Weakly semi-supervised cardiac MRI segmentation method based on pseudo-label frequency domain dynamic hybrid supervision

Abstract

The application discloses a kind of weak semi-supervised cardiac MRI segmentation methods based on pseudo-label frequency domain dynamic hybrid supervision, belong to medical image segmentation technical field, including the following steps: S1, a small amount of cardiac MRI image in training set is provided with scribble label;S2, establish main segmentation model and auxiliary segmentation model, output corresponding predicted value;S3, generate supervision pseudo-label using Haar wavelet transform and frequency domain dynamic hybrid method;S4, combined with partial cross-entropy loss constraint strategy establishes total loss function to optimize training three segmentation models, realize the segmentation of cardiac MRI image.The application generates reliable supervision signal by pseudo-label frequency domain dynamic hybrid supervision method, significantly improves the utilization efficiency of unmarked data using a small amount of scribble label and a large amount of unlabelled data;Using weak semi-supervised strategy, only a small amount of scribble label can realize high-quality segmentation, reduce the dependence on pixel-by-pixel accurate annotation, solve the problem of high medical image annotation cost.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of medical image segmentation, in particular to a weak semi-supervised cardiac MRI segmentation method based on pseudo-label frequency domain dynamic hybrid supervision. BACKGROUND

[0002] The economic development has significantly improved the material living standard of residents in China, but at the same time, it has also brought about excessive daily dietary sodium intake, excessive daily alcohol intake and other unhealthy factors, leading to a continuous increase in the prevalence of cardiovascular diseases. Statistics show that the mortality rate of cardiovascular diseases accounts for the first place in the composition ratio of disease deaths in China, which has brought a huge economic burden to the society. Therefore, how to improve the diagnosis rate and cure rate of cardiovascular diseases has become the research focus of the medical field.

[0003] Cardiac magnetic resonance imaging (MRI) as an advanced imaging examination method has the characteristics of clear imaging, high soft tissue resolution, etc., and can accurately display the cardiac anatomical structure in a non-invasive manner, providing important support for the diagnosis and treatment of cardiovascular diseases. The accurate diagnosis and treatment of cardiovascular diseases depend on the accurate measurement of cardiac function quantitative parameters, which has given rise to the demand for accurate segmentation of cardiac MRI images.

[0004] However, the existing cardiac MRI image segmentation method mainly relies on full supervision learning, which requires a large number of pixel-by-pixel labeled labels. Due to the characteristics of multiple slices and large data volume of cardiac MRI images, the cost of pixel-by-pixel labeling is high and difficult to obtain.

[0005] Therefore, how to realize accurate segmentation of cardiac MRI images under the condition of low labeling cost is a problem that needs to be solved by those skilled in the art. SUMMARY

[0006] The purpose of the present application is to provide a weak semi-supervised cardiac MRI segmentation method based on pseudo-label frequency domain dynamic hybrid supervision to solve the problems in the background art.

[0007] To achieve the above-mentioned purpose, the present application provides a weak semi-supervised cardiac MRI segmentation method based on pseudo-label frequency domain dynamic hybrid supervision, comprising the following steps:

[0008] S1, a group of cardiac MRI images are divided into a training set, a validation set and a test set, a small amount of cardiac MRI images in the training set are provided with scribble labels to obtain a labeled data pool and an unlabeled data pool in the training set;

[0009] S2, a main segmentation model, a first auxiliary segmentation model and a second auxiliary segmentation model are established, N samples are selected from the labeled data pool and the unlabeled data pool to form a paired sample set, and the paired sample set is input into the three segmentation models respectively, and the prediction values corresponding to each segmentation model are output;

[0010] S3, generating supervised pseudo labels by using a Haar wavelet transform and a frequency domain dynamic hybrid method for the prediction value obtained in S2;

[0011] S4, establishing a total loss function based on the supervised pseudo labels and combining a partial cross-entropy loss constraint strategy to optimize and train the three segmentation models, using an SGD optimizer to minimize the total loss function to obtain optimal cardiac MRI image backbone segmentation model parameters;

[0012] S5, based on the result of S4, inputting a test set into the optimized backbone segmentation model to obtain a cardiac MRI image segmentation result.

[0013] Preferably, in S2, the backbone segmentation model comprises a standard convolution module, a pooling layer and an activation layer.

[0014] Preferably, in S2, the first auxiliary segmentation model and the second auxiliary segmentation model have the same structure and each comprises a normalization layer, a full connection layer and a linear mapping layer.

[0015] Preferably, in S3, the specific process of generating the supervised pseudo labels is as follows:

[0016] S31, decomposing the prediction value of the backbone segmentation model and the prediction value of the first auxiliary segmentation model into low-frequency components and high-frequency components by using a Haar wavelet transform;

[0017] S32, respectively mixing the low-frequency components and the high-frequency components by using dynamic parameters;

[0018] S33, restoring the mixed low-frequency components and high-frequency components into the supervised pseudo labels by using a Haar wavelet inverse transform.

[0019] Preferably, the specific decomposition process of the Haar wavelet transform is as follows:

[0020] I 01 = 0.5 * I[0::2, :];

[0021] I 02 = 0.5 * I[1::2, :];

[0022] I1=I 01 [:,0::2];

[0023] I2=I 02 [:,0::2];

[0024] I3=I 01 [:,1::2];

[0025] I4=I 02 [:,1::2];

[0026] In the formula, I represents a two-dimensional image stored in matrix form, I[0::2,:] represents the even-numbered rows of I, I[1::2,:] represents the odd-numbered rows of I, and I1 and I2 are respectively the two rows of I. 01 and I 02 The even-numbered sequence, I3 and I4 are respectively I 01 and I 02 an odd sequence;

[0027] The frequency domain component corresponding to I is represented as:

[0028] LL I =I1+I2+I3+I4;

[0029] HL I = -I1-I2+I3+I4;

[0030] LH I = -I1+I2-I3+I4;

[0031] HH I =I1-I2-I3+I4;

[0032] In the formula, LL I HL represents the low-frequency component. I LH represents the horizontal high-frequency component. I HH represents the vertical high-frequency component. I This represents the diagonal high-frequency components.

[0033] Preferably, the restoration process of the Haar wavelet inverse transform is as follows:

[0034] I[0::2,0::2]=0.5×(LL I -HL I -LH I +HH I );

[0035] I[1::2,0::2]=0.5×(LL I -HL I +LH I -HH I );

[0036] I[0::2,1::2]=0.5×(LL I +HL I -LH I -HH I );

[0037] I[1::2,1::2]=0.5×(LL I +HL I +LH I +HHI ).

[0038] Preferably, in the S4, the total loss function is represented as:

[0039] L total = λ1·L fplms + λ2·L CE + λ3·L aux + λ4·L cross ;

[0040] In the formula, L total represents the total loss function, λ1-λ4 are hyperparameters, L fplms represents the pseudo-label frequency domain dynamic hybrid supervision loss, L CE represents the partial cross-entropy loss, L aux represents the auxiliary supervision loss, and L cross represents the mutual supervision loss.

[0041] Preferably, the pseudo-label frequency domain dynamic hybrid supervision loss is represented as:

[0042] L fplms = L Dice (PL1,Q c ) + L Dice (PL1,Q v ) ;

[0043] In the formula, L Dice (·) represents the Dice loss function, PL1 is the supervised pseudo-label, Q c is the predicted value of the backbone segmentation model, and Q v is the predicted value of the first auxiliary segmentation model.

[0044] The Dice loss function is defined as:

[0045]

[0046] where Y represents the supervised label and P represents the model predicted value.

[0047] Preferably, the partial cross-entropy loss is represented as:

[0048]

[0049] In the formula, L CE′ (·) represents the partial cross-entropy loss function, represents the predicted value of the backbone segmentation model f c (·) for the i-th labeled sample , represents the predicted value of the first auxiliary segmentation model f v (·) for the i-th labeled sample The predicted value, express Corresponding graffiti labels;

[0050] The partial cross-entropy loss function is defined as:

[0051]

[0052] Among them, Y i P represents the value of the i-th class of the graffiti label. i This represents the model prediction value corresponding to pixel i, and T represents the total number of segmentation categories.

[0053] Preferably, the auxiliary supervision loss is expressed as:

[0054]

[0055] In the formula, L MSE (·) represents the mean squared error loss function. The second auxiliary segmentation model f represents v′ (·) For the i-th unlabeled sample The predicted value, This indicates that the backbone segmentation model applies to the i-th unlabeled sample. The predicted value, This indicates that the first auxiliary segmentation model is applied to the i-th unlabeled sample. The predicted value;

[0056] The mean squared error loss function is defined as:

[0057]

[0058] Where S represents the total number of pixels, y i This represents the label value corresponding to pixel i.

[0059] Preferably, the mutual supervision loss is expressed as:

[0060]

[0061] In the formula, argmax(·) represents the maximum value index function.

[0062] Therefore, the weakly semi-supervised cardiac MRI segmentation method based on pseudo-label frequency domain dynamic hybrid supervision of the present invention has the following beneficial effects:

[0063] (1) By using a pseudo-label frequency domain dynamic hybrid supervision method, a reliable supervision signal is generated using a small number of graffiti labels and a large number of unlabeled data, which significantly improves the utilization efficiency of unlabeled data.

[0064] (2) By adopting a weak semi-supervised strategy, only a small number of graffiti labels are needed to effectively extract key tissues for quantifying cardiac function parameters, achieve high-quality segmentation, greatly reduce the dependence on pixel-by-pixel precise annotation, solve the problem of high cost of medical image annotation, and promote the application of data-driven algorithms in the field of cardiac medicine.

[0065] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0066] Figure 1 This is a schematic diagram of the overall architecture of an embodiment of the present invention;

[0067] Figure 2 This is a schematic diagram of the Unet backbone segmentation model according to an embodiment of the present invention;

[0068] Figure 3 This is a schematic diagram of the Swin-Unet-assisted segmentation model according to an embodiment of the present invention;

[0069] Figure 4 The following are comparison diagrams of segmentation results according to embodiments of the present invention; wherein, (a) is the original image; (b) is the ground truth label of the original image; (c) is the segmentation result of CPS; (d) is the segmentation result of URPC; (e) is the segmentation result of CBCT; (f) is the segmentation result of EVIL; (g) is the segmentation result obtained by CLB; and (h) is the segmentation result of the method of the present invention. Detailed Implementation

[0070] The technical solution of the present invention will be further described below with reference to the accompanying drawings and embodiments.

[0071] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are some embodiments of the present invention, but not all embodiments.

[0072] Example

[0073] like Figure 1 As shown, this invention provides a weakly semi-supervised cardiac MRI segmentation method based on pseudo-label frequency domain dynamic hybrid supervision. This method uses the PyTorch framework, with a hardware environment consisting of one Intel(R) Xeon(R) W-2175 CPU and one NVIDIA GeForce RTX 2080Ti GPU. The effectiveness of the proposed method is verified through the following steps:

[0074] S1, a set of cardiac MRI images is objectively divided into a training set, a validation set (used to adjust the hyperparameters of the model and evaluate the performance of the segmentation model during training) and a test set in a ratio of 7:1:2, a small number of cardiac MRI images in the training set D are provided with scribble labels, to obtain a labeled data pool and an unlabeled data pool in the training set, specifically:

[0075] The training set D is denoted as D={D L ,D U}, wherein D L represents the labeled data pool, D U represents the unlabeled data pool, and has:

[0076]

[0077] In the formula, m represents the number of samples in D L , n represents the number of samples in D U , and m<<n is satisfied; and respectively represent the i-th cardiac MRI image in D L and its corresponding scribble label, represents the j-th cardiac MRI image in D U . In this example, the training set contains 1312 cardiac MRI slices, the annotation ratio is set to 10%, there are 1181 unlabeled cardiac MRI slices, i.e. n=1181, and there are 131 labeled cardiac MRI slices, which include original images and corresponding scribble labels, i.e. m=131.

[0078] After annotation, each slice in the training set D is scaled to 224x224 size to ensure that all input data have the same dimensions for model processing and calculation.

[0079] S2, a main segmentation model f c (·), a first auxiliary segmentation model f v (·) and a second auxiliary segmentation model f v′ (·) are established; the main segmentation model f c (·) includes a standard convolution module, a pooling layer and an activation layer, in this example, the main segmentation model f c (·) adopts a Unet structure (as shown in Figure 2 ); the first auxiliary segmentation model and the second auxiliary segmentation model have the same structure, and both include a normalization layer, a fully connected layer and a linear mapping layer, in this embodiment, both adopt a Swin-Unet structure (as shown in Figure 3 );

[0080] respectively from the labeled data pool D L and the unlabeled data pool D UThe N sample sets are selected to form a paired sample set, and the paired sample set is input into three segmentation models to output prediction values corresponding to each segmentation model, specifically:

[0081] 1) The sample set selected from the labeled data pool is represented as:

[0082]

[0083] 2) The sample set selected from the unlabeled data pool is represented as:

[0084]

[0085] 3) The paired sample set is represented as:

[0086]

[0087] 4) The prediction values output by the model are represented as:

[0088]

[0089] In this embodiment, the batch size of the backbone segmentation model and the auxiliary segmentation model optimization training setting is 8 (8 images are processed by one GPU), that is, N = 4.

[0090] S3, for the prediction values obtained in S2, a supervised pseudo label is generated by using a Haar wavelet transform and a frequency domain dynamic hybrid method, and the specific process of generating the supervised pseudo label is:

[0091] S31, for the prediction value Q c of the backbone segmentation model and the prediction value Q v of the first auxiliary segmentation model, a Haar wavelet transform is used to decompose into low-frequency components {LL c , LL v} and high-frequency components {{LH c , HL c , HH c}, {LH v , HL v , HH v}}, and the specific decomposition process of the Haar wavelet transform is:

[0092] For any two-dimensional image I stored in a matrix form, let:

[0093] I 01 = 0.5 x I[0::2, :];

[0094] I 02 = 0.5 x I[1::2, :];

[0095] I1=I 01 [:,0::2];

[0096] I2 = I 02 [:,0::2];

[0097] I3 = I 01 [:,1::2];

[0098] I4 = I 02 [:,1::2];

[0099] where I[0::2,:] denotes even rows of I, I[1::2,:] denotes odd rows of I, I1, I2 are even columns of I 01 and I 02 , respectively, and I3, I4 are odd columns of I 01 and I 02 , respectively;

[0100] then the frequency domain components {LL I , HL I , LH I , HH I} corresponding to I are represented as:

[0101] LL I = I1 + I2 + I3 + I4;

[0102] HL I = -I1 - I2 + I3 + I4;

[0103] LH I = -I1 + I2 - I3 + I4;

[0104] HH I = I1 - I2 - I3 + I4;

[0105] where LL I denotes a low frequency component, HL I denotes a horizontal high frequency component, LH I denotes a vertical high frequency component, and HH I denotes a diagonal high frequency component.

[0106] S32, for low frequency components and high frequency components, dynamic parameters are respectively used for mixing, specifically:

[0107] 1) mixing LL c , LL v with dynamic parameter r1 ∈ (0, 1):

[0108] LL1 = r1 · LL c + (1 - r1) · LL v ;

[0109] 2) mixing {LHc HL c ,HH c}、{LH v HL v ,HH v}:

[0110] LH1=r2·LH c +(1-r2)·LH v ;

[0111] HL1 = r2·HL c +(1-r2)·HL v ;

[0112] HH1=r2·HH c +(1-r2)·HH v ;

[0113] S33. The mixed low-frequency component LL1 and high-frequency component {LH1,HL1,HH1} are restored to the supervision pseudo-label PL1 using the Haar wavelet inverse transform. The restoration process of the Haar wavelet inverse transform is as follows:

[0114] I[0::2,0::2]=0.5×(LL I -HL I -LH I +HH I );

[0115] I[1::2,0::2]=0.5×(LL I -HL I +LH I -HH I );

[0116] I[0::2,1::2]=0.5×(LL I +HL I -LH I -HH I );

[0117] I[1::2,1::2]=0.5×(LL I +HL I +LH I +HH I ).

[0118] S4. Based on supervised pseudo-labels and combined with a partial cross-entropy loss constraint strategy, establish the total loss function (including pseudo-label frequency domain dynamic hybrid supervised loss L). fplms Partial cross-entropy loss L CE , auxiliary supervision loss L aux Mutual supervision loss L cross) three segmentation models are optimized and trained, and the SGD optimizer is used to minimize the total loss function to obtain the optimal cardiac MRI image backbone segmentation model parameters, specifically:

[0119] 1) using f c (·) and f v (·) two types of prediction values Q c , Q v are calculated Pseudo-label frequency domain dynamic hybrid supervision loss L fplms , which is used to enrich the supervision form and improve the model's regional and boundary segmentation ability, L fplms is expressed as:

[0120] L fplms = L Dice (PL1, Q c ) + L Dice (PL1, Q v );

[0121] In the formula, L Dice (·) represents the Dice loss function, and L Dice (·) is defined as:

[0122]

[0123] Where Y represents the supervision label, and P represents the model prediction value.

[0124] 2) using f c (·) and f v (·) two types of labeled prediction values are calculated Partial cross-entropy loss L CE with the scribble label, so that the scribble label directly supervises the model training and ensures the basic segmentation ability of the model, L CE is expressed as:

[0125]

[0126] In the formula, L CE′ (·) represents the partial cross-entropy loss function, and L CE′ (·) is expressed as:

[0127]

[0128] Where Y i represents the i-th value of the scribble label, P i represents the model prediction value corresponding to pixel point i, and T represents the total number of segmentation categories.

[0129] 3) using f v′ (·) to obtain the unlabeled prediction value As pseudo-labels, calculate the pair f c (·), f v (·) Auxiliary supervision loss for unlabeled data L aux To achieve enhanced supervision of the model on unlabeled data, L aux Represented as:

[0130]

[0131] In the formula, L MSE (·) represents the mean squared error loss function, L MSE (·) is represented as:

[0132]

[0133] Where S represents the total number of pixels, y i This represents the label value corresponding to pixel i.

[0134] 4) Using f c (·), f v (·) Unlabeled predicted values ​​obtained Falsify labels and calculate the pair f v (·), f c (·) Mutual supervision loss of unlabeled data L cross To enable model supervision of unlabeled data, L cross Represented as:

[0135]

[0136] In the formula, argmax(·) represents the maximum value index function.

[0137] 5) The total loss L total Represented as:

[0138] L total =λ1·L fplms +λ2·L CE +λ3·L aux +λ4·L cross ;

[0139] In the formula, λ1 to λ4 are all hyperparameters. In this example, an exponential decay strategy is applied to adjust the learning rate, with the initial learning rate and number of iterations set to 0.01 and 30k, respectively. Hyperparameters λ1 and λ2 are set to 0.06 and 0.5, respectively, and λ3 is defined by the time-varying Gaussian warm-up function.

[0140]

[0141] Among them, t i t represents the current iteration number. totalwherein λ 3 represents the total number of iterations, and λ 4 = 7 · λ 3.

[0142] S5, based on the optimal cardiac MRI image backbone segmentation model parameters obtained in S4, inputting the test set into the optimized backbone segmentation model to obtain a cardiac MRI image segmentation result.

[0143] To effectively illustrate the effectiveness of the weak semi-supervised cardiac MRI image segmentation method based on the pseudo-label frequency domain dynamic hybrid supervision proposed in the present application, Table 1 shows the performance comparison with other mainstream methods on the test set, and the evaluation indexes are the Dice similarity coefficient (DSC), Jaccard coefficient (JC) and average surface distance (ASD), and the visualization segmentation results are as shown in Figure 4

[0144] Table 1 Comparison of cardiac MRI image segmentation performance of different models

[0145] Model DSC JC ASD CPS 48.85 32.32 4.21 URPC 46.73 30.49 2.27 CBCT 52.38 35.49 7.27 EVIL 63.42 46.44 3.71 CLB 62.17 45.10 3.40 The invention 82.57 70.31 1.59

[0146] From Figure 4 and Table 1, it can be seen that the present application is significantly better than the mainstream methods such as CPS, URPC and EVIL, can better realize the effective segmentation of the cardiac MRI image under low labeling cost, and improves the utilization rate of the unmarked data, verifies the clinical practical value of realizing high-precision segmentation under low labeling cost, and promotes the application of data-driven algorithms in the field of cardiac medical treatment.

[0147] Therefore, the weak semi-supervised cardiac MRI segmentation method based on the pseudo-label frequency domain dynamic hybrid supervision can generate reliable supervision signals by using a small amount of scribble labels and a large amount of unmarked data through the pseudo-label frequency domain dynamic hybrid supervision method, and significantly improves the utilization efficiency of the unmarked data.

[0148] By adopting the weak semi-supervised strategy, only a small amount of scribble labels are needed to effectively extract the key tissues of the cardiac functional parameter quantization, realize high-quality segmentation, greatly reduce the dependence on pixel-by-pixel accurate labeling, solve the problem of high medical image labeling cost, and promote the application of data-driven algorithms in the field of cardiac medical treatment.

[0149] Finally, it should be noted that: the above examples are only used to illustrate the technical solutions of the present application but not to limit them, although the present application has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present application can still be modified or replaced by equivalents, and these modifications or equivalent replacements cannot make the modified technical solutions deviate from the spirit and scope of the technical solutions of the present application.​

Claims

1. A weakly semi-supervised cardiac MRI segmentation method based on pseudo-label frequency domain dynamic hybrid supervision, characterized in that, Includes the following steps: S1. Divide a set of cardiac MRI images into a training set, a validation set, and a test set. Provide graffiti labels for a small number of cardiac MRI images in the training set to obtain labeled data pools and unlabeled data pools in the training set. S2. Establish a backbone segmentation model, a first auxiliary segmentation model, and a second auxiliary segmentation model. Select N samples from the labeled data pool and the unlabeled data pool respectively to form a paired sample set. Input the paired sample set into the three segmentation models respectively and output the predicted value corresponding to each segmentation model. S3. For the predicted values ​​obtained from the main segmentation model and the first auxiliary segmentation model in S2, the Haar wavelet transform and frequency domain dynamic hybrid method are used to generate supervised pseudo-labels. S4. Construct a pseudo-label frequency domain dynamic hybrid supervision loss based on supervised pseudo-labels, the predicted values ​​of the backbone segmentation model and the first auxiliary segmentation model, and construct a partial cross-entropy loss by supervising the training of the backbone segmentation model and the first auxiliary segmentation model based on graffiti labels. Construct auxiliary supervision losses based on the predicted values ​​of the backbone segmentation model, the first auxiliary segmentation model and the second auxiliary segmentation model for unlabeled samples, respectively. Construct mutual supervision losses based on the predicted values ​​of the backbone segmentation model and the first auxiliary segmentation model for unlabeled samples, respectively. Based on the above four losses, a total loss function is constructed, and the SGD optimizer is used to minimize the total loss function to obtain the optimal cardiac MRI image trunk segmentation model parameters. S5. Based on the results of S4, the test set is input into the optimized trunk segmentation model to obtain the cardiac MRI image segmentation results.

2. The weakly semi-supervised cardiac MRI segmentation method based on pseudo-label frequency domain dynamic hybrid supervision according to claim 1, characterized in that: In S2, the backbone segmentation model includes a standard convolutional module, a pooling layer, and an activation layer.

3. The weakly semi-supervised cardiac MRI segmentation method based on pseudo-label frequency domain dynamic hybrid supervision according to claim 1, characterized in that: In S2, the first auxiliary segmentation model and the second auxiliary segmentation model have the same structure, both including a normalization layer, a fully connected layer and a linear mapping layer.

4. The weakly semi-supervised cardiac MRI segmentation method based on pseudo-label frequency domain dynamic hybrid supervision according to claim 1, characterized in that, In step S3, the specific process of generating the supervised pseudo-label is as follows: S31. The predicted values ​​of the main segmentation model and the first auxiliary segmentation model are decomposed into low-frequency components and high-frequency components using Haar wavelet transform. S32. Dynamic parameters are used to mix the low-frequency and high-frequency components respectively; S33. The mixed low-frequency and high-frequency components are restored to the supervision pseudo-labels by using the Haar wavelet inverse transform.

5. The weakly semi-supervised cardiac MRI segmentation method based on pseudo-label frequency domain dynamic hybrid supervision according to claim 1, characterized in that, In S4, the total loss function is expressed as: ； In the formula, Represents the total loss function. ~ All of these are hyperparameters. This represents the frequency domain dynamic hybrid supervision loss of pseudo-labels. This represents partial cross-entropy loss. Indicates the loss of auxiliary supervision, This indicates losses due to mutual supervision.

6. The weakly semi-supervised cardiac MRI segmentation method based on pseudo-label frequency domain dynamic hybrid supervision according to claim 5, characterized in that, The pseudo-label frequency domain dynamic hybrid supervision loss is expressed as: ； In the formula, Represents the Dice loss function. To monitor for false labels, These are the predicted values ​​from the backbone segmentation model. This is the predicted value of the first auxiliary segmentation model.

7. The weakly semi-supervised cardiac MRI segmentation method based on pseudo-label frequency domain dynamic hybrid supervision according to claim 5, characterized in that, The partial cross-entropy loss is expressed as: ； In the formula, This represents the partial cross-entropy loss function. This represents the backbone segmentation model. For the A labeled sample The predicted value, Represents the first auxiliary segmentation model For the A labeled sample The predicted value, express The corresponding graffiti label.

8. The weakly semi-supervised cardiac MRI segmentation method based on pseudo-label frequency domain dynamic hybrid supervision according to claim 5, characterized in that, The auxiliary supervision loss is expressed as: ； In the formula, This represents the mean squared error loss function. Indicates the second auxiliary segmentation model For the Unlabeled samples The predicted value, This indicates that the backbone segmentation model applies to the first... Unlabeled samples The predicted value, This indicates that the first auxiliary segmentation model is applied to the first... Unlabeled samples The predicted value.

9. The weakly semi-supervised cardiac MRI segmentation method based on pseudo-label frequency domain dynamic hybrid supervision according to claim 5, characterized in that, The mutual supervision loss is expressed as: ； In the formula, This represents the maximum value index function.

Images