Heart AI system based on deep learning

Abstract

The invention discloses a heart AI system based on deep learning, and relates to the technical field of medical image processing, and the heart AI system comprises a two-dimensional fully-supervised left atrium segmentation algorithm module, a three-dimensional semi-supervised left atrium segmentation algorithm module, and a whole heart segmentation algorithm module. Meanwhile, through combination of Transform and a residual network, semantic information in the image is captured in a larger visual field, so that the segmentation precision of the network is improved; a two-dimensional fully-supervised left atrium segmentation algorithm module based on axial attention is provided from a two-dimensional angle by taking UNet as a basic model, the algorithm module firstly constructs a double-layer residual error convolution module, gradient information is transmitted through a plurality of paths, then a task-driven feature enhancement module is introduced to optimize feature representation, and finally, a task-driven left atrium segmentation algorithm is constructed. A multivariate feature fusion module is designed to integrate features, and finally a feature coupling module is designed, so that the limitation of traditional convolution operation is overcome.

Description

Technical Field

[0001] The present invention relates to the field of medical image processing technology, and in particular to a cardiac AI system based on deep learning. Background Art

[0002] The cardiac AI system is an artificial intelligence technology applied in the field of cardiac medicine. Its purpose is to predict, diagnose and treat heart diseases by analyzing and identifying medical imaging data. At present, many research institutions and companies at home and abroad are conducting relevant research and applications of cardiac AI systems. A deep learning-based cardiac segmentation method has been proposed, which can accurately extract various tissue structures in cardiac images, providing important basic data for subsequent analysis and diagnosis of heart diseases.

[0003] In the study of few-shot learning for cardiac segmentation tasks, whether it is using unsupervised learning to reuse data and learn the three-dimensional geometric relationship between the short axis and the standard axis of the cardiac image, or using an unsupervised teacher-student model in the semi-supervised learning stage, using unlabeled data, and establishing a shape constraint network to estimate the quality of pseudo labels to select reliable pseudo labels for the next stage of learning, the calculations of these methods are extremely complex, and the amount of calculations will expand again, making the calculations more difficult. Therefore, how to reduce the computational complexity of the model and reduce the computational cost has become an urgent problem to be solved. For this reason, we propose a cardiac AI system based on deep learning. Summary of the invention

[0004] The purpose of the present invention is to solve the shortcomings of the prior art and to propose a cardiac AI system based on deep learning.

[0005] In order to achieve the above object, the present invention adopts the following technical solutions:

[0006] A heart AI system based on deep learning, including a two-dimensional fully supervised left atrium segmentation algorithm module, a three-dimensional semi-supervised left atrium segmentation algorithm module, and a whole heart segmentation algorithm module. The two-dimensional fully supervised left atrium segmentation algorithm module is based on the UNet model and combines the Transformer and residual network to capture semantic information in the image in a larger field of view, thereby improving the segmentation accuracy of the network; the two-dimensional fully supervised left atrium segmentation algorithm module includes: a double-layer residual convolution module, a task-driven feature enhancement module, a multi-feature fusion module, and a feature coupling module;

[0007] The three-dimensional semi-supervised left atrium segmentation algorithm module can capture rich contextual information between adjacent slices; the three-dimensional semi-supervised left atrium segmentation algorithm module is based on VNet, and its network structure continues the encoder-decoder structure of UNet mentioned above, and draws on the jump connection idea of ​​UNet to supplement the lost spatial detail information;

[0008] The whole heart segmentation algorithm module is embedded in a weighted axial attention module to decompose the attention along the width axis and the height axis, and control its learning ability by adjusting the relative position encoding weights.

[0009] Preferably, the double-layer residual convolution module replaces the ordinary convolution in the original UNet model, enhancing the expression ability of the model so that it can better learn the complex features in the image.

[0010] Preferably, the task-driven feature enhancement module focuses on channel features to improve feature discrimination and further enhance the model's ability to distinguish left atrium-irrelevant features.

[0011] Preferably, the multi-feature fusion module realizes the integration of low-level features of the encoder and high-level features of the decoder, thereby fully mining local information and global information in the image.

[0012] Preferably, the feature coupling module overcomes the difference in features extracted by convolution and Transformer, and further optimizes the segmentation effect.

[0013] Preferably, the three-dimensional semi-supervised left atrium segmentation algorithm module is suitable for the semi-supervised left atrium segmentation task. The module uses Monte Carlo dropout to estimate uncertainty to retain pseudo-labels with high confidence, and at the same time enhances the module's learning ability for the overall characteristics of the left atrium by incorporating prior features. The introduction of boundary loss function and false positive rate loss function improves the module's segmentation accuracy for boundaries and fine structures, and uses consistency regularization based on combined enhancement to enhance the model's anti-interference ability.

[0014] Preferably, the three-dimensional semi-supervised left atrium segmentation algorithm module proposes an integrated framework comprising four stages, and the four stages of the algorithm module include: supervised pre-training of anatomical structure fusion, generation of high-quality pseudo-labels, consistency regularization based on combination enhancement, and high-quality pseudo-label optimization model.

[0015] Preferably, the whole heart segmentation algorithm module also includes a target domain. The pseudo labels required by the target domain are generated by an unsupervised teacher-student model. At the same time, the Hausdorff distance and D ice value formula and the newly proposed pseudo loss are used to control the quality of the pseudo labels actually participating in the training. Only pseudo labels that meet the conditions can participate in the training as shape constraints.

[0016] Preferably, the whole heart segmentation algorithm module freezes the parameters of the encoder and decoder parts, abandons the meta-learning method in the classifier part, and only uses cross entropy, D ice value and KL divergence for supervision; finally, the VAE module fits the segmentation prediction and evaluates the segmentation quality through the learned shape distribution information.

[0017] The beneficial effects of the present invention are:

[0018] 1. In the present invention, in view of the problems of poor detail processing ability of existing segmentation algorithms, scarce labeled data, inability of traditional convolutional networks to fully utilize global information and insufficient clinical applications, a two-dimensional fully supervised left atrium segmentation algorithm module based on axial attention is proposed based on the UNet-based model from a two-dimensional perspective based on a fully supervised deep learning method. The algorithm module first constructs a double-layer residual convolution module to transmit gradient information through multiple paths, and then introduces a task-driven feature enhancement module to optimize feature representation, designs a multi-feature fusion module to integrate features, and finally designs a feature coupling module to overcome the limitations of traditional convolution operations.

[0019] 2. In this invention, in view of the challenges of insufficient labeled data and complex three-dimensional medical image data, a three-dimensional semi-supervised left atrium segmentation algorithm module integrating anatomical structure is designed based on VNet as the basic model. The algorithm module first enhances the learning ability of the model on the overall structure of the left atrium by integrating anatomical structure to complete pre-training on a small number of labeled data sets, and then uses the pre-trained model to generate pseudo labels and prediction results on weakly enhanced and strongly enhanced unlabeled data, respectively, and screen out pseudo labels with high confidence. Finally, the joint supervision loss and consistency unsupervised loss are calculated, and iterative training is performed on labeled data and pseudo-label data to optimize the segmentation effect. Experimental analysis is performed on public data sets to prove the superiority of this method.

[0020] 2. In the present invention, from the two perspectives of few-sample learning method and reducing the number of parameters, the whole heart segmentation algorithm module is improved and the whole heart segmentation task of CT images is explored: a whole heart segmentation algorithm module for CT images based on semi-supervised learning is proposed and this algorithm module is improved, and a whole heart segmentation algorithm module for CT images based on unsupervised learning is proposed, which solves the problems of needing labeled samples during training, limited data sets, and insufficient use of shape information. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Figure 1 This is a framework diagram of a two-dimensional fully supervised left atrium segmentation algorithm module of a deep learning-based cardiac AI system of the present invention.

[0022] Figure 2 This is a framework diagram of a three-dimensional semi-supervised left atrium segmentation algorithm module of a deep learning-based cardiac AI system of the present invention.

[0023] Figure 3 This is a framework diagram of a whole heart segmentation algorithm module of a deep learning-based cardiac AI system of the present invention.

[0024] Figure 4This is a schematic diagram of the qualitative results of the comparison between a two-dimensional segmentation algorithm of a deep learning-based cardiac AI system of the present invention and five algorithms.

[0025] Figure 5 This is a schematic diagram of the segmentation results of an unsupervised whole heart segmentation algorithm module of a deep learning-based cardiac AI system of the present invention. DETAILED DESCRIPTION

[0026] The technical solutions in the embodiments of the present invention will be described clearly and completely below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments.

[0027] A heart AI system based on deep learning, including a two-dimensional fully supervised left atrium segmentation algorithm module, a three-dimensional semi-supervised left atrium segmentation algorithm module, and a whole heart segmentation algorithm module:

[0028] The two-dimensional fully supervised left atrium segmentation algorithm module is based on the UNet model and combines Transformer and residual networks to capture semantic information in the image in a larger field of view, thereby improving the segmentation accuracy of the network;

[0029] The two-dimensional fully supervised left atrium segmentation algorithm module includes: a double-layer residual convolution module, a task-driven feature enhancement module, a multi-feature fusion module and a feature coupling module; the double-layer residual convolution module replaces the ordinary convolution in the original UNet model to enhance the expression ability of the model so that it can better learn the complex features in the image; the task-driven feature enhancement module focuses on channel features to improve the feature discrimination and further enhance the model's ability to distinguish irrelevant features of the left atrium; the multi-feature fusion module realizes the integration of the low-level features of the encoder and the high-level features of the decoder, thereby fully mining the local and global information in the image; the feature coupling module overcomes the feature differences extracted by convolution and Transformer, and further optimizes the segmentation effect; the two-dimensional fully supervised left atrium segmentation algorithm module has achieved better results in the left atrium segmentation problem. The framework diagram of the algorithm module is shown in the attached figure. Figure 1 As shown;

[0030] The 3D semi-supervised left atrium segmentation algorithm module can capture rich contextual information between adjacent slices. The 3D semi-supervised left atrium segmentation algorithm module is based on VNet, and its network structure continues the encoder-decoder structure of UNet mentioned above, and draws on the idea of ​​jump connection of UNet to supplement the lost spatial detail information.

[0031] The three-dimensional semi-supervised left atrium segmentation algorithm module is suitable for the semi-supervised left atrium segmentation task. The module uses Monte Carlo dropout to estimate uncertainty to retain pseudo-labels with high confidence, and at the same time enhances the module's learning ability of the overall characteristics of the left atrium by incorporating prior features. The introduction of boundary loss function and false positive rate loss function improves the module's segmentation accuracy for boundaries and subtle structures, and uses consistency regularization based on combined enhancement to enhance the model's anti-interference ability; the three-dimensional semi-supervised left atrium segmentation algorithm module proposes an integrated framework consisting of four stages. The four stages of the algorithm module include: supervised pre-training that integrates anatomical structures, generating high-quality pseudo-labels, consistency regularization based on combined enhancement, and high-quality pseudo-label optimization model; the framework diagram of the three-dimensional semi-supervised left atrium segmentation algorithm module is shown in the attached figure. Figure 2 shown.

[0032] The whole heart segmentation algorithm module is embedded in a weighted axial attention module, which decomposes attention along the width axis and the height axis, and controls its learning ability by adjusting the relative position encoding weights;

[0033] The whole heart segmentation algorithm module also includes a target domain. The pseudo labels required by the target domain are generated by the unsupervised teacher-student model. At the same time, the Hausdorff distance and Dice value formulas as well as the newly proposed pseudo loss are used to control the quality of the pseudo labels actually involved in the training. Only pseudo labels that meet the conditions can participate in the training as shape constraints. The whole heart segmentation algorithm module freezes the parameters of the encoder and decoder parts, abandons the meta-learning method in the classifier part, and only uses cross entropy, Dice value and KL divergence for supervision. Finally, the VAE module fits the segmentation prediction and evaluates the segmentation quality through the learned shape distribution information. The whole heart segmentation algorithm module is shown in the attached figure. Figure 3 shown.

[0034] Experimental data

[0035] Left Atrium Segmentation Dataset:

[0036] The images in the dataset were acquired using a Siemens 1.5T or 3T scanner. The patients were scanned approximately 20-25 minutes after the injection of 0.1mmol / kg gadolinium contrast agent. High-resolution left atrium LGE-MRI image capture was achieved through three-dimensional respiratory navigation and inversion recovery prepared gradient echo pulse sequence;

[0037] The spatial resolution of the 3D LGE-MRI scan is 0.625×0.625×0.625mm 3 , the image size is 576×576×88 or 640×640×88 pixels, and each LGE-MRI scan contains 88 slices in the Z-axis direction;

[0038] Three experts in the field of cardiology performed manual segmentation to obtain labels, including the mitral valve, LA appendages, and pulmonary veins; the grayscale value range of the original MRI image was 0-255, and the segmentation label was binary (255 for positive samples and 0 for negative samples).

[0039] Provide 120 multi-modality whole-heart images from multiple centers, including 60 cardiac CT / CTA and 60 3D cardiac MRI covering the entire cardiac substructure;

[0040] The data were collected based on an in vivo clinical setting and used clinically; there were various image qualities in the datasets, some of which were relatively poor; however, when it came to actual clinical use, it was necessary to include these datasets to validate the robustness of the developed algorithms.

[0041] Cardiac CT / CTA data were acquired using conventional cardiac CT angiography; all data covered the entire heart from the upper abdomen to the aortic arch; slices were acquired in axial view with an in-plane resolution of approximately 0.78 × 0.78 mm and an average slice thickness of 1.60 mm; MRI data were acquired using a 3D balanced steady-state free precession (b-SSFP) sequence with an acquisition resolution of approximately 2 mm in each direction and reconstructed (resampled) to approximately 1 mm.

[0042] The dataset is divided into training (20 CT and 20 MRI representatives) and testing (40 CT and 40 MRI) datasets.

[0043] For the training dataset, manual segmentations of seven whole-heart substructures are provided:

[0044] (1) Left ventricular blood cavity (label value 500);

[0045] (2) right ventricular blood cavity (label value 600);

[0046] (3) left atrial blood cavity (label value 420);

[0047] (4) right atrial blood cavity (label value 550);

[0048] (5) left ventricular myocardium (reference numeral 205);

[0049] (6) ascending aorta (label value 820), defined as the aortic trunk from the aortic valve to the level of the atria; for example, the average length of this trunk measured in 25 healthy subjects was approximately 41.9 mm;

[0050] (7) Pulmonary artery (label value 850), defined as the main trunk between the pulmonary valve and the bifurcation, for example, the average length of this trunk measured in 25 healthy subjects is approximately 41.9 mm.

[0051] The large vessels of interest include the ascending aorta and pulmonary artery. Due to different fields of view, the coverage of large vessels will vary in different scans. In order to have a consistent definition between different subjects, the ascending aorta and pulmonary artery are therefore defined as above, while the verification of the whole heart segmentation can focus only on the main trunk of the vessel.

[0052] Experimental environment and training parameters

[0053] The PyTorch framework is used, and the cuda and cudnn libraries are used to accelerate the training process. The server hardware parameters used are as follows: the operating system is Ubuntu 20.04, the memory is 64G, the graphics card is NVIDIA GTX3090Ti, and the video memory is 12G.

[0054] In the preprocessing stage, the MRI images and their labels were randomly scaled with a scaling ratio of (0.5, 1.5), and then the images were flipped horizontally and vertically with a probability of 0.5, and then randomly cropped to a size of 256×256, and finally equalized by a contrast-limited adaptive histogram. The Adam stochastic gradient descent method was used for training, and the initial learning rate was set to 0.001. An exponential decay was performed every 5 iterations. The learning rate was adjusted, momentum was set to 0.9, weight_decay was set to 0.001, and batchsize was set to 4. In order to quantitatively evaluate the performance of the segmentation algorithm, the Dice coefficient, Jaccard coefficient, Hausdorff distance, and ASD distance between the model segmentation results and the labels were calculated as evaluation indicators.

[0055] The three-dimensional semi-supervised left atrium segmentation algorithm integrating anatomical structure used 100 LGE-MRI data sets and divided the data sets into 80 training data and 20 test data in a ratio of 8:2.

[0056] In the preprocessing stage, the original data is first randomly cropped to obtain a 3D block of size 112×112×80, and then the 3D blocks are randomly flipped 90 degrees, 180 degrees, or 270 degrees along the axial plane to enhance data diversity and reduce the risk of model overfitting. Finally, these 3D blocks are normalized to have zero mean and unit variance. The Adam optimizer is used in the training process, and the number of training iterations is 6K; among them, the momentum is set to 0.9, weight_decay is 0.001, the initial learning rate is 0.01, and the learning rate decays by 10% every 2.5K iterations.

[0057] The whole heart segmentation based on unsupervised learning is first optimized using the SGD optimizer with momentum set to 0.9 and weight decay to 1e-4.

[0058] Experimental results and analysis

[0059] The two-dimensional fully supervised left atrium segmentation algorithm module based on axial attention is superior to other comparison modules in terms of Dice and Jaccard indicators, as shown in Table 1. This shows that the module shows high accuracy in the left atrium segmentation task.

[0060] Table 1 Quantitative results of the comparison between the two-dimensional segmentation method and other methods

[0061]

[0062] Note: “—” indicates that the content is not found in the literature.

[0063] In the table, through a detailed comparison of four evaluation indicators, this method is significantly better than the existing classic segmentation network model in all aspects. Specifically, in terms of the Dice coefficient, this method is 5.4% higher than PSPNet, 3.7% higher than UNet, 4.87% higher than DeepLabV3, and 2.6% higher than SegNet.

[0064] Compared with the baseline model UNet, this method has achieved significant improvement, which is mainly due to the introduction of four modules and the optimized loss function. In addition, compared with the methods recently proposed on this dataset, although it is slightly inferior to the method of Uslu et al. in terms of ASD distance and HD distance, it performs well in Dice coefficient and Jaccard coefficient, which is 1.3% higher than the Dice coefficient of Uslu et al. and 1.9% higher than the Jaccard coefficient of Uslu et al., showing the superiority of this method in the left atrium segmentation task, and at the same time indicating that there is still room for further optimization of the model's boundary processing method.

[0065] As attached Figure 4 The segmentation results of the left atrium starting segment, main segment, and ending segment in the test set are shown. The two columns on the left are the left atrium starting segment, the two columns in the middle are the left atrium main segment, and the two columns on the right are the ending segment. The first row is the original image, the second row is the label, and the third to seventh rows show the segmentation results of different comparison models respectively. The last row is the segmentation result of the method in this chapter.

[0066] In the segmentation process of the starting segment of the left atrium, the differences shown by various models are not obvious, and the method in this chapter is closer to the label in shape. For the main segment of the left atrium, the left atrium is usually connected to structures such as the pulmonary veins and the mitral valve, which makes the shape of the left atrium more complex than the starting segment and the topological structure changes more dramatically. This method shows good ability in capturing structures such as the pulmonary veins, while other methods generally have over-segmentation or under-segmentation phenomena, making it difficult to accurately capture changes in the morphology of the left atrium. In the ending segment of the left atrium, this method shows good detail processing capabilities and achieves accurate tracking of the splitting and fusion of multiple parts. The advantages of the method proposed in this chapter in the segmentation of the starting segment, main segment, and ending segment of the left atrium may be due to the focus on global information, thus having higher accuracy in the segmentation of the pulmonary veins and other small structures.

[0067] In the three-dimensional semi-supervised left atrium segmentation algorithm module integrating anatomical structure, Table 2 shows the quantitative results of the comparison between our proposed three-dimensional semi-supervised left atrium segmentation algorithm module and the other five semi-supervised segmentation modules.

[0068] Table 2 Quantitative results of the comparison between the three-dimensional segmentation method and the five methods

[0069]

[0070]

[0071] Compared with other semi-supervised segmentation modules, this module achieves better results in all four metrics. In order to evaluate the performance of the model under different ratios of labeled data to unlabeled data, two groups of experiments were set up, 10% labeled data, 90% unlabeled data and 20% labeled data, 80% unlabeled data. With the increase of labeled data, the segmentation accuracy of the module is also improved accordingly. Under the setting of 20% labeled data and 80% unlabeled data, the model achieves the best results in terms of Dice coefficient and Jaccard coefficient.

[0072] The experimental results of the whole heart segmentation algorithm module based on unsupervised learning are shown in Table 3. Only the whole heart segmentation algorithm module proposed in the present invention is relatively accurate in the segmentation of the left ventricular myocardium Myo, left atrium LA and right atrium RA; the Dice value and Jaccard coefficient of the left ventricular myocardium are 0.765 and 0.701 respectively; the Dice value and Jaccard coefficient of the left atrium are 0.752 and 0.671 respectively; the Dice value and Jaccard coefficient of the right atrium RA are 0.723 and 0.633 respectively; in addition, the Dice value and Jaccard coefficient of UL-TransUNet in WHS are 0.694 and 0.632 respectively.

[0073] Table 3 Experimental results of the whole heart segmentation algorithm module dataset

[0074]

[0075] The experimental results prove the effectiveness and feasibility of the whole heart segmentation algorithm module proposed in the present invention. In addition, the unsupervised learning method reduces the demand for labeled samples in SSL-TransUNet, and to a certain extent reduces the dependence of the few-sample learning method on the task structure and data set due to meta-learning, thereby improving the accuracy and versatility of the model. In general, the improvement of the SSL-TransUNet network model made by the present invention has achieved excellent results.

[0076] The generated prediction results are processed using CRF in the present invention, as shown in the attached Figure 5 Some 2D final segmentation results for unsupervised whole heart segmentation;

[0077] The left image is the original heart image, the middle image is the manually annotated real label, and the right side is the segmentation result of the method of the present invention. The various substructures of the heart are represented by different colors by the present invention, where red represents LV, green represents MYO, light blue represents LA, yellow represents AORTA, purple represents RA, white represents PA, and dark blue represents RV. It can be seen from the segmentation results that the UL-TransUNet proposed by the present invention can achieve whole heart segmentation more accurately, further verifying the effectiveness of the research content of the present invention.

[0078] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can make equivalent replacements or changes according to the technical scheme and inventive concept of the present invention within the technical scope disclosed by the present invention, which should be covered by the protection scope of the present invention.

Claims

1. A cardiac AI system based on deep learning, comprising a two-dimensional fully supervised left atrium segmentation algorithm module, a three-dimensional semi-supervised left atrium segmentation algorithm module, and a whole heart segmentation algorithm module, characterized in that: The two-dimensional fully supervised left atrium segmentation algorithm module is based on the UNet model and combines the Transformer and residual network to capture the semantic information in the image in a larger field of view, thereby improving the segmentation accuracy of the network; the two-dimensional fully supervised left atrium segmentation algorithm module includes: a double-layer residual convolution module, a task-driven feature enhancement module, a multi-feature fusion module and a feature coupling module; The three-dimensional semi-supervised left atrium segmentation algorithm module can capture rich contextual information between adjacent slices; the three-dimensional semi-supervised left atrium segmentation algorithm module is based on VNet, and its network structure continues the encoder-decoder structure of UNet mentioned above, and draws on the jump connection idea of ​​UNet to supplement the lost spatial detail information; The whole heart segmentation algorithm module is embedded in a weighted axial attention module to decompose the attention along the width axis and the height axis, and control its learning ability by adjusting the relative position encoding weights.

2. A deep learning-based cardiac AI system according to claim 1, characterized in that: The double-layer residual convolution module replaces the ordinary convolution in the original UNet model, enhancing the expressive power of the model so that it can better learn the complex features in the image.

3. A deep learning-based cardiac AI system according to claim 1, characterized in that: The task-driven feature enhancement module focuses on channel features to improve the distinguishability of features and further enhances the model's ability to distinguish features irrelevant to the left atrium.

4. The deep learning-based cardiac AI system according to claim 1, characterized in that: The multi-feature fusion module realizes the integration of the low-level features of the encoder and the high-level features of the decoder, thereby fully mining the local information and global information in the image.

5. The deep learning-based cardiac AI system according to claim 1, characterized in that: The feature coupling module overcomes the difference in features extracted by convolution and Transformer, and further optimizes the segmentation effect.

6. A cardiac AI system based on deep learning according to claim 1, characterized in that: The three-dimensional semi-supervised left atrium segmentation algorithm module is suitable for the semi-supervised left atrium segmentation task. The module uses Monte Carlo dropout to estimate uncertainty to retain pseudo labels with high confidence, and at the same time enhances the module's learning ability for the overall characteristics of the left atrium by incorporating prior features. The introduction of boundary loss function and false positive rate loss function improves the module's segmentation accuracy for boundaries and subtle structures, and uses consistency regularization based on combined enhancement to enhance the model's anti-interference ability.

7. The deep learning-based cardiac AI system according to claim 1, characterized in that: The three-dimensional semi-supervised left atrium segmentation algorithm module proposes an integrated framework consisting of four stages. The four stages of the algorithm module include: supervised pre-training integrating anatomical structures, generating high-quality pseudo labels, consistency regularization based on combination enhancement, and high-quality pseudo label optimization model.

8. The deep learning-based cardiac AI system according to claim 1, characterized in that: The whole heart segmentation algorithm module also includes a target domain. The pseudo labels required by the target domain are generated by an unsupervised teacher-student model. At the same time, the Hausdorff distance and Dice value formulas and the newly proposed pseudo loss are used to control the quality of the pseudo labels actually participating in the training. Only pseudo labels that meet the conditions can participate in the training as shape constraints.

9. The deep learning-based cardiac AI system according to claim 1, characterized in that: The whole heart segmentation algorithm module freezes the parameters of the encoder and decoder parts, abandons the meta-learning method in the classifier part, and only uses cross entropy, Dice value and KL divergence for supervision; finally, the VAE module fits the segmentation prediction and evaluates the segmentation quality through the learned shape distribution information.

Images