Multi-Stage Segmentation Method for Multi-Modal MRI Heart Images

Abstract

The present invention discloses a multi-stage segmentation method based on multi-modal MRI cardiac images, mainly solving the problems that the prior art is not applicable to multi-modal and cannot effectively fuse features between cross-modalities. The method includes: 1) generating a cardiac contour annotation dataset from the original MRI data and performing normalization; 2) constructing a cardiac MRI contour segmentation model and training a deep learning model that fuses three-modal data using the normalized dataset to obtain a first-stage segmentation result; 3) multiplying the original data by the first-stage segmentation result matrix, and generating a second-stage dataset after channel stacking; 4) building a ventricle-myocardium segmentation model and training it using the second-stage dataset; 5) obtaining the second-stage segmentation result, i.e., the final segmentation result, through forward model inference. The present invention can effectively utilize the advantages of multi-modalities, improve the segmentation accuracy, and to a certain extent solve the dilemma of poor interpretability of the deep learning black box in the medical field.

Description

Technical Field

[0001] The present invention belongs to the technical field of image processing, and further relates to medical image segmentation technology. Specifically, it is a multi-stage segmentation method for multi-modal magnetic resonance imaging (MRI) cardiac images, which can be used for medical image processing. Background Art

[0002] Multi-functional modal magnetic resonance imaging (MRI) is an advanced imaging technology that provides richer tissue information by using multiple different imaging modes. For example, nuclear magnetic resonance T1W highlights the differences in tissue transverse relaxation, and nuclear magnetic resonance T2W highlights the differences in tissue longitudinal relaxation. Nuclear magnetic resonance imaging experts usually manually segment the target by visually observing the differences between the target to be segmented and the background, and combine multi-functional modal MRI information for tissue analysis. However, this method is not only time-consuming and laborious, but also requires very professional imaging knowledge. Deep learning methods have greatly alleviated this difficult situation. By building a neural network model and training a certain number of MRI data samples, they can quickly and accurately obtain the segmentation result after inputting the MRI data to be predicted. When existing MRI segmentation technologies involve multi-functional modalities, they usually only use the modality with the clearest contour of the target to be segmented for target segmentation. When separately segmenting the myocardium, left ventricle, and right ventricle of MRI cardiac images, they often can only use the functional modality image with the best image quality. Since MRI has rich functional modalities and existing segmentation technologies cannot effectively handle the highly non-linear relationships between various modalities, they cannot effectively utilize all MRI modality data, resulting in insufficient utilization of image information. Moreover, once multi-modal MRI segmentation involves multi-class segmentation, the situation becomes more complex.

[0003] Northwestern Polytechnical University disclosed a dual-modal brain tumor MRI segmentation method based on a clustering fusion algorithm in its patent document with the name "A Dual-Modal Brain Tumor MRI Segmentation Method Based on a Clustering Fusion Algorithm" (application number 201811414799.3, publication number CN 109685767A). By inputting different modal images of tumor patients, extracting their pixel points and features, then obtaining the clustering result through a comprehensive three-distance K-means clustering algorithm, and finally using the extraction of the largest connected region to output the final result. The disadvantage of this method is that the clustering algorithm used only establishes connections between various modalities through distance formulas and cannot effectively fuse cross-modal features. Moreover, using the extraction of the largest connected region as the final prediction result may eliminate many non-correct predictions in the free regions, affecting the accuracy.

[0004] The University of Electronic Science and Technology of China discloses a cardiac MRI segmentation method in its patent document titled "A Cardiac MRI Segmentation Method and System" (Application No. 202111394636.5, Publication No. CN 113902738 A). By constructing and training a monitoring model for the central points of the left and right ventricles in cardiac MRI, extracting the position information of the central points of the left and right ventricles in the cardiac MRI data, and extracting the distance map of the region of interest from the central point based on this position information, then constructing a cardiac MRI segmentation model and training it using the distance map, and connecting the two models in series to construct an automatic cardiac MRI segmentation model to obtain the final segmentation result. Although this method improves the segmentation performance to a certain extent, it does not utilize the advantages of multi-functional modalities. It only builds a model on a single-modal dataset, with a very low utilization rate of modalities, and has poor interpretability in the medical field. Summary of the Invention

[0005] The object of the present invention is to address the deficiencies of the above-mentioned existing technologies by proposing a multi-stage segmentation method for multi-modal MRI cardiac images, which is used to solve the problems of the single application scenario of the existing technologies, inapplicability to multi-modalities, and inability to effectively fuse features between cross-modalities, and at the same time overcome the problem of poor interpretability of deep learning in the field of medical image processing. The present invention uses a two-stage deep learning method for multi-modal MRI segmentation, combines various modal data to improve the target segmentation accuracy, and solves the problem of insufficient segmentation accuracy in the case of multi-modal nuclear magnetic resonance MRI.

[0006] The idea to implement the present invention is as follows: First, generate a cardiac contour annotation dataset from the original data, and use this dataset to train a deep learning model that fuses three modal data to segment the entire cardiac contour and obtain the first-stage segmentation result; then, multiply the original data by the segmentation result matrix of the first stage, generate a dataset without background after channel stacking, then build a ventricle-myocardium segmentation model and train it using this dataset, and finally obtain the second-stage segmentation result, that is, the final segmentation result, through forward model inference.

[0007] To achieve the above object, the technical solution of the present invention includes:

[0008] (1) Split both the original three-dimensional nuclear magnetic resonance imaging (MRI) data and its corresponding three-dimensional annotation data into two-dimensional image data; the original three-dimensional MRI data generates a dataset after splitting This dataset contains N samples, where the i-th sample contains three MRI modalities, corresponding to BSSFP, LGE, and T2 respectively; the three-dimensional annotation data generates a second-stage annotation dataset GT2 after splitting.

[0009] (2) The three regional values of the labeled data in dataset D1 are 1, 2, and 3 respectively, and the background value is 0. The three regions are the left ventricle, the right ventricle, and the myocardium. Set all pixel values greater than 0 in the corresponding labels of the data to 1 to obtain the labeled data of the heart contour. Let the i-th sample in dataset D1 The corresponding labeled data is y i , and obtain the labeled data of all samples in dataset D1 to form a first-stage labeled dataset GT1;

[0010] (3) Normalize each pixel value of the samples in dataset D1 to obtain the normalized data samples;

[0011] (4) Construct a heart MRI contour segmentation model:

[0012] (4.1) Improve the Unet model and build an MS-Unet network structure with one backbone encoder and one decoder. The backbone encoder contains five downsamplings. Before each downsampling, two convolutional operations are performed, and the ReLU function is used as the activation layer after each convolutional operation. The downsampling is implemented through a max pooling layer. The decoder contains five upsamplings. Before each upsampling, two convolutional operations are performed, and the ReLU function is used as the activation layer after each convolutional operation. The upsampling is implemented using the nearest neighbor interpolation algorithm. Before each downsampling, the feature maps obtained in the encoding path pass through two multi-scale feature extraction modules through skip connections to extract the multi-scale features of the heart MRI, and then are concatenated with the feature maps of the same resolution in the decoding path to restore the image information lost due to downsampling in the encoder;

[0013] (4.2) Use three parallel MS-Unet network structures, followed by two convolutional layers, to obtain a heart MRI contour segmentation model;

[0014] (5) Use the MRI images of the same subject as the input data of the model constructed in step (4). After forward propagation, obtain the first-stage prediction result Pred1. Calculate the loss with the samples in the first-stage labeled dataset GT1, and use the Adam algorithm optimizer to optimize the loss function to obtain the trained heart MRI contour segmentation model;

[0015] (6) Use the trained heart MRI contour segmentation model to predict the MRI to be measured of the same subject to obtain the predicted heart contour, that is, the first-stage segmentation result Pred1;

[0016] (7) Generate the MRI dataset D2 for second-stage training:

[0017] Multiply the three modal images of MRI with Pred1 respectively. In the multiplied results, only the heart part of the three-modal MRI data is visible. Stack the three-modal data in the channel dimension to generate the dataset D2 = [x1, x2,..., x i ,..., x N ;

[0018] (8) Build a ventricle-myocardium segmentation model and train it using the dataset D2:

[0019] Use an MS-Unet network structure to form a ventricle-myocardium segmentation model; take the dataset D2 as the model input, and after forward propagation, obtain the two-stage prediction result Pred2. Calculate the loss with the samples in the two-stage annotation dataset GT2, and use the Adam algorithm optimizer to optimize the loss function to obtain the trained ventricle-myocardium segmentation model;

[0020] (9) Input the data to be measured into the trained ventricle-myocardium segmentation model, and after forward inference, output the final segmentation result.

[0021] Compared with the existing technologies, the present invention has the following advantages:

[0022] First, since the present invention adopts a U-net model based on a residual multi-pooling module and a multi-scale convolution module, it can effectively extract image features of different scales, and the residual connection between the encoder and decoder of the model can prevent gradient explosion, ensuring the robustness and generalization ability of the algorithm;

[0023] Second, since the present invention uses a hierarchical fusion strategy to combine multiple MRI modalities for image segmentation, it overcomes the defect of the existing technology of MRI in the multi-modal case of maximizing the elimination of modal redundancy, enabling the present invention to improve the utilization rate of MRI modalities while obtaining consistent and highly robust segmentation results in multiple modalities;

[0024] Third, since the present invention adopts a multi-stage segmentation method, first, the contour of the whole heart is segmented in the first stage, and on this basis, the left ventricle, right ventricle, and myocardium of the heart region are further segmented in the second stage, which not only improves the segmentation accuracy, but also from a medical perspective, simplifies the segmentation of multi-category targets from simple to complex, and to a certain extent solves the dilemma of poor interpretability of the deep learning black box in the medical field. Description of the Drawings

[0025] Figure 1 is the implementation flowchart of the present invention;

[0026] Figure 2 is the structural diagram of the multi-stage segmentation framework in the present invention;

[0027] Figure 3 It is a schematic diagram of the encoding and decoding segmentation model in the present invention. Specific embodiments

[0028] The present invention will be further described below in conjunction with the accompanying drawings and embodiments.

[0029] Example 1: Referring to Figure 1 , the multi-stage segmentation method for multi-modal MRI cardiac images proposed by the present invention specifically includes the following steps:

[0030] Step 1. Split both the original three-dimensional nuclear magnetic resonance imaging (MRI) data and its corresponding three-dimensional annotation data into two-dimensional image data; among them, the original three-dimensional MRI data generates a dataset This dataset contains N samples, and the i-th sample contains three MRI modalities, corresponding to BSSFP, LGE, and T2 respectively; after splitting the three-dimensional annotation data, a two-stage annotation dataset GT2 is obtained;

[0031] Step 2. The three types of region values in the annotation data in dataset D1 are 1, 2, and 3 respectively, and the background value is 0. The three types of regions are the left ventricle, the right ventricle, and the myocardium; set all pixel values greater than 0 in the corresponding label of the data to 1 to obtain the annotation data of the cardiac contour. Let the annotation data corresponding to the i-th sample in dataset D1 be y i , and obtain the annotation data of all samples in dataset D1, and form a first-stage annotation dataset GT1;

[0032] Step 3. Normalize each pixel value of the samples in dataset D1 to obtain the normalized data samples, as follows:

[0033]

[0034] Among them, Y k represents the normalized pixel value of the k-th pixel in the input image, X k represents the pixel value of the k-th pixel in the input image, X min represents the minimum pixel value in the input image, and X max represents the maximum pixel value in the input image.

[0035] Step 4. Construct a cardiac MRI contour segmentation model:

[0036] (4.1) Improve the Unet model and build an MS-Unet network structure with a backbone encoder and a decoder; the backbone encoder includes five downsamplings. Before each downsampling, two convolutional operations are performed, and the ReLU function is used as the activation layer after each convolutional operation. The downsampling is implemented through a max-pooling layer; the decoder includes five upsamplings. Before each upsampling, two convolutional operations are performed, and the ReLU function is used as the activation layer after each convolutional operation. The upsampling is implemented using the nearest neighbor interpolation algorithm; before each downsampling, the feature maps obtained in the encoding path pass through two multi-scale feature extraction modules through skip connections to extract the multi-scale features of the cardiac MRI, and then are concatenated with the feature maps of the same resolution in the decoding path to restore the image information lost due to downsampling in the encoder. In each convolutional operation in this step of this embodiment, the convolutional kernel size is set to 3*3.

[0037] The above two multi-scale feature extraction modules are specifically a residual multi-pooling module and a multi-scale convolutional model. The former consists of multiple parallel pooling operations, which can detect objects of different sizes through multiple effective fields of view without additional parameter calculation. The latter consists of multiple dilated convolutions of different sizes, which use different receptive fields to broaden the structure and combine a residual connection mechanism to avoid gradient explosion and disappearance.

[0038] (4.2) Use three parallel MS-Unet network structures, followed by two convolutional layers, to obtain a cardiac MRI contour segmentation model; in this embodiment, the convolutional kernel sizes of the two convolutional layers used are both 1*1.

[0039] Step 5. Use the MRI images of the same subject as the input data of the model constructed in step (4). After forward propagation, obtain the first-stage prediction result Pred1, calculate the loss with the samples in the first-stage annotation dataset GT1, and use the Adam algorithm optimizer to optimize the loss function to obtain the trained cardiac MRI contour segmentation model;

[0040] The loss calculation is specifically to use the sum of the Dice loss function and the cross-entropy loss function as the cardiac MRI contour segmentation loss function value L:

[0041]

[0042] Among them, X is the predicted value of the model, Y is the GT1 annotation map, N represents the number of pixels in the prediction result, y c represents the label value, that is, 0 or 1, and p c represents the predicted probability after the softmax function.

[0043] The use of the Adam algorithm optimizer to optimize the loss function, where the learning rate of the optimizer adopts the following adjustment strategy:

[0044]

[0045] Among them, let the number of training times of the model be \(e\), and in this embodiment, \(e = [200, 300]\) is taken. The training period for obtaining the new learning rate is \(step\_size\), and \(step\_size\) is initialized to 1. One sample is put into the model for training in each training; \(epoch\) p represents the \(p\)-th training, where \(p = 1, 2,\cdots, e\); \(new\_lr\) represents the new learning rate obtained after every \(step\_size\) training, \(initial\_lr\) represents the initial learning rate, and \(\gamma\) represents the update factor with an initial value of 0.9.

[0046] Step 6. Use the trained cardiac MRI contour segmentation model to predict the MRI to be measured of the same subject, and obtain the predicted cardiac contour, that is, the one-stage segmentation result \(Pred1\);

[0047] Step 7. Generate the MRI dataset \(D2\) for two-stage training:

[0048] Perform multiplication operations on the three-modal images of the MRI respectively with \(Pred1\). In the multiplied results, only the cardiac part of the three-modal MRI data is visible. Stack the three-modal data in the channel dimension to generate the dataset \(D2 = [x1, x2,\cdots, x i ,\cdots, x N \), where the \(i\)-th sample \(x i is calculated by the following formula:

[0049]

[0050] where \(concat\) represents the stacking operation on the data in the channel dimension.

[0051] Step 8. Build a ventricle-myocardium segmentation model and train it using the dataset \(D2\):

[0052] Use an MS-Unet network structure to form a ventricle-myocardium segmentation model; take the dataset \(D2\) as the model input. After forward propagation, obtain the two-stage prediction result \(Pred2\). Calculate the loss with the samples in the two-stage annotation dataset \(GT2\), and use the Adam algorithm optimizer to optimize the loss function to obtain the trained ventricle-myocardium segmentation model. The specific implementation of the loss calculation and the optimization of the loss function using the Adam algorithm optimizer here is the same as in Step 5.

[0053] Step 9. Input the data to be measured into the trained ventricle-myocardium segmentation model, and after forward inference, output the final segmentation result.

[0054] Example 2: The overall implementation steps of this example are the same as those of Example 1. Now, with reference to Figure 2 , the implementation process of multi-stage segmentation in the present invention will be further described by giving specific examples in two stages:

[0055] The first stage: In order to improve the segmentation accuracy of the left ventricle, right ventricle, and myocardium, and make the best use of the three MRI modality data, a multi-modal segmentation model is first constructed, named the cardiac MRI contour segmentation model;

[0056] Step 1.1). Obtain a multi-modal cardiac MRI dataset.

[0057] In this example, the cardiac MRI dataset MS-CMRSeg2019, which is publicly available on the network and has expert annotations, is used. This dataset contains cardiac MRI data samples of 45 myocardial subjects, and each subject sample contains three MRI functional modalities: BSSFP, LGE, and T2. Among them, bSSFP is a balanced steady-state free precession sequence, LGE is a T1-weighted, inversion recovery, gradient echo sequence, and T2 is a T2-weighted, black blood spectral presaturation attenuation inversion recovery sequence. In this example, 30 subject samples are used as the training set, and 15 subject samples are used as the test set. Since all samples are three-dimensional and the number of sample slices is small, in this example, the three-dimensional MRI data is split into two-dimensional image data to generate the dataset This dataset contains N samples, where the i-th sample contains three MRI modalities, corresponding to BSSFP, LGE, and T2 respectively; after splitting the three-dimensional annotation data, a two-stage annotation dataset GT2 is obtained;

[0058] Step 1.2). Generate cardiac contour label data.

[0059] Since the myocardium, left ventricle, and right ventricle in cardiac MRI are connected, the three types of regions in the corresponding label data are 1, 2, and 3 respectively, and the background value is 0. The proposed method sets all pixel values greater than 0 in the corresponding label of the data to 1, and the background 0 value remains unchanged, so as to obtain the annotation data of the cardiac contour. Let the annotation data corresponding to the i-th sample in the dataset D1 be y i , and obtain the annotation data of all samples in the dataset D1, and form a first-stage annotation dataset GT1;

[0060] Step 1.3). Data normalization processing.

[0061] Normalize each pixel value in the image, which can reduce the loss of the neural network and effectively accelerate the convergence speed of the model.

[0062] Step 1.4). Build a cardiac MRI contour segmentation model;

[0063] To make the most of the advantages of MRI multi-modalities and improve the segmentation accuracy of the cardiac contour, the present invention constructs a cardiac MRI contour segmentation model, as Figure 2 shown;

[0064] The cardiac MRI contour segmentation model mainly consists of three improved Unet models, and this improved Unet model is named MS-Unet (multi-scale Unet), as Figure 3 shown. The MS-Unet network structure has a backbone encoder and a decoder, and the VGG16 classifier network is used as the encoder of MP-Unet. Before each downsampling, the feature maps obtained in the encoding path pass through two multi-scale feature extraction modules through skip connections to extract the multi-scale features of cardiac MRI, and then are concatenated with the feature maps of the same resolution in the decoding path to recover the image information lost due to downsampling in the encoder.

[0065] The encoder includes five downsamplings. Before each downsampling, two convolutional operations are performed, and the size of each convolutional kernel is set to 3*3. After each convolutional operation, the ReLU function is used as the activation layer, and the downsampling is implemented through the max pooling layer.

[0066] The decoder includes five upsamplings. Before each upsampling, two convolutional operations are performed, and the size of each convolutional kernel is set to 3*3. After each convolutional operation, the ReLU function is used as the activation layer, and the upsampling is implemented using the nearest neighbor interpolation algorithm.

[0067] The two multi-scale feature extraction modules: the residual multi-pooling module and the multi-scale convolutional model. The former consists of multiple parallel pooling operations, and different-sized objects are detected through multiple effective fields of view without additional parameter calculations. The latter consists of multiple dilated convolutions of different sizes, and different receptive fields are used to widen the structure, and the residual connection mechanism is combined to avoid gradient explosion and disappearance;

[0068] Three modalities of cardiac MRI, BSSFP, LGE, and T2, are respectively input into three MS-Unet models. At the end of the model, the output feature maps enter two convolutional layers through the concat operation to output the final segmentation result;

[0069] Step 1.5). Train the cardiac MRI contour segmentation model.

[0070] During the training process, the cardiac MRI contour segmentation model inputs three different modality MRI images of the same subject respectively. After forward propagation, the predicted result Pred1 is obtained and the loss is calculated with the label data GT1. The model uses the sum of the Dice loss function and the cross-entropy loss function as the cardiac MRI contour segmentation loss function. The model training uses the Adam algorithm optimizer to optimize the loss, and the learning rate in the optimizer adopts the adjustment strategy of the following formula:

[0071]

[0072] where the number of training times of the model is 300, the training period for obtaining the new learning rate is step_size, step_size is initialized to 1, and 1 sample is put into training in each training of the model; epoch p represents the p-th training, p = 1, 2,..., e; new_lr represents the new learning rate obtained after every step_size training, and new_lr generates new values in each round of training, and the learning rate changes dynamically; initial_lr represents the initial learning rate, and γ represents the update factor with an initial value of 0.9.

[0073] The second stage: Based on the first stage, according to the cardiac contour predicted by the cardiac MRI contour segmentation model, further segmentation is performed to obtain the complete myocardium, left ventricle, and right ventricle.

[0074] Step 2.1). Generate the MRI dataset for the second-stage training.

[0075] In the first stage, the cardiac MRI contour segmentation model segmented the cardiac contour of the subject and obtained the predicted result Pred1. The present invention performs multiplication operations on the three modality images of the MRI with Pred1 respectively, so as to remove the interference of the background in the original data. Because the label value of the entire heart part in Pred1 is 1 and the background label value is 0, in the result of multiplying Pred1 by the original data, only the heart part of the three modality MRI data is visible. Subsequently, the three modality data are stacked in the channel dimension and input into the second-stage segmentation model MS-Unet.

[0076] To improve the segmentation accuracy of the second stage, it is necessary to regenerate a new dataset D2 = [x1, x2,..., x N , where the i-th sample is calculated by the following formula:

[0077]

[0078] where are the three modality data of the i-th sample in the dataset D1 respectively, and concat represents stacking the data in the channel dimension;

[0079] Step 2.2). Build a ventricular-myocardial segmentation model.

[0080] Since the boundaries of the left ventricle, right ventricle, and myocardium in the whole heart are relatively distinct, and the heart contour has been identified through the prediction of the first-stage cardiac MRI contour segmentation model with less interference from the background, the segmentation task in the second stage is simpler than that in the first stage, and only one MS-Unet is needed.

[0081] Step 2.3). Train the ventricular-myocardial segmentation model.

[0082] During the training process, the ventricular-myocardial segmentation model inputs the D2 dataset samples. After forward propagation, the prediction result Pred2 is obtained and the loss is calculated with the label data GT2. The loss function and training strategy of the model are similar to those in the first stage and will not be elaborated here.

[0083] After the model training is completed, for the prediction of new samples (including three modalities), first, the images are normalized, and after inputting into the model, the segmentation result is obtained through forward inference.

[0084] The parts not detailed in the present invention belong to the common general knowledge of those skilled in the art.

[0085] The above are only the preferred embodiments of the present invention and are not intended to limit the present invention. Obviously, for professionals in the field, after understanding the content and principle of the present invention, various modifications and changes in form and details may be made without departing from the principle and structure of the present invention. However, these modifications and changes based on the idea of the present invention are still within the protection scope of the claims of the present invention.

Claims

1. A multi-stage segmentation method for multi-modal MRI cardiac images, characterized in that, Construct multiple models with multi-scale feature extraction capabilities, and adopt an inter-layer fusion strategy to fuse the multi-modal features extracted by them. In addition, adopt a multi-stage segmentation method to obtain the final segmentation result. The implementation steps are as follows: (1) Split both the original three-dimensional nuclear magnetic resonance imaging (MRI) data and its corresponding three-dimensional annotation data into two-dimensional image data; the original three-dimensional MRI data generates a dataset after splitting This dataset contains N samples, where the i-th sample contains three MRI modalities, corresponding to BSSFP, LGE, and T2 respectively; after splitting the three-dimensional annotation data, a two-stage annotation dataset GT2 is obtained. (2) In the dataset D1, the three regional values of the labeled data are 1, 2, and 3 respectively, and the background value is 0. The three regions are the left ventricle, the right ventricle, and the myocardium. Set all pixel values greater than 0 in the corresponding labels of the data to 1 to obtain the labeled data of the heart contour. Let the i-th sample in the dataset D1 have the corresponding labeled data y i , and obtain the corresponding labeled data of all samples in the dataset D1 to form the first-stage labeled dataset GT1; (3) Normalize each pixel value of the samples in dataset D1 to obtain the normalized data samples; (4) Construct a cardiac MRI contour segmentation model: (4.1) Improve the Unet model and build an MS-Unet network structure with one backbone encoder and one decoder. The backbone encoder contains five downsamplings. Before each downsampling, two convolutional operations are performed, and the ReLU function is used as the activation layer after each convolutional operation. The downsampling is implemented through a max-pooling layer. The decoder contains five upsamplings. Before each upsampling, two convolutional operations are performed, and the ReLU function is used as the activation layer after each convolutional operation. The upsampling is implemented using the nearest neighbor interpolation algorithm. Before each downsampling, the feature maps obtained in the encoding path pass through two multi-scale feature extraction modules through skip connections to extract the multi-scale features of the cardiac MRI, and then are concatenated with the feature maps of the same resolution in the decoding path to restore the image information lost due to downsampling in the encoder; (4.2) Use three parallel MS-Unet network structures, followed by two convolutional layers, to obtain a cardiac MRI contour segmentation model; (5) Use the MRI images of the same subject as the input data of the model constructed in step (4). After forward propagation, obtain the first-stage prediction result Pred1. Calculate the loss with the samples in the first-stage annotation dataset GT1, and use the Adam algorithm optimizer to optimize the loss function to obtain the trained cardiac MRI contour segmentation model; (6) Use the trained cardiac MRI contour segmentation model to predict the MRI to be tested of the same subject to obtain the predicted cardiac contour, that is, the first-stage segmentation result Pred1; (7) Generate the MRI dataset D2 for the second-stage training: Multiply the three modal images of MRI with Pred1 respectively. In the result of the multiplication, only the heart part of the three-modal MRI data is visible. Stack the three-modal data in the channel dimension to generate the dataset D2 = [x1, x2,..., x i ,..., x N ; (8) Build a ventricle-myocardium segmentation model and train it using dataset D2: Use one MS-Unet network structure to form a ventricle-myocardium segmentation model. Use dataset D2 as the model input. After forward propagation, obtain the second-stage prediction result Pred2. Calculate the loss with the samples in the second-stage annotation dataset GT2, and use the Adam algorithm optimizer to optimize the loss function to obtain the trained ventricle-myocardium segmentation model; (9) Input the data to be tested into the trained ventricle-myocardium segmentation model, and after forward inference, output the final segmentation result.

2. The method according to claim 1, wherein: In step (3), the normalization of each pixel value of the samples in dataset D1 is implemented as follows: Among them, Y k represents the normalized pixel value of the k-th pixel in the input image, X k represents the pixel value of the k-th pixel in the input image, X min represents the minimum pixel value in the input image, X max represents the maximum pixel value in the input image.

3. The method according to claim 1, wherein: In the convolutional operation described in step (4.1), the size of each convolutional kernel is set to 3*3; the convolutional kernel sizes of the two convolutional layers described in step (4.2) are both 1*1.

4. The method according to claim 1, characterized in that: The two multi-scale feature extraction modules described in step (4.1) are specifically a residual multi-pooling module and a multi-scale convolutional model. The former consists of multiple parallel pooling operations, which can detect objects of different sizes through multiple effective fields of view without additional parameter calculations. The latter consists of multiple dilated convolutions with different sizes, uses different receptive fields to widen the structure, and combines a residual connection mechanism to avoid gradient explosion and disappearance.

5. The method according to claim 1, characterized in that: The loss calculation described in steps (5) and (8) is specifically to use the sum of the Dice loss function and the cross-entropy loss function as the value L of the heart MRI contour segmentation loss function: Where X is the predicted value of the model, Y is the GT1 annotation map, N represents the number of pixels of the prediction result, and y c represents the label value, that is, 0 or 1, and p c represents the predicted probability after the softmax function.

6. The method according to claim 5, wherein: The Adam algorithm optimizer is used to optimize the loss function in steps (5) and (8). The learning rate of the optimizer adopts the following adjustment strategy: Among them, let the number of training times of the model be \(e\), the training period for obtaining the new learning rate be \(step\_size\), initialize \(step\_size\) to 1, and the model puts 1 sample for training in each training; \(epoch\) p represents the \(p\)-th training, where \(p = 1, 2,\cdots, e\); \(new\_lr\) represents the new learning rate obtained after every \(step\_size\) training, \(initial\_lr\) represents the initial learning rate, and \(\gamma\) represents the update factor with an initial value of 0.

9.

7. The method according to claim 1, wherein: In step (7), the generated dataset D2 = [x1, x2,..., x i ,..., x N , where the i-th sample x i is calculated by the following formula: Among them, concat means to perform a stacking operation on the data in the channel dimension.

Images