CN107909588A - Partition system under MRI cortex based on three-dimensional full convolutional neural networks - Google Patents
Partition system under MRI cortex based on three-dimensional full convolutional neural networks Download PDFInfo
- Publication number
- CN107909588A CN107909588A CN201710620454.2A CN201710620454A CN107909588A CN 107909588 A CN107909588 A CN 107909588A CN 201710620454 A CN201710620454 A CN 201710620454A CN 107909588 A CN107909588 A CN 107909588A
- Authority
- CN
- China
- Prior art keywords
- module
- convolution
- fcnn
- layer
- training
- 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.)
- Pending
Links
- 238000013527 convolutional neural network Methods 0.000 title claims abstract description 31
- 238000005192 partition Methods 0.000 title claims abstract description 15
- 238000012549 training Methods 0.000 claims abstract description 45
- 210000003625 skull Anatomy 0.000 claims abstract description 21
- 238000012805 post-processing Methods 0.000 claims abstract description 15
- 238000007781 pre-processing Methods 0.000 claims abstract description 14
- 238000000265 homogenisation Methods 0.000 claims abstract description 13
- 238000012937 correction Methods 0.000 claims abstract description 12
- 238000013461 design Methods 0.000 claims abstract description 6
- 238000012545 processing Methods 0.000 claims abstract description 5
- 238000000034 method Methods 0.000 claims description 11
- 238000009795 derivation Methods 0.000 claims description 6
- 230000004913 activation Effects 0.000 claims description 4
- 238000004364 calculation method Methods 0.000 claims description 4
- 238000003708 edge detection Methods 0.000 claims description 4
- 238000001514 detection method Methods 0.000 claims description 3
- 230000011218 segmentation Effects 0.000 claims description 3
- 239000000284 extract Substances 0.000 claims description 2
- 230000005540 biological transmission Effects 0.000 claims 1
- 238000002203 pretreatment Methods 0.000 abstract description 5
- 238000009826 distribution Methods 0.000 description 7
- 238000003475 lamination Methods 0.000 description 4
- 238000005516 engineering process Methods 0.000 description 3
- 230000007935 neutral effect Effects 0.000 description 3
- 230000009467 reduction Effects 0.000 description 3
- 230000002739 subcortical effect Effects 0.000 description 3
- 238000013459 approach Methods 0.000 description 2
- 238000004422 calculation algorithm Methods 0.000 description 2
- 210000001159 caudate nucleus Anatomy 0.000 description 2
- 210000001653 corpus striatum Anatomy 0.000 description 2
- 238000010586 diagram Methods 0.000 description 2
- 230000006870 function Effects 0.000 description 2
- 238000007689 inspection Methods 0.000 description 2
- 239000011159 matrix material Substances 0.000 description 2
- 238000010606 normalization Methods 0.000 description 2
- 230000008569 process Effects 0.000 description 2
- 241000208340 Araliaceae Species 0.000 description 1
- 241001566735 Archon Species 0.000 description 1
- 208000012902 Nervous system disease Diseases 0.000 description 1
- 235000005035 Panax pseudoginseng ssp. pseudoginseng Nutrition 0.000 description 1
- 235000003140 Panax quinquefolius Nutrition 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 230000008901 benefit Effects 0.000 description 1
- 210000003710 cerebral cortex Anatomy 0.000 description 1
- 230000007423 decrease Effects 0.000 description 1
- 201000010099 disease Diseases 0.000 description 1
- 208000037265 diseases, disorders, signs and symptoms Diseases 0.000 description 1
- 230000000694 effects Effects 0.000 description 1
- 239000000686 essence Substances 0.000 description 1
- 238000000605 extraction Methods 0.000 description 1
- 235000008434 ginseng Nutrition 0.000 description 1
- 238000003384 imaging method Methods 0.000 description 1
- 230000006872 improvement Effects 0.000 description 1
- 230000003902 lesion Effects 0.000 description 1
- 238000010339 medical test Methods 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 210000000478 neocortex Anatomy 0.000 description 1
- 238000011176 pooling Methods 0.000 description 1
- 238000002360 preparation method Methods 0.000 description 1
- 238000011160 research Methods 0.000 description 1
- 238000005096 rolling process Methods 0.000 description 1
- 238000003860 storage Methods 0.000 description 1
- 239000000126 substance Substances 0.000 description 1
- 210000001103 thalamus Anatomy 0.000 description 1
- 210000001519 tissue Anatomy 0.000 description 1
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/10—Segmentation; Edge detection
- G06T7/12—Edge-based segmentation
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
- G06N3/084—Backpropagation, e.g. using gradient descent
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T17/00—Three dimensional [3D] modelling, e.g. data description of 3D objects
- G06T17/30—Polynomial surface description
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T5/00—Image enhancement or restoration
- G06T5/50—Image enhancement or restoration using two or more images, e.g. averaging or subtraction
-
- 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
- G06T2207/10081—Computed x-ray tomography [CT]
-
- 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
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Software Systems (AREA)
- Mathematical Physics (AREA)
- Computing Systems (AREA)
- Artificial Intelligence (AREA)
- Data Mining & Analysis (AREA)
- Evolutionary Computation (AREA)
- General Health & Medical Sciences (AREA)
- Molecular Biology (AREA)
- Biophysics (AREA)
- General Engineering & Computer Science (AREA)
- Biomedical Technology (AREA)
- Computational Linguistics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Health & Medical Sciences (AREA)
- Algebra (AREA)
- Mathematical Analysis (AREA)
- Mathematical Optimization (AREA)
- Pure & Applied Mathematics (AREA)
- Computer Graphics (AREA)
- Geometry (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Magnetic Resonance Imaging Apparatus (AREA)
Abstract
The present invention discloses partition system under a kind of MRI cortex based on three-dimensional full convolutional neural networks,Including training module,Convolution pre-processing module,FCNN modules and convolution post-processing module,The training module and convolution pre-processing module are connected with FCNN modules,FCNN modules are connected with convolution post-processing module,Training module is used to be trained the training set of input,Convolution pre-processing module is used to carry out gray scale homogenization to image,The pretreatment that correction of inhomogeneous fields and skull are peeled off,It is sent into FCNN modules and is identified after completion convolution pre-treatment,By 9 layers of convolutional layer,After 3 layers of full articulamentum and last classification layer,Obtain nine density profiles,Convolution post processing by convolution post-processing module,Complete the processing that nine density profiles are isolated with small subregion,And nine density profiles are merged,An obtained design sketch,So as to improve the accuracy and efficiency of subregion under cortex.
Description
Technical field
The present invention relates to field of medical technology, particularly relates under a kind of MRI cortex based on three-dimensional full convolutional neural networks
Partition system.
Background technology
To the accurate subregion of cerebral cortex lower structure, the research for many nervous system diseases is significant.Such as
The diseased region of Parkinson's is likely to occur in infracortical black substance, and the lesion of caudate nucleus, volt core and lenticular nucleus may be with oneself
The generation for closing disease is related.Corpus straitum isocortex lower structure, can be checked by MRI and obtain three-dimensional structure, and be tied under these cortex
The accurate subregion of structure is still not easy accurately to accomplish.Clinician people work area is depended on to the subregion of subcortical structures at present
Point, and manually distinguish there are many limitations, such as time-consuming and laborious, limited accuracy, the shadow for being easily subject to subjective experience
Ring.
Convolutional neural networks (CNN) have the identification that precedent is applied to medical image before this, but CNN mainly should before this
For the identification to two dimensional image, used wave filter (convolution kernel) is all two-dimentional.Two-dimentional CNN compared to three-dimensional CNN,
Its maximum advantage is that operand is smaller, to the of less demanding of operational capability and storage capacity, but medical test skill at present
There are a large amount of 3 D medical images (such as CT, MRI) in art field, and then abandoned using two-dimentional CNN in vertical direction merely
Important information, has some impact on accuracy of identification.Although big using three-dimensional CNN operands, can preserve well vertical
The upward information of Nogata.
Therefore, it is necessary to partition system under a kind of new MRI cortex based on three-dimensional full convolutional neural networks is designed, with solution
Certainly above-mentioned technical problem.
The content of the invention
For problem present in background technology, the object of the present invention is to provide one kind based on three-dimensional full convolutional neural networks
MRI cortex under partition system, using three-dimensional full convolutional neural networks, by a large amount of head mri images by mark into
Row training, and the MRI image to newly importing completes subregion, so as to improve the accuracy and efficiency of subregion under cortex.
The technical proposal of the invention is realized in this way:A kind of lower point of MRI cortex based on three-dimensional full convolutional neural networks
Sound zone system, including training module, convolution pre-processing module, FCNN modules and convolution post-processing module, the training module and
Convolution pre-processing module is connected with FCNN modules, and FCNN modules are connected with convolution post-processing module, wherein, the training module
Calculated using back-propagation algorithm, the training set of input is trained, trained image first passes around forward-propagating approach, warp
Cross activation functional operation and obtain a value, then the weight of last layer is tried to achieve by the progress chain type derivation of this value, by changing for multilayer
Generation, chain type derivation obtain the weight of each layer, after completing training using training set, obtain a series of suitable for Subcortex point
The weight in area is to FCNN modules;The convolution pre-processing module:Comprising gray scale homogenization module, correction of inhomogeneous fields module and
Skull removes module, and pretreatment is peeled off for carrying out corresponding gray scale homogenization, correction of inhomogeneous fields and skull to image;Institute
State FCNN modules:For being identified to pretreated image line, by 9 layers of convolutional layer, 3 layers of full articulamentum and last point
After class layer, nine density profiles are obtained;The convolution post-processing module:For the place to isolating small subregion in nine density maps
Manage and nine density maps are merged, obtain a design sketch.
In the above-mentioned technical solutions, it is described before training module is trained, it is necessary to the weight of presetting each layer, to pre-
The weight of setting is calculated, and method is the weight for each layer, nlRepresent connection number of this layer for each unit, and
The weight of this layer is then set as that meeting variance isGaussian Profile a series of values.
In the above-mentioned technical solutions, the training module is before training, it is necessary to set multiple parameters, including training every time
Amount of images, iterations, whole the training set number, study momentum, the initial learning rate that are trained, according to the ginseng of setting
Number, training module are trained the training set of input.
In the above-mentioned technical solutions, the gray scale homogenization module is used to be distributed unbalanced place to image pixel intensity
Reason, makes the grey value profile of pixel between 0-255.
In the above-mentioned technical solutions, the correction of inhomogeneous fields module uses curved surface fitting method, to the pixel of picture
Classify, then extract the pixel for reflecting non homogen field variation tendency in image in all pixels point, try to achieve non homogen field
The parameter of surface fitting, is then fitted whole curved surface using these pixels, so as to complete to correct.
In the above-mentioned technical solutions, the skull removes module using edge detection operator detection skull edge, and utilizes
Automatic segmentation is realized at edge connection operator connection edge, is removed the skull in image, is reduced image size and calculation amount.
Locate under MRI cortex of the present invention based on three-dimensional full convolutional neural networks before partition system, including training module, convolution
Module, FCNN modules and convolution post-processing module are managed, the training module and convolution pre-processing module connect with FCNN modules
Connect, FCNN modules are connected with convolution post-processing module, and training module is used to be trained the training set of input, convolution pre-treatment
Module is used for the pretreatment that image is carried out gray scale homogenization, correction of inhomogeneous fields and skull and peeled off, and completes convolution pre-treatment
It is sent into FCNN modules and is identified afterwards, after 9 layers of convolutional layer, 3 layers of full articulamentum and last classification layer, obtains nine
Density profile, the convolution post processing by convolution post-processing module, completes the place that nine density profiles are isolated with small subregion
Reason, and nine density profiles are merged, obtain a design sketch, so as to improve the accuracy and effect of subregion under cortex
Rate.
Brief description of the drawings
Fig. 1 is structure diagram of the present invention;
Fig. 2 is FCNN structure charts of the present invention;
Fig. 3 is instantiation design sketch of the present invention.
Embodiment
Below in conjunction with the attached drawing in the embodiment of the present invention, the technical solution in the embodiment of the present invention is carried out clear, complete
Site preparation describes, it is clear that described embodiment is only part of the embodiment of the present invention, instead of all the embodiments.Based on this
Embodiment in invention, the every other reality that those of ordinary skill in the art are obtained without creative efforts
Example is applied, belongs to the scope of protection of the invention.
Partition system under a kind of MRI cortex based on three-dimensional full convolutional neural networks of the present invention, key point are
Three-dimensional full convolutional neural networks and its structure, the structure include three groups of convolutional layers, three layers of full articulamentum (being expressed as convolutional layer) and
One layer of classification layer, wherein, three component class layers respectively include three-layer coil lamination, and three layers in each group of convolutional layer all use same number
Convolution kernel.The system includes four modules, after being respectively training module, convolution pre-processing module, FCNN modules and convolution
Processing module, connection relation are the detailed description to above-mentioned each module below as shown in Fig. 1.
(1) training module:
Core used in training module is that backpropagation (BP) algorithm calculates.Before being trained, need first
Want the weight of presetting each layer.If initialize weight using changeless standard deviation, then calculate by backpropagation
The possible convergence of weight that training is drawn is not high.Therefore, it is necessary to calculate presetting weight, method is for each layer
Weight, nlConnection number of this layer for each unit is represented, and the weight in this layer is then set as that meeting variance isHeight
A series of values of this distribution.The weight so set convergence after training is higher.
Before training, user also needs to each and every one more parameters of setting, including batch size, i.e., the picture number trained every time
Amount;Iteration, iterations;Epoch, the number that whole training set is trained;Learn momentum;Initial learning rate etc..Root
According to the parameter of setting, training module is trained the training set of input.
Training process is mainly calculated using backpropagation.Trained image first passes around forward-propagating approach, by activation
Functional operation obtains a value, then the weight of last layer is tried to achieve by the progress chain type derivation of this value.By the iteration of multilayer, chain type
Derivation obtains the weight of each layer.After completing training using training set, just obtain a series of suitable for Subcortex subregion
Weight (i.e. wave filter).
(2) convolution pre-processing module:
Since different instrument or imaging system use different agreements or different parameters, the image of acquisition is in the presence of poor
It is different to use FCNN that image is identified, it is necessary to carry out pretreatment.In convolution pre-processing module, three are primarily present certainly
Module, is that gray scale homogenization module, correction of inhomogeneous fields module and skull remove module respectively.
A. gray scale homogenization module:There may be image pixel intensity to be distributed unbalanced situation for the image obtained, this
The image of sample there may be it is overall partially bright or overall partially dark the problem of, be not easy to subsequent treatment.Therefore, gray scale uniforms module just
Be the grey value profile for making pixel between 0-255, its method is:
Y=(x-MinValue) (MaxValue-MinValue)
Wherein, x is the value before normalization, and y is the value after normalization, and MaxValue is the maximum of gradation of image,
MinValue is the minimum value of gradation of image.
B. correction of inhomogeneous fields module:Non homogen field refers to that average, the variance of the different regional areas of identical tissue have
Larger deviation.The presence of non homogen field can reduce the accuracy of subregion, it is therefore desirable to which non homogen field is corrected.
Correction to non homogen field can use curved surface fitting method, it is necessary first to classifies to the pixel of picture,
Then the pixel of non homogen field variation tendency can be reflected by extracting in all pixels point in image, try to achieve non homogen field curved surface plan
The parameter of conjunction, is then fitted whole curved surface using these pixels, so as to complete to correct.
C. skull removes module:Skull belongs to extraneous areas in head mri image, is not acted as subregion under cortex
With image size can be reduced by removing skull, while reduce unnecessary calculation amount.The removal of skull depends on edge
The method of detection, mainly detects skull edge using edge detection operator, and is realized certainly using edge connection operator connection edge
Dynamic segmentation.
In skull removes module, main edge detection operator to be used is mainly Sobel operators, including two group 3 × 3
Matrix, it is made planar convolution with image can obtain brightness difference approximation.
(3) FCNN modules:
There are full articulamentum in traditional CNN neutral nets, for obtained characteristic pattern to be classified.And in full convolution
In neutral net, full articulamentum is expressed as convolutional layer, therefore all layers are all convolutional layers, because of referred to herein as full convolutional neural networks
(FCNN)。
In this FCNN neutral nets, there are three groups of convolutional layers, each group of convolutional layer after input layer three layers
Convolutional layer.In three layers of first group of convolutional layer, each layer has 25 characteristic patterns, corresponding to (equivalent to 25 filters of 25 groups of weights
Ripple device carries out convolution in last layer and obtains).And each layer has 50 characteristic layers in three layers of second group of convolutional layer;3rd group of volume
Each layer has 75 characteristic layers in three layers of lamination.
In this three groups of convolutional layers, used convolution kernel is all 3 × 3 × 3 convolution kernel.Seen from the above description, roll up
Lamination has had reached 9 layers, and 9 layers of convolutional layer can all cause the reduction of characteristic pattern resolution ratio after each layer is undergone, and lose one
A little local messages, the reduction of resolution ratio can make the accuracy of scoring area, and since convolution kernel is bigger, the reduction of resolution ratio is faster,
Therefore we use less convolutional layer in convolutional layer, so that the decline of resolution ratio slows down.And use less convolution
Core, by the Level Expand of FCNN to 9 layers so that the level of FCNN is deeper, can capture deeper feature.
It is three layers of full articulamentum for being expressed as convolutional layer after three groups of convolutional layers are passed through.Used in full articulamentum
Convolution kernel size be 1 × 1 × 1, by using 1 × 1 × 1 convolution kernel, full articulamentum can be converted to convolutional layer, and
And the spatial information in keeping characteristics figure, and cause full articulamentum to possess information in this layer of characteristic pattern of study to greatest extent
Ability.Three-layer coil lamination has 400,200 and 150 characteristic patterns respectively.
In FCNN, used activation primitive is ReLu functions, its expression formula is:
F (x)=max (0, x)
After three layers of full articulamentum, last layer is classification layer.The matrix diagram of full articulamentum is categorized as 9 by classification layer
Characteristic pattern, wherein eight correspond to different subcortical structures subregions, respectively left and right thalamus, lenticular nucleus, caudate nucleus and grey
Archon, remaining one is Background.9 characteristic patterns of classification layer generation are represented calculates different points by softmax functions
The probability distribution of area's homogenization, is actually probability distribution density figure.Probability calculation formula for the n-th class is:
Division result under cortex is finally obtained by this 9 density maps.
(4) convolution operation post-processing module:
In nine density maps obtained after FCNN, it is understood that there may be separated fritter isolated area.Judge fritter isolated area
The distance between main Density Distribution area relation, is given up if main Density Distribution area is significantly away from, if close to
Then it is merged with main Density Distribution area in main Density Distribution area.Calculating process by calculate two subregions between apart from area ratio
Realize.Wherein, shown in the structure chart 2 of FCNN.
To sum up, a three-dimensional head mri image file of input, system first carry out it convolution pre-treatment, including ash
Degree homogenization, correction of inhomogeneous fields and skull are peeled off.It is sent into FCNN and is identified after completion convolution pre-treatment, by 9 layer
After convolutional layer, 3 layers of full articulamentum and last classification layer, nine density profiles are obtained.Post-processed by convolution, completion pair
After isolating the processing of small subregion, nine density profiles are merged, obtain a design sketch, as shown under Fig. 3.
Compared with prior art, partition system under MRI cortex of the present invention based on three-dimensional full convolutional neural networks, have with
Lower beneficial effect:
1. in convolutional neural networks, often go after carrying out convolution every time and the resolution ratio of each column can all reduce convolution kernel size
Subtract 1, therefore used convolution kernel is bigger, the resolution ratio of characteristic pattern declines faster after convolution, the local feature of loss
Also it is more.And the convolution kernel used in the system used in FCNN be 3 × 3 × 3, than 7 used in general CNN ×
7 × 7 is much smaller, therefore still can keep higher resolution ratio even across the convolution of multilayer, so as to keep one in original image
A little fine features.And given up pooling layers of all max in FCNN, all full articulamentums are also all changed into rolling up
The convolutional layer that product core is 1 × 1 × 1, so as to save the space characteristics of characteristic pattern.
2. when carrying out head mri inspection, due to the presence of disturbing factor in inspection environment and instrument and equipment itself
Influence, the possible heterogeneity of gray scale in the image generated, and there are non-uniform field, be unfavorable for the subsequent treatment to image.Cause
This is provided with convolution pre-processing module in the present system, and intensity profile on the one hand is reset to 0-255 so as to complete to ash
The homogenization of degree, is on the other hand corrected the inhomogenous non-uniform field of regional by the method for surface fitting, so that
Improve the identifiability of image;Also the skull in image is peeled off at the same time, avoids having an impact subcortical structures subregion,
Reduce the information of redundancy.
3. flat image can only be identified in two dimension CNN, and at present the medical image of many types be all it is three-dimensional, such as
CT、MRI.If carry out image recognition using ordinary two dimensional CNN, then the letter in vertical direction in these 3-D views can be lost
Breath, so as to be had some impact on to the accuracy of image recognition.And used in the present system is three-dimensional FCNN, the volume that uses
Product core is three-dimensional dimension, therefore can efficiently use the information in 3-dimensional image in vertical direction, improves the utilization rate of information.Together
When in this kind of FCNN the convolution number of plies greatly increase to 9 layers, with the number of plies it is incremental can extraction feature also become increasingly complex,
It the deeper feature such as can catch.
The foregoing is merely illustrative of the preferred embodiments of the present invention, is not intended to limit the invention, all essences in the present invention
With within principle, any modification, equivalent replacement, improvement and so on, should all be included in the protection scope of the present invention god.
Claims (6)
- A kind of 1. partition system under MRI cortex based on three-dimensional full convolutional neural networks, it is characterised in that:Including training module, Convolution pre-processing module, FCNN modules and convolution post-processing module, the training module and convolution pre-processing module with FCNN modules connect, and FCNN modules are connected with convolution post-processing module, wherein,The training module:Calculated using backpropagation, the training set of input is trained, trained image first passes around just To route of transmission, a value is obtained by activation primitive computing, then the weight of last layer is tried to achieve by the progress chain type derivation of this value, By the iteration of multilayer, chain type derivation obtains the weight of each layer, after completing training using training set, obtains a series of be suitable for The weight of Subcortex subregion is to FCNN modules;The convolution pre-processing module:Module is removed comprising gray scale homogenization module, correction of inhomogeneous fields module and skull, is used Pretreatment is peeled off in carrying out corresponding gray scale homogenization, correction of inhomogeneous fields and skull to image;The FCNN modules:For to pretreated image line identify, by 9 layers of convolutional layer, 3 layers of full articulamentum and finally Classification layer after, obtain nine density profiles;The convolution post-processing module:For to isolating the processing of small subregion in nine density maps and nine density maps being melted Close, an obtained design sketch.
- 2. partition system under the MRI cortex according to claim 1 based on three-dimensional full convolutional neural networks, its feature exist In:, it is necessary to which the weight of presetting each layer, calculates presetting weight before training module is trained, method is For each layer of weight, nl represents connection number of this layer for each unit, and the weight in this layer is then set as the side of meeting Difference isGaussian Profile a series of values.
- 3. partition system under the MRI cortex according to claim 1 based on three-dimensional full convolutional neural networks, its feature exist In:The training module is including amount of images trained every time, iterations, whole before training, it is necessary to set multiple parameters Number that a training set is trained, study momentum, initial learning rate, according to the parameter of setting, instruction of the training module to input Practice collection to be trained.
- 4. partition system under the MRI cortex according to claim 1 based on three-dimensional full convolutional neural networks, its feature exist In:The gray scale homogenization module is used to be distributed unbalanced processing to image pixel intensity, the grey value profile of pixel is existed Between 0-255.
- 5. partition system under the MRI cortex according to claim 1 based on three-dimensional full convolutional neural networks, its feature exist In:The correction of inhomogeneous fields module uses curved surface fitting method, classifies to the pixel of picture, then extracts in image Reflect the pixel of non homogen field variation tendency in all pixels point, try to achieve the parameter of non homogen field surface fitting, then utilize These pixels are fitted whole curved surface, so as to complete to correct.
- 6. partition system under the MRI cortex according to claim 1 based on three-dimensional full convolutional neural networks, its feature exist In:The skull removes module using edge detection operator detection skull edge, and real using edge connection operator connection edge Now automatic segmentation, removes the skull in image, reduces image size and calculation amount.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201710620454.2A CN107909588A (en) | 2017-07-26 | 2017-07-26 | Partition system under MRI cortex based on three-dimensional full convolutional neural networks |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201710620454.2A CN107909588A (en) | 2017-07-26 | 2017-07-26 | Partition system under MRI cortex based on three-dimensional full convolutional neural networks |
Publications (1)
Publication Number | Publication Date |
---|---|
CN107909588A true CN107909588A (en) | 2018-04-13 |
Family
ID=61840203
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201710620454.2A Pending CN107909588A (en) | 2017-07-26 | 2017-07-26 | Partition system under MRI cortex based on three-dimensional full convolutional neural networks |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN107909588A (en) |
Cited By (7)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN109242849A (en) * | 2018-09-26 | 2019-01-18 | 上海联影智能医疗科技有限公司 | Medical image processing method, device, system and storage medium |
CN109222902A (en) * | 2018-08-27 | 2019-01-18 | 上海铱硙医疗科技有限公司 | Parkinson's prediction technique, system, storage medium and equipment based on nuclear magnetic resonance |
CN109509177A (en) * | 2018-10-22 | 2019-03-22 | 杭州依图医疗技术有限公司 | A kind of method and device of brain phantom identification |
CN109741339A (en) * | 2018-12-14 | 2019-05-10 | 强联智创(北京)科技有限公司 | A kind of partition method and system |
CN109872328A (en) * | 2019-01-25 | 2019-06-11 | 腾讯科技(深圳)有限公司 | A kind of brain image dividing method, device and storage medium |
CN110533639A (en) * | 2019-08-02 | 2019-12-03 | 杭州依图医疗技术有限公司 | A kind of key independent positioning method and device |
CN111462055A (en) * | 2020-03-19 | 2020-07-28 | 沈阳先进医疗设备技术孵化中心有限公司 | Skull detection method and device |
Citations (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN106650913A (en) * | 2016-12-31 | 2017-05-10 | 中国科学技术大学 | Deep convolution neural network-based traffic flow density estimation method |
WO2017096125A1 (en) * | 2015-12-02 | 2017-06-08 | The Cleveland Clinic Foundation | Automated lesion segmentation from mri images |
-
2017
- 2017-07-26 CN CN201710620454.2A patent/CN107909588A/en active Pending
Patent Citations (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
WO2017096125A1 (en) * | 2015-12-02 | 2017-06-08 | The Cleveland Clinic Foundation | Automated lesion segmentation from mri images |
CN106650913A (en) * | 2016-12-31 | 2017-05-10 | 中国科学技术大学 | Deep convolution neural network-based traffic flow density estimation method |
Non-Patent Citations (2)
Title |
---|
DOLZ J: ""3D fully convolutional networks for subcortical segmentation in MRI: A large-scale study"", 《ARXIV》 * |
林建忠: "《金融信息分析》", 30 April 2015 * |
Cited By (10)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN109222902A (en) * | 2018-08-27 | 2019-01-18 | 上海铱硙医疗科技有限公司 | Parkinson's prediction technique, system, storage medium and equipment based on nuclear magnetic resonance |
CN109242849A (en) * | 2018-09-26 | 2019-01-18 | 上海联影智能医疗科技有限公司 | Medical image processing method, device, system and storage medium |
CN109509177A (en) * | 2018-10-22 | 2019-03-22 | 杭州依图医疗技术有限公司 | A kind of method and device of brain phantom identification |
CN109741339A (en) * | 2018-12-14 | 2019-05-10 | 强联智创(北京)科技有限公司 | A kind of partition method and system |
CN109872328A (en) * | 2019-01-25 | 2019-06-11 | 腾讯科技(深圳)有限公司 | A kind of brain image dividing method, device and storage medium |
CN109872328B (en) * | 2019-01-25 | 2021-05-07 | 腾讯科技(深圳)有限公司 | Brain image segmentation method, device and storage medium |
US11748889B2 (en) | 2019-01-25 | 2023-09-05 | Tencent Technology (Shenzhen) Company Limited | Brain image segmentation method and apparatus, network device, and storage medium |
CN110533639A (en) * | 2019-08-02 | 2019-12-03 | 杭州依图医疗技术有限公司 | A kind of key independent positioning method and device |
CN111462055A (en) * | 2020-03-19 | 2020-07-28 | 沈阳先进医疗设备技术孵化中心有限公司 | Skull detection method and device |
CN111462055B (en) * | 2020-03-19 | 2024-03-08 | 东软医疗***股份有限公司 | Skull detection method and device |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN107909588A (en) | Partition system under MRI cortex based on three-dimensional full convolutional neural networks | |
Kollem et al. | A review of image denoising and segmentation methods based on medical images | |
CN108198184B (en) | Method and system for vessel segmentation in contrast images | |
CN105957063B (en) | CT image liver segmentation method and system based on multiple dimensioned weighting similarity measure | |
Gao et al. | Classification of CT brain images based on deep learning networks | |
CN109584251A (en) | A kind of tongue body image partition method based on single goal region segmentation | |
CN108053417B (en) | lung segmentation device of 3D U-Net network based on mixed rough segmentation characteristics | |
CN105069746B (en) | Video real-time face replacement method and its system based on local affine invariant and color transfer technology | |
CN109584254A (en) | A kind of heart left ventricle's dividing method based on the full convolutional neural networks of deep layer | |
CN108257135A (en) | The assistant diagnosis system of medical image features is understood based on deep learning method | |
CN107437092A (en) | The sorting algorithm of retina OCT image based on Three dimensional convolution neutral net | |
CN108805209A (en) | A kind of Lung neoplasm screening method based on deep learning | |
CN107169974A (en) | It is a kind of based on the image partition method for supervising full convolutional neural networks more | |
CN102136142B (en) | Nonrigid medical image registration method based on self-adapting triangular meshes | |
CN104268873B (en) | Breast tumor partition method based on nuclear magnetic resonance images | |
CN106934761A (en) | A kind of method for registering of three-dimensional non-rigid optical coherence tomographic image | |
CN109886965B (en) | Retina layer segmentation method and system combining level set with deep learning | |
CN107730507A (en) | A kind of lesion region automatic division method based on deep learning | |
CN110232389A (en) | A kind of stereoscopic vision air navigation aid based on green crop feature extraction invariance | |
CN107424145A (en) | The dividing method of nuclear magnetic resonance image based on three-dimensional full convolutional neural networks | |
CN108053398A (en) | A kind of melanoma automatic testing method of semi-supervised feature learning | |
CN107563434A (en) | A kind of brain MRI image sorting technique based on Three dimensional convolution neutral net, device | |
CN110533683A (en) | A kind of image group analysis method merging traditional characteristic and depth characteristic | |
CN108428225A (en) | Image department brain image fusion identification method based on multiple dimensioned multiple features | |
CN109902682A (en) | A kind of mammary gland x line image detection method based on residual error convolutional neural networks |
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 | ||
RJ01 | Rejection of invention patent application after publication |
Application publication date: 20180413 |
|
RJ01 | Rejection of invention patent application after publication |