A left atrium and scar segmentation method based on Deep U-Net model

Abstract

The present invention belongs to the field of medical image processing technology, and specifically relates to a left atrium and scar segmentation method based on the DeepU-Net model. It comprises: 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 the left atrium segmentation model; step 3: preparing the left atrium scar segmentation model data; step 4: training the left atrium scar segmentation model; step 5: left atrium scar testing and post-processing. The present invention adopts a weighted loss function and a distance-based regularization term for constraints, which can achieve more accurate segmentation of the left atrial scar. The present invention adopts a two-stage segmentation method, which on the one hand reduces the influence of complex background images on the segmentation of the left atrial scar, and on the other hand, the trained left atrium model can also provide convenience for other medical diagnosis scenarios that require observation of the left atrium. The present invention can more effectively deal with the problem of data domain offset.

Description

Technical Field

[0001] The present invention belongs to the technical field of medical image processing, and in particular relates to a left atrium and scar segmentation method based on a Deep U-Net model. Background Art

[0002] Atrial fibrillation (AF) is the most common arrhythmia in clinical medical practice, with an incidence rate of up to 1%. The incidence rate increases rapidly with the patient's age. As my country gradually enters an aging society, the number of AF patients is increasing rapidly.

[0003] The most common treatment for atrial fibrillation is radiofrequency ablation (RFA) with pulmonary vein isolation. This procedure, which creates scars by ablating the pulmonary vein walls, blocks the flow of electrical current from the pulmonary veins into the left atrium, preventing electrical current disturbances within the left atrium and thus reducing the incidence of atrial fibrillation. The patient's left atrial health before and after RFA, as well as the location and number of scars left atrial after the RFA trial, provide crucial information for diagnosing atrial fibrillation and evaluating surgical outcomes. Therefore, visualization of the left atrium and scars on cardiac imaging is crucial.

[0004] Late Gadolinium Enhancement Magnetic Resonance Imaging (LGE MRI) is an effective technique for visualizing and quantifying atrial scars. However, due to the poor quality of LGE cardiac images, manually achieving accurate segmentation of the left atrium and scar requires significant time and resources. Therefore, computer-assisted segmentation has been a key research area in cardiac imaging. There are two main requirements for left atrium and scar segmentation methods: high accuracy and strong domain generalization.

[0005] For left atrial scar segmentation, the segmentation method needs to accurately segment the scar on the LGE MRI image with blurred boundaries. Since the scar target on the left atrium is small and complexly distributed, it is very difficult to directly segment it. G Yang et al. proposed a multi-view dual-task recursive attention Unet in the paper "G Yang, J Chen, Z Gao, S Li, H Ni, E Angelini, and J Keegan, Simultaneous left atrium anatomy and scar segmentations via deep learning in multiview information with attention Future GeneratiorComputerSystems.107.215-228.2020". It can depict the atrial scar while segmenting the left atrium. This method achieves the simultaneous segmentation of cardiac structure and scar, but fails to fully learn the spatial relationship between the two, so the scar segmentation accuracy is not high.

[0006] To improve the domain generalization performance of segmentation methods, segmentation algorithms must maintain stable performance when segmenting images from different data domains. Different cardiac imaging modalities or images generated by different imaging devices belong to different data domains. When using a segmentation model trained on data from a single or small number of data domains to segment data from an unknown data domain, the model's segmentation performance often degrades.

[0007] In the paper "Xu, Qinwei, et al. A fourier-based framework for domain generalization [C]. Proceedings of the IEEE / CVF Conference on Computer Vision and Pattern Recognition. 2021", Xu et al. used the characteristics of the phase after Fourier transform of the image to preserve high-level semantic information and not be easily affected by data offset, and adopted a data enhancement method based on Fourier transform to enhance the data, so that the model can learn more phase information. Although this method improves the domain generalization performance of the model to a certain extent by transferring the image to the frequency domain for data enhancement, it lacks consideration of the image spatial domain data enhancement. Therefore, the trained model is not strong enough in extracting information about spatial domain image features, and thus fails to cope well with the task of heart segmentation in the spatial domain. In the paper "Zhou, K., Yang, Y., Qiao, Y., & Xiang, T. Domain Generalization with MixStyle. In International Conference on Learning Representations," Zhou et al. proposed a method for probabilistically mixing instance-level feature statistics across source domains. This method generates a new data domain by mixing data domains of different styles, increasing the domain diversity of the training data and thus improving the domain generalization performance of the trained model. This method focuses on style fusion between image domains and lacks enhancement of the images in a single data domain, thus failing to fully address the domain shift problem.

[0008] This paper discloses a method for left atrium and scar segmentation based on a Deep U-Net model. This method addresses the limitations of existing methods, such as low accuracy for left atrial scar segmentation and poor adaptability to unknown domains. Specifically, the present invention addresses the following technical issues in the prior art:

[0009] 1. Existing heart segmentation methods have low segmentation accuracy for the left atrium of the heart, and the segmented left atrium structure is incomplete or erroneous segmentation occurs.

[0010] 2. Existing cardiac scar segmentation methods are difficult to segment tiny left atrial scars that are small and have complex spatial distribution.

[0011] 3. Existing heart segmentation methods have poor segmentation effects on images with unknown distributions and are difficult to use for learning transfer. Summary of the Invention

[0012] The present invention proposes a Deep U-Net-based left atrium and scar segmentation algorithm for medical magnetic resonance imaging. It uses a two-stage progressive segmentation method of left atrium segmentation and left atrial scar segmentation to overcome the limitations of existing medical image left atrium and scar segmentation methods.

[0013] In order to achieve the above object, the technical solution adopted by the present invention is as follows:

[0014] A left atrium and scar segmentation method based on a Deep U-Net model, comprising:

[0015] Step 1: Enhance the training data set for training the left atrial segmentation model and the training data set for training the left atrial scar segmentation model;

[0016] Step 2: Use the training data set of the left atrium segmentation model enhanced in step 1 to train the left atrium segmentation model;

[0017] Step 3: Use the left atrial segmentation model trained in step 2 to segment all data in the left atrial scar training dataset enhanced in step 1, and then process it to obtain a training dataset for training the left atrial scar segmentation model;

[0018] Step 4: Use the training data obtained in step 3 to train the left atrial scar segmentation model;

[0019] Step 5: Use the left atrial scar segmentation model trained in step 4 to perform left atrial scar testing and post-processing.

[0020] Preferably, the step 1 comprises:

[0021] All data in the dataset are preprocessed in the same way. The dataset consists of four parts: training dataset D1 for left atrium segmentation model training, testing dataset D2 for left atrium segmentation model testing, training dataset D3 for left atrial scar segmentation model training, and testing dataset D4 for left atrial scar segmentation model testing. Each data in the dataset is a data sequence consisting of multiple slices.

[0022] Each image data in the training datasets D1, D2, D3 and D4 undergoes the following preprocessing operations in sequence:

[0023] Step 1.1: Calculate the pixel histogram of each image data in the training data sets D1, D2, D3 and D4, crop the pixel histogram, and retain the grayscale values with a frequency greater than 20;

[0024] Step 1.2: Normalize the pixel values of each slice in each data sequence processed in step 1.1;

[0025] Step 1.3: Based on step 1.2, each slice in each image data in the training datasets D1 and D3 is randomly flipped with a probability of 0.5;

[0026] Step 1.4: Based on step 1.3, each slice in each data sequence is randomly shifted in both the horizontal and vertical directions by an offset of (0.1, 0.1);

[0027] Step 1.5: Based on step 1.4, each slice in each data sequence is randomly scaled in both horizontal and vertical directions with parameters of (0.7, 1.3);

[0028] Step 1.6: Based on step 1.5, perform elastic deformation on each slice in each data sequence with a probability of 0.3;

[0029] Step 1.7: Based on step 1.6, perform random contrast enhancement with a gamma parameter in the range of (0.7, 1.3) on each slice in each data sequence;

[0030] Step 1.8: Based on step 1.7, each slice in each data sequence is randomly rotated within the interval of (-30°, 30°) with a probability of 0.5;

[0031] Step 1.9: Based on step 1.8, add Gaussian noise with mean 0 and variance (0, 0.1) to each slice in each data sequence with a probability of 0.15;

[0032] Step 1.10: Based on step 1.9, perform elastic deformation and random motion on each slice in each data sequence with a probability of 0.3.

[0033] Preferably, the step 2 is specifically as follows:

[0034] The deep learning segmentation method Deep U-Net is selected as the segmentation framework, using hierarchical feature representation and symmetrical encoding and decoding paths, including a 6-layer convolutional network structure, with the maximum number of channels set to 512, and a weighted cross entropy loss function L weightedCE and dice loss function L Dice and L seg =L Dice +L weightedCE The total loss function is used to constrain the training process. The optimizer is the stochastic gradient descent optimizer SGD, the batch size is 16, the learning rate is first set to 1e-4, and the cosine annealing strategy is adopted. All data of the left atrium segmentation training dataset D1 preprocessed in step 1 are used for training. After 600 iterations, the model M for segmenting the left atrium of cardiac images is obtained.la .

[0035] Preferably, the step 3 is specifically as follows:

[0036] The left atrium segmentation model M obtained by training in step 2 la Segment all data in the enhanced left atrial scar training dataset D3 to obtain the label L corresponding to the left atrium la , and then the obtained left atrium label L la The corresponding original training data images in D3 are spliced according to the channel dimension to obtain the training data set D5 for training the left atrial scar segmentation model, which is used for training the left atrial scar segmentation model.

[0037] Preferably, the step 4 is specifically as follows:

[0038] The deep learning segmentation algorithm Deep U-Net is selected as the segmentation framework. The left atrial scar segmentation model is trained using the training dataset D5 obtained in step 3. The loss function of the training process is weighted dice loss. Weighted cross entropy loss L weightedCE and distance loss The sum of the three: L seg =L Dice +L weightedCE +0.1*L Distance The optimizer is the stochastic gradient descent optimizer SGD, the batch size is 16, the learning rate is first set to 1e-4, the cosine annealing strategy is adopted, and the model is trained by the 4-fold cross-validation training method. It is terminated after 600 iterations. Finally, the left atrial scar segmentation model M is obtained through training. scari , i=1,2,3,4.

[0039] Preferably, the step 5 is specifically as follows:

[0040] The left atrial scar segmentation model M obtained by training in step 4 scari All data in the scar segmentation test dataset D4 are tested, and the segmentation label L of the left atrial scar is obtained by the model voting integration method. scar ,Then the binary erosion and dilation operations with a kernel radius of 6 are used for post-processing to obtain the final left atrial scar segmentation result.

[0041] Compared with the prior art, the present invention has the following beneficial effects:

[0042] 1. Higher accuracy for left atrial scar segmentation. The present invention uses a weighted loss function and a distance-based regularization term to constrain the scar segmentation model training process, so that the trained model can achieve more accurate segmentation of left atrial scars.

[0043] 2. Broader application scenarios. The present invention adopts a two-stage segmentation method, first segmenting the left atrium in cardiac images, and then using the segmented left atrium labels as training data to train the left atrial scar segmentation model. This not only reduces the complexity of left atrial scar segmentation due to complex background images, but also facilitates the application of the trained left atrial model in other medical diagnostic scenarios requiring observation of the left atrium.

[0044] The present invention enhances the model training data based on the characteristics of left atrial scar segmentation, and uses the enhanced data to train a segmentation model with excellent domain generalization performance, which can more effectively address the problem of data domain offset. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] The accompanying drawings are used to provide further understanding of the present invention and constitute a part of the specification. They are used to explain the present invention together with the embodiments of the present invention and do not constitute a limitation of the present invention.

[0046] In the attached figure:

[0047] Figure 1 This is a flow chart of the left atrium and scar segmentation of the present invention.

[0048] Figure 2 This is a two-stage segmentation framework diagram for left atrial scar segmentation in the present invention.

[0049] Figure 3 Left atrium segmentation visualization results. DETAILED DESCRIPTION

[0050] The preferred embodiments of the present invention are described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are only used to illustrate and explain the present invention, and are not used to limit the present invention.

[0051] Example:

[0052] A left atrium and scar segmentation method based on a Deep U-Net model, comprising:

[0053] Step 1: Enhance the training dataset for the left atrial segmentation model and the training dataset for the left atrial scar segmentation model. Specifically:

[0054] All data in the dataset are preprocessed in the same way. The dataset consists of four parts: training dataset D1 for left atrium segmentation model training, testing dataset D2 for left atrium segmentation model testing, training dataset D3 for left atrial scar segmentation model training, and testing dataset D4 for left atrial scar segmentation model testing. Each data in the dataset is a data sequence consisting of multiple slices.

[0055] Each image data in the training datasets D1, D2, D3 and D4 undergoes the following preprocessing operations in sequence:

[0056] Step 1.1: Calculate the pixel histogram of each image data in the training data sets D1, D2, D3 and D4, crop the pixel histogram, and retain the grayscale values with a frequency greater than 20;

[0057] Step 1.2: Normalize the pixel values of each slice in each data sequence processed in step 1.1;

[0058] Step 1.3: Based on step 1.2, each slice in each image data in the training datasets D1 and D3 is randomly flipped with a probability of 0.5;

[0059] Step 1.4: Based on step 1.3, each slice in each data sequence is randomly shifted in both the horizontal and vertical directions by an offset of (0.1, 0.1);

[0060] Step 1.5: Based on step 1.4, each slice in each data sequence is randomly scaled in both horizontal and vertical directions with parameters of (0.7, 1.3);

[0061] Step 1.6: Based on step 1.5, perform elastic deformation on each slice in each data sequence with a probability of 0.3;

[0062] Step 1.7: Based on step 1.6, perform random contrast enhancement with a gamma parameter in the range of (0.7, 1.3) on each slice in each data sequence;

[0063] Step 1.8: Based on step 1.7, each slice in each data sequence is randomly rotated within the interval of (-30°, 30°) with a probability of 0.5;

[0064] Step 1.9: Based on step 1.8, add Gaussian noise with mean 0 and variance (0, 0.1) to each slice in each data sequence with a probability of 0.15;

[0065] Step 1.10: Based on step 1.9, perform elastic deformation and random motion on each slice in each data sequence with a probability of 0.3.

[0066] Step 2: Use the training data set of the left atrium segmentation model enhanced in step 1 to train the left atrium segmentation model. Specifically:

[0067] The deep learning segmentation method DeepU-Net is selected as the segmentation framework, using hierarchical feature representation and symmetrical encoding and decoding paths, including a 6-layer convolutional network structure, with the maximum number of channels set to 512, and a weighted cross entropy loss function L weightedCE and dice loss function L Dice and L seg =L Dice +L weightedCE The total loss function is used to constrain the training process. The optimizer is the stochastic gradient descent optimizer SGD, the batch size is 16, the learning rate is first set to 1e-4, and the cosine annealing strategy is adopted. All data of the left atrium segmentation training dataset D1 preprocessed in step 1 are used for training. After 600 iterations, the model M for segmenting the left atrium of cardiac images is obtained. la .

[0068] Step 3: Use the left atrial segmentation model trained in step 2 to segment all the data in the left atrial scar training dataset enhanced in step 1, and then process it to obtain a training dataset for training the left atrial scar segmentation model. Specifically:

[0069] The left atrium segmentation model M obtained by training in step 2 la Segment all data in the enhanced left atrial scar training dataset D3 to obtain the label L corresponding to the left atrium la , and then the obtained left atrium label L la The corresponding original training data images in D3 are spliced according to the channel dimension to obtain the training data set D5 for training the left atrial scar segmentation model, which is used for training the left atrial scar segmentation model.

[0070] Step 4: Use the training data obtained in step 3 to train the left atrial scar segmentation model. Specifically, select the deep learning segmentation algorithm Deep U-Net as the segmentation framework, use the training dataset D5 obtained in step 3 to train the left atrial scar segmentation model, and use the weighted dice loss as the loss function during the training process. Weighted cross entropy loss L weightedCE and distance loss The sum of the three: L seg =L Dice +L weightedCE +0.1*L Distance The optimizer is the stochastic gradient descent optimizer SGD, the batch size is 16, the learning rate is first set to 1e-4, the cosine annealing strategy is adopted, and the model is trained by the 4-fold cross-validation training method. It is terminated after 600 iterations. Finally, the left atrial scar segmentation model M is obtained through training. scari , i=1,2,3,4.

[0071] Step 5: Use the left atrial scar segmentation model trained in step 4 to perform left atrial scar testing and post-processing. Specifically:

[0072] The left atrial scar segmentation model M obtained by training in step 4 scari All data in the scar segmentation test dataset D4 are tested, and the segmentation label L of the left atrial scar is obtained by the model voting integration method. scar ,Then the binary erosion and dilation operations with a kernel radius of 6 are used for post-processing to obtain the final left atrial scar segmentation result.

[0073] Simulation analysis

[0074] 1. Simulation conditions

[0075] The present invention uses the Pytorch framework for simulation on an Intel(R) i9-10900X 3.7GHz CPU, 125G memory, two Nvidia GTX1080 GPUs, and an Ubuntu20.04 operating system.

[0076] The test dataset used in the experiment comes from the Left Atrium and Scar Segmentation Challenge (LAScarQS 2022), which was proposed by Li et al. in the paper "Li, L., Zimmer, VA, Schnabel, JA, & Zhuang, X. (2022). AtrialJSQnet: A New framework for joint segmentation and quantification of left atrium and scars incorporating spatial and shape information. Medical image analysis, 76, 102303." The dataset contains 60 left atrial scar segmentation data and 130 left atrial segmentation data from different medical devices in 3 schools and medical institutions. It is open source for the accuracy and generalization performance research of the left atrial scar algorithm.

[0077] 2. Simulation content

[0078] The accuracy and domain generalization performance of the present invention for segmentation of the left atrium and scar in medical images were tested.

[0079] In order to verify the improvement of the model domain generalization ability through data augmentation in the present invention, the domain generalization method (FourierAug) mentioned by Xu et al. in the document “Xu, Qinwei, et al. A fourier-based framework for domain generalization [C]. Proceedings of the IEEE / CVF Conference on Computer Vision and Pattern Recognition. 2021.”, the method (MixStyle) proposed by Zhou et al. in the document “Zhou, K., Yang, Y., Qiao, Y., & Xiang, T. Domain Generalization with MixStyle. In International Conference on Learning Representations.” and the method (Swin) proposed by Hatamizadeh et al. in the document “Hatamizadeh A, Nath V, Tang Y, et al. Swin unetr: Swin transformers for semantic segmentation of brain tumors in mri images [C]. International MICCAIBrainlesion Workshop. Springer, Cham, 2022: 272-284.” were selected. UNETR), in the left atrium segmentation experiment, the dice coefficient, average surface distance (ASD) and Hausdorff distance (HD) of the segmentation results of the present invention and the control method and the ground truth were calculated as evaluation indicators, and the comparison results are shown in Table 1 (the dominant results are marked in bold). The left atrium segmentation results of the present invention and the control method are visualized, and the comparison of the visualization results is shown in the figure below. Figure 3 shown.

[0080] Table 1. Left atrium segmentation model performance verification results

[0081]

[0082] In order to verify the two-stage segmentation method of the present invention and the Dice loss function based on distance weighting and the distance loss function as a regularization term To investigate the effectiveness of left atrial scar segmentation, ablation experiments were designed using single-stage segmentation, single-stage segmentation with added spatial information, and two-stage segmentation with added spatial information. The Dice results of the four-fold cross-validation of the scar segmentation validation data using different experimental methods are shown in Table 2 (the dominant results are marked in bold).

[0083] Table 2. Performance verification results of left atrial scar segmentation model

[0084]

[0085] From the comparison of the results of left atrium segmentation between the present invention and the comparative method in Table 1, it can be seen that the present invention outperforms the Fourier Aug method in terms of Dice score, average surface distance (ASD), and Hausdorff distance (HD). This shows that the method of domain generalization through data enhancement has better and more stable left atrium segmentation performance than the comparative method. Figure 3 The visualization comparison results of the left atrium segmentation of the present invention and the comparison method can more intuitively show that the segmentation result of the left atrium structure of the heart of the present invention is closer to the left atrium Ground Truth, and there is no obvious wrong segmentation.

[0086] The second column of Table 2 shows the results of the single-stage left atrial scar segmentation method. The third column shows the results of the single-stage left atrial scar segmentation method using the distance-weighted Dice loss function and the distance loss function as a regularization term. The fourth column shows the results of the two-stage left atrial scar segmentation method using the distance-weighted Dice loss function and the distance loss function as a regularization term. It can be seen that the two-stage left atrial scar segmentation method using the distance-weighted Dice loss function and the distance loss function as a regularization term has better overall performance than the comparison method.

[0087] The basic principles, main features, and advantages of the present invention are shown and described above. Those skilled in the art should understand that the present invention is not limited to the foregoing embodiments. The foregoing embodiments and descriptions are merely illustrative of the principles of the present invention. Various changes and modifications may be made to the present invention without departing from the spirit and scope of the present invention. Such changes and modifications are intended to fall within the scope of the present invention. The scope of protection claimed in the present invention is defined by the appended claims and their equivalents.

Claims

1. A left atrium and scar segmentation method based on a Deep U-Net model, characterized by: include: Step 1: Enhance the training data set for training the left atrial segmentation model and the training data set for training the left atrial scar segmentation model; Step 2: Use the training data set of the left atrium segmentation model enhanced in step 1 to train the left atrium segmentation model; Step 3: Use the left atrial segmentation model trained in step 2 to segment all data in the left atrial scar training dataset enhanced in step 1, and then process it to obtain a training dataset for training the left atrial scar segmentation model; The step 3 is specifically as follows: The left atrium segmentation model M obtained by training in step 2 la Segment all data in the enhanced left atrial scar training dataset D3 to obtain the label L corresponding to the left atrium la , and then the obtained left atrium label L la The training dataset D5 for training the left atrial scar segmentation model is obtained by splicing the corresponding original training data images in D3 according to the channel dimension. Step 4: Use the training data obtained in step 3 to train the left atrial scar segmentation model; The step 4 is specifically as follows: The deep learning segmentation algorithm Deep U-Net is selected as the segmentation framework. The left atrial scar segmentation model is trained using the training dataset D5 obtained in step 3. The loss function of the training process is weighted dice loss. Weighted cross entropy loss L weightedCE and distance loss The sum of the three: L seg =L Dice +L weightedCE +0.1*L Distance The optimizer is the stochastic gradient descent optimizer SGD, the batch size is 16, the learning rate is first set to 1e-4, the cosine annealing strategy is adopted, and the model is trained by the 4-fold cross-validation training method. It is terminated after 600 iterations. Finally, the left atrial scar segmentation model M is obtained through training. scari , i=1,2,3,4; Step 5: Use the left atrial scar segmentation model trained in step 4 to perform left atrial scar testing and post-processing.

2. The method for left atrium and scar segmentation based on the Deep U-Net model according to claim 1, characterized in that: The step 1 comprises: All data in the dataset are preprocessed in the same way. The dataset consists of four parts: training dataset D1 for left atrium segmentation model training, testing dataset D2 for left atrium segmentation model testing, training dataset D3 for left atrial scar segmentation model training, and testing dataset D4 for left atrial scar segmentation model testing. Each data in the dataset is a data sequence consisting of multiple slices. Each image data in the training datasets D1, D2, D3 and D4 undergoes the following preprocessing operations in sequence: Step 1.1: Calculate the pixel histogram of each image data in the training data sets D1, D2, D3 and D4, crop the pixel histogram, and retain the grayscale values with a frequency greater than 20; Step 1.2: Normalize the pixel values of each slice in each data sequence processed in step 1.1; Step 1.3: Based on step 1.2, each slice in each image data in the training datasets D1 and D3 is randomly flipped with a probability of 0.5; Step 1.4: Based on step 1.3, each slice in each data sequence is randomly shifted in both the horizontal and vertical directions by an offset of (0.1, 0.1); Step 1.5: Based on step 1.4, each slice in each data sequence is randomly scaled in both horizontal and vertical directions with parameters of (0.7, 1.3); Step 1.6: Based on step 1.5, perform elastic deformation on each slice in each data sequence with a probability of 0.3; Step 1.7: Based on step 1.6, perform random contrast enhancement with a gamma parameter in the range of (0.7, 1.3) on each slice in each data sequence; Step 1.8: Based on step 1.7, each slice in each data sequence is randomly rotated within the interval of (-30°, 30°) with a probability of 0.5; Step 1.9: Based on step 1.8, add Gaussian noise with mean 0 and variance (0, 0.1) to each slice in each data sequence with a probability of 0.15; Step 1.10: Based on step 1.9, perform elastic deformation and random motion on each slice in each data sequence with a probability of 0.

3.

3. The method for left atrium and scar segmentation based on the Deep U-Net model according to claim 2, characterized in that: The step 2 is specifically as follows: The deep learning segmentation method Deep U-Net is selected as the segmentation framework, using hierarchical feature representation and symmetrical encoding and decoding paths, including a 6-layer convolutional network structure, with the maximum number of channels set to 512, and a weighted cross entropy loss function L weightedCE and dice loss function L Dice and L seg =L Dice +L weightedCE The total loss function is used to constrain the training process. The optimizer is the stochastic gradient descent optimizer SGD, the batch size is 16, the learning rate is first set to 1e-4, and the cosine annealing strategy is adopted. All data of the left atrium segmentation training dataset D1 preprocessed in step 1 are used for training. After 600 iterations, the model M for segmenting the left atrium of cardiac images is obtained. la .

4. The method for left atrium and scar segmentation based on a Deep U-Net model according to claim 3, characterized in that: The step 5 is specifically as follows: The left atrial scar segmentation model M obtained by training in step 4 scari All data in the scar segmentation test dataset D4 are tested, and the segmentation label L of the left atrial scar is obtained by the model voting integration method. scar ,Then the binary erosion and dilation operations with a kernel radius of 6 are used for post-processing to obtain the final left atrial scar segmentation result.

Images