stripping), utilize existing image segmentation skill
Art carries out automatic image segmentation to the hippocampus head of hippocampus, fimbria of hippocampus, CA1, CA2, CA3, CA4/ dentate fascia and hippocampus tail.Will
The hippocampus head of hippocampus, fimbria of hippocampus, the 3-D view of CA1, CA2, CA3, CA4/ dentate fascia and hippocampus tail are converted to two dimensional image,
91 two dimension slicings of the size that thickness is voxel, the chi of above-mentioned voxel are got on the direction with plane-parallel
Very little size can arbitrarily be set according to the quality of the image of Magnetic resonance imaging, then according to the size of two dimension slicing,
The less section of area of hippocampus head, fimbria of hippocampus, CA1, CA2, CA3, CA4/ dentate fascia and the hippocampus tail of hippocampus is filtered out, is stayed
Lower 62 two dimension slicings, zoom to 96 × 96 size, that is, be divided into 96 × 96 mutual nonoverlapping regions.Then enter
Row normalizing operation, makes above-mentioned two dimension slicing be transformed into conventional coordinates.Data enhancement operations are carried out to image, are such as rotated,
Height displacement, scaling, width displacement etc..The 91 of the size that thickness is voxel are got on the direction parallel with coronal-plane
Individual two dimension slicing, the size of above-mentioned voxel can arbitrarily be set according to the quality of the image of Magnetic resonance imaging, so
Afterwards according to the size of two dimension slicing, the less section of brain area is filtered out, 62 two dimension slicings are left, zoom to 96 ×
96 size.Then operation is standardized, above-mentioned two dimension slicing is transformed into conventional coordinates.Data increasing is carried out to image
Strong operation, such as rotates, height displacement, scaling, width displacement etc..Thickness is got on the direction parallel with sagittal plane for voxel
Size 91 two dimension slicings, the size of above-mentioned voxel can be according to the quality of the image of Magnetic resonance imaging
Arbitrarily set, then according to the size of two dimension slicing, filter out the less section of brain area, leave 62 two dimensions and cut
Piece, zoom to 96 × 96 size.Then operation is standardized, above-mentioned two dimension slicing is transformed into conventional coordinates.It is right
Image carries out data enhancement operations, such as rotates, height displacement, scaling, width displacement etc..
The pretreatment of the image of the nuclear magnetic resonance of the alzheimer disease neuroimaging plan (ADNI) in the U.S. and two dimension are cut
Piece obtain can by the various Medical Imagings of nibabel or other increasing income or business software realize, can also use
The programming languages such as matlab are according to image denoising, the principle programming realization of image segmentation and image registration.
To above-mentioned training set, all above-mentioned two dimension slicing marks to each subject from Elderly people group
The label of upper Elderly people group, to each from all of the subject for being clinically diagnosed as Amnestic mild cognitive impairment
Above-mentioned two dimension slicing mark on be clinically diagnosed as the label of Amnestic mild cognitive impairment, derived to each in clinic
On be diagnosed as on all above-mentioned two dimension slicings marks of the subject of alzheimer disease and be clinically diagnosed as A Erzi seas
The label for disease of writing from memory,
The depth of embodiment 1 is realized using the Computational frame of python language and the keras deep learning increased income
The programming of the software section of habit, hardware is used as using common personal computer or high-performance computer.
Scikit is the storehouse increased income of a machine learning, and Scikit OneHotEncoder functions can realize only heat
Coding, one-hot coding is also referred to as an efficient coding, utilizes the OneHotEncoder function pairs in the storehouse of scikit machine learning
Above-mentioned label realizes one-hot coding.The function of one-hot coding is to carry out one-hot coding by the label to subject, will be tested
The value of the label of person is converted into the vector of three-dimensional, and every one-dimensional value of the three-dimensional vector is 0 or 1.For example, (1,0,
0) Elderly people group is represented, (0,1,0) represents Amnestic mild cognitive impairment, (0,0,1) represents alzheimer disease.For by
The label of examination person, if it has 3 possible values (alzheimer disease, Amnestic mild cognitive impairment, Elderly people group),
So after one-hot coding, the value of label is converted to 3 binary features.Also, this 3 binary feature mutual exclusions, if
It is alzheimer disease, is not just Amnestic mild cognitive impairment, nor Elderly people group;If forgetting type mild cognitive
Infringement, is not just alzheimer disease, nor Elderly people group;It is not just alzheimer disease if Elderly people group,
Nor Amnestic mild cognitive impairment.By realizing one-hot coding to label, it is easy to the deep learning in two-dimentional self-encoding encoder
Training of the middle realization to neutral net.
Pair above-mentioned three group two dimension slicing parallel with sagittal plane with horizontal plane, coronal-plane is trained respectively, obtains phase
The model of the Computer assisted identification for the two-dimentional self-encoding encoder answered.
Above-mentioned each two dimension slicing is divided into size identical nonoverlapping 96 × 96=9216 region mutually, often
The numerical value of the intensity of the gray scale of the image in individual region as each element of 96 × 96 matrix numerical value, by above-mentioned 96 × 96
The element of matrix be used as the input of the two-dimentional self-encoding encoder of above-mentioned artificial intelligence, the two-dimentional own coding of above-mentioned artificial intelligence
The first layer of device is the convolutional layer with 128 core, and the size of core is 3 × 3, and activation primitive is relu, above-mentioned artificial intelligence
The second layer of two-dimentional self-encoding encoder be maximum pond layer, the size of core is 2 × 2, the two-dimentional own coding of above-mentioned artificial intelligence
The third layer of device is the convolutional layer for having 64 core, and the size of core is 3 × 3, and activation primitive is relu, the two of above-mentioned artificial intelligence
It is pond layer to tie up the 4th layer of self-encoding encoder, and the size of core is 2 × 2, the 5th of the two-dimentional self-encoding encoder of above-mentioned artificial intelligence the
Layer is the convolutional layer with 32 core, and the size of core is 3 × 3, and activation primitive is relu, above-mentioned artificial intelligence it is two-dimentional self-editing
The layer 6 of code device is pond layer, and the size of core is 2 × 2, each step of above-mentioned two-dimentional self-encoding encoder in the way of convolution
The coding step of realization have it is corresponding realized in the way of deconvolution decoding the step of, pass through convolution and batch normalization
The processing that (batch normalization) is encoded, passes through deconvolution and batch normalization (batch
Normalization the processing) decoded, pond index refers to the size of the core of pond layer, and up-sampling refers to and pond
The opposite operation of effect.The output of above-mentioned layer 6 is launched into one-dimension array, the two dimension of above-mentioned artificial intelligence is connected to certainly
First hidden layer of encoder, above-mentioned first hidden layer of the two-dimentional self-encoding encoder of above-mentioned artificial intelligence has 200 nerves single
Member, activation primitive is Dropout, and the output of above-mentioned first hidden layer is connected into the two-dimentional self-encoding encoder of above-mentioned artificial intelligence
The second hidden layer, above-mentioned second hidden layer of the two-dimentional self-encoding encoder of above-mentioned artificial intelligence has 200 neural units, then will
The output of above-mentioned second hidden layer is connected to the output layer of the two-dimentional self-encoding encoder of above-mentioned artificial intelligence, and above-mentioned is artificial
The activation primitive of the output layer of the two-dimentional self-encoding encoder of intelligence is softmax.Each layer of neuron is to next layer of neuron
Transformation parameter be train come parameter, all parameters trained are the models of the two-dimentional self-encoding encoder trained, will
The parameter for representing the model of this neutral net is stored in local hard disk.
After the training of above-mentioned two-dimentional self-encoding encoder is completed, with the model of the two-dimentional self-encoding encoder trained to testing
Card collection is verified, and test set is tested.
During checking, the model pair of the Computer assisted identification of the above-mentioned two-dimentional self-encoding encoder trained is used
62 parallel and parallel with sagittal plane with plane-parallel, with coronal-plane obtained according to above-mentioned method of each subject
Each two dimension slicing of × 3=186 two dimension slicing is differentiated that the differentiation result of each two dimension slicing is that the two dimension is cut
Piece is the probability for three kinds of situations for belonging to alzheimer disease, Amnestic mild cognitive impairment and Elderly people group, by this 186
Differentiate result be averaged, it is possible to learn the subject belong to alzheimer disease, Amnestic mild cognitive impairment and normally
Any probability in three kinds of situations of old age group.
That is, using above-mentioned artificial intelligence deep learning two-dimentional self-encoding encoder to all above-mentioned subjects
The morphological feature of above-mentioned two dimension slicing be identified, obtained according to the identification to above-mentioned morphological feature for above-mentioned
Two dimension slicing to Elderly people group, be clinically diagnosed as Amnestic mild cognitive impairment and be clinically diagnosed as A Erzi
Write from memory three kinds of situations of disease of sea carry out the sorting techniques of three classification, using it is above-mentioned for above-mentioned two dimension slicing to Elderly people
Group, three kinds of situations for being clinically diagnosed as Amnestic mild cognitive impairment and being clinically diagnosed as alzheimer disease are carried out
The sorting technique of three classification is verified to above-mentioned checking collection, and the structure of above-mentioned two-dimentional self-encoding encoder is adjusted in the verification and is had
The parameter of pass, until the sorting technique to above-mentioned three classification reaches the accuracy rate of the differentiation between any two of three kinds of above-mentioned situations
More than 85% is reached, the accuracy rate of the differentiation between any two of three kinds of above-mentioned situations is reached in the sorting technique of above-mentioned three classification
Reach in the case of more than 85%, advisory opinion can be provided for clinical diagnosis.
Embodiment 5
Embodiment 5 utilizes alzheimer disease neuroimaging plan (Alzheimer ' the s Disease in the U.S.
Neuroimaging Initiative, abbreviation ADNI) data realize the present invention the Magnetic resonance imaging to head hippocampus
The hippocampus head of body, fimbria of hippocampus, the view data of CA1, CA2, CA3, CA4/ dentate fascia and hippocampus tail carry out area of computer aided auxiliary
The technical scheme of identification.From the official website of the alzheimer disease neuroimaging plan in the U.S. download by clinical definite Ah
Wurz sea is silent sick (AD), the image of the nuclear magnetic resonance of the head of Amnestic mild cognitive impairment (MCI) and Elderly people group (NC)
Data,
The data of the alzheimer disease neuroimaging plan in the U.S., 3013 scanning results (scans), 321 are tested
Person (subject), can divide training set, checking collection and test set
The situation of training set
Verify the situation of collection
The situation of test set
The nuclear magnetic resonance image number of above-mentioned head is obtained from the website of the alzheimer disease neuroimaging plan in the U.S.
According to full brain three-dimensional data.The pretreatment of view data is first carried out to above-mentioned three-dimensional data, eddy current is mainly carried out
Calibration (Eddy Current Correction), remove skull (Skull Stripping), utilize existing image segmentation skill
Art carries out automatic image segmentation to the hippocampus head of hippocampus, fimbria of hippocampus, CA1, CA2, CA3, CA4/ dentate fascia and hippocampus tail.Will
The hippocampus head of hippocampus, fimbria of hippocampus, the 3-D view of CA1, CA2, CA3, CA4/ dentate fascia and hippocampus tail are converted to two dimensional image,
91 two dimension slicings of the size that thickness is voxel, the chi of above-mentioned voxel are got on the direction with plane-parallel
Very little size can arbitrarily be set according to the quality of the image of Magnetic resonance imaging, then according to the size of two dimension slicing,
The less section of area of hippocampus head, fimbria of hippocampus, CA1, CA2, CA3, CA4/ dentate fascia and the hippocampus tail of hippocampus is filtered out, is stayed
Lower 62 two dimension slicings, are divided into 256 × 256 regions.Then operation is standardized, above-mentioned two dimension slicing is transformed into mark
Conventional coordinates.Data enhancement operations are carried out to image, are such as rotated, height displacement, scaling, width displacement etc..Flat with coronal-plane
91 two dimension slicings of the size that thickness is voxel are got on capable direction, the size of above-mentioned voxel can root
Arbitrarily set according to the quality of the image of Magnetic resonance imaging, then according to the size of two dimension slicing, filter out brain face
The less section of product, leaves 62 two dimension slicings, is divided into 256 × 256 regions.Then operation is standardized, is made above-mentioned
Two dimension slicing is transformed into conventional coordinates.Data enhancement operations are carried out to image, are such as rotated, height displacement, scaling, width displacement
Deng.91 two dimension slicings of the size that thickness is voxel, above-mentioned voxel are got on the direction parallel with sagittal plane
Size can arbitrarily be set according to the quality of the image of Magnetic resonance imaging, it is then big according to the area of two dimension slicing
It is small, the less section of brain area is filtered out, 62 two dimension slicings are left, is divided into 256 × 256 regions.Then standard is carried out
Change operation, above-mentioned two dimension slicing is transformed into conventional coordinates.Data enhancement operations are carried out to image, are such as rotated, height position
Move, scaling, width displacement etc..
The pretreatment of the image of the nuclear magnetic resonance of the alzheimer disease neuroimaging plan (ADNI) in the U.S. and two dimension are cut
Piece obtain can by the various Medical Imagings of nibabel or other increasing income or business software realize, can also use
The programming languages such as matlab are according to image denoising, the principle programming realization of image segmentation and image registration.
To above-mentioned training set, all above-mentioned two dimension slicing marks to each subject from Elderly people group
The label of upper Elderly people group, to each from all of the subject for being clinically diagnosed as Amnestic mild cognitive impairment
Above-mentioned two dimension slicing mark on be clinically diagnosed as the label of Amnestic mild cognitive impairment, derived to each in clinic
On be diagnosed as on all above-mentioned two dimension slicings marks of the subject of alzheimer disease and be clinically diagnosed as A Erzi seas
The label for disease of writing from memory,
The depth of embodiment 1 is realized using the Computational frame of python language and the keras deep learning increased income
The programming of the software section of habit, hardware is used as using common personal computer or high-performance computer.
Scikit is the storehouse increased income of a machine learning, and Scikit OneHotEncoder functions can realize only heat
Coding, one-hot coding is also referred to as an efficient coding, utilizes the OneHotEncoder function pairs in the storehouse of scikit machine learning
Above-mentioned label realizes one-hot coding.The function of one-hot coding is to carry out one-hot coding by the label to subject, will be tested
The value of the label of person is converted into the vector of three-dimensional, and every one-dimensional value of the three-dimensional vector is 0 or 1.For example, (1,0,
0) Elderly people group is represented, (0,1,0) represents Amnestic mild cognitive impairment, (0,0,1) represents alzheimer disease.For by
The label of examination person, if it has 3 possible values (alzheimer disease, Amnestic mild cognitive impairment, Elderly people group),
So after one-hot coding, the value of label is converted to 3 binary features.Also, this 3 binary feature mutual exclusions, if
It is alzheimer disease, is not just Amnestic mild cognitive impairment, nor Elderly people group;If forgetting type mild cognitive
Infringement, is not just alzheimer disease, nor Elderly people group;It is not just alzheimer disease if Elderly people group,
Nor Amnestic mild cognitive impairment.By realizing one-hot coding to label, it is easy to the deep learning in two-dimentional self-encoding encoder
Training of the middle realization to neutral net.
Pair above-mentioned three group two dimension slicing parallel with sagittal plane with horizontal plane, coronal-plane is trained respectively, obtains phase
The model of the Computer assisted identification for the two-dimentional self-encoding encoder answered.
Above-mentioned each two dimension slicing is divided into size identical nonoverlapping 96 × 96=9216 region mutually, often
The numerical value of the intensity of the gray scale of the image in individual region as each element of 96 × 96 matrix numerical value, by above-mentioned 96 × 96
The element of matrix be used as the input of the two-dimentional self-encoding encoder of above-mentioned artificial intelligence, the two-dimentional own coding of above-mentioned artificial intelligence
The first layer of device is the convolutional layer with 128 core, and the size of core is 3 × 3, and activation primitive is relu, above-mentioned artificial intelligence
The second layer of two-dimentional self-encoding encoder be maximum pond layer, the size of core is 2 × 2, the two-dimentional own coding of above-mentioned artificial intelligence
The third layer of device is the convolutional layer for having 64 core, and the size of core is 3 × 3, and activation primitive is relu, the two of above-mentioned artificial intelligence
It is pond layer to tie up the 4th layer of self-encoding encoder, and the size of core is 2 × 2, the 5th of the two-dimentional self-encoding encoder of above-mentioned artificial intelligence the
Layer is the convolutional layer with 32 core, and the size of core is 3 × 3, and activation primitive is relu, above-mentioned artificial intelligence it is two-dimentional self-editing
The layer 6 of code device is pond layer, and the size of core is 2 × 2, each step of above-mentioned two-dimentional self-encoding encoder in the way of convolution
The coding step of realization have it is corresponding realized in the way of deconvolution decoding the step of, pass through convolution and batch normalization
The processing that (batch normalization) is encoded, passes through deconvolution and batch normalization (batch
Normalization the processing) decoded, pond index refers to the size of the core of pond layer, and up-sampling refers to and pond
The opposite operation of effect.The output of above-mentioned layer 6 is launched into one-dimension array, the two dimension of above-mentioned artificial intelligence is connected to certainly
First hidden layer of encoder, above-mentioned first hidden layer of the two-dimentional self-encoding encoder of above-mentioned artificial intelligence has 200 nerves single
Member, activation primitive is Dropout, and the output of above-mentioned first hidden layer is connected into the two-dimentional self-encoding encoder of above-mentioned artificial intelligence
The second hidden layer, above-mentioned second hidden layer of the two-dimentional self-encoding encoder of above-mentioned artificial intelligence has 200 neural units, then will
The output of above-mentioned second hidden layer is connected to the output layer of the two-dimentional self-encoding encoder of above-mentioned artificial intelligence, and above-mentioned is artificial
The activation primitive of the output layer of the two-dimentional self-encoding encoder of intelligence is softmax.Each layer of neuron is to next layer of neuron
Transformation parameter be train come parameter, all parameters trained are the models of the two-dimentional self-encoding encoder trained, will
The parameter for representing the model of this neutral net is stored in local hard disk.
After the training of above-mentioned two-dimentional self-encoding encoder is completed, with the model of the two-dimentional self-encoding encoder trained to testing
Card collection is verified, and test set is tested.
During checking, the model pair of the Computer assisted identification of the above-mentioned two-dimentional self-encoding encoder trained is used
62 parallel and parallel with sagittal plane with plane-parallel, with coronal-plane obtained according to above-mentioned method of each subject
Each two dimension slicing of × 3=186 two dimension slicing is differentiated that the differentiation result of each two dimension slicing is that the two dimension is cut
Piece belongs to the probability of three kinds of situations of alzheimer disease, Amnestic mild cognitive impairment and Elderly people group, and this 186 are sentenced
Other result is averaged, it is possible to learn that the subject belongs to alzheimer disease, Amnestic mild cognitive impairment and normal old
Any probability in three kinds of situations that year is organized.
That is, using above-mentioned artificial intelligence deep learning two-dimentional self-encoding encoder to all above-mentioned subjects
The morphological feature of above-mentioned two dimension slicing be identified, obtained according to the identification to above-mentioned morphological feature for above-mentioned
Two dimension slicing to Elderly people group, be clinically diagnosed as Amnestic mild cognitive impairment and be clinically diagnosed as A Erzi
Write from memory three kinds of situations of disease of sea carry out the sorting techniques of three classification, using it is above-mentioned for above-mentioned two dimension slicing to Elderly people
Group, three kinds of situations for being clinically diagnosed as Amnestic mild cognitive impairment and being clinically diagnosed as alzheimer disease are carried out
The sorting technique of three classification is verified to above-mentioned checking collection, and the structure of above-mentioned two-dimentional self-encoding encoder is adjusted in the verification and is had
The parameter of pass, until the sorting technique to above-mentioned three classification reaches the accuracy rate of the differentiation between any two of three kinds of above-mentioned situations
More than 85% is reached, the accuracy rate of the differentiation between any two of three kinds of above-mentioned situations is reached in the sorting technique of above-mentioned three classification
Reach in the case of more than 85%, advisory opinion can be provided for clinical diagnosis.
Embodiment 6
Embodiment 6 is realized using the two dimension slicing of the influence data of the Magnetic resonance imaging of clinical plain scan head
The hippocampus head of hippocampus of the image data of the Magnetic resonance imaging to head of the present invention, fimbria of hippocampus, CA1, CA2, CA3,
The view data of CA4/ dentate fascias and hippocampus tail carries out the technical scheme that area of computer aided assists in identifying.Using passing through clinic
The nuclear-magnetism of the head of the alzheimer disease (AD) made a definite diagnosis, Amnestic mild cognitive impairment (MCI) and Elderly people group (NC) is total to
The two dimension slicing of the image data shaken completing to train, verify and test.
20 two dimension slicings with getting the size that thickness is voxel on the direction of the plane-parallel of scanning, on
The size for the voxel stated can arbitrarily be set according to the quality of the image of Magnetic resonance imaging, be divided into 96 × 96 areas
Domain.Then operation is standardized, above-mentioned two dimension slicing is transformed into conventional coordinates.Data enhancing behaviour is carried out to image
Make, such as rotate, height displacement, scaling, width displacement etc..Hippocampus head, sea using existing image Segmentation Technology to hippocampus
Horse umbrella, CA1, CA2, CA3, CA4/ dentate fascia and hippocampus tail carry out automatic image segmentation.
To above-mentioned training set, all above-mentioned two dimension slicing marks to each subject from Elderly people group
The label of upper Elderly people group, to each from all of the subject for being clinically diagnosed as Amnestic mild cognitive impairment
Above-mentioned two dimension slicing mark on be clinically diagnosed as the label of Amnestic mild cognitive impairment, derived to each in clinic
On be diagnosed as on all above-mentioned two dimension slicings marks of the subject of alzheimer disease and be clinically diagnosed as A Erzi seas
The label for disease of writing from memory,
The depth of embodiment 1 is realized using the Computational frame of python language and the keras deep learning increased income
The programming of the software section of habit, hardware is used as using common personal computer or high-performance computer.
Scikit is the storehouse increased income of a machine learning, and Scikit OneHotEncoder functions can realize only heat
Coding, one-hot coding is also referred to as an efficient coding, utilizes the OneHotEncoder function pairs in the storehouse of scikit machine learning
Above-mentioned label realizes one-hot coding.The function of one-hot coding is to carry out one-hot coding by the label to subject, will be tested
The value of the label of person is converted into the vector of three-dimensional, and every one-dimensional value of the three-dimensional vector is 0 or 1.For example, (1,0,
0) Elderly people group is represented, (0,1,0) represents Amnestic mild cognitive impairment, (0,0,1) represents alzheimer disease.For by
The label of examination person, if it has 3 possible values (alzheimer disease, Amnestic mild cognitive impairment, Elderly people group),
So after one-hot coding, the value of label is converted to 3 binary features.Also, this 3 binary feature mutual exclusions, if
It is alzheimer disease, is not just Amnestic mild cognitive impairment, nor Elderly people group;If forgetting type mild cognitive
Infringement, is not just alzheimer disease, nor Elderly people group;It is not just alzheimer disease if Elderly people group,
Nor Amnestic mild cognitive impairment.By realizing one-hot coding to label, it is easy to the deep learning in two-dimentional self-encoding encoder
Training of the middle realization to neutral net.
Above-mentioned each two dimension slicing is divided into size identical nonoverlapping 96 × 96=9216 region mutually, often
The numerical value of the intensity of the gray scale of the image in individual region as each element of 96 × 96 matrix numerical value, by above-mentioned 96 × 96
The element of matrix be used as the input of the two-dimentional self-encoding encoder of above-mentioned artificial intelligence, the two-dimentional own coding of above-mentioned artificial intelligence
The first layer of device is the convolutional layer with 128 core, and the size of core is 3 × 3, and activation primitive is relu, above-mentioned artificial intelligence
The second layer of two-dimentional self-encoding encoder be maximum pond layer, the size of core is 2 × 2, the two-dimentional own coding of above-mentioned artificial intelligence
The third layer of device is the convolutional layer for having 64 core, and the size of core is 3 × 3, and activation primitive is relu, the two of above-mentioned artificial intelligence
It is pond layer to tie up the 4th layer of self-encoding encoder, and the size of core is 2 × 2, the 5th of the two-dimentional self-encoding encoder of above-mentioned artificial intelligence the
Layer is the convolutional layer with 32 core, and the size of core is 3 × 3, and activation primitive is relu, above-mentioned artificial intelligence it is two-dimentional self-editing
The layer 6 of code device is pond layer, and the size of core is 2 × 2, each step of above-mentioned two-dimentional self-encoding encoder in the way of convolution
The coding step of realization have it is corresponding realized in the way of deconvolution decoding the step of, pass through convolution and batch normalization
The processing that (batch normalization) is encoded, passes through deconvolution and batch normalization (batch
Normalization the processing) decoded, pond index refers to the size of the core of pond layer, and up-sampling refers to and pond
The opposite operation of effect.The output of above-mentioned layer 6 is launched into one-dimension array, the two dimension of above-mentioned artificial intelligence is connected to certainly
First hidden layer of encoder, above-mentioned first hidden layer of the two-dimentional self-encoding encoder of above-mentioned artificial intelligence has 200 nerves single
Member, activation primitive is Dropout, and the output of above-mentioned first hidden layer is connected into the two-dimentional self-encoding encoder of above-mentioned artificial intelligence
The second hidden layer, above-mentioned second hidden layer of the two-dimentional self-encoding encoder of above-mentioned artificial intelligence has 200 neural units, then will
The output of above-mentioned second hidden layer is connected to the output layer of the two-dimentional self-encoding encoder of above-mentioned artificial intelligence, and above-mentioned is artificial
The activation primitive of the output layer of the two-dimentional self-encoding encoder of intelligence is softmax.Each layer of neuron is to next layer of neuron
Transformation parameter be train come parameter, all parameters trained are the models of the two-dimentional self-encoding encoder trained, will
The parameter for representing the model of this neutral net is stored in local hard disk.
Above-mentioned each two dimension slicing is divided into size identical nonoverlapping 96 × 96=9216 region mutually, often
The numerical value of the intensity of the gray scale of the image in individual region as each element of 96 × 96 matrix numerical value, by above-mentioned 96 × 96
The element of matrix be used as the input of the two-dimentional self-encoding encoder of above-mentioned artificial intelligence, the two-dimentional own coding of above-mentioned artificial intelligence
The first layer of device is the convolutional layer with 32 core, and the size of core is 3 × 3, and activation primitive is relu, above-mentioned artificial intelligence
The second layer of two-dimentional self-encoding encoder is maximum pond layer, and the size of core is 2 × 2, the two-dimentional self-encoding encoder of above-mentioned artificial intelligence
Third layer be the convolutional layer for having 64 core, the size of core is 3 × 3, and activation primitive is relu, the two dimension of above-mentioned artificial intelligence
The 4th layer of self-encoding encoder is pond layer, and the size of core is 2 × 2, the layer 5 of the two-dimentional self-encoding encoder of above-mentioned artificial intelligence
For the convolutional layer with 128 core, the size of core is 3 × 3, and activation primitive is relu, above-mentioned artificial intelligence it is two-dimentional self-editing
The layer 6 of code device is pond layer, and the size of core is 2 × 2, and the output of above-mentioned layer 6 is launched into one-dimension array, is connected to
First hidden layer of the two-dimentional self-encoding encoder for the artificial intelligence stated, above-mentioned the first of the two-dimentional self-encoding encoder of above-mentioned artificial intelligence
Hidden layer has 200 neural units, and activation primitive is Dropout, and the output of above-mentioned first hidden layer is connected into above-mentioned people
Second hidden layer of the two-dimentional self-encoding encoder of work intelligence, above-mentioned second hidden layer of the two-dimentional self-encoding encoder of above-mentioned artificial intelligence
There are 200 neural units, then the output of above-mentioned second hidden layer is connected to the two-dimentional own coding of above-mentioned artificial intelligence
The output layer of device, the activation primitive of the output layer of the two-dimentional self-encoding encoder of above-mentioned artificial intelligence is softmax.Each layer of god
It is to train the parameter come through member to the transformation parameter of next layer of neuron, all parameters trained are two trained
The model of self-encoding encoder is tieed up, the parameter that will represent the model of this neutral net is stored in local hard disk.
After the training of above-mentioned two-dimentional self-encoding encoder is completed, with the model of the two-dimentional self-encoding encoder trained to testing
Card collection is verified, and test set is tested.
During checking, the model pair of the Computer assisted identification of the above-mentioned two-dimentional self-encoding encoder trained is used
Each two dimension slicing of 20 two dimension slicings with plane-parallel obtained according to above-mentioned method of each subject
Differentiated, the differentiation result of each two dimension slicing is that the two dimension slicing is to belong to alzheimer disease, forgetting type slightly to recognize
Know the probability of three kinds of situations of infringement and Elderly people group, this 20 differentiation results are averaged, it is possible to learn that this is tested
Any probability that person belongs in three kinds of situations of alzheimer disease, Amnestic mild cognitive impairment and Elderly people group.
That is, using above-mentioned artificial intelligence deep learning two-dimentional self-encoding encoder to all above-mentioned subjects
The morphological feature of above-mentioned two dimension slicing be identified, obtained according to the identification to above-mentioned morphological feature for above-mentioned
Two dimension slicing to Elderly people group, be clinically diagnosed as Amnestic mild cognitive impairment and be clinically diagnosed as A Erzi
Write from memory three kinds of situations of disease of sea carry out the sorting techniques of three classification, using it is above-mentioned for above-mentioned two dimension slicing to Elderly people
Group, three kinds of situations for being clinically diagnosed as Amnestic mild cognitive impairment and being clinically diagnosed as alzheimer disease are carried out
The sorting technique of three classification is verified to above-mentioned checking collection, and the structure of above-mentioned two-dimentional self-encoding encoder is adjusted in the verification and is had
The parameter of pass, until the sorting technique to above-mentioned three classification reaches the accuracy rate of the differentiation between any two of three kinds of above-mentioned situations
More than 85% is reached, the accuracy rate of the differentiation between any two of three kinds of above-mentioned situations is reached in the sorting technique of above-mentioned three classification
Reach in the case of more than 85%, advisory opinion can be provided for clinical diagnosis.
Embodiment 7
Embodiment 7 utilizes alzheimer disease neuroimaging plan (Alzheimer ' the s Disease in the U.S.
Neuroimaging Initiative, abbreviation ADNI) data realize the present invention the Magnetic resonance imaging to head image
The hippocampus head of the hippocampus of data, fimbria of hippocampus, the view data of CA1, CA2, CA3, CA4/ dentate fascia and hippocampus tail are calculated
The technical scheme that machine assists in identifying.Downloaded from the official website of the alzheimer disease neuroimaging plan in the U.S. by facing
The nuclear-magnetism of the head of the alzheimer disease (AD) that bed is made a definite diagnosis, Amnestic mild cognitive impairment (MCI) and Elderly people group (NC)
The image data of resonance,
The data of the alzheimer disease neuroimaging plan in the U.S., 3013 scanning results (scans), 321 are tested
Person (subject), can divide training set, checking collection and test set
The situation of training set
Verify the situation of collection
The situation of test set
The nuclear magnetic resonance image number of above-mentioned head is obtained from the website of the alzheimer disease neuroimaging plan in the U.S.
According to full brain three-dimensional data.The pretreatment of view data is first carried out to above-mentioned three-dimensional data, eddy current is mainly carried out
Calibration (Eddy Current Correction), remove skull (Skull Stripping), utilize existing image segmentation skill
Art carries out automatic image segmentation to the hippocampus head of hippocampus, fimbria of hippocampus, CA1, CA2, CA3, CA4/ dentate fascia and hippocampus tail.Will
The hippocampus head of hippocampus, fimbria of hippocampus, the 3-D view of CA1, CA2, CA3, CA4/ dentate fascia and hippocampus tail are converted to two dimensional image,
91 two dimension slicings of the size that thickness is voxel, the chi of above-mentioned voxel are got on the direction with plane-parallel
Very little size can arbitrarily be set according to the quality of the image of Magnetic resonance imaging, then according to the size of two dimension slicing,
The less section of area of hippocampus head, fimbria of hippocampus, CA1, CA2, CA3, CA4/ dentate fascia and the hippocampus tail of hippocampus is filtered out, is stayed
Lower 62 two dimension slicings, zoom to 96 × 96 size, that is, be divided into 96 × 96 mutual nonoverlapping regions.Then enter
Row normalizing operation, makes above-mentioned two dimension slicing be transformed into conventional coordinates.Data enhancement operations are carried out to image, are such as rotated,
Height displacement, scaling, width displacement etc..
The pretreatment of the image of the nuclear magnetic resonance of the alzheimer disease neuroimaging plan (ADNI) in the U.S. and two dimension are cut
Piece obtain can by the various Medical Imagings of nibabel or other increasing income or business software realize, can also use
The programming languages such as matlab are according to image denoising, the principle programming realization of image segmentation and image registration.
To above-mentioned training set, all above-mentioned two dimension slicing marks to each subject from Elderly people group
The label of upper Elderly people group, to each from all of the subject for being clinically diagnosed as Amnestic mild cognitive impairment
Above-mentioned two dimension slicing mark on be clinically diagnosed as the label of Amnestic mild cognitive impairment, derived to each in clinic
On be diagnosed as on all above-mentioned two dimension slicings marks of the subject of alzheimer disease and be clinically diagnosed as A Erzi seas
The label for disease of writing from memory,
The depth of embodiment 1 is realized using the Computational frame of python language and the keras deep learning increased income
The programming of the software section of habit, hardware is used as using common personal computer or high-performance computer.
Scikit is the storehouse increased income of a machine learning, and Scikit OneHotEncoder functions can realize only heat
Coding, one-hot coding is also referred to as an efficient coding, utilizes the OneHotEncoder function pairs in the storehouse of scikit machine learning
Above-mentioned label realizes one-hot coding.The function of one-hot coding is to carry out one-hot coding by the label to subject, will be tested
The value of the label of person is converted into the vector of three-dimensional, and every one-dimensional value of the three-dimensional vector is 0 or 1.For example, (1,0,
0) Elderly people group is represented, (0,1,0) represents Amnestic mild cognitive impairment, (0,0,1) represents alzheimer disease.For by
The label of examination person, if it has 3 possible values (alzheimer disease, Amnestic mild cognitive impairment, Elderly people group),
So after one-hot coding, the value of label is converted to 3 binary features.Also, this 3 binary feature mutual exclusions, if
It is alzheimer disease, is not just Amnestic mild cognitive impairment, nor Elderly people group;If forgetting type mild cognitive
Infringement, is not just alzheimer disease, nor Elderly people group;It is not just alzheimer disease if Elderly people group,
Nor Amnestic mild cognitive impairment.By realizing one-hot coding to label, it is easy to the deep learning in two-dimentional self-encoding encoder
Training of the middle realization to neutral net.
From hippocampus head, fimbria of hippocampus, CA1, CA2, CA3, CA4/ dentate fascia and the hippocampus tail of the hippocampus of each subject
Magnetic resonance imaging three-dimensional voxel image in choose with the specified quantity of coronal-plane, sagittal plane or plane-parallel two
Section group is tieed up, each two dimension slicing group has the continuous adjacent with coronal-plane, sagittal plane or plane-parallel of given quantity
Thickness be the two dimension slicing of the size of voxel, the size of above-mentioned voxel can be according to the figure of Magnetic resonance imaging
The quality of picture is arbitrarily set, and two layers adjacent of the two dimension slicing being parallel to each other refers in two two dimension slicings being parallel to each other
Any one voxel of any one layer of two dimension slicing in addition to fringe region is another with the two two dimension slicings being parallel to each other
One voxel of one layer of two dimension slicing is adjacent, that is, two layers adjacent of the two dimension slicing being parallel to each other refers to this two layers
There is no other voxel between two dimension slicing, to one layer of the two dimension slicing in the centre of each above-mentioned two dimension slicing group not
All voxels of 3 to 9 voxels including edge carry out the extraction of textural characteristics, the extracting method of above-mentioned textural characteristics
It is that the numerical value change in a different direction of the intensity level of gray scale according to adjacent voxel is realized, will be above-mentioned each
The texture of all voxels in addition to 3 to 6 voxels at edge of one layer of two dimension slicing of the centre of individual two dimension slicing group
Feature as two-dimentional self-encoding encoder input, using artificial intelligence deep learning two-dimentional self-encoding encoder to above-mentioned two dimension slicing
The textural characteristics of group are identified, and are obtained according to the identification to above-mentioned textural characteristics for above-mentioned two dimension slicing resistance to normal
Old group, be clinically diagnosed as Amnestic mild cognitive impairment and be clinically diagnosed as three kinds of situations of alzheimer disease
Carry out the sorting techniques of three classification, using it is above-mentioned for above-mentioned two dimension slicing group to Elderly people group, clinically true
Examine the classification for the classification of three kinds of situations progress three for for Amnestic mild cognitive impairment and being clinically diagnosed as alzheimer disease
Method is verified to above-mentioned checking collection, the structure and relevant parameter of above-mentioned two-dimentional self-encoding encoder is adjusted in the verification, directly
To above-mentioned three classification sorting techniques reach three kinds of above-mentioned situations differentiation between any two rate of accuracy reached to 85% with
On, above-mentioned three classification sorting techniques reach three kinds of above-mentioned situations differentiation between any two rate of accuracy reached to 85% with
In the case of upper, the image offer that above-mentioned three sorting techniques classified are used for clinical nuclear magnetic resonance is referred into diagnostic comments.
Above-mentioned each two dimension slicing is divided into size identical nonoverlapping 96 × 96=9216 region mutually, often
The numerical value of the intensity of the gray scale of the image in individual region as each element of 96 × 96 matrix numerical value, by above-mentioned 96 × 96
The element of matrix be used as the input of the two-dimentional self-encoding encoder of above-mentioned artificial intelligence, the two-dimentional own coding of above-mentioned artificial intelligence
The first layer of device is the convolutional layer with 128 core, and the size of core is 3 × 3, and activation primitive is relu, above-mentioned artificial intelligence
The second layer of two-dimentional self-encoding encoder be maximum pond layer, the size of core is 2 × 2, the two-dimentional own coding of above-mentioned artificial intelligence
The third layer of device is the convolutional layer for having 64 core, and the size of core is 3 × 3, and activation primitive is relu, the two of above-mentioned artificial intelligence
It is pond layer to tie up the 4th layer of self-encoding encoder, and the size of core is 2 × 2, the 5th of the two-dimentional self-encoding encoder of above-mentioned artificial intelligence the
Layer is the convolutional layer with 32 core, and the size of core is 3 × 3, and activation primitive is relu, above-mentioned artificial intelligence it is two-dimentional self-editing
The layer 6 of code device is pond layer, and the size of core is 2 × 2, each step of above-mentioned two-dimentional self-encoding encoder in the way of convolution
The coding step of realization have it is corresponding realized in the way of deconvolution decoding the step of, pass through convolution and batch normalization
The processing that (batch normalization) is encoded, passes through deconvolution and batch normalization (batch
Normalization the processing) decoded, pond index refers to the size of the core of pond layer, and up-sampling refers to and pond
The opposite operation of effect.The output of above-mentioned layer 6 is launched into one-dimension array, the two dimension of above-mentioned artificial intelligence is connected to certainly
First hidden layer of encoder, above-mentioned first hidden layer of the two-dimentional self-encoding encoder of above-mentioned artificial intelligence has 200 nerves single
Member, activation primitive is Dropout, and the output of above-mentioned first hidden layer is connected into the two-dimentional self-encoding encoder of above-mentioned artificial intelligence
The second hidden layer, above-mentioned second hidden layer of the two-dimentional self-encoding encoder of above-mentioned artificial intelligence has 200 neural units, then will
The output of above-mentioned second hidden layer is connected to the output layer of the two-dimentional self-encoding encoder of above-mentioned artificial intelligence, and above-mentioned is artificial
The activation primitive of the output layer of the two-dimentional self-encoding encoder of intelligence is softmax.Each layer of neuron is to next layer of neuron
Transformation parameter be train come parameter, all parameters trained are the models of the two-dimentional self-encoding encoder trained, will
The parameter for representing the model of this neutral net is stored in local hard disk.
After the training of above-mentioned two-dimentional self-encoding encoder is completed, with the model of the two-dimentional self-encoding encoder trained to testing
Card collection is verified, and test set is tested.
During checking, the model pair of the Computer assisted identification of the above-mentioned two-dimentional self-encoding encoder trained is used
The textural characteristics of each the two dimension slicing group obtained according to above-mentioned method of each subject are differentiated that each two
The differentiation result for tieing up section group is that the two dimension slicing group belongs to alzheimer disease, Amnestic mild cognitive impairment and Elderly people
Group three kinds of situations probability, by these differentiation results be averaged, it is possible to learn the subject belong to alzheimer disease,
Any probability in three kinds of situations of Amnestic mild cognitive impairment and Elderly people group.
That is, using above-mentioned artificial intelligence deep learning two-dimentional self-encoding encoder to all above-mentioned subjects
The textural characteristics of above-mentioned two dimension slicing group be identified, obtained being directed to above-mentioned two according to the identification to above-mentioned textural characteristics
Dimension section group to Elderly people group, be clinically diagnosed as Amnestic mild cognitive impairment and be clinically diagnosed as A Erzi
Write from memory three kinds of situations of disease of sea carry out the sorting techniques of three classification, using it is above-mentioned for above-mentioned two dimension slicing group to normal old
Year group, three kinds of situations for being clinically diagnosed as Amnestic mild cognitive impairment and being clinically diagnosed as alzheimer disease are entered
Row three classify sorting technique above-mentioned checking collection is verified, adjust in the verification above-mentioned two-dimentional self-encoding encoder structure and
Relevant parameter, until the sorting technique to above-mentioned three classification reaches the accurate of the differentiation between any two of three kinds of above-mentioned situations
Rate reaches more than 85%, above-mentioned three classification sorting techniques reach three kinds of above-mentioned situations differentiation between any two it is accurate
Rate is reached in the case of more than 85%, can provide advisory opinion for clinical diagnosis.
Claims (10)
1. a kind of nuclear magnetic resonance image processing unit based on self-encoding encoder, the device includes, pre-processed for storage image
The storage medium of the view data of the Magnetic resonance imaging of program, the program of the deep learning of artificial intelligence and head, utilizes people
Main frame or meter that the deep learning of work intelligence is handled and analyzed to the view data of the Magnetic resonance imaging of head
Calculation machine cluster, the program pre-processed for display image, the program of the deep learning of artificial intelligence, the nuclear magnetic resonance of various heads
The display device of the running and operation result of the view data of imaging and various programs,
It is characterized in that:Said apparatus can realize Elderly people group, the forgetting type mild cognitive damage to any given quantity
The process image of the view data of the Magnetic resonance imaging of the three-dimensional head of the subject of evil and alzheimer disease is split
To hippocampus substructure any given quantity two dimension slicing realize artificial intelligence deep learning self-encoding encoder
Supervised learning, according to above-mentioned Elderly people group, Amnestic mild cognitive impairment and A Erzi to any given quantity
The two dimension slicing of the view data of the Magnetic resonance imaging of the three-dimensional head of the subject of the silent disease in sea carries out the depth of artificial intelligence
The neural network model of deep learning obtained by the supervised learning for the self-encoding encoder for spending study, utilizes any given quantity
The two dimension slicing of any given quantity of the view data of the Magnetic resonance imaging of the three-dimensional head of subject to be identified
Three classification of Elderly people group, Amnestic mild cognitive impairment and alzheimer disease are carried out to above-mentioned subject to be identified
Division identification.
2. a kind of nuclear magnetic resonance image processing method based on self-encoding encoder, the device that the above method is used includes, for storing
The storage of the view data of the Magnetic resonance imaging of the program of image preprocessing, the program of the deep learning of artificial intelligence and head
Medium, the calculating that the view data of the Magnetic resonance imaging of head is handled and analyzed using the deep learning of artificial intelligence
Machine host or computer cluster, the program pre-processed for display image, the program of the deep learning of artificial intelligence, various heads
The view data of Magnetic resonance imaging and the running of various programs and the display device of operation result,
It is characterized in that:The above method can realize Elderly people group, the forgetting type mild cognitive damage to any given quantity
The process image of the view data of the Magnetic resonance imaging of the three-dimensional head of the subject of evil and alzheimer disease is split
To hippocampus substructure any given quantity two dimension slicing realize artificial intelligence deep learning self-encoding encoder
Supervised learning, according to above-mentioned Elderly people group, Amnestic mild cognitive impairment and A Erzi to any given quantity
The two dimension slicing of the view data of the Magnetic resonance imaging of the three-dimensional head of the subject of the silent disease in sea carries out the depth of artificial intelligence
The neural network model of deep learning obtained by the supervised learning for the self-encoding encoder for spending study, utilizes any given quantity
The two dimension slicing of any given quantity of the view data of the Magnetic resonance imaging of the three-dimensional head of subject to be identified
Three classification of Elderly people group, Amnestic mild cognitive impairment and alzheimer disease are carried out to above-mentioned subject to be identified
Division identification.
3. method according to claim 2, it is characterised in that:The nuclear-magnetism of the subject of the Elderly people group of specified quantity is total to
Shake the view data of imaging, the nuclear-magnetism of the subject for being clinically diagnosed as Amnestic mild cognitive impairment of specified quantity
The nuclear-magnetism of the view data of resonance image-forming and the subject for being clinically diagnosed as alzheimer disease of specified quantity are total to
Shake imaging view data make artificial intelligence deep learning two-dimentional self-encoding encoder supervised learning training set, will specify
The view data of the Magnetic resonance imaging of the subject of the Elderly people group of quantity, specified quantity are clinically diagnosed as
The view data of the Magnetic resonance imaging of the subject of Amnestic mild cognitive impairment and specified quantity it is clinically true
Examine the Magnetic resonance imaging of the subject for alzheimer disease view data make artificial intelligence deep learning two dimension
The checking collection of the supervised learning of self-encoding encoder, by the image of the Magnetic resonance imaging of the subject of the Elderly people group of specified quantity
The figure of data, the Magnetic resonance imaging of the subject for being clinically diagnosed as Amnestic mild cognitive impairment of specified quantity
As data and the image of the Magnetic resonance imaging of the subject for being clinically diagnosed as alzheimer disease of specified quantity
Data make the test set of the supervised learning of the two-dimentional self-encoding encoder of the deep learning of artificial intelligence, record each subject's
ID or sequence number, build the two-dimentional self-encoding encoder of the deep learning of artificial intelligence, utilize the two of the deep learning of artificial intelligence
Tie up the form that self-encoding encoder realizes the hippocampus head based on hippocampus, fimbria of hippocampus, CA1, CA2, CA3, CA4/ dentate fascia and hippocampus tail
Learn the Computer assisted identification of the deep learning of feature and/or realized using the two-dimentional self-encoding encoder of the deep learning of artificial intelligence
The deep learning of the textural characteristics of hippocampus head, fimbria of hippocampus, CA1, CA2, CA3, CA4/ dentate fascia and hippocampus tail based on hippocampus
Computer assisted identification, the shape of hippocampus head, fimbria of hippocampus, CA1, CA2, CA3, CA4/ dentate fascia and hippocampus tail based on hippocampus
Assisting in identifying for the deep learning of the artificial intelligence of state feature refers to:From the hippocampus head of each subject, fimbria of hippocampus, CA1,
Chosen and coronal-plane, sagittal plane in the image of the three-dimensional voxel of the Magnetic resonance imaging of CA2, CA3, CA4/ dentate fascia and hippocampus tail
With the specified quantity of plane-parallel and the size using voxel specified Spacing as the two dimension slicing of thickness, above-mentioned voxel
Size can arbitrarily be set according to the quality of the image of Magnetic resonance imaging, to above-mentioned training set, to it is each come
The label of all upper Elderly people groups of above-mentioned two dimension slicing mark of the subject of Elderly people group is come from, is derived to each
Clinically it is diagnosed as clinically true on all above-mentioned two dimension slicing marks of the subject of Amnestic mild cognitive impairment
The label for Amnestic mild cognitive impairment is examined, to each from the subject's for being clinically diagnosed as alzheimer disease
The label of alzheimer disease is clinically diagnosed as on all above-mentioned two dimension slicing marks, by above-mentioned each two dimension slicing
Size identical nonoverlapping multiple regions mutually are divided into, will be anti-based on the numerical value of the intensity of the gray scale in each region
Reflect the region image information feature as the two-dimentional self-encoding encoder of the deep learning of artificial intelligence input, using above-mentioned
Morphology of the two-dimentional self-encoding encoder of the deep learning of artificial intelligence to the above-mentioned two dimension slicing of all above-mentioned subjects
Feature is identified, according to identification to above-mentioned morphological feature obtain for above-mentioned two dimension slicing to Elderly people group,
Three kinds of situations for being clinically diagnosed as Amnestic mild cognitive impairment and being clinically diagnosed as alzheimer disease carry out three
The sorting technique of classification, using it is above-mentioned for above-mentioned two dimension slicing to Elderly people group, be clinically diagnosed as forgetting
Type mild cognitive impairment and be clinically diagnosed as alzheimer disease three kinds of situations carry out three classification sorting techniques to upper
The checking collection stated is verified, the structure and relevant parameter of above-mentioned two-dimentional self-encoding encoder is adjusted in the verification, until to above-mentioned
The sorting technique of three classification reaches the rate of accuracy reached of the differentiation between any two of three kinds of above-mentioned situations to more than 85%, above-mentioned
The sorting techniques of three classification reach situation of the rate of accuracy reached of the differentiation between any two of three kinds of above-mentioned situations to more than 85%
Under, the image offer that above-mentioned three sorting techniques classified are used for clinical nuclear magnetic resonance is referred into diagnostic comments;
The artificial intelligence of the textural characteristics of full brain based on hippocampus head, fimbria of hippocampus, CA1, CA2, CA3, CA4/ dentate fascia and hippocampus tail
The assisting in identifying for deep learning of energy refers to:From hippocampus head, fimbria of hippocampus, CA1, CA2, CA3, CA4/ dentation of each subject
Return and hippocampus tail Magnetic resonance imaging three-dimensional voxel image in choose and coronal-plane, sagittal plane or plane-parallel finger
The two dimension slicing group of fixed number amount, each two dimension slicing group is with given quantity and coronal-plane, sagittal plane or plane-parallel
Continuous adjacent the size using voxel as the two dimension slicing of thickness, two layers adjacent of the two dimension slicing being parallel to each other refer to
Any one voxel of any one layer of two dimension slicing in addition to fringe region in two two dimension slicings for being parallel to each other and this
Two voxels of another layer of two dimension slicing of two dimension slicing being parallel to each other are adjacent, that is, the phase being parallel to each other
Two layers adjacent of two dimension slicing refers to do not have other voxel between this two layers of two dimension slicing, in each above-mentioned two dimension slicing group
All voxels of 3 to 9 voxels for not including edge of one layer of two dimension slicing of centre carry out carrying for textural characteristics
Take, the extracting method of above-mentioned textural characteristics be the numerical value of the intensity level of the gray scale according to adjacent voxel in a different direction
Change realize, by one layer of two dimension slicing of the centre of each above-mentioned two dimension slicing group except 3 to 6 of edge
The textural characteristics of all voxels beyond voxel utilize the deep learning of artificial intelligence as the input of two-dimentional self-encoding encoder
The textural characteristics of the above-mentioned two dimension slicing group of all above-mentioned subjects are identified two-dimentional self-encoding encoder, according to above-mentioned
The identification of textural characteristics obtain for above-mentioned two dimension slicing hinder to Elderly people group, to be clinically diagnosed as forgetting type slight
Cognitive impairment and be clinically diagnosed as alzheimer disease three kinds of situations carry out three classification sorting techniques, using above-mentioned
For above-mentioned two dimension slicing group to Elderly people group, be clinically diagnosed as Amnestic mild cognitive impairment and clinically
The sorting techniques for being diagnosed as the classification of three kinds of situations progress three of alzheimer disease are verified to above-mentioned checking collection, in checking
The structure and relevant parameter of the middle above-mentioned two-dimentional self-encoding encoder of adjustment, it is above-mentioned until being reached to above-mentioned three sorting techniques classified
The rate of accuracy reached of the differentiation between any two of three kinds of situations reaches above-mentioned to more than 85% in the sorting technique of above-mentioned three classification
The rate of accuracy reached of the differentiation between any two of three kinds of situations uses above-mentioned three sorting techniques classified in the case of more than 85%
There is provided in the image to clinical nuclear magnetic resonance and refer to diagnostic comments.
4. method according to claim 3, it is characterised in that:The construction method of the two-dimentional self-encoding encoder of above-mentioned artificial intelligence
Can be:Any form of the deep learning of artificial intelligence, any structure can be used and appoint and mode training method two
Tie up self-encoding encoder.
5. method according to claim 3, it is characterised in that:The image of above-mentioned each two dimension slicing is divided into equal in magnitude
With multiple regions of non-overlapping copies, by the numerical value of the intensity of the gray scale of the pixel of the image in each region in above-mentioned multiple regions
It is used as the input of above-mentioned two-dimentional self-encoding encoder.
6. method according to claim 3, it is characterised in that:The image of above-mentioned each two dimension slicing is divided into equal in magnitude
With multiple regions of non-overlapping copies, the numerical value of the intensity of the gray scale of the pixel of the image in each region in above-mentioned multiple regions is
The primitive character of image, passes through special from the numerical value of the intensity of the gray scale of the pixel of the image in each region in above-mentioned multiple regions
Levy conversion and obtain the advanced features that can be classified by machine learning, by the above-mentioned height that can be classified by machine learning
Level feature as above-mentioned two-dimentional self-encoding encoder input.
7. method according to claim 3, it is characterised in that:The construction method of the two-dimentional self-encoding encoder of above-mentioned artificial intelligence
It is:The image of above-mentioned each two dimension slicing is divided into 96 × 96 regions, the number of the intensity of the gray scale of the image in each region
Value regard the element of 96 × 96 above-mentioned matrix as above-mentioned artificial intelligence as the numerical value of the element of 96 × 96 matrix
The input of two-dimentional self-encoding encoder, the first layer of the two-dimentional self-encoding encoder of above-mentioned artificial intelligence is the convolutional layer with 128 core,
The size of core is 3 × 3, and activation primitive is relu, and the second layer of the two-dimentional self-encoding encoder of above-mentioned artificial intelligence is maximum pond
Layer, the size of core is 2 × 2, and the third layer of the two-dimentional self-encoding encoder of above-mentioned artificial intelligence is the convolutional layer for having 64 core, core
Size is 3 × 3, and activation primitive is relu, and the 4th layer of the two-dimentional self-encoding encoder of above-mentioned artificial intelligence is pond layer, core it is big
Small is 2 × 2, and the layer 5 of the two-dimentional self-encoding encoder of above-mentioned artificial intelligence is the convolutional layer with 32 core, and the size of core is 3
× 3, activation primitive is relu, and the layer 6 of the two-dimentional self-encoding encoder of above-mentioned artificial intelligence is pond layer, the size of core is 2 ×
2, the coding step realized in the way of convolution of each step of above-mentioned two-dimentional self-encoding encoder has corresponding with deconvolution
The step of decoding that mode is realized, the output of above-mentioned layer 6 is launched into one-dimension array, is connected to above-mentioned artificial intelligence
First hidden layer of two-dimentional self-encoding encoder, above-mentioned first hidden layer of the two-dimentional self-encoding encoder of above-mentioned artificial intelligence has 200
Neural unit, activation primitive is Dropout, and the output of above-mentioned first hidden layer is connected into the two dimension of above-mentioned artificial intelligence certainly
Second hidden layer of encoder, above-mentioned second hidden layer of the two-dimentional self-encoding encoder of above-mentioned artificial intelligence has 200 nerves single
Member, then the output of above-mentioned second hidden layer is connected to the output layer of the two-dimentional self-encoding encoder of above-mentioned artificial intelligence, on
The activation primitive of the output layer of the two-dimentional self-encoding encoder for the artificial intelligence stated is softmax;Each layer of neuron is to next layer
The transformation parameter of neuron be to train the parameter come, the set of all parameter trained represent the two dimension that trains from
The model of encoder.
8. method according to claim 3, it is characterised in that:The construction method of the two-dimentional self-encoding encoder of above-mentioned artificial intelligence
It is:The image of above-mentioned each two dimension slicing is divided into 256 × 256 region, the intensity of the gray scale of the image in each region
Numerical value as the element of 256 × 256 matrix numerical value, using the element of 256 × 256 above-mentioned matrix as above-mentioned artificial
The input of the two-dimentional self-encoding encoder of intelligence, the first layer of the two-dimentional self-encoding encoder of above-mentioned artificial intelligence is with 128 cores
Convolutional layer, the size of core is 3 × 3, and activation primitive is relu, and the second layer of the two-dimentional self-encoding encoder of above-mentioned artificial intelligence is most
Great Chiization layer, the size of core is 2 × 2, and the third layer of the two-dimentional self-encoding encoder of above-mentioned artificial intelligence is the convolution for having 64 core
Layer, the size of core is 3 × 3, and activation primitive is relu, and the 4th layer of the two-dimentional self-encoding encoder of above-mentioned artificial intelligence is pond
Layer, the size of core is 2 × 2, and the layer 5 of the two-dimentional self-encoding encoder of above-mentioned artificial intelligence is the convolutional layer with 32 core, core
Size be 3 × 3, activation primitive is relu, and the layer 6 of the two-dimentional self-encoding encoder of above-mentioned artificial intelligence is pond layer, core
Size is 2 × 2, the coding step realized in the way of convolution of each step of above-mentioned two-dimentional self-encoding encoder have it is corresponding with
The step of decoding that the mode of deconvolution is realized, the output of above-mentioned layer 6 is launched into one-dimension array, is connected to above-mentioned people
First hidden layer of the two-dimentional self-encoding encoder of work intelligence, above-mentioned first hidden layer of the two-dimentional self-encoding encoder of above-mentioned artificial intelligence
There are 200 neural units, activation primitive is Dropout, the output of above-mentioned first hidden layer is connected to above-mentioned artificial intelligence
Two-dimentional self-encoding encoder the second hidden layer, above-mentioned second hidden layer of the two-dimentional self-encoding encoder of above-mentioned artificial intelligence has 200
Individual neural unit, then the output of above-mentioned second hidden layer is connected to above-mentioned artificial intelligence two-dimentional self-encoding encoder it is defeated
Go out layer, the activation primitive of the output layer of the two-dimentional self-encoding encoder of above-mentioned artificial intelligence is softmax;Each layer of neuron is arrived
The transformation parameter of next layer of neuron is to train the parameter come, and the set of all parameters trained represents what is trained
The model of two-dimentional self-encoding encoder.
9. method according to claim 3, it is characterised in that:Above-mentioned texture feature can be with appointing in image processing field
The feature of the texture of what reflection image.
10. method according to claim 3, it is characterised in that:Above-mentioned texture feature includes any in herein below
Any several combination in one or herein below:The energy value of the gray scale of each voxel, each voxel and with this
The contrast and the contrast of energy value of the gray scale of 26 closest voxels of voxel, the numerical value of the gray scale of voxel along from certain
Two to the line on the center of one voxel to the direction of the line at the center of 26 nearest voxels with the voxel
The gradient of the change of the voxel of the continuous adjacent of the number along given quantity of opposite direction extension, above-mentioned given quantity
Number can be 3,5,7 or 9.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201710283992.7A CN107067396B (en) | 2017-04-26 | 2017-04-26 | Nuclear magnetic resonance image processing device and method based on self-encoder |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201710283992.7A CN107067396B (en) | 2017-04-26 | 2017-04-26 | Nuclear magnetic resonance image processing device and method based on self-encoder |
Publications (2)
Publication Number | Publication Date |
---|---|
CN107067396A true CN107067396A (en) | 2017-08-18 |
CN107067396B CN107067396B (en) | 2021-11-16 |
Family
ID=59603927
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201710283992.7A Active CN107067396B (en) | 2017-04-26 | 2017-04-26 | Nuclear magnetic resonance image processing device and method based on self-encoder |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN107067396B (en) |
Cited By (7)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN107506796A (en) * | 2017-08-25 | 2017-12-22 | 电子科技大学 | A kind of alzheimer disease sorting technique based on depth forest |
CN108009520A (en) * | 2017-12-21 | 2018-05-08 | 东南大学 | A kind of finger vein identification method and system based on convolution variation self-encoding encoder neutral net |
CN108765338A (en) * | 2018-05-28 | 2018-11-06 | 西华大学 | Spatial target images restored method based on convolution own coding convolutional neural networks |
WO2019078923A1 (en) * | 2017-10-19 | 2019-04-25 | General Electric Company | Deep learning architecture for automated image feature extraction |
CN110391014A (en) * | 2018-04-18 | 2019-10-29 | 西门子医疗有限公司 | Utilize the medical image acquisition for the sequence prediction for using deep learning |
US10460440B2 (en) | 2017-10-24 | 2019-10-29 | General Electric Company | Deep convolutional neural network with self-transfer learning |
WO2020087838A1 (en) * | 2018-10-29 | 2020-05-07 | 深圳先进技术研究院 | Blood vessel wall plaque recognition device, system and method, and storage medium |
Citations (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN104851101A (en) * | 2015-05-25 | 2015-08-19 | 电子科技大学 | Brain tumor automatic segmentation method based on deep learning |
CN104881680A (en) * | 2015-05-25 | 2015-09-02 | 电子科技大学 | Alzheimer's disease and mild cognitive impairment identification method based on two-dimension features and three-dimension features |
CN106096636A (en) * | 2016-05-31 | 2016-11-09 | 安徽工业大学 | A kind of Advancement Type mild cognition impairment recognition methods based on neuroimaging |
-
2017
- 2017-04-26 CN CN201710283992.7A patent/CN107067396B/en active Active
Patent Citations (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN104851101A (en) * | 2015-05-25 | 2015-08-19 | 电子科技大学 | Brain tumor automatic segmentation method based on deep learning |
CN104881680A (en) * | 2015-05-25 | 2015-09-02 | 电子科技大学 | Alzheimer's disease and mild cognitive impairment identification method based on two-dimension features and three-dimension features |
CN106096636A (en) * | 2016-05-31 | 2016-11-09 | 安徽工业大学 | A kind of Advancement Type mild cognition impairment recognition methods based on neuroimaging |
Non-Patent Citations (2)
Title |
---|
EHSAN HOSSEINI-ASL 等: "ALZHEIMER’S DISEASE DIAGNOSTICS BY A DEEPLY SUPERVISED ADAPTABLE 3D CONVOLUTIONAL NETWORK", 《ARXIV:1607.00556V1》 * |
闫瑾: "多类型磁共振影像特征融合算法用于阿尔兹海默病早期诊断", 《中国优秀硕士学位论文全文数据库 医药卫生科技辑》 * |
Cited By (14)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN107506796A (en) * | 2017-08-25 | 2017-12-22 | 电子科技大学 | A kind of alzheimer disease sorting technique based on depth forest |
US10607135B2 (en) | 2017-10-19 | 2020-03-31 | General Electric Company | Training an auto-encoder on a single class |
WO2019078923A1 (en) * | 2017-10-19 | 2019-04-25 | General Electric Company | Deep learning architecture for automated image feature extraction |
CN111433785A (en) * | 2017-10-19 | 2020-07-17 | 通用电气公司 | Deep learning architecture for automated image feature extraction |
US10796221B2 (en) | 2017-10-19 | 2020-10-06 | General Electric Company | Deep learning architecture for automated image feature extraction |
US11093820B2 (en) | 2017-10-19 | 2021-08-17 | General Electric Company | Image analysis using deviation from normal data |
CN111433785B (en) * | 2017-10-19 | 2023-11-28 | 通用电气公司 | Deep learning architecture for automated image feature extraction |
US10460440B2 (en) | 2017-10-24 | 2019-10-29 | General Electric Company | Deep convolutional neural network with self-transfer learning |
US11074687B2 (en) | 2017-10-24 | 2021-07-27 | General Electric Company | Deep convolutional neural network with self-transfer learning |
CN108009520A (en) * | 2017-12-21 | 2018-05-08 | 东南大学 | A kind of finger vein identification method and system based on convolution variation self-encoding encoder neutral net |
CN110391014A (en) * | 2018-04-18 | 2019-10-29 | 西门子医疗有限公司 | Utilize the medical image acquisition for the sequence prediction for using deep learning |
CN110391014B (en) * | 2018-04-18 | 2024-03-22 | 西门子医疗有限公司 | Medical image acquisition method and system using sequence prediction using deep learning |
CN108765338A (en) * | 2018-05-28 | 2018-11-06 | 西华大学 | Spatial target images restored method based on convolution own coding convolutional neural networks |
WO2020087838A1 (en) * | 2018-10-29 | 2020-05-07 | 深圳先进技术研究院 | Blood vessel wall plaque recognition device, system and method, and storage medium |
Also Published As
Publication number | Publication date |
---|---|
CN107067396B (en) | 2021-11-16 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN107067395A (en) | A kind of nuclear magnetic resonance image processing unit and method based on convolutional neural networks | |
CN107145727A (en) | The medical image processing devices and method of a kind of utilization convolutional neural networks | |
CN107067396A (en) | A kind of nuclear magnetic resonance image processing unit and method based on self-encoding encoder | |
CN107133461A (en) | A kind of medical image processing devices and method based on self-encoding encoder | |
Farooq et al. | A deep CNN based multi-class classification of Alzheimer's disease using MRI | |
Cheng et al. | CNNs based multi-modality classification for AD diagnosis | |
Basheera et al. | A novel CNN based Alzheimer’s disease classification using hybrid enhanced ICA segmented gray matter of MRI | |
CN110522448B (en) | Brain network classification method based on atlas neural network | |
Li et al. | Alzheimer's disease classification based on combination of multi-model convolutional networks | |
Gupta et al. | Natural image bases to represent neuroimaging data | |
Gupta et al. | Fibernet 2.0: an automatic neural network based tool for clustering white matter fibers in the brain | |
CN112508953B (en) | Meningioma rapid segmentation qualitative method based on deep neural network | |
Tian et al. | Hierarchical organization of the human subcortex unveiled with functional connectivity gradients | |
CN115393269A (en) | Extensible multi-level graph neural network model based on multi-modal image data | |
CN110532907A (en) | Based on face as the Chinese medicine human body constitution classification method with tongue picture bimodal feature extraction | |
CN117218453B (en) | Incomplete multi-mode medical image learning method | |
CN107169955A (en) | A kind of intelligentized medical image processing devices and method | |
CN104463885B (en) | A kind of Multiple Sclerosis lesions region segmentation method | |
Zaina et al. | An exemplar pyramid feature extraction based Alzheimer disease classification method | |
Alagarsamy et al. | Identification of Brain Tumor using Deep Learning Neural Networks | |
Li et al. | Alzheimer's level classification by 3D PMNet using PET/MRI multi-modal images | |
CN110930349A (en) | Medical image processing device and method using convolutional neural network | |
CN116309615A (en) | Multi-mode MRI brain tumor image segmentation method | |
CN113786185B (en) | Static brain network feature extraction method and system based on convolutional neural network | |
Pallawi et al. | Study of Alzheimer’s disease brain impairment and methods for its early diagnosis: a comprehensive survey |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
PB01 | Publication | ||
PB01 | Publication | ||
SE01 | Entry into force of request for substantive examination | ||
SE01 | Entry into force of request for substantive examination | ||
GR01 | Patent grant | ||
GR01 | Patent grant |