CN111951286A - Ventricular scar segmentation method based on U-shaped network - Google Patents
Ventricular scar segmentation method based on U-shaped network Download PDFInfo
- Publication number
- CN111951286A CN111951286A CN202010820645.5A CN202010820645A CN111951286A CN 111951286 A CN111951286 A CN 111951286A CN 202010820645 A CN202010820645 A CN 202010820645A CN 111951286 A CN111951286 A CN 111951286A
- Authority
- CN
- China
- Prior art keywords
- segmentation
- data
- scar
- training
- ventricular
- 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
- 231100000241 scar Toxicity 0.000 title claims abstract description 55
- 230000011218 segmentation Effects 0.000 title claims abstract description 53
- 230000002861 ventricular Effects 0.000 title claims abstract description 43
- 238000000034 method Methods 0.000 title claims abstract description 21
- 238000012549 training Methods 0.000 claims abstract description 37
- 238000012360 testing method Methods 0.000 claims abstract description 17
- 230000002107 myocardial effect Effects 0.000 claims abstract description 9
- 238000002474 experimental method Methods 0.000 claims abstract description 8
- 230000000694 effects Effects 0.000 claims abstract description 7
- 238000010606 normalization Methods 0.000 claims abstract description 7
- 238000007781 pre-processing Methods 0.000 claims abstract description 6
- 238000012545 processing Methods 0.000 claims abstract description 5
- 238000011156 evaluation Methods 0.000 claims abstract description 4
- 238000013135 deep learning Methods 0.000 claims description 4
- 230000006870 function Effects 0.000 claims description 3
- 238000012795 verification Methods 0.000 claims description 3
- 238000013528 artificial neural network Methods 0.000 description 6
- 238000003709 image segmentation Methods 0.000 description 6
- 206010000891 acute myocardial infarction Diseases 0.000 description 4
- 230000000747 cardiac effect Effects 0.000 description 4
- 238000013527 convolutional neural network Methods 0.000 description 4
- 230000009286 beneficial effect Effects 0.000 description 2
- 238000011160 research Methods 0.000 description 2
- 208000024172 Cardiovascular disease Diseases 0.000 description 1
- 206010019280 Heart failures Diseases 0.000 description 1
- 208000001871 Tachycardia Diseases 0.000 description 1
- 230000032683 aging Effects 0.000 description 1
- 206010003119 arrhythmia Diseases 0.000 description 1
- 230000006793 arrhythmia Effects 0.000 description 1
- 238000011161 development Methods 0.000 description 1
- 230000018109 developmental process Effects 0.000 description 1
- 238000003745 diagnosis Methods 0.000 description 1
- 235000006694 eating habits Nutrition 0.000 description 1
- 208000019622 heart disease Diseases 0.000 description 1
- 238000003062 neural network model Methods 0.000 description 1
- 230000033764 rhythmic process Effects 0.000 description 1
- 230000000630 rising effect Effects 0.000 description 1
- 230000006794 tachycardia Effects 0.000 description 1
- 238000012800 visualization Methods 0.000 description 1
Images
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
- G06T7/11—Region-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
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/20—Special algorithmic details
- G06T2207/20081—Training; Learning
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/20—Special algorithmic details
- G06T2207/20084—Artificial neural networks [ANN]
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/20—Special algorithmic details
- G06T2207/20112—Image segmentation details
- G06T2207/20132—Image cropping
-
- 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
Abstract
The ventricular scar segmentation method based on the U-shaped network sequentially comprises the following steps: (1) selecting a data set and a basic model: because the collection of ventricular scar data is difficult and the data is difficult to obtain in large quantity, an LGE-MRI data set is selected for an experiment, and the Unet network model has a good segmentation effect on medical images; (2) preprocessing the acquired experimental data set: firstly, reserving a myocardial part by using a mask image, then cutting according to the position and the size of a mask, filling the size of the mask to 160 multiplied by 160 according to a center filling mode, and finally performing normalization processing on the myocardial part; (3) data division: dividing the preprocessed image data into a training set and a test set; (4) the learning rate is subjected to polynomial attenuation in the training process; (5) evaluation indexes of ventricular scar segmentation; (6) improving an original network; (7) performing ventricular scar segmentation; the method provided by the invention improves the accuracy of ventricular scar segmentation, and is simple and convenient and easy to implement.
Description
Technical Field
The invention belongs to the technical field of medical image processing, and particularly relates to a ventricular scar segmentation method based on a U-shaped neural network model.
Background
Acute Myocardial Infarction (AMI) is a very common critical picture in clinic, social aging, faster and faster life rhythm, psychological problems caused by life pressure and dietary habits with higher oil and sugar which are changed along with the improvement of living standard, and the incidence rate of the acute myocardial infarction in China also shows a rising trend. Acute myocardial infarction is the major cause of many other heart diseases, such as arrhythmia, heart failure, etc. The ventricular scar segmentation result is important for cardiovascular disease diagnosis, and the tachycardia can be calculated by using the ventricular scar segmentation result. On the other hand, the myocardial morphology can be judged by using the segmentation result. Because ventricular scar boundaries are not obvious in cardiac magnetic resonance images, ventricular scar forms vary greatly from patient to patient and from slice to slice. This makes it difficult for the conventional medical image segmentation method to ensure good robustness under the condition of good segmentation effect, and the conventional medical image segmentation method is often performed with the help of experienced doctors, which is time-consuming and labor-consuming. In addition, the personal level of a doctor can also influence the experiment, so that it is very meaningful for the doctor and a patient to find a ventricular scar segmentation method with high segmentation precision and good robustness.
With the advent of large-scale labeled data and the development of computers, the realization of automatic segmentation of the heart using deep learning algorithms has become a hot spot of current research. Among them, Olaf Ronneberger proposed a U-Net model for medical image segmentation based on the Full Convolutional Network (FCN) model in 2015. Both U-Net and FCN have classical encoding-decoding topologies, but U-Net has a symmetric network structure and hopping connections, and the results of U-Net are superior to FCN in segmentation of cardiac images. Aiming at the problem that the heart image segmentation precision is difficult to improve, the improvement research performed by researchers on the basis of U-Net can be roughly divided into two types: improvement study based on 2D U-Net framework, improvement study based on 3D U-Net framework. Although the 3D network makes full use of the three-dimensional features of the MR images, the network is difficult to train from scratch due to the large number of parameters in the training network, and there is a problem in that the consumption of the GPU, the memory, and the like is excessive. Therefore, 2D networks are favored by researchers for ventricular scar segmentation tasks. The invention provides a ventricular scar segmentation method based on a U-shaped network, which is combined with the method and aims to improve the segmentation precision, and is mainly used for solving the problem of low precision of ventricular scar boundary segmentation.
Disclosure of Invention
The invention aims to provide a ventricular scar segmentation method based on a U-shaped network, which is high in segmentation accuracy.
In order to solve the technical problems, the invention provides the following technical scheme: the ventricular scar segmentation method based on the U-shaped network sequentially comprises the following steps:
(1) selecting a data set and a basic model: because the collection of ventricular scar data is difficult and the data is difficult to obtain in large quantity, an LGE-MRI data set is selected for an experiment, and the Unet network model has a good segmentation effect on medical images;
(2) preprocessing the acquired experimental data set: firstly, reserving a myocardial part by using a mask image, then cutting according to the position and the size of a mask, filling the size of the mask to 160 multiplied by 160 according to a center filling mode, and finally performing normalization processing on the myocardial part;
(3) data division: dividing the preprocessed image data into a training set and a test set;
(4) carrying out maximum and minimum normalization pretreatment on the pixel values of the acquired experimental data according to a formula (1);
where x represents the pixel value at a location, min and max represent the minimum and maximum values of the pixel value in the image, resulting in x*Is in the range of 0 to 1;
(5) the learning rate is subjected to polynomial attenuation in the training process, the attenuation formula is shown as (2),
wherein base _ lr is 0.01, iter is the current iteration number, max _ iter is the maximum iteration number, and power is 0.5;
(6) evaluation indexes of ventricular scar segmentation;
1) dice score (Dice): wherein A represents the region enclosed by the segmentation result, and M represents the region enclosed by the expert label;
2) accuracy (Accuracy): wherein TP represents true positive, meaning pixels correctly classified as scar; TN indicates true negative, meaning pixels correctly classified as background; FP represents false positive, meaning pixels misclassified as scar; FN indicates false negative, meaning pixels misclassified as background;
3) jaccard: wherein A represents the scar region in the algorithm segmentation result, and B represents the scar region in the mark;
(7) ventricular scar segmentation was performed as follows:
1) a training stage: in the training stage, the DFU-Net network provided by the invention is subjected to parameter training. The Keras deep learning framework is used for learning model parameters of the training set, the RMSprop algorithm is adopted, the initial learning rate (lr) is set to be 0.001, polynomial attenuation is carried out by using a formula (2-8), and the attenuation rate of the moving average (rho) of the square of the gradient is set to be 0.9. Training 500 epochs, and selecting a weight coefficient when the verification set has minimum loss as a final training weight of a training stage by using a Dice loss function in an experiment to provide a weight parameter for a subsequent testing stage;
the main innovation of the DFU-Net network provided by the invention is as follows:
a. introducing a dense connection network: dense connected networks may enhance the propagation of features in neural networks. The features in the underlying neural network are better multiplexed in the upper layer, and the number of parameters is also reduced. Meanwhile, in order to enlarge the receptive field, the extended convolution is used for improving the receptive field of the neural network.
b. Introduction of CRF structure: the conditional random field can overcome the problem of lack of edge constraint brought by a convolutional neural network, the solving process of the mean algorithm of the Dense CRF is changed into an RNN form, and the RNN form is combined with the CNN to form an end-to-end network, so that a better segmentation effect is finally achieved.
2) And (3) a testing stage: the data set contains 15 data, 11 of the ventricular scar data sets are randomly selected as training sets, and 4 of the ventricular scar data sets are selected as testing sets. The image sizes are 480 × 480 and 512 × 512, respectively, the number of slices varies from 10 to 50, and the format is NRRD. And performing ventricular scar segmentation on the test image by using the final weight parameters obtained in the training stage to finally obtain a segmented ventricular scar result graph.
Through the technical scheme, the invention has the beneficial effects that: an improved DFU-Net cardiac image segmentation algorithm is provided. By utilizing the extended convolution, the receptive field is expanded, the resolution is kept unchanged, the depth of the network is improved, a dense connection network and a conditional random field are introduced, the parameter quantity is reduced, and finally a trainable end-to-end segmentation algorithm is formed. Compared with the classic U-Net segmentation algorithm, the algorithm has a finer structure, can effectively solve the problems of over-segmentation and under-segmentation of ventricular scar segmentation, and is high in segmentation result precision.
Drawings
FIG. 1 is a flow chart of data preprocessing according to the present invention.
FIG. 2 is a flow chart of the present invention.
Detailed Description
As shown in fig. 1, a ventricular scar segmentation method based on a U-shaped network sequentially includes the following steps:
(1) selecting a data set and a basic model: because the collection of ventricular scar data is difficult and the data is difficult to obtain in large quantity, an LGE-MRI data set is selected for an experiment, and the Unet network model has a good segmentation effect on medical images;
(2) preprocessing the acquired experimental data set: firstly, reserving a myocardial part by using a mask image, then cutting according to the position and the size of a mask, filling the size of the mask to 160 multiplied by 160 according to a center filling mode, and finally performing normalization processing on the myocardial part;
(3) data division: the preprocessed image data is divided into a training set and a testing set, so that the image data is conveniently used for training and testing;
(4) preprocessing the acquired experimental data according to a formula (1), and using maximum and minimum normalization on the original DICOM image to enable the gray scale range to fall between 0 and 255 for visualization of results;
where x represents the pixel value at a location, min and max represent the minimum and maximum values of the pixel value in the image, resulting in x*Is in the range of 0 to 1.
(5) The learning rate is subjected to polynomial attenuation in the training process, the attenuation formula is shown as (2),
wherein base _ lr is 0.01, iter is the current iteration number, max _ iter is the maximum iteration number, and power is 0.5;
(6) evaluation indexes of ventricular scar segmentation;
1) dice score (Dice): wherein A represents the region enclosed by the segmentation result, and M represents the region enclosed by the expert label;
2) accuracy (Accuracy): wherein TP represents true positive, meaning pixels correctly classified as scar; TN indicates true negative, meaning pixels correctly classified as background; FP represents false positive, meaning pixels misclassified as scar; FN indicates false negative, meaning pixels of the background that were misclassified Wie;
3) jaccard: wherein A represents the scar region in the algorithm segmentation result, and B represents the scar region in the mark;
(7) ventricular scar segmentation was performed as follows:
1) a training stage: in the training stage, the DFU-Net network provided by the invention is subjected to parameter training. The Keras deep learning framework is used for learning model parameters of the training set, the RMSprop algorithm is adopted, the initial learning rate (lr) is set to be 0.001, polynomial attenuation is carried out by using a formula (2-8), and the attenuation rate of the moving average (rho) of the square of the gradient is set to be 0.9. Training 500 epochs, and selecting a weight coefficient when the verification set has minimum loss as a final training weight of a training stage by using a Dice loss function in an experiment to provide a weight parameter for a subsequent testing stage;
the main innovation of the DFU-Net network provided by the invention is as follows:
c. introducing a dense connection network: dense connected networks may enhance the propagation of features in neural networks. The features in the underlying neural network are better multiplexed in the upper layer, and the number of parameters is also reduced. Meanwhile, in order to enlarge the receptive field, the extended convolution is used for improving the receptive field of the neural network.
d. Introduction of CRF structure: the conditional random field can overcome the problem of lack of edge constraint brought by a convolutional neural network, the solving process of the mean algorithm of the Dense CRF is changed into an RNN form, and the RNN form is combined with the CNN to form an end-to-end network, so that a better segmentation effect is finally achieved;
2) and (3) a testing stage: the data set contains 15 data, 11 of the ventricular scar data sets are randomly selected as training sets, and 4 of the ventricular scar data sets are selected as testing sets. The image sizes are 480 × 480 and 512 × 512, respectively, the number of slices varies from 10 to 50, and the format is NRRD. And performing ventricular scar segmentation on the test image by using the final weight parameters obtained in the training stage to finally obtain a segmented ventricular scar result graph.
Through the technical scheme, the invention has the beneficial effects that: an improved DFU-Net cardiac image segmentation algorithm is provided. By utilizing the extended convolution, the receptive field is expanded, the resolution is kept unchanged, the depth of the network is improved, a dense connection network and a conditional random field are introduced, the parameter quantity is reduced, and finally a trainable end-to-end segmentation algorithm is formed. Compared with the classic U-Net segmentation algorithm, the algorithm has a finer structure, can effectively solve the problems of over-segmentation and under-segmentation of ventricular scar segmentation, and is high in segmentation result precision.
Claims (1)
1. The ventricular scar segmentation method based on the U-shaped network is characterized by comprising the following steps: the method sequentially comprises the following steps:
(1) selecting a data set and a basic model: because the collection of ventricular scar data is difficult and the data is difficult to obtain in large quantity, an LGE-MRI data set is selected for an experiment, and the Unet network model has a good segmentation effect on medical images;
(2) preprocessing the acquired experimental data set: firstly, reserving a myocardial part by using a mask image, then cutting according to the position and the size of a mask, filling the size of the mask to 160 multiplied by 160 according to a center filling mode, and finally performing normalization processing on the myocardial part;
(3) data division: dividing the preprocessed image data into a training set and a test set;
(4) carrying out maximum and minimum normalization pretreatment on the pixel values of the acquired experimental data according to a formula (1);
where x denotes the pixel value at a certain position, min and max denote graphsMinimum and maximum values of pixel values in a pixel, resulting in x*Is in the range of 0 to 1;
(5) the learning rate is subjected to polynomial attenuation in the training process, the attenuation formula is shown as (2),
wherein base _ lr is 0.01, iter is the current iteration number, max _ iter is the maximum iteration number, and power is 0.5;
(6) evaluation indexes of ventricular scar segmentation;
1) dice score (Dice): wherein A represents the region enclosed by the segmentation result, and M represents the region enclosed by the expert label;
2) accuracy (Accuracy): wherein TP represents true positive, meaning pixels correctly classified as scar; TN indicates true negative, meaning pixels correctly classified as background; FP represents false positive, meaning pixels misclassified as scar; FN indicates false negative, meaning pixels misclassified as background;
3) jaccard: wherein A represents the scar region in the algorithm segmentation result, and B represents the scar region in the mark;
(7) ventricular scar segmentation was performed as follows:
1) a training stage: in the training stage, the DFU-Net network provided by the invention is subjected to parameter training; learning model parameters of a training set by using a Keras deep learning framework, setting an initial learning rate (lr) to be 0.001 by using an RMSprop algorithm, performing polynomial attenuation by using a formula (2), and setting an attenuation rate of a moving average (rho) of gradient squares to be 0.9; training 500 epochs, and selecting a weight coefficient when the verification set has minimum loss as a final training weight of a training stage by using a Dice loss function in an experiment to provide a weight parameter for a subsequent testing stage;
2) and (3) a testing stage: the data set comprises 15 data, 11 ventricular scar data sets are randomly selected as training sets, and 4 ventricular scar data sets are selected as testing sets; the sizes of the images are respectively 480 multiplied by 480 and 512 multiplied by 512, the number of slices is different from 10 to 50, and the format is NRRD; and performing ventricular scar segmentation on the test image by using the final weight parameters obtained in the training stage to finally obtain a segmented ventricular scar result graph.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202010820645.5A CN111951286A (en) | 2020-08-14 | 2020-08-14 | Ventricular scar segmentation method based on U-shaped network |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202010820645.5A CN111951286A (en) | 2020-08-14 | 2020-08-14 | Ventricular scar segmentation method based on U-shaped network |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN111951286A true CN111951286A (en) | 2020-11-17 |
Family
ID=73342430
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202010820645.5A Pending CN111951286A (en) | 2020-08-14 | 2020-08-14 | Ventricular scar segmentation method based on U-shaped network |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN111951286A (en) |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN114972165A (en) * | 2022-03-24 | 2022-08-30 | 中山大学孙逸仙纪念医院 | Method and device for measuring time-average shearing force |
-
2020
- 2020-08-14 CN CN202010820645.5A patent/CN111951286A/en active Pending
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN114972165A (en) * | 2022-03-24 | 2022-08-30 | 中山大学孙逸仙纪念医院 | Method and device for measuring time-average shearing force |
| CN114972165B (en) * | 2022-03-24 | 2024-03-15 | 中山大学孙逸仙纪念医院 | Method and device for measuring time average shearing force |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN109325942B (en) | Fundus image structure segmentation method based on full convolution neural network | |
| CN107230206B (en) | Multi-mode data-based 3D pulmonary nodule segmentation method for hyper-voxel sequence lung image | |
| CN101763644B (en) | Pulmonary nodule three-dimensional segmentation and feature extraction method and system thereof | |
| CN111340819A (en) | Image segmentation method, device and storage medium | |
| CN106097305B (en) | The intratracheal tree dividing method that two-pass region growing combining form is rebuild | |
| CN109727253A (en) | Divide the aided detection method of Lung neoplasm automatically based on depth convolutional neural networks | |
| CN102496156A (en) | Medical image segmentation method based on quantum-behaved particle swarm cooperative optimization | |
| CN111179237A (en) | Image segmentation method and device for liver and liver tumor | |
| KR20120041468A (en) | System for detection of interstitial lung diseases and method therefor | |
| CN112819747A (en) | Method for automatically diagnosing benign and malignant nodules based on lung tomography image | |
| CN110070531A (en) | For detecting the model training method of eyeground picture, the detection method and device of eyeground picture | |
| CN115082493A (en) | 3D (three-dimensional) atrial image segmentation method and system based on shape-guided dual consistency | |
| CN112163570A (en) | SVM (support vector machine) electrocardiosignal identification method based on improved Husky algorithm optimization | |
| CN113487568B (en) | Liver surface smoothness measuring method based on differential curvature | |
| CN111951286A (en) | Ventricular scar segmentation method based on U-shaped network | |
| CN114692698A (en) | One-dimensional electrocardiogram data classification method based on residual error network | |
| CN110288581A (en) | A kind of dividing method based on holding shape convexity Level Set Models | |
| WO2020007026A1 (en) | Segmentation model training method and apparatus, and computer-readable storage medium | |
| CN112950611A (en) | Liver blood vessel segmentation method based on CT image | |
| CN112634308A (en) | Nasopharyngeal carcinoma target area and endangered organ delineation method based on different receptive fields | |
| CN112634291A (en) | Automatic burn wound area segmentation method based on neural network | |
| CN111784690A (en) | Heart right ventricle segmentation method based on U-shaped neural network model | |
| CN110610502A (en) | Automatic aortic arch region positioning and segmentation method based on CT image | |
| CN105956587A (en) | Method for automatically extracting meniscus from knee-joint magnetic resonance image sequence based on shape constraint | |
| CN116310525A (en) | Pathological image classification method based on contrast representation distillation and output distillation |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| WD01 | Invention patent application deemed withdrawn after publication |
Application publication date: 20201117 |
|
| WD01 | Invention patent application deemed withdrawn after publication |