CN110163877A - A kind of method and system of MRI ventricular structure segmentation - Google Patents
A kind of method and system of MRI ventricular structure segmentation Download PDFInfo
- Publication number
- CN110163877A CN110163877A CN201910446810.2A CN201910446810A CN110163877A CN 110163877 A CN110163877 A CN 110163877A CN 201910446810 A CN201910446810 A CN 201910446810A CN 110163877 A CN110163877 A CN 110163877A
- Authority
- CN
- China
- Prior art keywords
- feature
- image
- mri
- layer
- carries out
- 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
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/25—Fusion techniques
- G06F18/253—Fusion techniques of extracted features
-
- 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
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/10—Segmentation; Edge detection
- G06T7/11—Region-based segmentation
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/20—Image preprocessing
- G06V10/32—Normalisation of the pattern dimensions
-
- 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
- G06T2207/30—Subject of image; Context of image processing
- G06T2207/30004—Biomedical image processing
- G06T2207/30048—Heart; Cardiac
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- General Engineering & Computer Science (AREA)
- Life Sciences & Earth Sciences (AREA)
- Artificial Intelligence (AREA)
- Evolutionary Computation (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Biophysics (AREA)
- Biomedical Technology (AREA)
- Computing Systems (AREA)
- General Health & Medical Sciences (AREA)
- Computational Linguistics (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Molecular Biology (AREA)
- Health & Medical Sciences (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Evolutionary Biology (AREA)
- Multimedia (AREA)
- Magnetic Resonance Imaging Apparatus (AREA)
Abstract
The present disclosure discloses a kind of method and systems of MRI ventricular structure segmentation, comprising: inputs cardiac MRI to be split;The cardiac MRI to be split of input is pre-processed;Pretreated cardiac MRI to be split is input to preparatory trained neural network model;Complete ventricular structure segmentation, output left ventricle, right ventricle and myocardium segmentation result.It is whether accurate that the resulting left ventricle of segmentation, right ventricle and cardiac muscle are evaluated with Huo Siduofu distance by Dice coefficient.Evaluation result shows that the method can obtain high-precision segmented image, and then can be improved diagnosis efficiency.
Description
Technical field
This disclosure relates to field of medical image processing, more particularly to a kind of method and system of MRI ventricular structure segmentation.
Background technique
The statement of this part is only to refer to background technique relevant to the disclosure, not necessarily constitutes the prior art.
In implementing the present disclosure, following technical problem exists in the prior art in inventor:
Heart is the important organ of human body, meanwhile, heart disease is also one of highest disease of lethality.For heart disease
The diagnosis and treatment of disease always are the problem of medical field.Before medical image means are applied in clinical diagnosis, doctor mainly passes through
The electrocardiogram of patient is measured, and the state of an illness of observation patient shows to diagnose to the state of an illness.This method usually will appear disease
Situations such as disease diagnoses not in time, Estimation About Patient's Condition mistake.And with the development of medical image technology, non-intruding can be used in doctor
Means obtain the image data of organ in patient body.Medical Imaging Technology relies on its safe ready, the injury to human body as a result,
Small advantage is widely applied in modern medicine diagnosis and treatment.So how to handle the medical image of acquisition, make it more
Medical worker is served well, the diagnosis for carrying out the state of an illness to patient is particularly important with estimation.
Magnetic resonance image (Magnetic Resonance Imaging, MRI) is the electromagnetic wave using certain frequency to putting
The substance being placed in magnetic field, which carries out excitation, makes its atom generate resonance, then converts this resonance signal by certain mathematical method
For a kind of technology of image output.Magnetic resonance imaging shows the different tissue of people's cylinder water content with different gray levels,
Have the following characteristics that (1) without ionising radiation, does not generate damage to human body.(2) good soft tissue contrast.(3) have and appoint
The ability of meaning planar imaging, this is that other scanning modes cannot accomplish.(4) blood vessel and blood flow imaging be can be carried out.It (5) can be with
It carries out multiple groups and knits parametric imaging.With the continuous development of mr techniques, including phased coil technology and ecg-gating technology
(Eeg-triggered imaging) etc., the image taking speed and image quality of MRI have obtained huge raising.With other imagings
Mode is good to the imaging effect of soft tissue organs in human body compared to MRI, this, which makes MRI gradually, becomes clinically cardiac function measurement
Goldstandard.
The ventricle of heart is muscle pump, mainly completes Ejection function, is played an important role in cardiac function.Meanwhile
Ventricle is also the region that lesion is easy in heart, therefore ventricular morphology and dyskinesia are considered as cardiovascular clinical diagnosis
Important evidence.To help patient to carry out the diagnosis of cardiovascular disease, doctor is dedicated to determining that the ventricle of patient holds according to cardiac MRI
The variation of product, myocardial wall thickness, ejection fraction and tube wall thickening property.For the above-mentioned cardiac function index of determination, doctor needs ventricle
The correct segmentation of cardiac muscle, therefore, left ventricle, right ventricle and being segmented in cardiac MRI for cardiac muscle have received widespread attention.
Currently, calculating clinical indices dependent on doctor accurately manual segmentation left and right ventricles and cardiac muscle, but manual segmentation
It is time-consuming and laborious, and different doctor's segmentation result is inconsistent, the same doctor may also in segmentation twice result it is different.Cause
This, the diagnosis that quickly accurate MRI ventricular structure automatic division method will help doctor preferably to carry out cardiovascular disease has
Great value and significance.
Currently, has the document divided automatically about ventricle on a small quantity.
For example, the method that heart CT is divided automatically is realized in existing patent proposition with V-Net.This method is based on 3D
Convolution divides 3D rendering end to end, uses Dice coefficient as objective function.Since this method is to 3D number
According to being divided automatically, so that there are calculation amounts is excessive, the disadvantages of parameter is excessive.And the method is for CT image, in the heart
CT image will be inferior to MRI on dirty this soft tissue imaging capabilities.
Existing patent proposes the method for dividing left ventricle with level set.It needs to handle image firstly, choosing, adopts later
Left ventriculography picture is pre-processed with optimization Mean Shift clustering algorithm.Then it uses and improves hough transform circle detection algorithm,
Left cardiac lumen is positioned, interior outer membrane segmentation initial profile is obtained, finally with outer membrane in double horizontal model segmented images, and with pair
Level Set Models energy function formula convergence inspection.By above step, obtain dividing interior outer membrane contours segmentation effect picture.This
Kind method solves the problems, such as that segmentation initial position is artificially arranged, but due to having used this method of level set, can exist by parameter
Influence excessive, the disadvantages of splitting speed is slower.
Existing document proposes that carrying out to divide automatically left ventricle by using region-growing method includes the isostructural blood pool of papillary muscle
Region.The method passes through the window that user intercepts cardiac image centre of slice dirty district domain by hand first, chooses in left room blood pool
Pixel calculates the mean value and variance of pixel in blood pool as seed point.This method can effectively be partitioned into left ventricle, but
It is to need manual selected seed point, is not full automatic dividing method.
In conclusion current ventricular segmentation algorithm is to be only limited to segmentation left ventricle, but also there is part and need mostly
The semi-automatic segmentation algorithm intervened manually.The rare algorithm that can divide left ventricle, right ventricle and cardiac muscle simultaneously is still directed to
CT image.
Summary of the invention
In order to solve the deficiencies in the prior art, present disclose provides a kind of method and systems of MRI ventricular structure segmentation;This
Kind method can divide left ventricle, right ventricle and cardiac muscle in MRI automatically.This method is highly effective to MRI segmentation, has realization
Simply, Medical Image Processing is enriched and developed to high accuracy for examination.
In a first aspect, present disclose provides MRI ventricular structure dividing methods;
MRI ventricular structure dividing method, comprising:
Input cardiac MRI to be split;
The cardiac MRI to be split of input is pre-processed;
Trained neural network model is called, pretreated cardiac MRI is split, completes ventricular structure point
It cuts, output left ventricle, right ventricle and myocardium segmentation result.
Second aspect, the disclosure additionally provide MRI ventricular structure segmenting system;
MRI ventricular structure segmenting system, comprising:
Input module is configured as input to cardiac MRI to be split;
Preprocessing module is configured as pre-processing the Cardiac Magnetic Resonance Images MRI to be split of input;
Divide module, is configured as calling trained neural network model, pretreated cardiac MRI is carried out
Ventricular structure segmentation, output left ventricle, right ventricle and myocardium segmentation result are completed in segmentation.
Compared with prior art, the beneficial effect of the disclosure is:
1 building neural network model, realization are divided end to end, can solve the low problem of current accuracy using this method,
Technical support is provided for computer-aided diagnosis.
2 can divide left ventricle, right ventricle and cardiac muscle simultaneously, provide convenience for clinical diagnosis.
3 are added to residual error connection in neural network, can effectively inhibit gradient disappearance problem.
4 on this kind of soft-tissue imaging of heart the imaging effect of MRI be better than CT.Therefore, the left heart is carried out for MRI
The full-automatic partition method of room, right ventricle and cardiac muscle can not only free doctor from the segmentation work of magnanimity, and
Objective, accurate ventricular segmentation result can be obtained.To the diagnosis efficiency for improving doctor, reduces misdiagnosis rate and be of great significance.
Detailed description of the invention
The accompanying drawings constituting a part of this application is used to provide further understanding of the present application, and the application's shows
Meaning property embodiment and its explanation are not constituted an undue limitation on the present application for explaining the application.
Fig. 1 is the method flow diagram of one embodiment;
Fig. 2 is neural network structure figure constructed by the present invention.
Specific embodiment
It is noted that described further below be all exemplary, it is intended to provide further instruction to the application.Unless another
It indicates, all technical and scientific terms used herein has usual with the application person of an ordinary skill in the technical field
The identical meanings of understanding.
It should be noted that term used herein above is merely to describe specific embodiment, and be not intended to restricted root
According to the illustrative embodiments of the application.As used herein, unless the context clearly indicates otherwise, otherwise singular
Also it is intended to include plural form, additionally, it should be understood that, when in the present specification using term "comprising" and/or " packet
Include " when, indicate existing characteristics, step, operation, device, component and/or their combination.
Embodiment one present embodiments provides MRI ventricular structure dividing method;
As shown in Figure 1, MRI ventricular structure dividing method, comprising:
Input cardiac MRI to be split;
The cardiac MRI to be split of input is pre-processed;
Trained neural network model is called, pretreated cardiac MRI is split, completes ventricular structure point
It cuts, output left ventricle, right ventricle and myocardium segmentation result.
As one or more embodiments, the pretreatment, comprising:
The cardiac MRI to be split of input is sliced frame by frame, by three-dimensional cardiac MRI to be split be converted to several two
Tie up image;
By the different all two dimensional images of size, size normalised processing, unified image size are carried out;
For size normalised treated image, data set is expanded by way of image rotation;
To the image after expansion, gray value standard processing is carried out respectively.
It is to be understood that the data that cardiac function index only needs systolic duration and diastolic interval are calculated in clinic, because
This segmentation is three-dimensional data with MRI data.It is that can reduce model instruction by the benefit that 3-D image is converted to two dimensional image processing
Parameter when practicing, training for promotion speed, improving operational speed.
As one or more embodiments, the cardiac MRI to be split of the three-dimensional is led in the environment of Python3.5
Enter NiBabel data packet, by the three-dimensional MRI data for the nifty format that NiBabel data packet acquires, later again to
The nifty formatted data of reading is extracted frame by frame, and then converts 2-D data for three-dimensional data.
It is to be understood that clinically MRI acquires equipment or the difference of acquisition protocols will lead to the size of MRI data not
One, the data of uniform sizes are inputted in the form of batch into neural network, can be with the training of accelerans network, training for promotion
Efficiency increases the wide usage of this patent.
The different all two dimensional images of size are subjected to size normalised processing as one or more embodiments,
Unified image size;Specific steps include:
It is discontented with the image of 256*256 for size, is filled with using its image pixel minimum value to 256*256.For
Size is greater than the image of 256*256, is cut to it so that its size is down to 256*256.
It is to be understood that the benefit expanded by the way of image rotation data set is that over-fitting can be effectively prevented
Phenomenon occurs.
It is to be understood that building neural metwork training parted pattern belongs to supervised learning, being for supervised learning label must not
It can lack.It is exactly Ground Truth in the label of image segmentation field supervised learning.The Ground Truth of medical image is needed
It wants the doctor of profession to mark manually, is often difficult to largely to obtain, this has also resulted in training data and has been not enough.In nerve
If training data is not enough in network model training, over-fitting will result in.The i.e. resulting model of training of over-fitting is to instruction
Practice data overfitting, is merely able to Accurate Segmentation training set, Accurate Segmentation cannot be accomplished for other data.
It is described for size normalised treated image as one or more embodiments, pass through the side of image rotation
Formula expands data set, and specific steps include:
By it is size normalised treated image rotates respectively 0 °, 15 °, 30 °, 45 °, 60 °, 75 °, 90 °, 105 ° and
120°。
It is to be understood that different MRI acquisition instruments and different acquisition protocols can make MRI data generation gray scale unbalanced
The case where, it generates the high or extremely low gray value of gray value and peels off phenomenon.To solve this problem, two-dimensional image data is carried out
Grey scale.To there is the Outlier Data beyond value range to have relatively good effect.
As one or more embodiments, gray value standardization is handled using Z-SCORE model.
As one or more embodiments, is handled, is specifically included using Z-SCORE model:
The mean value of image all pixels gray value is calculated, the standard deviation of image all pixels gray value is calculated, using each
The gray value of pixel subtracts mean value and obtains difference, the ash by obtained difference compared with standard deviation, after obtaining gray value standard
Angle value.
It is to be understood that the formula of Z-SCORE model are as follows:
Wherein, y indicates that the gray value Jing Guo grey scale, x indicate original gray value.μ indicates all gray values
Mean value, σ indicate the standard deviation of all gray values.By being standardized to two-dimensional image data, gray value is inhibited to peel off phenomenon.
As one or more embodiments, the neural network model, comprising:
Sequentially connected input layer, the first convolutional layer, the first pond layer, the second convolutional layer, the second pond layer, third convolution
Layer, third pond layer, Volume Four lamination, the 4th pond layer, the 5th convolutional layer, the 5th pond layer, the first warp lamination, second are instead
Convolutional layer, third warp lamination, the 4th warp lamination, the 5th warp lamination, the 6th warp lamination, the 6th convolutional layer and output
Layer.
Wherein, the input layer, the first convolutional layer, the first pond layer, the second convolutional layer, the second pond layer, third convolution
Layer, third pond layer, Volume Four lamination, the 4th pond layer, the 5th convolutional layer, the 5th pond layer are to adopt under neural network model
Sample channel;
Wherein, the first warp lamination, the second warp lamination, third warp lamination, the 4th warp lamination, the 5th warp
Lamination, the 6th warp lamination, the 6th convolutional layer and output layer are the up-sampling channel of neural network model.
As one or more embodiments, the working principle of the neural network model, comprising:
Input layer receives pretreated cardiac MRI to be split, and received image is sent into the first convolutional layer and is carried out
Process of convolution, then the output valve of the first convolutional layer is sent into the first pond layer and carries out pondization operation, obtains fisrt feature figure;
The output valve fisrt feature figure of first pondization operation is sent into the second convolutional layer and carries out process of convolution, then volume Two
The output valve of lamination is sent into the second pond layer and carries out pondization operation, obtains second feature figure;
The output valve second feature figure of second pondization operation is sent into third convolutional layer and carries out process of convolution, and then third is rolled up
The output valve of lamination is sent into third pond layer and carries out pondization operation, obtains third feature figure;
The output valve third feature figure of third pondization operation is sent into Volume Four lamination and carries out process of convolution, then Volume Four
The output valve of lamination is sent into the 4th pond layer and carries out pondization operation, obtains fourth feature figure;
The output valve fourth feature figure of 4th pondization operation, feeding the 5th convolutional layer progress process of convolution, then volume five
The output valve of lamination is sent into the 5th pond layer and carries out pondization operation, obtains fifth feature figure;
The output valve fifth feature figure of 5th pondization operation is sent into the first warp lamination and carries out deconvolution processing, to expand
Characteristic pattern size;The sixth feature figure that first warp lamination is handled and the fourth feature figure of the 4th pond layer output carry out
Fusion, obtains the first fusion feature figure;
Second warp lamination carries out deconvolution processing to the first warp lamination output valve, i.e., carries out to the first fusion feature figure
Deconvolution processing;The third feature figure for seventh feature figure and third pond the layer output that second warp lamination is handled carries out
Fusion, obtains the second fusion feature figure;
Third warp lamination carries out deconvolution processing to the second warp lamination output valve, i.e., carries out to the second fusion feature figure
Deconvolution processing, obtains eighth feature figure;
4th warp lamination carries out deconvolution processing to the fisrt feature figure that the first pond layer exports, and obtains ninth feature
Figure;
5th warp lamination carries out deconvolution processing to the second feature figure that the second pond layer exports, and obtains tenth feature
Figure;
6th warp lamination carries out deconvolution processing to the third feature figure that third pond layer exports, and obtains the 11st feature
Figure;
The characteristic pattern of eight, the nine, ten and 11 is subjected to cascade operation, obtains the 12nd characteristic pattern;
6th convolutional layer is carrying out convolution operation to the 12nd characteristic pattern, is realizing the classification to pixel, obtaining segmentation feature
Figure, is divided into left ventricle, four class of right ventricle, cardiac muscle and non-targeted object;
Argmax processing is being carried out to segmentation characteristic pattern, is obtaining final segmented image.
As one or more embodiments, the training step of the trained neural network model in advance, comprising:
Construct training set, the training set, comprising: several pretreated Cardiac Magnetic Resonance Images, each pre- place
Heart left ventricle, right ventricle and the myocardial boundary frame that Cardiac Magnetic Resonance Images after reason have medical image expert to mark manually
Pixel coordinate set;
Training set is input in neural network model, neural network model is trained, when neural network model
When loss function is restrained, deconditioning obtains trained neural network model.
As one or more embodiments, the working principle of the neural network model, comprising:
Step 5-1: it will read in by pretreated two dimension MRI, be inputted with batch size=5 quantity into institute's structure first
Among the modified form neural network model built.Image initially enters convolutional layer, carries out convolution operation.Convolution kernel size is 3 × 3,
Step-length is 1, padding=same.Pass through ReLU function again later.It is 256 that characteristic pattern by the output of this convolutional layer, which leads to dimension,
×256×64。
Step 5-2: it will be input among maximum pond layer by the characteristic pattern of convolutional layer, carry out pondization operation.Pondization step
A length of 2.Halve in this way by the characteristic pattern size of this pond layer, it is 128 × 128 × 64 that the characteristic pattern of output, which leads to dimension,.
Step 5-3: carrying out convolution operation, parameter such as step 5-1, the characteristic pattern of output lead to dimension be 128 × 128 ×
128。
Step 5-4: pondization operation is carried out, for parameter such as step 5-2, it is 64 × 64 × 128 that the characteristic pattern of output, which leads to dimension,.
Step 5-5: convolution operation is carried out, for parameter such as step 5-1, it is 64 × 64 × 256 that the characteristic pattern of output, which leads to dimension,.
Step 5-6: pondization operation is carried out, for parameter such as step 5-2, it is 32 × 32 × 256 that the characteristic pattern of output, which leads to dimension,.
Step 5-7: convolution operation is carried out, for parameter such as step 5-1, it is 32 × 32 × 512 that the characteristic pattern of output, which leads to dimension,.
Step 5-8: pondization operation is carried out, for parameter such as step 5-2, it is 16 × 16 × 512 that the characteristic pattern of output, which leads to dimension,.
Step 5-9: carrying out convolution operation, parameter such as step 5-1, the characteristic pattern of output lead to dimension be 16 × 16 ×
1024。
Step 5-10: pondization operation is carried out, for parameter such as step 5-2, it is 8 × 8 × 1024 that the characteristic pattern of output, which leads to dimension,.
Step 5-11: to step 5-10 output characteristic pattern carry out 2 times of deconvolution make its export characteristic pattern dimension 16 ×
16 × 512, it is big with the 4th layer of pond layer output characteristic pattern size of step 5-8 etc., then the spy that it is exported with the 4th layer of pond layer
Sign figure is merged in pixel scale, obtains new characteristic pattern.
Step 5-11: to step 5-11 output characteristic pattern carry out 2 times of deconvolution make its export characteristic pattern dimension 32 ×
32 × 256, it is big with step 5-6 third layer pond layer output characteristic pattern size etc..Again by the spy of itself and the output of third layer pond layer
Sign figure is merged in pixel scale, obtains new characteristic pattern.
Step 5-12: 8 times of deconvolution are carried out to the characteristic pattern of step 5-11 output, obtain 256 × 256 × 4 characteristic pattern.
Step 5-13: 2 times of deconvolution are carried out to the characteristic pattern of step 5-2 first layer pond layer output, obtain 256 × 256
× 4 characteristic pattern.
Step 5-14: 4 times of deconvolution are carried out to the characteristic pattern of step 5-4 second layer pond layer output, obtain 256 × 256
× 4 characteristic pattern.
Step 5-15: 8 times of deconvolution are carried out to the characteristic pattern of step 5-6 third layer pond layer output, obtain 256 × 256
× 4 characteristic pattern.
Step 5-16: by step 5-12, step 5-13, step 5-14, the characteristic pattern that step 5-15 is exported is at last
Dimension, that is, this dimension of channel are cascaded, and 256 × 256 × 12 new feature figure is obtained.Cascade in this step is
Residual error connection.It not only include neural network contextual information abundant by the new feature figure that residual error connects, but also can also be effective
Inhibit gradient disappearance problem.
Step 5-17: convolution operation is carried out again to the characteristic pattern of step 5-16 output, convolution kernel is having a size of 1 × 1, step-length
1, padding=same.This convolutional layer plays a pixel classifications.Finally obtain 256 × 256 × 4 segmentation characteristic pattern.
Step 5-18: resulting to step 5-17 segmentation characteristic pattern carries out argmax operation, by argmax, obtain
256 × 256 × 1 final segmentation figure.
Because cardiac MRI has intensity profile unevenness, characteristics of image is complicated, the problems such as artifact occurs, with traditional volume
Product neural network can generate situations such as parameter is excessive, and feature application is not enough.For MRI, this patent constructs a kind of improvement
Version neural network model is divided for ventricular structure, and this neural network model can divide left ventricle, right ventricle and the heart simultaneously
Flesh.
Convolutional layer is to carry out convolution operation to image, and its purpose is to extract image validity feature.Pond layer be in order to
The feature extracted to convolutional layer is concentrated, and extracts more effective feature, and reduce parameter, improves model training efficiency.
In this patent, we compare other pond modes, maximum pondization can be saved more effectively by the way of maximum pond
The texture information of image, in order to divide in high precision.
After convolutional layer, using ReLU as activation primitive.The main function of activation primitive is to provide the non-thread of network
Property modeling ability.If the network can only express Linear Mapping without activation primitive, even if there is again more hide at this time
Layer, whole network are also of equal value with monolayer neural networks.After only joined activation primitive, deep neural network just has
For the Nonlinear Mapping learning ability of layering.
The formula of ReLU are as follows:
ReLU=max (0, x)
Use ReLU as activation primitive, can contain gradient disappearance problem to a certain extent.Compared to other activation
Function, ReLU also have calculating speed fast, so that neural network restrains fast advantage, this is to the nerve of depth constructed by this patent
Network significance is outstanding.
After multilayer convolutional layer and pond layer, characteristic pattern can be smaller and smaller.By up-sampling channel enlarging property
Figure size can be realized the fusion of characteristic pattern, and then more fully utilize multilayer contextual information.
The deconvolution that 2 times are carried out to the characteristic pattern of layer 5 pond layer output obtains and the rulers such as the 4th layer of pond layer output
Very little characteristic pattern, is merged later.The deconvolution for carrying out 2 times to the characteristic pattern by fusion again later, with third layer pond
The output of layer is merged, then the deconvolution to merging resulting characteristic pattern for the second time and carrying out 8 times, can be obtained by this way with it is defeated
Enter the big characteristic pattern such as image.
Being possible to by the characteristic pattern in down-sampling channel can be as multilayer maximum pond filters useful feature.In order to
Characteristics of image is more fully utilized, this patent is added to residual error connection after up-sampling channel.Residual error connects so that nerve
Network can make full use of features at different levels during training, and is able to suppress gradient present in deep neural network and disappears
Mistake problem.The connection of residual error applied by this patent is to first layer, and the output of the second layer and third layer pond layer carries out deconvolution
Operation allows them to obtain and the big sizes such as input picture obtains later by the output cascade of they and up-sampling channel
New characteristic pattern.This characteristic pattern includes multi-layer image information.
Finally, carrying out 1 × 1 convolution to the output of residual error connection, this convolution behaviour is equivalent to pixel classifications.Later again by this
The output of convolutional layer carries out argmax operation, can obtain segmented image.Specific neural network structure figure is as shown in Figure 2.
The neural network structure application multiclass cross entropy of this patent institute framework is loss function, using Adam optimizer come excellent
Change loss function.
This patent is using average Dice coefficient and Huo Siduofu distance come evaluation test data segmentation effect.
The formula of Dice are as follows:
Wherein SRRepresent segmentation figure, SGTRepresent Ground Truth.This index primary evaluation pixel Duplication.
The formula of Huo Siduofu distance are as follows:
Wherein CRIndicate the profile of segmentation figure, CGTIndicate that Ground Truth profile, d (x, y) indicate the distance of point-to-point transmission.
Huo Siduofu distance indicates to be an o'clock longest distance from a profile to another profile point.Segmentation effect evaluation such as table 1
Shown, this neural network structure can get high-precision segmented image.
1 segmentation effect evaluation table of table
Embodiment two, the present embodiment additionally provide MRI ventricular structure segmenting system;
MRI ventricular structure segmenting system, comprising:
Input module is configured as input to cardiac MRI to be split;
Preprocessing module is configured as pre-processing the cardiac MRI to be split of input;
Divide module, is configured as calling trained neural network model, pretreated cardiac MRI is carried out
Ventricular structure segmentation, output left ventricle, right ventricle and myocardium segmentation result are completed in segmentation.
The foregoing is merely preferred embodiment of the present application, are not intended to limit this application, for the skill of this field
For art personnel, various changes and changes are possible in this application.Within the spirit and principles of this application, made any to repair
Change, equivalent replacement, improvement etc., should be included within the scope of protection of this application.
Claims (10)
1.MRI ventricular structure dividing method, characterized in that include:
Input cardiac MRI to be split;
The cardiac MRI to be split of input is pre-processed;
Trained neural network model in advance is called, is split to by the cardiac MRI with processing, ventricular structure point is completed
It cuts, output left ventricle, right ventricle and myocardium segmentation result.
2. the method as described in claim 1, characterized in that the pretreatment, comprising:
The cardiac MRI to be split of input is sliced frame by frame, three-dimensional cardiac MRI to be split is converted to several X-Y schemes
Picture;
By the different all two dimensional images of size, size normalised processing, unified image size are carried out;
For size normalised treated image, data set is expanded by way of image rotation;
To the image after expansion, gray value standard processing is carried out respectively.
3. method according to claim 2, characterized in that the cardiac MRI to be split of the three-dimensional is in Python3.5
NiBabel data packet is imported under environment, by the three-dimensional MRI data for the nifty format that NiBabel data packet acquires, it
The nifty formatted data read is extracted again frame by frame afterwards, and then converts 2-D data for three-dimensional data.
4. method according to claim 2, characterized in that by the different all two dimensional images of size, carry out size mark
Quasi-ization processing, unified image size;Specific steps include:
It is discontented with the image of 256*256 for size, is filled with using its image pixel minimum value to 256*256;For size
Image greater than 256*256 cuts it so that its size is down to 256*256.
5. method according to claim 2, characterized in that it is described for size normalised treated image, pass through image
The mode of rotation expands data set, and specific steps include:
Treated that image rotates 0 °, 15 °, 30 °, 45 °, 60 °, 75 °, 90 °, 105 ° and 120 ° respectively by size normalised.
6. method according to claim 2, characterized in that gray value standardization is handled using Z-SCORE model;Using
Z-SCORE model is handled, and is specifically included:
The mean value of image all pixels gray value is calculated, the standard deviation of image all pixels gray value is calculated, utilizes each pixel
Gray value subtract mean value and obtain difference, the gray value by obtained difference compared with standard deviation, after obtaining gray value standard.
7. the method as described in claim 1, characterized in that the neural network model, comprising:
Sequentially connected input layer, the first convolutional layer, the first pond layer, the second convolutional layer, the second pond layer, third convolutional layer,
Third pond layer, Volume Four lamination, the 4th pond layer, the 5th convolutional layer, the 5th pond layer, the first warp lamination, the second warp
Lamination, third warp lamination, the 4th warp lamination, the 5th warp lamination, the 6th warp lamination, the 6th convolutional layer and output layer.
8. the method for claim 7, characterized in that the working principle of the neural network model, comprising:
Input layer receives pretreated cardiac MRI to be split, and received image is sent into the first convolutional layer and carries out convolution
Processing, then the output valve of the first convolutional layer is sent into the first pond layer and carries out pondization operation, obtains fisrt feature figure;
The output valve fisrt feature figure of first pondization operation is sent into the second convolutional layer and carries out process of convolution, then the second convolutional layer
Output valve be sent into the second pond layer carry out pondization operate, obtain second feature figure;
The output valve second feature figure of second pondization operation is sent into third convolutional layer and carries out process of convolution, then third convolutional layer
Output valve be sent into third pond layer carry out pondization operate, obtain third feature figure;
The output valve third feature figure of third pondization operation is sent into Volume Four lamination and carries out process of convolution, then Volume Four lamination
Output valve be sent into the 4th pond layer carry out pondization operate, obtain fourth feature figure;
The output valve fourth feature figure of 4th pondization operation is sent into the 5th convolutional layer and carries out process of convolution, then the 5th convolutional layer
Output valve be sent into the 5th pond layer carry out pondization operate, obtain fifth feature figure;
The output valve fifth feature figure of 5th pondization operation is sent into the first warp lamination and carries out deconvolution processing, to expand feature
Figure size;The fourth feature figure that the sixth feature figure that first warp lamination is handled is exported with the 4th pond layer is melted
It closes, obtains the first fusion feature figure;
Second warp lamination carries out deconvolution processing to the first warp lamination output valve, i.e., carries out warp to the first fusion feature figure
Product processing;The third feature figure that the seventh feature figure that second warp lamination is handled is exported with third pond layer is melted
It closes, obtains the second fusion feature figure;
Third warp lamination carries out deconvolution processing to the second warp lamination output valve, i.e., carries out warp to the second fusion feature figure
Product processing, obtains eighth feature figure;
4th warp lamination carries out deconvolution processing to the fisrt feature figure that the first pond layer exports, and obtains ninth feature figure;
5th warp lamination carries out deconvolution processing to the second feature figure that the second pond layer exports, and obtains tenth feature figure;
6th warp lamination carries out deconvolution processing to the third feature figure that third pond layer exports, and obtains the 11st characteristic pattern;
The characteristic pattern of eight, the nine, ten and 11 is subjected to cascade operation, obtains the 12nd characteristic pattern;
6th convolutional layer is carrying out convolution operation to the 12nd characteristic pattern, is realizing the classification to pixel, obtaining segmentation characteristic pattern,
It is divided into left ventricle, four class of right ventricle, cardiac muscle and non-targeted object;
Argmax processing is being carried out to segmentation characteristic pattern, is obtaining final segmented image.
9. the method as described in claim 1, characterized in that the training step of the trained neural network model in advance,
Include:
Construct training set, the training set, comprising: several pretreated Cardiac Magnetic Resonance Images, after each pretreatment
The Cardiac Magnetic Resonance Images heart left ventricle, right ventricle and the myocardial boundary frame pixel that there is medical image expert to mark manually
Point coordinate set;
Training set is input in neural network model, neural network model is trained, when the loss of neural network model
When function convergence, deconditioning obtains trained neural network model.
10.MRI ventricular structure segmenting system, characterized in that include:
Input module is configured as input to cardiac MRI to be split;
Preprocessing module is configured as pre-processing the cardiac MRI to be split of input;
Divide module, be configured as calling trained neural network model in advance, to by being carried out with the cardiac MRI of processing
Ventricular structure segmentation, output left ventricle, right ventricle and myocardium segmentation result are completed in segmentation.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910446810.2A CN110163877A (en) | 2019-05-27 | 2019-05-27 | A kind of method and system of MRI ventricular structure segmentation |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910446810.2A CN110163877A (en) | 2019-05-27 | 2019-05-27 | A kind of method and system of MRI ventricular structure segmentation |
Publications (1)
Publication Number | Publication Date |
---|---|
CN110163877A true CN110163877A (en) | 2019-08-23 |
Family
ID=67629375
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201910446810.2A Pending CN110163877A (en) | 2019-05-27 | 2019-05-27 | A kind of method and system of MRI ventricular structure segmentation |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN110163877A (en) |
Cited By (7)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110807463A (en) * | 2019-09-17 | 2020-02-18 | 珠海格力电器股份有限公司 | Image segmentation method and device, computer equipment and storage medium |
CN111144486A (en) * | 2019-12-27 | 2020-05-12 | 电子科技大学 | Heart nuclear magnetic resonance image key point detection method based on convolutional neural network |
CN111275720A (en) * | 2020-01-20 | 2020-06-12 | 浙江大学 | Full end-to-end small organ image identification method based on deep learning |
CN111466894A (en) * | 2020-04-07 | 2020-07-31 | 上海尽星生物科技有限责任公司 | Ejection fraction calculation method and system based on deep learning |
CN112734770A (en) * | 2021-01-06 | 2021-04-30 | 中国人民解放军陆军军医大学第二附属医院 | Multi-sequence fusion segmentation method for cardiac nuclear magnetic images based on multilayer cascade |
WO2022052300A1 (en) * | 2020-09-08 | 2022-03-17 | 浙江大学山东工业技术研究院 | Electrocardio parameter calculation method based on deep learning |
CN114240964A (en) * | 2021-12-09 | 2022-03-25 | 电子科技大学 | Cardiac magnetic resonance image myocardial region segmentation method and system |
Citations (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US8781552B2 (en) * | 2011-10-12 | 2014-07-15 | Siemens Aktiengesellschaft | Localization of aorta and left atrium from magnetic resonance imaging |
CN108898606A (en) * | 2018-06-20 | 2018-11-27 | 中南民族大学 | Automatic division method, system, equipment and the storage medium of medical image |
-
2019
- 2019-05-27 CN CN201910446810.2A patent/CN110163877A/en active Pending
Patent Citations (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US8781552B2 (en) * | 2011-10-12 | 2014-07-15 | Siemens Aktiengesellschaft | Localization of aorta and left atrium from magnetic resonance imaging |
CN108898606A (en) * | 2018-06-20 | 2018-11-27 | 中南民族大学 | Automatic division method, system, equipment and the storage medium of medical image |
Non-Patent Citations (1)
Title |
---|
宋宝玉: "基于深度学习的心脏MRI双心室自动分割与心脏病分类", 《中国优秀硕士学位论文全文数据库》 * |
Cited By (12)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110807463A (en) * | 2019-09-17 | 2020-02-18 | 珠海格力电器股份有限公司 | Image segmentation method and device, computer equipment and storage medium |
CN111144486A (en) * | 2019-12-27 | 2020-05-12 | 电子科技大学 | Heart nuclear magnetic resonance image key point detection method based on convolutional neural network |
CN111144486B (en) * | 2019-12-27 | 2022-06-10 | 电子科技大学 | Heart nuclear magnetic resonance image key point detection method based on convolutional neural network |
CN111275720A (en) * | 2020-01-20 | 2020-06-12 | 浙江大学 | Full end-to-end small organ image identification method based on deep learning |
CN111275720B (en) * | 2020-01-20 | 2022-05-17 | 浙江大学 | Full end-to-end small organ image identification method based on deep learning |
CN111466894A (en) * | 2020-04-07 | 2020-07-31 | 上海尽星生物科技有限责任公司 | Ejection fraction calculation method and system based on deep learning |
CN111466894B (en) * | 2020-04-07 | 2023-03-31 | 上海深至信息科技有限公司 | Ejection fraction calculation method and system based on deep learning |
WO2022052300A1 (en) * | 2020-09-08 | 2022-03-17 | 浙江大学山东工业技术研究院 | Electrocardio parameter calculation method based on deep learning |
CN112734770A (en) * | 2021-01-06 | 2021-04-30 | 中国人民解放军陆军军医大学第二附属医院 | Multi-sequence fusion segmentation method for cardiac nuclear magnetic images based on multilayer cascade |
CN112734770B (en) * | 2021-01-06 | 2022-11-25 | 中国人民解放军陆军军医大学第二附属医院 | Multi-sequence fusion segmentation method for cardiac nuclear magnetic images based on multilayer cascade |
CN114240964A (en) * | 2021-12-09 | 2022-03-25 | 电子科技大学 | Cardiac magnetic resonance image myocardial region segmentation method and system |
CN114240964B (en) * | 2021-12-09 | 2023-04-07 | 电子科技大学 | Cardiac magnetic resonance image myocardial region segmentation method and system |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN110163877A (en) | A kind of method and system of MRI ventricular structure segmentation | |
Xu et al. | DW-Net: A cascaded convolutional neural network for apical four-chamber view segmentation in fetal echocardiography | |
Lombaert et al. | Human atlas of the cardiac fiber architecture: study on a healthy population | |
CN109598722B (en) | Image analysis method based on recurrent neural network | |
CN103781416B (en) | Use the comprehensive cardiovascular analysis of body phase-contrast MRI | |
CN110517238B (en) | AI three-dimensional reconstruction and human-computer interaction visualization network system for CT medical image | |
CN108257135A (en) | The assistant diagnosis system of medical image features is understood based on deep learning method | |
CN107590809A (en) | Lung dividing method and medical image system | |
CN105469419A (en) | Pulmonary nodule detection method based on CT image | |
CN110599499B (en) | MRI image heart structure segmentation method based on multipath convolutional neural network | |
Xu et al. | Convolutional-neural-network-based approach for segmentation of apical four-chamber view from fetal echocardiography | |
CN110363772B (en) | Heart MRI segmentation method and system based on antagonistic network | |
CN109902682A (en) | A kind of mammary gland x line image detection method based on residual error convolutional neural networks | |
Uccheddu et al. | 3D printing of cardiac structures from medical images: an overview of methods and interactive tools | |
Li et al. | Automatic quantification of epicardial adipose tissue volume | |
CN115830016A (en) | Medical image registration model training method and equipment | |
Yin et al. | Automatic breast tissue segmentation in MRIs with morphology snake and deep denoiser training via extended Stein’s unbiased risk estimator | |
CN104915989B (en) | Blood vessel three-dimensional dividing method based on CT images | |
CN111429404B (en) | Imaging system and method for detecting cardiovascular and cerebrovascular vessels | |
CN111986216B (en) | RSG liver CT image interactive segmentation algorithm based on neural network improvement | |
Smith et al. | Automated torso contour extraction from clinical cardiac MR slices for 3D torso reconstruction | |
CN109741439A (en) | A kind of three-dimensional rebuilding method of two dimension MRI fetus image | |
CN110570416A (en) | Method for visualization and 3D printing of multi-modal cardiac images | |
CN109741330A (en) | A kind of medical image cutting method of mixed filtering strategy and fuzzy C-mean algorithm | |
CN113205472A (en) | Cross-modal MR image mutual generation method based on cyclic generation countermeasure network cycleGAN model |
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 | ||
RJ01 | Rejection of invention patent application after publication |
Application publication date: 20190823 |