CN116704184A - Left atrium and scar segmentation method based on Deep U-Net model - Google Patents
Left atrium and scar segmentation method based on Deep U-Net model Download PDFInfo
- Publication number
- CN116704184A CN116704184A CN202310596653.XA CN202310596653A CN116704184A CN 116704184 A CN116704184 A CN 116704184A CN 202310596653 A CN202310596653 A CN 202310596653A CN 116704184 A CN116704184 A CN 116704184A
- Authority
- CN
- China
- Prior art keywords
- left atrium
- training
- scar
- segmentation
- model
- 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
- 230000011218 segmentation Effects 0.000 title claims abstract description 161
- 210000005246 left atrium Anatomy 0.000 title claims abstract description 113
- 231100000241 scar Toxicity 0.000 title claims abstract description 110
- 238000000034 method Methods 0.000 title claims abstract description 70
- 238000012549 training Methods 0.000 claims abstract description 119
- 230000001746 atrial effect Effects 0.000 claims abstract description 41
- 238000012360 testing method Methods 0.000 claims abstract description 21
- 238000012545 processing Methods 0.000 claims abstract description 6
- 230000002708 enhancing effect Effects 0.000 claims abstract description 3
- 230000006870 function Effects 0.000 claims description 20
- 230000008569 process Effects 0.000 claims description 10
- 241000885184 Scaris Species 0.000 claims description 6
- 238000000137 annealing Methods 0.000 claims description 6
- 230000005489 elastic deformation Effects 0.000 claims description 6
- 238000007781 pre-processing Methods 0.000 claims description 6
- 238000013135 deep learning Methods 0.000 claims description 5
- 238000002790 cross-validation Methods 0.000 claims description 3
- 230000003628 erosive effect Effects 0.000 claims description 3
- 230000010354 integration Effects 0.000 claims description 3
- 238000010606 normalization Methods 0.000 claims description 3
- 238000012805 post-processing Methods 0.000 claims description 3
- 238000013519 translation Methods 0.000 claims description 3
- 238000003745 diagnosis Methods 0.000 abstract description 2
- 208000032544 Cicatrix Diseases 0.000 description 8
- 230000037387 scars Effects 0.000 description 8
- 206010003658 Atrial Fibrillation Diseases 0.000 description 7
- 230000000747 cardiac effect Effects 0.000 description 6
- 238000002595 magnetic resonance imaging Methods 0.000 description 4
- 238000007674 radiofrequency ablation Methods 0.000 description 4
- 229910052688 Gadolinium Inorganic materials 0.000 description 3
- UIWYJDYFSGRHKR-UHFFFAOYSA-N gadolinium atom Chemical compound [Gd] UIWYJDYFSGRHKR-UHFFFAOYSA-N 0.000 description 3
- 210000003492 pulmonary vein Anatomy 0.000 description 3
- 230000000007 visual effect Effects 0.000 description 3
- 208000003174 Brain Neoplasms Diseases 0.000 description 2
- 238000012733 comparative method Methods 0.000 description 2
- 230000003111 delayed effect Effects 0.000 description 2
- 230000000694 effects Effects 0.000 description 2
- 238000003384 imaging method Methods 0.000 description 2
- 230000001965 increasing effect Effects 0.000 description 2
- 238000013508 migration Methods 0.000 description 2
- 230000005012 migration Effects 0.000 description 2
- 238000004088 simulation Methods 0.000 description 2
- 230000009466 transformation Effects 0.000 description 2
- 206010051290 Central nervous system lesion Diseases 0.000 description 1
- 238000005481 NMR spectroscopy Methods 0.000 description 1
- 241000125258 Scandix pecten-veneris Species 0.000 description 1
- 230000032683 aging Effects 0.000 description 1
- 238000004458 analytical method Methods 0.000 description 1
- 238000013459 approach Methods 0.000 description 1
- 206010003119 arrhythmia Diseases 0.000 description 1
- 230000006793 arrhythmia Effects 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 238000010586 diagram Methods 0.000 description 1
- 208000037265 diseases, disorders, signs and symptoms Diseases 0.000 description 1
- 208000035475 disorder Diseases 0.000 description 1
- 238000002474 experimental method Methods 0.000 description 1
- 238000000605 extraction Methods 0.000 description 1
- 230000004927 fusion Effects 0.000 description 1
- 230000003862 health status Effects 0.000 description 1
- 238000010191 image analysis Methods 0.000 description 1
- 238000003709 image segmentation Methods 0.000 description 1
- 230000006872 improvement Effects 0.000 description 1
- 238000002955 isolation Methods 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 230000000750 progressive effect Effects 0.000 description 1
- 238000011002 quantification Methods 0.000 description 1
- 238000011160 research Methods 0.000 description 1
- 238000010200 validation analysis Methods 0.000 description 1
- 238000012795 verification Methods 0.000 description 1
Classifications
-
- 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
- G06V10/26—Segmentation of patterns in the image field; Cutting or merging of image elements to establish the pattern region, e.g. clustering-based techniques; Detection of occlusion
-
- 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
- G06N3/0464—Convolutional networks [CNN, ConvNet]
-
- 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
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
- G06V10/77—Processing image or video features in feature spaces; using data integration or data reduction, e.g. principal component analysis [PCA] or independent component analysis [ICA] or self-organising maps [SOM]; Blind source separation
- G06V10/774—Generating sets of training patterns; Bootstrap methods, e.g. bagging or boosting
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
- G06V10/82—Arrangements for image or video recognition or understanding using pattern recognition or machine learning using neural networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V2201/00—Indexing scheme relating to image or video recognition or understanding
- G06V2201/03—Recognition of patterns in medical or anatomical images
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
- Y02T—CLIMATE CHANGE MITIGATION TECHNOLOGIES RELATED TO TRANSPORTATION
- Y02T10/00—Road transport of goods or passengers
- Y02T10/10—Internal combustion engine [ICE] based vehicles
- Y02T10/40—Engine management systems
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Evolutionary Computation (AREA)
- Health & Medical Sciences (AREA)
- General Health & Medical Sciences (AREA)
- Artificial Intelligence (AREA)
- Computing Systems (AREA)
- Software Systems (AREA)
- Multimedia (AREA)
- Life Sciences & Earth Sciences (AREA)
- Medical Informatics (AREA)
- Databases & Information Systems (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Computational Linguistics (AREA)
- Data Mining & Analysis (AREA)
- Molecular Biology (AREA)
- General Engineering & Computer Science (AREA)
- Mathematical Physics (AREA)
- Image Analysis (AREA)
Abstract
The invention belongs to the technical field of medical image processing, and particularly relates to a left atrium and scar segmentation method based on a deep U-Net model. Comprising the following steps: step 1: enhancing the training data for training the left atrium segmentation model and the training data for training the left atrium scar segmentation model; step 2: training a left atrium segmentation model; step 3: preparing left atrium scar segmentation model data; step 4: training a left atrium scar segmentation model; step 5: left atrial scar testing and post-treatment. The invention adopts the weighting loss function and the regularization term based on distance to carry out constraint, thereby realizing more accurate segmentation of the left atrial scar. The invention adopts a two-stage segmentation method, on one hand, the image of the left atrium scar segmentation subject to the complex background is reduced, and on the other hand, the left atrium model obtained by training can also provide convenience for other medical diagnosis scenes needing to observe the left atrium. The invention can more effectively solve the problem of data domain offset.
Description
Technical Field
The invention belongs to the technical field of medical image processing, and particularly relates to a left atrium and scar segmentation method based on a Deep U-Net model.
Background
Atrial fibrillation (Atrial Fibrillation, AF), abbreviated as atrial fibrillation, is the most common arrhythmia in medical clinical practice, has a rate of occurrence up to 1%, and the rate of occurrence rises rapidly with the age of patients, and the number of patients suffering from atrial fibrillation increases rapidly with the gradual entry of China into an aging society.
The most common method for treating atrial fibrillation is a radio frequency ablation operation for pulmonary vein isolation, and the occurrence of atrial fibrillation is reduced by performing radio frequency ablation on the wall of a pulmonary vein to generate scars to block the current from the pulmonary vein from flowing into the left atrium and prevent the current disorder in the left atrium. The health status of the left atrium of the patient before and after the radio frequency ablation operation and the position and the number of scars generated on the left atrium after the radio frequency ablation test provide important information for diagnosing atrial fibrillation and evaluating the operation effect, so the method is important for the visual operation of the left atrium and the scars of the heart image.
Delayed gadolinium enhanced magnetic resonance imaging (Late Gadolinium Enhancement Magnetic Resonance Imaging, LGE MRI) technology is an effective technique for visualizing and quantifying atrial scars, and the use of computer-aided segmentation has long been an important research direction for medical image cardiac segmentation because of the poor quality of magnetic resonance delayed gadolinium enhanced cardiac medical images and the significant time and resources required to achieve accurate segmentation of the left atrium and scar manually. There are two main requirements for the method of segmenting the left atrium and scar of medical image, one is that the segmentation method must have high accuracy, and the other is that the segmentation method must have a strong domain generalization performance.
For left atrial scar segmentation work, the segmentation method requires precise segmentation of the scar on the blurred-boundary LGE MRI image. Because the scar targets on the left atrium are tiny and complex in distribution, the direct segmentation of the scar targets has high segmentation difficulty. GYang et al in the literature "GYang, J Chen, Z Gao, S Li, H Ni, EAngenli, and J Keegan, simultaneous left atrium anatomyand scar segmentations via deep learning in multiview information with attention Future GeneratiorComputer systems.107.215-228.2020" propose a multi-view, dual-task, recursive attention Unet for the simultaneous segmentation of the left atrium and the delineation of atrial scars, which method achieves simultaneous segmentation of cardiac structures and scars, but does not learn sufficiently the spatial relationship between the two, and therefore the scar segmentation accuracy is not high.
For the work of increasing the domain generalization performance of the segmentation method, the segmentation algorithm needs to ensure that the performance is kept stable when images of different data domains are segmented. Images generated by different cardiac imaging modes or different imaging devices belong to different data fields, and when a segmentation model obtained by training data from a single or a small number of data fields is used for segmenting data in an unknown data field, the segmentation performance of the model is often reduced.
Xu et al in document "Xu, qinwei, et al A fourier-based framework for domain generalization [ C ]. Proceedings ofthe IEEE/CVF Conference on ComputerVision and Pattern recording.2021" uses the feature that the phase of an image after fourier transformation can preserve advanced semantic information and is not susceptible to data migration, and adopts a data enhancement method based on fourier transformation to enhance data, so that a model learns more phase information. According to the method, although the domain generalization performance of the model is improved to a certain extent by transferring the image to the frequency domain for data enhancement, consideration of image space domain data enhancement is lacking, so that the information extraction capability of the trained model on the space domain image features is not strong enough, and cardiac segmentation tasks in the space domain cannot be well dealt with. The method for carrying out the cross-source domain example-level feature statistics based on probability is proposed in the documents of Zhou, K, yang, Y, qiao, Y, & Xiang, T.Domain general utilization MixStyle.In International Conference on Learning representational, and different types of data domains are mixed to obtain a new data domain, so that domain diversity of training data is increased, and domain generalization performance of a model obtained through training is improved. This approach focuses on style fusion between individual image fields, lacks enhancement to the individual data field images themselves, and thus does not adequately address the field offset problem.
The invention discloses a left atrium and scar segmentation method based on a Deep U-Net model, which mainly solves the limitations of the existing method that the segmentation accuracy of left atrium scars of the heart is not high and the adaptability to unknown domains is poor. Specifically, the invention solves the following technical problems existing in the prior art:
1. the existing heart segmentation method has low segmentation precision on the left atrium of the heart, and the segmented left atrium is not full in structure or has wrong segmentation.
2. The existing cardiac scar segmentation method is difficult to segment the tiny left atrium micro-scar with complex spatial distribution.
3. The existing heart segmentation method has poor image segmentation effect on unknown distribution, and is difficult to use for learning and migration.
Disclosure of Invention
The invention provides a medical nuclear magnetic resonance image left atrium and scar segmentation algorithm based on Deep U-Net, which uses a two-stage progressive segmentation mode of left atrium segmentation and left atrium scar segmentation to solve the limitations of the existing medical image left atrium and scar segmentation method.
In order to achieve the above purpose, the technical scheme adopted by the invention is as follows:
a left atrium and scar segmentation method based on Deep U-Net model, comprising:
step 1: enhancing a training data set for training the left atrium segmentation model and a training data set for training the left atrium scar segmentation model;
step 2: training a left atrium segmentation model by using the training data set of the left atrium segmentation model enhanced in the step 1;
step 3: dividing all data in the left atrium scar training data set enhanced in the step 1 by using the left atrium segmentation model obtained by training in the step 2, and then processing to obtain a training data set for training the left atrium scar segmentation model;
step 4: training the left atrium scar segmentation model by using the training data obtained in the step 3;
step 5: and (5) performing left atrial scar testing and post-processing by using the left atrial scar segmentation model trained in the step (4).
Preferably, the step 1 includes:
all data in the data set are subjected to the same preprocessing, the data set consists of four parts, namely a training data set D1 for training a left atrium segmentation model, a test data set D2 for testing the left atrium segmentation model, a training data set D3 for training a left atrium scar segmentation model and a test data set D4 for testing the left atrium scar segmentation model, and each data of the data set is a data sequence formed by a plurality of slices;
each image data in the training data sets D1, D2, D3 and D4 is subjected to the following preprocessing operations in order:
step 1.1: calculating a pixel histogram of each image data in the training data sets D1, D2, D3 and D4, clipping the pixel histogram, and reserving gray values with the frequency being more than 20;
step 1.2: carrying out pixel value normalization on each slice in each data sequence processed in the step 1.1;
step 1.3: on the basis of step 1.2, randomly flipping each slice in each image data in the training data sets D1 and D3 with a probability of 0.5;
step 1.4: on the basis of step 1.3, each slice in each data sequence is shifted (0.1) in both horizontal and vertical directions, respectively, by random translation;
step 1.5: on the basis of step 1.4, carrying out random scaling with a parameter (0.7,1.3) on each slice in each data sequence in both horizontal and vertical directions;
step 1.6: on the basis of the step 1.5, carrying out elastic deformation with the probability of 0.3 on each slice in each data sequence;
step 1.7: performing random contrast enhancement with gamma parameters ranging from (0.7,1.3) for each slice in each data sequence based on step 1.6;
step 1.8: on the basis of step 1.7, randomly rotating each slice in each data sequence with a probability of 0.5 within an interval (-30 °,30 °);
step 1.9: on the basis of step 1.8, adding a mean value of 0 to each slice in each data sequence with a probability of 0.15, the variance obeys (0,0.1) uniformly distributed Gaussian noise;
step 1.10: on the basis of step 1.9, each slice in each data sequence is subjected to elastic deformation and random motion with a probability of 0.3.
Preferably, the step 2 specifically includes:
deep U-Net is selected as a segmentation framework, hierarchical feature representation and symmetrical coding and decoding paths are adopted, a 6-layer convolution network structure is expanded, and the maximum channel number is set to 512 so as to weight a cross entropy loss function L weightedCE And the dice loss function L Dice Sum L of seg ＝L Dice +L weightedCE The training process is restrained as a total loss function of training, the optimizer is a random gradient descent optimizer SGD, the batch size is 16, the learning rate is firstly set to be 1e-4, a cosine annealing strategy is adopted, training is carried out by using all data of the left atrium segmentation training data set D1 preprocessed in the step 1, and the training process is terminated after 600 iterations, so that a model M for segmenting the left atrium of a heart image is obtained la 。
Preferably, the step 3 specifically includes:
left atrium segmentation model M trained by step 2 la Segmenting all data in the enhanced left atrium scar training data set D3 to obtain a label L corresponding to the left atrium la The left atrium label L is then obtained la Spelling corresponding original training data image in D3 according to channel dimensionThen, a training data set D5 for training the left atrial scar segmentation model is obtained for training the left atrial scar segmentation model.
Preferably, the step 4 specifically includes:
selecting Deep learning segmentation algorithm Deep U-Net as segmentation frame, training the left atrial scar segmentation model by using training data set D5 obtained in step 3, wherein the loss function in the training process is weighted dice lossWeighted cross entropy loss L weightedCE And distance lossThe sum of the three components:
L seg ＝L Dice +L weightedCE +0.1*L Distance the optimizer is a random gradient descent optimizer SGD, the batch size is 16, the learning rate is firstly set to be 1e-4, a cosine annealing strategy is adopted, the model is trained through a training method of 4-fold cross validation, the model is terminated after 600 iterations, and finally the left atrium scar segmentation model M is obtained through training scari ，i＝1,2,3,4。
Preferably, the step 5 specifically includes:
left atrium scar segmentation model M trained by step 4 scari All data in the scar segmentation test data set D4 are tested, and a segmentation label L of the left atrial scar is obtained through a model voting integration method scar And then performing post-treatment by using binary erosion and expansion operation with the nuclear radius of 6 to obtain a final left atrial scar segmentation result.
Compared with the prior art, the invention has the beneficial effects that:
1. the accuracy of the scar segmentation for the left atrium is higher. The invention adopts the weighting loss function and the regularization term based on distance to carry out constraint on the process of training the scar segmentation model, so that the model obtained by training can realize more accurate segmentation of the left atrial scar.
2. Has wider application scene. The invention adopts a two-stage segmentation method, namely, the left atrium segmentation is firstly carried out on the heart image, then the left atrium label obtained by segmentation is used as training data to train the left atrium scar segmentation model, on one hand, the image of the left atrium scar segmentation subject to complex background is reduced, and on the other hand, the left atrium model obtained by training can also provide convenience for other medical diagnosis scenes needing to observe the left atrium.
The domain generalization capability is stronger. The invention enhances model training data aiming at the segmentation characteristics of the left atrium scar, and obtains the segmentation model with excellent domain generalization performance by the enhanced data training, thereby being capable of more effectively solving the problem of data domain deviation.
Drawings
The accompanying drawings are included to provide a further understanding of the invention and are incorporated in and constitute a part of this specification, illustrate the invention and together with the embodiments of the invention, serve to explain the invention.
In the drawings:
FIG. 1 is a flow chart of the left atrium and scar segmentation of the present invention.
Fig. 2 is a 2-stage segmentation frame diagram of the left atrial scar segmentation of the present invention.
Fig. 3 visualizes the left atrial segmentation.
Detailed Description
The preferred embodiments of the present invention will be described below with reference to the accompanying drawings, it being understood that the preferred embodiments described herein are for illustration and explanation of the present invention only, and are not intended to limit the present invention.
Examples:
a left atrium and scar segmentation method based on Deep U-Net model, comprising:
step 1: the training data set for training the left atrium segmentation model and the training data set for training the left atrium scar segmentation model are enhanced. The method comprises the following steps:
all data in the data set are subjected to the same preprocessing, the data set consists of four parts, namely a training data set D1 for training a left atrium segmentation model, a test data set D2 for testing the left atrium segmentation model, a training data set D3 for training a left atrium scar segmentation model and a test data set D4 for testing the left atrium scar segmentation model, and each data of the data set is a data sequence formed by a plurality of slices;
each image data in the training data sets D1, D2, D3 and D4 is subjected to the following preprocessing operations in order:
step 1.1: calculating a pixel histogram of each image data in the training data sets D1, D2, D3 and D4, clipping the pixel histogram, and reserving gray values with the frequency being more than 20;
step 1.2: carrying out pixel value normalization on each slice in each data sequence processed in the step 1.1;
step 1.3: on the basis of step 1.2, randomly flipping each slice in each image data in the training data sets D1 and D3 with a probability of 0.5;
step 1.4: on the basis of step 1.3, each slice in each data sequence is shifted (0.1) in both horizontal and vertical directions, respectively, by random translation;
step 1.5: on the basis of step 1.4, carrying out random scaling with a parameter (0.7,1.3) on each slice in each data sequence in both horizontal and vertical directions;
step 1.6: on the basis of the step 1.5, carrying out elastic deformation with the probability of 0.3 on each slice in each data sequence;
step 1.7: performing random contrast enhancement with gamma parameters ranging from (0.7,1.3) for each slice in each data sequence based on step 1.6;
step 1.8: on the basis of step 1.7, randomly rotating each slice in each data sequence with a probability of 0.5 within an interval (-30 °,30 °);
step 1.9: on the basis of step 1.8, adding a mean value of 0 to each slice in each data sequence with a probability of 0.15, the variance obeys (0,0.1) uniformly distributed Gaussian noise;
step 1.10: on the basis of step 1.9, each slice in each data sequence is subjected to elastic deformation and random motion with a probability of 0.3.
Step 2: training the left atrium segmentation model using the training data set of the enhanced left atrium segmentation model of step 1. The method comprises the following steps:
deep learning segmentation method deep U-Net is selected as segmentation framework, hierarchical characteristic representation and symmetrical coding and decoding paths are adopted, 6-layer convolution network structure is covered and expanded, the maximum channel number is set to 512, and a weighted cross entropy loss function L is used weightedCE And the dice loss function L Dice Sum L of seg ＝L Dice +L weightedCE The training process is restrained as a total loss function of training, the optimizer is a random gradient descent optimizer SGD, the batch size is 16, the learning rate is firstly set to be 1e-4, a cosine annealing strategy is adopted, training is carried out by using all data of the left atrium segmentation training data set D1 preprocessed in the step 1, and the training process is terminated after 600 iterations, so that a model M for segmenting the left atrium of a heart image is obtained la 。
Step 3: and (3) segmenting all data in the left atrium scar training data set enhanced in the step (1) by using the left atrium segmentation model obtained by training in the step (2), and then processing to obtain a training data set for training the left atrium scar segmentation model. The method comprises the following steps:
left atrium segmentation model M trained by step 2 la Segmenting all data in the enhanced left atrium scar training data set D3 to obtain a label L corresponding to the left atrium la The left atrium label L is then obtained la And the corresponding original training data images in the step D3 are spliced according to the channel dimension to obtain a training data set D5 for training the left atrial scar segmentation model, and the training data set D5 is used for training the left atrial scar segmentation model.
Step 4: and (3) training the left atrial scar segmentation model by using the training data obtained in the step (3). The method comprises the following steps: selecting Deep learning segmentation algorithm Deep U-Net as segmentation frame, training the left atrial scar segmentation model by using training data set D5 obtained in step 3, wherein the loss function in the training process is weighted dice lossWeighted cross entropy loss L weightedCE And distance lossThe sum of the three components: l (L) seg ＝L Dice +L weightedCE +0.1*L Distance The optimizer is a random gradient descent optimizer SGD, the batch size is 16, the learning rate is firstly set to be 1e-4, a cosine annealing strategy is adopted, the model is trained through a training method of 4-fold cross validation, the model is terminated after 600 iterations, and finally the left atrium scar segmentation model M is obtained through training scari ，i＝1,2,3,4。
Step 5: and (5) performing left atrial scar testing and post-processing by using the left atrial scar segmentation model trained in the step (4). The method comprises the following steps:
left atrium scar segmentation model M trained by step 4 scari All data in the scar segmentation test data set D4 are tested, and a segmentation label L of the left atrial scar is obtained through a model voting integration method scar And then performing post-treatment by using binary erosion and expansion operation with the nuclear radius of 6 to obtain a final left atrial scar segmentation result.
Simulation analysis
1. Simulation conditions
The invention uses Pytorch framework to simulate on the central processing unit which is Intel (R) i9-10900X 3.7GHz CPU, internal memory 125G, two Nvidia GTX1080 GPU and Ubuntu20.04 operating systems.
The test data set used in the experiment was from the left atrium on scar segmentation challenge race (lascar qs 2022), which was developed by Li et al in literature "Li, l., zimmer, v.a., schnabel, j.a., zhuang, x. (2022). AtrialJSQnet: ANew framework forjoint segmentation and quantification of left atrium and scars incorporating spatial and shape information. Medical image analysis,76,102303," which contains 60 left atrial scar segmentation data and 130 left atrial segmentation data, from different medical devices of 3 schools and medical institutions, was developed for the accuracy and generalization performance study of the left atrial scar algorithm.
2. Emulation content
The invention was tested for accuracy and generalization performance of segmentation of the left atrium and scar of medical images.
In order to verify the improvement of the model domain generalization capability through data enhancement, a domain generalization method (Fourier rAbg) mentioned in Xu et al in the literature "Xu, qinwei, et al A four-based framework for domain generalization [ C ]. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recgntion.2021 ], and a method (Swou K., yang, Y., qiao, Y., & Xiang, T.Domain Generalization with MixStyle.In International Conference on Learning RepresentationA method (MixStyle) mentioned in the literature" xumizadeh A, nath V, tang Y, et al Swin transformers for semantic segmentation of brain tumors in mri images [ C ]. International MICCAI Brainlesion Workshop. Springer, m, 2022:272-284) were selected, and a distance between the font and the font (three-dimensional) was calculated as a comparison result by the left heart and contrast table (three-dimensional) as a rough index (visual index) shown in the literature "Hatamizadeh V, tan Y, et al, swin transformers for semantic segmentation of brain tumors in mri images [ C ]. 35 79 w-SPringer, m, 2022:272-284'. The left atrium segmentation results of the present invention and the comparison method are visualized, and a comparison chart of the visualized results is shown in fig. 3.
TABLE 1 left atrial segmentation model Performance validation results
To verify the 2-stage segmentation method of the present invention, a distance-weighted-based Dice loss function
TABLE 2 results of Performance verification of left atrial scar segmentation model
From a comparison of the results of left atrial segmentation performed by the inventive and comparative methods of Table 1, it can be seen that the inventive method is superior to the ForierAb method in terms of both the Dice score, average Surface Distance (ASD) and Hausdorff Distance (HD) Fourier indices. The method of domain generalization through data enhancement is illustrated to have more excellent and stable left atrium segmentation performance than the comparative method. The left atrial segmentation visual comparison results of the present invention and the comparison method in fig. 3 can more intuitively show that the left atrial structure segmentation result of the present invention is closer to the left atrium groudtruth, and no obvious erroneous segmentation occurs.
Table 2 second column shows the results of the single-stage left atrial scar segmentation method, third column shows the results of the segmentation using the distance-weighted Dice loss function and the single-stage left atrial scar segmentation method as the distance loss function of the regularization term, and fourth column shows the results of the segmentation using the distance-weighted Dice loss function and the two-stage left atrial scar segmentation method as the distance loss function of the regularization term. It can be seen that the overall performance of the two-stage left atrial scar segmentation method using a distance-weighted based Dice loss function and a distance loss function as a regularization term is superior to the comparison method.
The foregoing has shown and described the basic principles, principal features and advantages of the invention. It will be understood by those skilled in the art that the present invention is not limited to the embodiments described above, and that the above embodiments and descriptions are merely illustrative of the principles of the present invention, and various changes and modifications may be made without departing from the spirit and scope of the invention, which is defined in the appended claims. The scope of the invention is defined by the appended claims and equivalents thereof.
Claims (6)
1. A left atrium and scar segmentation method based on Deep U-Net model is characterized in that: comprising the following steps:
step 1: enhancing a training data set for training the left atrium segmentation model and a training data set for training the left atrium scar segmentation model;
step 2: training a left atrium segmentation model by using the training data set of the left atrium segmentation model enhanced in the step 1;
step 3: dividing all data in the left atrium scar training data set enhanced in the step 1 by using the left atrium segmentation model obtained by training in the step 2, and then processing to obtain a training data set for training the left atrium scar segmentation model;
step 4: training the left atrium scar segmentation model by using the training data obtained in the step 3;
step 5: and (5) performing left atrial scar testing and post-processing by using the left atrial scar segmentation model trained in the step (4).
2. The left atrium and scar segmentation method based on Deep U-Net model according to claim 1, wherein: the step 1 comprises the following steps:
all data in the data set are subjected to the same preprocessing, the data set consists of four parts, namely a training data set D1 for training a left atrium segmentation model, a test data set D2 for testing the left atrium segmentation model, a training data set D3 for training a left atrium scar segmentation model and a test data set D4 for testing the left atrium scar segmentation model, and each data of the data set is a data sequence formed by a plurality of slices;
each image data in the training data sets D1, D2, D3 and D4 is subjected to the following preprocessing operations in order:
step 1.1: calculating a pixel histogram of each image data in the training data sets D1, D2, D3 and D4, clipping the pixel histogram, and reserving gray values with the frequency being more than 20;
step 1.2: carrying out pixel value normalization on each slice in each data sequence processed in the step 1.1;
step 1.3: on the basis of step 1.2, randomly flipping each slice in each image data in the training data sets D1 and D3 with a probability of 0.5;
step 1.4: on the basis of step 1.3, each slice in each data sequence is shifted (0.1) in both horizontal and vertical directions, respectively, by random translation;
step 1.5: on the basis of step 1.4, carrying out random scaling with a parameter (0.7,1.3) on each slice in each data sequence in both horizontal and vertical directions;
step 1.6: on the basis of the step 1.5, carrying out elastic deformation with the probability of 0.3 on each slice in each data sequence;
step 1.7: performing random contrast enhancement with gamma parameters ranging from (0.7,1.3) for each slice in each data sequence based on step 1.6;
step 1.8: on the basis of step 1.7, randomly rotating each slice in each data sequence with a probability of 0.5 within an interval (-30 °,30 °);
step 1.9: on the basis of step 1.8, adding a mean value of 0 to each slice in each data sequence with a probability of 0.15, the variance obeys (0,0.1) uniformly distributed Gaussian noise;
step 1.10: on the basis of step 1.9, each slice in each data sequence is subjected to elastic deformation and random motion with a probability of 0.3.
3. The left atrium and scar segmentation method based on Deep U-Net model according to claim 2, wherein: the step 2 specifically comprises the following steps:
deep U-Net is selected as a segmentation framework, hierarchical feature representation and symmetrical coding and decoding paths are adopted, a 6-layer convolution network structure is expanded, and the maximum channel number is set to 512 so as to weight a cross entropy loss function L weightedCE And the dice loss function L Dice Sum L of seg ＝L Dice +L weightedCE The training process is restrained as a total loss function of training, the optimizer is a random gradient descent optimizer SGD, the batch size is 16, the learning rate is firstly set to be 1e-4, a cosine annealing strategy is adopted, training is carried out by using all data of the left atrium segmentation training data set D1 preprocessed in the step 1, and the training process is terminated after 600 iterations, so that a model M for segmenting the left atrium of a heart image is obtained la 。
4. The left atrium and scar segmentation method based on Deep U-Net model according to claim 3, wherein: the step 3 specifically comprises the following steps:
left atrium segmentation model M trained by step 2 la Segmenting all data in the enhanced left atrium scar training data set D3 to obtain a label L corresponding to the left atrium la The left atrium label L is then obtained la And the corresponding original training data images in the step D3 are spliced according to the channel dimension to obtain a training data set D5 for training the left atrial scar segmentation model, and the training data set D5 is used for training the left atrial scar segmentation model.
5. The method for segmenting the left atrium and the scar based on the Deep U-Net model according to claim 4, wherein the method comprises the following steps: the step 4 specifically comprises the following steps:
selecting Deep learning segmentation algorithm Deep U-Net as segmentation frame, and using training data set D5 obtained in step 3 to segment left atriumThe scar segmentation model is trained, and the loss function in the training process is weighted dice lossWeighted cross entropy loss L weightedCE And distance lossThe sum of the three components:
L seg ＝L Dice +L weightedCE +0.1*L Distance the optimizer is a random gradient descent optimizer SGD, the batch size is 16, the learning rate is firstly set to be 1e-4, a cosine annealing strategy is adopted, the model is trained through a training method of 4-fold cross validation, the model is terminated after 600 iterations, and finally the left atrium scar segmentation model M is obtained through training scari ，i＝1,2,3,4。
6. The method for segmenting the left atrium and the scar based on the Deep U-Net model according to claim 5, wherein the method comprises the following steps: the step 5 specifically comprises the following steps:
left atrium scar segmentation model M trained by step 4 scari All data in the scar segmentation test data set D4 are tested, and a segmentation label L of the left atrial scar is obtained through a model voting integration method scar And then performing post-treatment by using binary erosion and expansion operation with the nuclear radius of 6 to obtain a final left atrial scar segmentation result.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202310596653.XA CN116704184A (en) | 2023-05-25 | 2023-05-25 | Left atrium and scar segmentation method based on Deep U-Net model |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202310596653.XA CN116704184A (en) | 2023-05-25 | 2023-05-25 | Left atrium and scar segmentation method based on Deep U-Net model |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN116704184A true CN116704184A (en) | 2023-09-05 |
Family
ID=87833210
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202310596653.XA Pending CN116704184A (en) | 2023-05-25 | 2023-05-25 | Left atrium and scar segmentation method based on Deep U-Net model |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN116704184A (en) |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN117011315A (en) * | 2023-10-07 | 2023-11-07 | 中国人民解放军总医院第二医学中心 | LGE MRI left atrial scar segmentation system and method |
-
2023
- 2023-05-25 CN CN202310596653.XA patent/CN116704184A/en active Pending
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN117011315A (en) * | 2023-10-07 | 2023-11-07 | 中国人民解放军总医院第二医学中心 | LGE MRI left atrial scar segmentation system and method |
| CN117011315B (en) * | 2023-10-07 | 2023-12-29 | 中国人民解放军总医院第二医学中心 | LGE MRI left atrial scar segmentation system and method |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN109191476B (en) | Novel biomedical image automatic segmentation method based on U-net network structure | |
| WO2022121100A1 (en) | Darts network-based multi-modal medical image fusion method | |
| Chen et al. | MRI brain tissue classification using unsupervised optimized extenics-based methods | |
| CN114266939B (en) | Brain extraction method based on ResTLU-Net model | |
| CN113191968B (en) | Method for establishing three-dimensional ultrasonic image blind denoising model and application thereof | |
| CN113256749B (en) | Rapid magnetic resonance imaging reconstruction algorithm based on high-dimensional correlation prior information | |
| CN110827232A (en) | Cross-modal MRI (magnetic resonance imaging) synthesis method based on morphological feature GAN (gain) | |
| CN116704184A (en) | Left atrium and scar segmentation method based on Deep U-Net model | |
| CN111932512B (en) | Intracranial hemorrhage detection method applied to CT image based on CNN and NLSTM neural network | |
| CN117437420A (en) | Cross-modal medical image segmentation method and system | |
| CN118261822B (en) | Self-supervision magnetic resonance diffusion weighted image denoising method based on voxel replacement | |
| CN112489048A (en) | Deep network-based automatic optic nerve segmentation method | |
| Fan et al. | Magnetic-resonance image segmentation based on improved variable weight multi-resolution Markov random field in undecimated complex wavelet domain | |
| CN113689441B (en) | Left ventricle ultrasonic dynamic segmentation method based on DeepLabV network | |
| Joseph et al. | Optimised CNN based brain tumour detection and 3D reconstruction | |
| CN112164447B (en) | Image processing method, device, equipment and storage medium | |
| Singla et al. | Speckle and shadows: ultrasound-specific physics-based data augmentation for kidney segmentation | |
| Singla et al. | Speckle and shadows: ultrasound-specific physics-based data augmentation applied to kidney segmentation | |
| CN106709921B (en) | Color image segmentation method based on space Dirichlet mixed model | |
| CN115311135A (en) | 3 DCNN-based isotropic MRI resolution reconstruction method | |
| CN111402357A (en) | PET reconstruction method based on anatomical prior information | |
| Fang et al. | Combining multiple style transfer networks and transfer learning for lge-cmr segmentation | |
| Guo et al. | Semi‐supervised super‐resolution of diffusion‐weighted images based on multiple references | |
| Fan et al. | TR-TransGAN: Temporal Recurrent Transformer Generative Adversarial Network for Longitudinal MRI Dataset Expansion | |
| He et al. | Gating feature dense network for single anisotropic mr image super-resolution |
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 |