A cardiac magnetic resonance image semi-supervised segmentation method and device based on multi-view consistency, equipment and storage medium

Abstract

The present application relates to the technical field of medical image segmentation and semi-supervised learning, in particular to a cardiac magnetic resonance image semi-supervised segmentation method, device and equipment based on multi-view consistency and storage medium, comprising the following steps: step one: establishing a segmentation model, encoding the source image from the forward and reverse double-view to extract image features, obtaining the forward feature map and the reverse feature map; step two: using a view fusion module to fuse the image features under the forward and reverse double-view to obtain the edge information enhanced fusion feature map; the present application can effectively segment the cardiac magnetic resonance image, can provide more accurate left ventricle, right ventricle and myocardial region segmentation images for doctors, and assist doctors in diagnosing heart diseases.

Description

Technical Field

[0001] The present invention relates to the technical field of medical image segmentation and semi-supervised learning, and specifically to a method, apparatus, device and storage medium for semi-supervised segmentation of cardiac magnetic resonance images based on multi-view consistency. Background Art

[0002] Cardiac magnetic resonance imaging (MRI) is the most direct and reliable method for diagnosing cardiovascular diseases. However, manual interpretation of images is time-consuming, labor-intensive, and inefficient. Therefore, automatic segmentation of the left ventricle, right ventricle, and myocardial regions in MRI images can help assess cardiac function, such as measuring ventricular contraction and relaxation, which is crucial for diagnosing a variety of cardiovascular diseases, such as heart failure, coronary artery disease, and valvular disease.

[0003] The advantage of deep learning in medical image segmentation is that it can automatically extract rich features from the source image, including low-level features and high-level abstract features, without the need for manually designed complex feature extractors. It is also good at processing complex and nonlinear image features, which is conducive to the automatic segmentation of cardiac organs with complex and variable shapes, sizes and internal structures. However, the segmentation performance of traditional deep convolutional neural network models depends on a large amount of high-quality data labels, but manual labeling of cardiac images is time-consuming, labor-intensive and costly, resulting in too little labeled data available for clinical use. Currently, semi-supervised learning technology can effectively solve the problem of medical image segmentation where clinical labeled data is scarce. Semi-supervised segmentation methods can be roughly divided into two categories, namely pseudo-label-based methods and consistency regularization methods. The former effectively expands training data through high-confidence labels, while the latter improves model generalization performance by constraining the model's prediction results.

[0004] Existing semi-supervised segmentation methods only consider learning the consistency of prediction results from the same perspective. In other words, consistency regularization focuses on extracting the same features to obtain two similar predictions. This training method can easily lead to the decoder being biased towards incorrect predictions.

[0005] In summary, how to design a segmentation method based on semi-supervised learning technology, adopt multi-view consistency learning ideas, construct a semi-supervised semantic segmentation model for cardiac magnetic resonance images, improve the utilization of unlabeled data under the conditions of a small number of labeled images and a large number of unlabeled images, so as to reduce the error bias problem of model learning and realize accurate segmentation of the heart part in magnetic resonance images has become an urgent problem to be solved. Summary of the Invention

[0006] The object of the present invention is to provide a method, apparatus, device and storage medium for semi-supervised segmentation of cardiac magnetic resonance images based on multi-view consistency to solve the above problems.

[0007] In order to achieve the above objectives, the following technical solutions are provided:

[0008] A semi-supervised segmentation method for cardiac magnetic resonance images based on multi-view consistency includes the following steps:

[0009] Step 1: Establish a segmentation model, encode the source image from the forward and reverse dual perspectives to extract image features, and obtain forward feature maps and reverse feature maps;

[0010] Step 2: Use the perspective fusion module to fuse the image features under the forward and reverse perspectives to obtain a fused feature map with enhanced edge information;

[0011] Step 3: Use different decoding strategies to decode the forward feature map, reverse feature map and fusion feature map respectively to obtain the segmentation prediction map;

[0012] Step 4: Obtain a high-confidence segmentation pseudo mask for the unlabeled image based on directional consistency;

[0013] Step 5: Train model parameters based on codec feature consistency constraints and multi-view feature supervision;

[0014] Step 6: Input the image into the trained segmentation model. The model automatically learns image features from multiple perspectives and outputs a segmentation visualization result of the heart area in the magnetic resonance image.

[0015] Preferably,

[0016] In step 1, feature extraction uses the Unet encoding structure, and the features extracted from the forward perspective are , the features extracted from the reverse perspective are .

[0017] Preferably,

[0018] Step 2 is as follows:

[0019] S1: In the view fusion module, the features extracted from the forward view are reversed through the REV operation, and then the features extracted from the reverse view are multiplied by a dot product operation;

[0020] The REV operation is as follows:

[0021] The formula for calculating the reverse feature map under the forward perspective is as follows (1):

[0022] =Softmax(- ) (1)

[0023] Where, is the feature extracted from the forward view, The Softmax function is a normalized exponential function for the feature map after reverse for the forward view angle;

[0024] S2: subtracting the point multiplication result feature map of and pixel by pixel to obtain the fused feature for improving the performance of edge segmentation, as shown in the following formula (2):

[0025] (2)

[0026] In the formula, the fused feature, is the feature extracted from the reverse view angle.

[0027] Preferably, the Unet decoding structure is used for decoding in step three to obtain the segmentation prediction maps under three view angles of forward, reverse and fusion, i.e. the forward segmentation prediction map , the reverse segmentation prediction map , and the fusion segmentation prediction map .

[0028] Preferably,

[0029] Step four is specifically:

[0030] The prediction probabilities of each pixel point in the prediction maps and output by the two decoders of the forward and reverse view angles are compared with the threshold value at the same time, and when , the prediction probabilities of the same pixel point belonging to the segmentation region are both greater than the threshold value, it is the consistent part, otherwise it is the inconsistent part; when the result is the consistent part, the pixel directly participates in the unsupervised loss calculation, and when the result is the inconsistent part, the direction consistency loss function is used to further decouple the data;

[0031] In the formula, the formulas of the segmentation mask and the threshold value are shown in formulas (3)-(6):

[0032] (3)

[0033] (4)

[0034] (5)

[0035] (6)

[0036] In the formula, the segmentation mask of the consistent part, ​Represents the segmentation mask of the inconsistent part, γ is the threshold, λ is the weight size, , t is the number of iterations, B represents the number of training images, C represents the number of channels; P represents the forward and reverse prediction images, 、 Represent the thresholds for forward and reverse prediction graph calculations, respectively;

[0037] Among them, the directional consistency loss function further decouples the operations of inconsistent data as follows:

[0038] Calculate the positive and negative high confidence areas in the inconsistent parts 、 and positive and reverse low confidence regions , , the formula is shown in formula (7)-formula (14):

[0039] (7)

[0040] (8)

[0041] (9)

[0042] (10)

[0043] (11)

[0044] (12)

[0045] (13)

[0046] (14)

[0047] Where, Segmentation masks representing high confidence regions where the forward decoder prediction map is inconsistent, Segmentation masks representing high confidence regions where the backward decoder prediction map is inconsistent, Represents the segmentation prediction probability map of the positive inconsistent part, The segmentation prediction probability map representing the reverse inconsistent part;

[0048] The direction consistency loss is defined as follows (15):

[0049] (15)

[0050] Where L2 is the Euclidean distance loss function, and the detach function prevents high-confidence regions from participating in gradient backpropagation, that is, it ensures that high-confidence regions do not move and only moves low-confidence regions toward high-confidence regions.

[0051] Preferably,

[0052] Step 5 is as follows:

[0053] Channel compression is performed on the feature map to reduce the pressure on model parameter training. The formula is as follows (16):

[0054] (16)

[0055] Where, represents the feature map of the mth layer, express The i-th slice in the channel dimension, represents the corresponding mapping result, C represents the number of channels; p is a hyperparameter, and p=2 is taken in this formula;

[0056] Encoder-decoder feature consistency loss The definition is as shown in the following formula (17):

[0057] (17)

[0058] Where N is The number of pixels, n is the number of network layers, and represents the i-th pixel in the m-th feature map of the encoder and decoder respectively;

[0059] Multi-view feature supervision loss The definition is as shown in the following formula (18):

[0060] (18)

[0061] in, 、 、 They are the predicted segmentation images after forward, reverse and fusion view feature decoding, is the label image corresponding to the source image, and the loss function of the model training adopts the Dice loss function;

[0062] The definition of the Dice loss function is as follows (19):

[0063] (19)

[0064] in, , is the set of true values, , is the set of segmentation prediction results.

[0065] The present invention also provides a device for semi-supervised segmentation of cardiac magnetic resonance images based on multi-view consistency, which is applied to the aforementioned method for semi-supervised segmentation of cardiac magnetic resonance images based on multi-view consistency.

[0066] include:

[0067] A model building unit, used to build a segmentation model;

[0068] A training unit, used to train the segmentation model;

[0069] Output unit, used to output visualization results.

[0070] The present invention also provides a semi-supervised segmentation device for cardiac magnetic resonance images based on multi-view consistency.

[0071] The device includes: a processor and a memory storing computer program instructions;

[0072] When the processor executes the computer program instructions, the semi-supervised segmentation method of cardiac magnetic resonance images based on multi-view consistency is implemented.

[0073] The present invention also provides a computer-readable storage medium,

[0074] The computer-readable storage medium stores computer program instructions, which, when executed by a processor, implement the method for semi-supervised segmentation of cardiac magnetic resonance images based on multi-view consistency.

[0075] The beneficial effects of the present invention are:

[0076] 1. The present invention first encodes the source image from the forward and reverse dual perspectives to extract image features to obtain a dual-perspective feature map, then uses a perspective fusion module to fuse the image features under the forward and reverse dual perspectives to further obtain a feature map with enhanced edge information, adopts different decoding strategies to decode the feature maps obtained under the forward, reverse and fusion perspectives respectively and obtain a segmentation prediction map, obtains a high-confidence segmentation pseudo-mask for the unlabeled image based on directional consistency, trains model parameters based on encoding and decoding feature consistency constraints and multi-perspective feature supervision, and finally loads the trained model and parameters to achieve accurate segmentation of the heart part in the magnetic resonance image; the present invention can effectively segment cardiac magnetic resonance images, and can provide doctors with more accurate segmentation images of the left ventricle, right ventricle and myocardial area, to assist doctors in diagnosing heart diseases. BRIEF DESCRIPTION OF THE DRAWINGS

[0077] Figure 1 is a flow chart of the method of the present invention;

[0078] Figure 2 It is a network structure diagram of the present invention;

[0079] Figure 3 Comparison of the segmentation results of the present invention (MVNet) after training the model with 10% labels on the ACDC dataset and other segmentation methods (DCNet). DETAILED DESCRIPTION

[0080] The following will be combined with the embodiments of the present invention to clearly and completely describe the technical solutions of the schematic diagrams in the embodiments of the present invention. Obviously, the embodiments described are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention. Example

[0081] like Figure 1-3 As shown:

[0082] This embodiment provides a semi-supervised segmentation method for cardiac magnetic resonance images based on multi-view consistency, comprising the following steps:

[0083] Step 1: Establish a segmentation model, encode the source image from the forward and reverse dual perspectives to extract image features, and obtain forward feature maps and reverse feature maps;

[0084] Step 2: Use the perspective fusion module to fuse the image features under the forward and reverse perspectives to obtain a fused feature map with enhanced edge information;

[0085] Step 3: Use different decoding strategies to decode the forward feature map, reverse feature map and fusion feature map respectively to obtain the segmentation prediction map;

[0086] Step 4: Obtain a high-confidence segmentation pseudo mask for the unlabeled image based on directional consistency;

[0087] Step 5: Train model parameters based on codec feature consistency constraints and multi-view feature supervision;

[0088] Step 6: Input the image into the trained segmentation model. The model automatically learns image features from multiple viewpoints and outputs a segmentation visualization of the heart in the MRI image.

[0089] In some embodiments,

[0090] In step 1, feature extraction uses the Unet encoding structure, which is divided into forward (predicting target class) and reverse (predicting background class) perspectives. The features extracted from the forward perspective are , the features extracted from the reverse perspective are ;

[0091] In some embodiments,

[0092] Step 2 is as follows:

[0093] S1: In the view fusion module, the features extracted from the forward view are reversed through the REV operation, and then the features extracted from the reverse view are multiplied by a dot product operation;

[0094] In step 2, a perspective fusion module is used to fuse the image features from both the forward and reverse perspectives to obtain a feature map with enhanced edge information. The perspective fusion operation is designed to enhance the model's ability to extract features from the heart boundary and improve segmentation accuracy at the boundary. In the perspective fusion module, the features extracted from the forward perspective are first reversed through the REV operation, and then dot-multiplied with the features extracted from the reverse perspective, thereby enhancing the consistent features extracted from both the forward and reverse perspectives and weakening the inconsistent features.

[0095] The REV operation is as follows:

[0096] The formula for calculating the reverse feature map under the forward perspective is as follows (1):

[0097] =Softmax(- ) (1)

[0098] Where, is the feature extracted from the forward view, is the feature map after inversion from the forward perspective, and the Softmax function is a normalized exponential function;

[0099] S2: and and The pixel-by-pixel subtraction of the dot multiplication result feature map is performed. Under this operation, the influence of the consistent part in the feature map will be weakened, while the inconsistent part, that is, the edge part, will be retained, thereby improving the performance of edge segmentation. The fused feature is calculated and the formula is shown in the following formula (2):

[0100] (2)

[0101] Where, Represents the fused features, Features extracted for reverse viewpoints;

[0102] In some embodiments,

[0103] In step 3, the Unet decoding structure is used for decoding to obtain the segmentation prediction maps under the three perspectives of forward, reverse and fusion, namely the forward segmentation prediction map , reverse segmentation prediction map , fusion segmentation prediction map ;

[0104] In some embodiments,

[0105] Step four is specifically:

[0106] The prediction probabilities of each pixel point in the corresponding prediction maps of the two decoders of the forward and reverse view angles are compared with a threshold value simultaneously, and when , the prediction probabilities of the same pixel point belonging to the segmentation region are both greater than the threshold value, it is a consistent part, otherwise it is an inconsistent part; when the result is the consistent part, the pixel directly participates in the unsupervised loss calculation, and when the result is the inconsistent part, the direction consistency loss function is used to further decouple the data. ,

[0107] In step four, the high-confidence segmentation pseudo mask of the unlabeled image is obtained based on the direction consistency, and the prediction probabilities of each pixel point in the corresponding prediction maps of the two decoders of the forward and reverse view angles are compared with a threshold value simultaneously, and when , the prediction probabilities of the same pixel point belonging to the segmentation region are both greater than the threshold value, it is a consistent part, otherwise it is an inconsistent part; the consistent part represents that the prediction probability of the pixel is highly correct and can directly participate in the unsupervised loss calculation, and the inconsistent part represents that the prediction result of the part is unreliable, so for the unreliable part, the method uses the direction consistency loss function to further decouple the data, and selects the part close to the high-density region with relatively high confidence in the unreliable inconsistent part, so that the low-confidence part that is easy to appear in the decision boundary is close to the high-confidence part, so that the decision boundary is more clear.

[0108] Wherein, the formulas of the segmentation mask and the threshold value are shown in equations (3)-(6):

[0109] (3)

[0110] (4)

[0111] (5)

[0112] (6)

[0113] In the formula, the segmentation mask of the consistent part is represented by , the segmentation mask of the inconsistent part is represented by , γ is the threshold value, λ is the weight size, , t is the iteration number, B represents the number of training images, C represents the channel number; P represents the forward and reverse prediction maps, ,​ Represent the thresholds for forward and reverse prediction graph calculations, respectively;

[0114] Among them, the directional consistency loss function further decouples the operations of inconsistent data as follows:

[0115] Calculate the positive and negative high confidence areas in the inconsistent parts 、 and positive and reverse low confidence regions , , the formula is shown in formula (7)-formula (14):

[0116] (7)

[0117] (8)

[0118] (9)

[0119] (10)

[0120] (11)

[0121] (12)

[0122] (13)

[0123] (14)

[0124] Where, Segmentation masks representing high confidence regions where the forward decoder prediction map is inconsistent, Segmentation masks representing high confidence regions where the backward decoder prediction map is inconsistent, Represents the segmentation prediction probability map of the positive inconsistent part, The segmentation prediction probability map representing the reverse inconsistent part;

[0125] The direction consistency loss is defined as follows (15):

[0126] (15)

[0127] Where L2 is the Euclidean distance loss function, and the detach function prevents high-confidence regions from participating in gradient backpropagation, that is, it ensures that high-confidence regions do not move, and only moves low-confidence regions toward high-confidence regions.

[0128] In some embodiments,

[0129] Step 5 is as follows:

[0130] Perform channel compression on feature maps to reduce the pressure on model parameter training;

[0131] Its main purpose is to constrain the consistency of encoding and decoding features, reduce the pressure of model parameter training, and perform channel compression on feature maps;

[0132] The formula is as follows (16):

[0133] (16)

[0134] Where, represents the feature map of the mth layer, express The i-th slice in the channel dimension, represents the corresponding mapping result, C represents the number of channels; p is a hyperparameter, and p=2 is taken in this formula;

[0135] Encoder-decoder feature consistency loss The definition is as shown in the following formula (17):

[0136] (17)

[0137] Where N is The number of pixels, n is the number of network layers, and represents the i-th pixel in the m-th feature map of the encoder and decoder respectively;

[0138] Multi-view feature supervision loss The definition is as shown in the following formula (18):

[0139] (18)

[0140] in, 、 、 They are the predicted segmentation images after forward, reverse and fusion view feature decoding, is the label image corresponding to the source image, and the loss function of the model training adopts the Dice loss function;

[0141] The definition of the Dice loss function is as follows (19):

[0142] (19)

[0143] in, , is the set of true values, , is the set of segmentation prediction results;

[0144] The present invention also provides a device for semi-supervised segmentation of cardiac magnetic resonance images based on multi-view consistency, which is applied to the above-mentioned method for semi-supervised segmentation of cardiac magnetic resonance images based on multi-view consistency.

[0145] include:

[0146] A model building unit, used to build a segmentation model;

[0147] A training unit, used to train the segmentation model;

[0148] Output unit, used to output visualization results;

[0149] The present invention also provides a semi-supervised segmentation device for cardiac magnetic resonance images based on multi-view consistency.

[0150] The above-mentioned device includes: a processor and a memory storing computer program instructions;

[0151] The processor implements the aforementioned method for semi-supervised segmentation of cardiac magnetic resonance images based on multi-view consistency when executing the computer program instructions;

[0152] The present invention also provides a computer-readable storage medium,

[0153] The computer-readable storage medium stores computer program instructions, which, when executed by a processor, implement the aforementioned method for semi-supervised segmentation of cardiac magnetic resonance images based on multi-view consistency;

[0154] Method comparison:

[0155] In order to verify the effectiveness and superiority of this method, a method comparison was conducted. SupOnly, URPC, CPS, DTC, MCNet, SSNet, DCNet and other methods were selected for parallel comparison. Each method was segmented on 10% labeled images of the ACDC dataset for quantitative comparison. Dice, Jaccard, 95HD and ASD of each method were calculated respectively. The segmentation results are shown in the figure below. Figure 3 The comparison results are shown in Table 1. As can be seen from the chart, the Dice and Jaccard indicators of the method in this embodiment are the highest among the methods of similar comparison. It can be seen that the predicted results and the true results of the segmentation method in this embodiment have the highest overlap and the highest similarity. The 95HD and ASD indicators of the method in this embodiment are the lowest among the methods of similar comparison. It can be seen that the distance between the predicted result set and the true result set of the segmentation method in this embodiment is the smallest and the similarity is the highest.

[0156] Table 1. Comparison results of image segmentation methods

[0157] .

Claims

1. A semi-supervised segmentation method for cardiac magnetic resonance images based on multi-view consistency, characterized in that: The steps include: Step 1: Establish a segmentation model, encode the source image from the forward and reverse dual perspectives to extract image features, and obtain forward feature maps and reverse feature maps; Step 2: Use the perspective fusion module to fuse the image features under the forward and reverse perspectives to obtain a fused feature map with enhanced edge information; Step 3: Use different decoding strategies to decode the forward feature map, reverse feature map and fusion feature map respectively to obtain the segmentation prediction map; Step 4: Obtain a high-confidence segmentation pseudo mask for the unlabeled image based on directional consistency; Step 5: Train model parameters based on codec feature consistency constraints and multi-view feature supervision; Step 6: Input the image into the trained segmentation model. The model automatically learns image features from multiple perspectives and outputs a segmentation visualization result of the heart area in the magnetic resonance image.

2. The method for semi-supervised segmentation of cardiac magnetic resonance images based on multi-view consistency according to claim 1, characterized in that: In step 1, feature extraction uses the Unet encoding structure, and the features extracted from the forward perspective are , the features extracted from the reverse perspective are .

3. The method for semi-supervised segmentation of cardiac magnetic resonance images based on multi-view consistency according to claim 2, characterized in that: Step 2 is as follows: S1: In the view fusion module, the features extracted from the forward view are reversed through the REV operation, and then the features extracted from the reverse view are multiplied by a dot product operation; The REV operation is as follows: The formula for calculating the reverse feature map under the forward perspective is as follows (1): =Softmax(- ) (1) Where, is the feature extracted from the forward view, is the feature map after inversion from the forward perspective, and the Softmax function is a normalized exponential function; S2: and and The feature map of the dot product result is subtracted pixel by pixel to calculate the fused features to improve the performance of edge segmentation. The formula is shown in the following formula (2): （2） Where, Represents the fused features, Features extracted for reverse viewpoint.

4. The method for semi-supervised segmentation of cardiac magnetic resonance images based on multi-view consistency according to claim 3, characterized in that: In step 3, the Unet decoding structure is used for decoding to obtain the segmentation prediction maps under the three perspectives of forward, reverse and fusion, namely the forward segmentation prediction map , reverse segmentation prediction map , fusion segmentation prediction map .

5. The method for semi-supervised segmentation of cardiac magnetic resonance images based on multi-view consistency according to claim 4, characterized in that: Step 4 is as follows: The predicted images output by the two decoders of the forward and reverse view and The predicted probability of each pixel in the corresponding is compared with the threshold at the same time. , If the probability of predicting the same pixel belonging to the segmented area is greater than the threshold, it is a consistent part, otherwise it is an inconsistent part; when the result is a consistent part, the pixel is directly involved in the unsupervised loss calculation, and when the result is an inconsistent part, the direction consistency loss function is used to further decouple the data; Among them, the formulas for segmentation mask and threshold are shown in Equation (3) to Equation (6): （3） （4） （5） （6） Where, The segmentation mask representing the consistent part, Represents the segmentation mask of the inconsistent part, γ is the threshold, λ is the weight size, , t is the number of iterations, B represents the number of training images, C represents the number of channels; P represents the forward and reverse prediction images, 、 Represent the thresholds for forward and reverse prediction graph calculations, respectively; Among them, the directional consistency loss function further decouples the operations of inconsistent data as follows: Calculate the positive and negative high confidence areas in the inconsistent parts 、 and positive and reverse low confidence regions , , the formula is shown in formula (7)-formula (14): （7） （8） （9） （10） （11） （12） （13） （14） Where, Segmentation masks representing high confidence regions where the forward decoder prediction map is inconsistent, Segmentation masks representing high confidence regions where the backward decoder prediction map is inconsistent, Represents the segmentation prediction probability map of the positive inconsistent part, The segmentation prediction probability map representing the reverse inconsistent part; The direction consistency loss is defined as follows (15): （15） Where L2 is the Euclidean distance loss function, and the detach function prevents high-confidence regions from participating in gradient backpropagation, that is, it ensures that high-confidence regions do not move and only moves low-confidence regions toward high-confidence regions.

6. The method for semi-supervised segmentation of cardiac magnetic resonance images based on multi-view consistency according to claim 5, characterized in that: Step 5 is as follows: Channel compression is performed on the feature map to reduce the pressure on model parameter training. The formula is as follows (16): （16） Where, represents the feature map of the mth layer, express The i-th slice in the channel dimension, represents the corresponding mapping result, C represents the number of channels; p is a hyperparameter, and p=2 is taken in this formula; Encoder-decoder feature consistency loss The definition is as shown in the following formula (17): （17） Where N is The number of pixels, n is the number of network layers, and represents the i-th pixel in the m-th feature map of the encoder and decoder respectively; Multi-view feature supervision loss The definition is as shown in the following formula (18): （18） in, 、 、 They are the predicted segmentation images after forward, reverse and fusion view feature decoding, is the label image corresponding to the source image, and the loss function of the model training adopts the Dice loss function; The definition of the Dice loss function is as follows (19): （19） in, , is the set of true values, , is the set of segmentation prediction results.

7. A device for semi-supervised segmentation of cardiac magnetic resonance images based on multi-view consistency, applied to the method for semi-supervised segmentation of cardiac magnetic resonance images based on multi-view consistency according to claim 6, characterized in that: include: A model building unit, used to build a segmentation model; A training unit, used to train the segmentation model; Output unit, used to output visualization results.

8. A device for semi-supervised segmentation of cardiac magnetic resonance images based on multi-view consistency, characterized in that: The device includes: a processor and a memory storing computer program instructions; When the processor executes the computer program instructions, it implements the semi-supervised segmentation method of cardiac magnetic resonance images based on multi-view consistency as described in any one of claims 1 to 6.

9. A computer-readable storage medium, characterized in that The computer-readable storage medium stores computer program instructions, which, when executed by a processor, implement the semi-supervised segmentation method for cardiac magnetic resonance images based on multi-view consistency as described in any one of claims 1 to 6.

Images