Medical image segmentation method based on multi-teacher network and pseudo-label contrast generation
By employing a multi-teacher network and a pseudo-label comparison generation method, the problem of uneven training signal quality in medical image segmentation is solved. Through adaptive processing of pseudo-labels and confidence matrices, the training process is optimized, thereby improving the model's segmentation performance.
Patent Information
- Application Number
- CN202410123550.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-01-29
- Publication Date
- 2024-11-08
- Estimated Expiration
- 2044-01-29
AI Technical Summary
Existing consistency-based medical image segmentation methods generate training signals of inconsistent quality, leading to confirmation bias in the model during training and preventing optimal performance.
A multi-teacher network and pseudo-label comparison generation method are adopted. Three medical image segmentation models with the same structure are semi-supervised for training. The prediction results of the teacher models are compared and analyzed to generate more accurate pseudo-labels and confidence matrices. The pseudo-labels and confidence matrices are updated by adaptive pruning and mixing. The model is trained by combining supervised and unsupervised loss.
The training signal quality of the medical image segmentation model was improved, the training process was optimized, and the model's performance was enhanced.
Smart Images

Figure CN118037651B_ABST
Abstract
Description
Technical Field
[0001] The present application belongs to the field of image processing technology, and specifically relates to a medical image segmentation method based on multi-teacher network and pseudo-label comparison generation. Background Art
[0002] Medical image segmentation, as an important branch of computer vision, aims to accurately divide and locate structural details in medical images, thereby providing a more detailed and comprehensive understanding of the patient's condition. In medical practice, image segmentation technology plays a key role in early detection, quantitative analysis and formulation of personalized treatment plans for diseases.
[0003] However, medical image segmentation faces a series of challenges, the most notable of which is the complexity and diversity of data. Medical images often have high noise, deformation and contrast differences, making traditional image processing methods difficult to cope with. Deep learning makes full use of large amounts of data and can achieve relatively high-quality results. However, since medical image data usually needs to be annotated by professionals, the annotation cost is high, resulting in relatively small data sets for medical images, which limits the generalization ability of deep learning models. Semi-supervised learning has attracted widespread attention for its ability to train models using unlabeled data while alleviating the annotation pressure and retaining good training results.
[0004] In the field of semi-supervised learning, consistency-based methods have attracted widespread attention. This method introduces a consistency loss in model prediction, forcing the model to produce consistent outputs for the same input under different views or data augmentation. This consistency enforcement is crucial to improving the efficiency of the model's utilization of unlabeled data. By combining the consistency loss with the supervised loss of labeled data in model training, unlabeled data can be effectively utilized to improve the generalization ability of the model. However, the quality of the training signals generated by these traditional consistency-based methods is uneven, and the model produces confirmation bias during training and cannot achieve optimal performance. Summary of the invention
[0005] The purpose of this application is to provide a medical image segmentation method based on a multi-teacher network and pseudo-label comparison generation, so as to overcome the problems of uneven quality of training signals provided by existing traditional consistency-based methods, confirmation bias in the model during training, and inability to achieve optimal performance.
[0006] In order to achieve the above purpose, the technical solution of this application is as follows:
[0007] A medical image segmentation method based on multi-teacher network and pseudo-label contrast generation, comprising:
[0008] Obtain medical images and perform preprocessing to generate training data sets, which include labeled data sets and unlabeled data sets;
[0009] Use labeled datasets to pre-train three medical image segmentation models with the same structure;
[0010] Semi-supervised training is performed on the three pre-trained medical image segmentation models. In each semi-supervised training round, one model is selected from the three medical image segmentation models as the student model, and the other two are used as teacher models.
[0011] Sampling labeled and unlabeled data from the training dataset allows the student model to make predictions on both unlabeled and labeled data, while the teacher model only makes predictions on unlabeled data;
[0012] Compare and analyze the prediction results of the two teacher models to obtain the unlabeled data pseudo labels and the unlabeled data confidence matrix;
[0013] For the unlabeled data pseudo labels and unlabeled data confidence matrix obtained through comparative analysis, regional masks are adaptively generated, adaptive cropping and blending are performed, and the unlabeled data pseudo labels and unlabeled data confidence matrix are updated;
[0014] Calculate the supervised loss and unsupervised loss, and sum them up to get the overall loss. Perform backpropagation to update the parameters of the student model. Based on the parameters of the student model, use the exponential sliding average method to update the parameters of the teacher model.
[0015] After the three medical image segmentation models are trained, any trained medical image segmentation model is used to segment the medical image to be segmented.
[0016] Preferably, in each semi-supervised training round, one model is selected from three medical image segmentation models as a student model, and the other two are selected as teacher models, including:
[0017] For the i-th training round, the model with the remainder of i divided by 3 is used as the student model, and the other two models are used as teacher models.
[0018] Preferably, the comparative analysis of the prediction results of the two teacher models to obtain unlabeled data pseudo labels and unlabeled data confidence matrices includes:
[0019] The prediction results obtained by the two teacher models are divided into foreground and background according to whether the value of each pixel is greater than the threshold, and the corresponding preliminary pseudo labels are obtained;
[0020] Intersect the two preliminary pseudo labels to obtain the unlabeled data pseudo labels;
[0021] Calculate the preliminary confidence matrix corresponding to the prediction results of the two teacher models;
[0022] The first region is obtained by taking the intersection of the unlabeled data pseudo labels and the preliminary pseudo labels of the two teacher models; the second region is obtained by taking the difference of the unlabeled data pseudo labels and the preliminary pseudo labels of the two teacher models; the other regions are taken as the third region, and then the values of the three regions are merged to obtain the third confidence matrix;
[0023] Multiply the third confidence matrix and the preliminary confidence matrices of the two teacher models pixel by pixel to obtain the unlabeled data confidence matrix.
[0024] Preferably, for the unlabeled data pseudo labels and the unlabeled data confidence matrix obtained by comparative analysis, adaptively generating regional masks, adaptively cropping and blending are performed respectively, and updating the unlabeled data pseudo labels and the unlabeled data confidence matrix includes:
[0025] Take each unlabeled data pseudo label obtained in this round as a pseudo label to be processed, and determine a different cropped pseudo label for it;
[0026] Generate a random region within the geometric range of the pseudo-label to be processed;
[0027] Calculate the average confidence of the cropped pseudo-labels in random regions;
[0028] Determine whether the average value is less than the threshold. If so, no mixing is performed and the pseudo-label to be processed remains unchanged. Otherwise, the content of the random area in the cropped pseudo-label is used to replace the content of the random area in the pseudo-label to be processed.
[0029] Preferably, the calculation of the supervised loss and the unsupervised loss and the weighted summation to obtain the overall loss includes:
[0030] Calculate the supervised loss based on the student model’s predictions and annotations for the labeled data;
[0031] Calculate the unsupervised loss based on the prediction results of the student model on the unlabeled data, the unlabeled data pseudo labels and the unlabeled data confidence matrix obtained by comparative analysis of the teacher model;
[0032] The supervised loss and the unsupervised loss are weighted together to get the overall loss.
[0033] Preferably, the supervised loss includes BCE loss and IoU loss.
[0034] The present application provides a medical image segmentation method based on multi-teacher networks and pseudo-label comparison generation, using multiple teacher networks to generate pseudo-labels, and by comparing and analyzing multiple pseudo-labels, more accurate pseudo-labels and confidence matrices are obtained. Thereby generating more effective training signals to guide the model for training, and obtaining a medical image segmentation model with better performance. The present application proposes a multi-teacher guided training framework to improve the reliability of pseudo-labels and thus improve model performance. Comparative analysis is performed on pseudo-labels generated by multiple teacher networks to provide accurate training signals, optimize the training process, and improve model performance. BRIEF DESCRIPTION OF THE DRAWINGS
[0035] Figure 1 A flowchart of the medical image segmentation method based on multi-teacher network and pseudo-label comparison generation for this application;
[0036] Figure 2 This is a schematic diagram of comparative analysis of prediction results of the teacher model in the embodiment of the present application;
[0037] Figure 3 This is a schematic diagram of cropping and mixing according to an embodiment of the present application. DETAILED DESCRIPTION
[0038] In order to make the purpose, technical solution and advantages of the present application more clearly understood, the present application is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.
[0039] An embodiment of the present application provides a medical image segmentation method based on multi-teacher networks and pseudo-label comparison generation. The method makes full use of the pseudo-labels generated by multiple teacher networks to improve the effectiveness of training signals and uses a semi-supervised method to train a high-performance model.
[0040] like Figure 1 As shown, the medical image segmentation method based on multi-teacher network and pseudo-label comparison generation in this embodiment includes:
[0041] Step S1: Acquire medical images and perform preprocessing to generate a training data set, where the training data set includes a labeled data set and an unlabeled data set.
[0042] For example, the Kvasir-seg public intestinal polyp dataset is used to divide 1000 medical images into 100 labeled training sets and 800 unlabeled training sets.
[0043] The training set data is normalized, the pixel values are adjusted to a fixed range, and then random cropping, random scaling, and random flipping are used for data enhancement, and all images are resized to 512 × 512. Then, strong data enhancement methods such as color jittering, Gaussian blur, and contrast enhancement are used on these data to further enhance the diversity of the images.
[0044] Step S2: pre-train three constructed medical image segmentation models with the same structure using labeled data sets.
[0045] In this example, commonly used medical image segmentation models such as Deeplab_v3 or Unet are used. Deeplab_v3 is used as an example in this embodiment. Different random number seeds are used to initialize the parameters of three Deeplab_v3. Using 100 labeled training set images, the three models are trained for 10 rounds respectively to obtain three models that can perform preliminary segmentation of medical images.
[0046] Step S3: perform semi-supervised training on the three pre-trained medical image segmentation models. In each semi-supervised training round, one model is selected from the three medical image segmentation models as a student model, and the other two are selected as teacher models.
[0047] In this example, the three models are labeled as model 0, model 1, and model 2. For the i-th training round, the model with the remainder of i divided by 3 is used as the student model, and the other two models are used as teacher models. This setting allows the three models to take turns as student models or teacher models, and each model is selected as a student model almost the same number of times.
[0048] It should be noted that this application does not limit how to select the student model and the teacher model. A person skilled in the art can also randomly generate a random number sequence with the same number of 0, 1, or 2, and select a student model with the corresponding number according to the sequence.
[0049] Step S4: Sample labeled and unlabeled data from the training data set and let the student model predict the unlabeled data and the labeled data, while the teacher model only predicts the unlabeled data.
[0050] In this example, within a training batch, some labeled images are sampled from the labeled dataset, and some unlabeled images are sampled from the unlabeled dataset. The student model is asked to make predictions on both the unlabeled and labeled data, while the teacher model only makes predictions on the unlabeled data.
[0051] For example, sample 2 labeled images from the labeled dataset and 16 unlabeled images from the unlabeled dataset. Let the student model predict the 16 unlabeled images and 2 labeled images, while the teacher model only predicts the 16 unlabeled images.
[0052] Step S5: Compare and analyze the prediction results of the two teacher models to obtain the unlabeled data pseudo labels and the unlabeled data confidence matrix.
[0053] This step performs pseudo-label reasoning analysis, such as Figure 2 As shown in the figure, the prediction results of the two teacher models are used to calculate the label and the confidence of the label respectively. By comparing the prediction results of the two different models, a more accurate model and label confidence are obtained. The depth of the color in the confidence matrix indicates the confidence level of the area.
[0054] Specifically, the following steps are included:
[0055] S5.1. The prediction results obtained by the two teacher models are divided into foreground and background according to whether the value of each pixel is greater than the threshold, and the corresponding preliminary pseudo-labels are obtained.
[0056] The prediction results obtained by the two teacher models are expressed as and According to whether the value of each pixel is greater than the threshold, it is divided into foreground and background, and a preliminary pseudo label Label is obtained. 1 tea and Label 2 tea The expression is as follows:
[0057]
[0058] in, is the value of a pixel in the pseudo label, θ is the set threshold, ω represents the pixel coordinates, and Ω represents the geometric range of the prediction result H×W. That is, each pixel is divided into foreground and background according to whether its value is greater than the threshold. If it is greater than or equal to the threshold, it is considered to be the foreground and the value is set to 1, otherwise it is considered to be the background and the value is set to 0. i is used to distinguish different teacher models.
[0059] S5.2. Intersect the two preliminary pseudo labels to obtain the pseudo labels for the unlabeled data.
[0060] In this step, two preliminary pseudo labels are and Perform intersection processing to obtain the final unlabeled data pseudo label Label tea , the expression is as follows:
[0061]
[0062] That is, the part that is considered as the foreground by both preliminary pseudo-labels continues to be the foreground of the final pseudo-label, otherwise it is set as the background.
[0063] S5.3. Calculate the preliminary confidence matrix corresponding to the prediction results of the two teacher models.
[0064] The prediction results of this step for the two teacher models and The preliminary confidence matrix is calculated according to the following method The expression is as follows:
[0065]
[0066] The confidence matrix describes the confidence of the pseudo-label. For each pixel of the pseudo-label, there is a value corresponding to its confidence in the range of [0,1]. For the prediction results of the model, the more the prediction value deviates from the decision boundary (0.5), the higher the confidence of the model in the prediction results.
[0067] In this embodiment, the preliminary confidence matrix corresponding to the prediction results of the two teacher models is calculated. For each predicted value of the prediction result, the square of the distance from the predicted value to the decision boundary is taken as the confidence. Then, this result is adjusted to the interval [0,1] by multiplying by 4.
[0068] S5.4. Take the intersection of the unlabeled data pseudo-labels and the preliminary pseudo-labels of the two teacher models to obtain the first region; take the difference of the unlabeled data pseudo-labels and the preliminary pseudo-labels of the two teacher models to obtain the second region; take the other regions as the third region, and then merge the values of the three regions to obtain the third confidence matrix.
[0069] In this embodiment, the unlabeled data pseudo labels are respectively combined with the two preliminary pseudo labels. Perform intersection processing to obtain region O1, and set the value to 1; the unlabeled data pseudo labels are respectively combined with the two preliminary pseudo labels Perform the difference processing to obtain region O2, with the value set to 0.1; the other regions are used as O3, with the value set to 0.6. Finally, the values of the three regions are combined to obtain the third confidence matrix The expression is as follows:
[0070]
[0071] In this embodiment, the confidence level of the area that is identified as foreground by both preliminary pseudo labels is high and is set to 1. The confidence level of the area where the two preliminary pseudo labels disagree is low and is set to 0.1. The confidence level of the area that is identified as background by both pseudo labels may still contain missed targets, so it is set to 0.6.
[0072] S5.5. Multiply the third confidence matrix and the preliminary confidence matrices of the two teacher models pixel by pixel to obtain the unlabeled data confidence matrix.
[0073] This embodiment will Multiply pixel by pixel to get the confidence matrix M of the pseudo label that highlights the foreground con , which is the final unlabeled data confidence matrix. The calculation formula is expressed as:
[0074]
[0075] in, It represents element-by-element multiplication. The three matrices are multiplied element-by-element to obtain the final unlabeled data confidence matrix.
[0076] Step S6: For the unlabeled data pseudo-labels and the unlabeled data confidence matrix obtained through comparative analysis, adaptively generate regional masks, perform adaptive cropping and blending respectively, and update the unlabeled data pseudo-labels and the unlabeled data confidence matrix.
[0077] This embodiment adaptively crops and mixes, and adaptively generates regional masks based on the confidence of the unlabeled data pseudo labels and the unlabeled data confidence matrix obtained in this round of training, and updates the unlabeled data pseudo labels and the unlabeled data confidence matrix.
[0078] In a specific embodiment, for the unlabeled data pseudo labels and the unlabeled data confidence matrix obtained by comparative analysis, a regional mask is adaptively generated, adaptive cropping and mixing are performed respectively, and the unlabeled data pseudo labels and the unlabeled data confidence matrix are updated, including:
[0079] S6.1. Take each unlabeled data pseudo-label obtained in this round as a pseudo-label to be processed, and determine a different cropped pseudo-label for it.
[0080] In this embodiment, for the pseudo label Label to be processed tea k , you need to determine a cropped pseudo label Label tea j , to process the pseudo label Label tea k Perform adaptive cropping and mixing, where the pseudo label Label is cropped tea j It is the final pseudo-label after this round of training that is different from Label tea k A final pseudo-label for .
[0081] For example, where j=k+1, that is, Label tea kThe subsequent final pseudo-label Label te a j is the cropped pseudo-label. If k + 1 exceeds the range, the first final pseudo-label is determined to be the cropped pseudo-label.
[0082] S6.2. Generate a random region within the geometric range of the pseudo-label to be processed.
[0083] Specifically, within the geometric range of the pseudo-label Label tea k a random region is generated, where k represents the k-th unlabeled data in this round.
[0084] S6.3. Calculate the average confidence level of the cropped pseudo-label in the random region.
[0085] Specifically, calculate the average value P of the confidence matrix of Label tea j in the random region Ω rand .
[0086] S6.4. Determine whether the average value is less than the threshold. If so, no mixing is performed and the pseudo-label to be processed remains unchanged. Otherwise, replace the content of the random region in the pseudo-label to be processed with the content of the random region in the cropped pseudo-label.
[0087] In this embodiment, the threshold is a random number R between 0 and 1. If the average confidence level P < R, it indicates that the confidence level of this region is low and no mixing is performed; otherwise, use the content of the random region Ω tea j of the cropped pseudo-label Label rand to update the content of this region in the pseudo-label to be processed Label tea k .
[0088] The above steps are as Figure 3 shown, with pseudo-label 1 and its confidence matrix 1 as the pseudo-label to be processed, and the determined cropped pseudo-label being pseudo-label 2, and the confidence matrix of pseudo-label 2 being confidence matrix 2.
[0089] For the mixed cropping of pseudo-labels, a random region is generated from pseudo-label 1 ( Figure 3 the region mask in
[0090] ). If the average confidence level of pseudo-label 2 in the random region is lower than the random value, then pseudo-label 1 remains unchanged. Otherwise, use the content of pseudo-label 2 in the random region to replace the corresponding region content in pseudo-label 1. Figure 3In the middle area mask), if the average value of the confidence of confidence matrix 2 in the random area is lower than the random value, the confidence matrix 1 is kept unchanged, otherwise the content of the confidence matrix 2 in the random area is used to replace the content of the corresponding area in the confidence matrix 1.
[0091] Step S7, calculate the supervised loss and the unsupervised loss, and sum them up to get the overall loss, perform back propagation to update the parameters of the student model, and update the parameters of the teacher model using the exponential sliding average method based on the parameters of the student model.
[0092] In this example, in order to make the model fully learn both the detail information and structural information of the image, the combination of BCE loss function and IoU loss is used as the loss function of the supervised part. On this basis, in order to give the model accurate and effective training signals on unlabeled data, the generated pseudo-labels and pseudo-label confidence matrices are fully utilized to construct effective unsupervised loss calculation.
[0093] Specifically, the following steps are included.
[0094] S7.1. Prediction results of labeled data based on the student model stu and label Y, and calculate the supervised loss.
[0095] The expression is as follows:
[0096]
[0097] For the needs of accurate segmentation and structure, loss(Y,Pred stu ) is the combination of BCE loss function and IoU loss function, and the expression is as follows:
[0098]
[0099]
[0100] loss(Y,Pred stu )=L bce (Y,Pred stu )+L iou (Y,Pred stu );
[0101] L sup (*,*) indicates a supervised loss function, Y is the label in the labeled data, Pred stu is the prediction result of the student model for the labeled data, ω represents the pixel coordinate, and Ω represents the geometric range of the prediction result H×W. bce (*,*) represents the BCE loss function, L iou(*,*) represents the IoU loss function, loss(Y,Pred stu ) represents the total loss function.
[0102] S7.2. Prediction results of unlabeled data based on the student model And the unlabeled data pseudo label Label obtained by comparative analysis of the teacher model tea And the unlabeled data confidence matrix M con , calculate the unsupervised loss.
[0103] The expression is as follows:
[0104]
[0105] Among them, L unsup (*,*) represents the unsupervised loss function, Label tea M represents the pseudo-label of unlabeled data obtained by comparative analysis of the teacher model. con represents the confidence matrix of unlabeled data obtained by comparative analysis of the teacher model. ω represents the pixel coordinates, and Ω represents the geometric range of the prediction result H×W.
[0106] S7.3. Add the weighted supervised loss and unsupervised loss to get the overall loss L total .
[0107] The expression is as follows:
[0108] L total =L sup +λL unsup ;
[0109] Here, λ is a weight that increases uniformly from 0 to 4 with each training round.
[0110] Finally, backpropagation is performed based on the overall loss to update the parameters θ of the student model stu According to the new parameters of the student model, the parameters of the teacher model are updated using the exponential sliding average method. and The expression is as follows:
[0111]
[0112] Here, α is a parameter that controls the update and is set to 0.001 in this example.
[0113] By repeating steps S3-S7 until a preset number of training rounds is reached, the training of the three medical image segmentation models is completed. In this example, 80 rounds of training are performed.
[0114] Step S8: After the three medical image segmentation models are trained, use any trained medical image segmentation model to segment the medical image to be segmented.
[0115] After the three medical image segmentation models are trained, any trained medical image segmentation model can be selected to segment the medical image to be segmented, thereby realizing the segmentation processing of the medical image.
[0116] In this example, a multi-teacher-guided semi-supervised medical image segmentation training framework is proposed, using multiple teacher networks to generate pseudo labels. On this basis, a comparative analysis of pseudo labels is performed to obtain pseudo labels and confidence matrices that can improve the accuracy of training signals. Furthermore, the confidence level is adaptively determined to determine whether to crop and mix with other images to reduce the impact of erroneous training signals caused by low-confidence areas. Using weighted loss calculation, full use of unlabeled data can be made to provide the model with effective training signals and improve the performance of the model.
[0117] The above-mentioned embodiments only express several implementation methods of the present application, and the descriptions thereof are relatively specific and detailed, but they cannot be understood as limiting the scope of the invention patent. It should be pointed out that, for a person of ordinary skill in the art, several variations and improvements can be made without departing from the concept of the present application, and these all belong to the protection scope of the present application. Therefore, the protection scope of the patent of the present application shall be subject to the attached claims.
Claims
1. A medical image segmentation method based on multi-teacher network and pseudo-label contrast generation, characterized in that: The medical image segmentation method based on multi-teacher network and pseudo-label comparison generation includes: Obtain medical images and perform preprocessing to generate training data sets, which include labeled data sets and unlabeled data sets; Use labeled datasets to pre-train three medical image segmentation models with the same structure; Semi-supervised training is performed on the three pre-trained medical image segmentation models. In each semi-supervised training round, one model is selected from the three medical image segmentation models as the student model, and the other two are used as teacher models. Sampling labeled and unlabeled data from the training dataset allows the student model to make predictions on both unlabeled and labeled data, while the teacher model only makes predictions on unlabeled data; Compare and analyze the prediction results of the two teacher models to obtain the unlabeled data pseudo labels and the unlabeled data confidence matrix; For the unlabeled data pseudo labels and unlabeled data confidence matrix obtained through comparative analysis, regional masks are adaptively generated, adaptive cropping and blending are performed, and the unlabeled data pseudo labels and unlabeled data confidence matrix are updated; Calculate the supervised loss and unsupervised loss, and sum them up to get the overall loss. Perform backpropagation to update the parameters of the student model. Based on the parameters of the student model, use the exponential sliding average method to update the parameters of the teacher model. After the three medical image segmentation models are trained, the medical image to be segmented is segmented using any trained medical image segmentation model; The prediction results of the two teacher models are compared and analyzed to obtain unlabeled data pseudo labels and unlabeled data confidence matrices, including: The prediction results obtained by the two teacher models are divided into foreground and background according to whether the value of each pixel is greater than the threshold, and the corresponding preliminary pseudo labels are obtained; Intersect the two preliminary pseudo labels to obtain the unlabeled data pseudo labels; Calculate the preliminary confidence matrix corresponding to the prediction results of the two teacher models; The first region is obtained by taking the intersection of the unlabeled data pseudo labels and the preliminary pseudo labels of the two teacher models; the second region is obtained by taking the difference of the unlabeled data pseudo labels and the preliminary pseudo labels of the two teacher models; the other regions are taken as the third region, and then the values of the three regions are merged to obtain the third confidence matrix; Multiply the third confidence matrix and the preliminary confidence matrices of the two teacher models pixel by pixel to obtain the unlabeled data confidence matrix; The method of adaptively generating a regional mask for the unlabeled data pseudo-label and the unlabeled data confidence matrix obtained by comparative analysis, adaptively cropping and blending the unlabeled data pseudo-label and the unlabeled data confidence matrix, respectively, includes: Take each unlabeled data pseudo label obtained in this round as a pseudo label to be processed, and determine a different cropped pseudo label for it; Generate a random region within the geometric range of the pseudo-label to be processed; Calculate the average confidence of the cropped pseudo-labels in random regions; Determine whether the average value is less than the threshold. If so, no mixing is performed and the pseudo-label to be processed remains unchanged. Otherwise, the content of the random area in the cropped pseudo-label is used to replace the content of the random area in the pseudo-label to be processed.
2. The medical image segmentation method based on multi-teacher network and pseudo-label comparison generation according to claim 1, characterized in that: In each semi-supervised training round, one model is selected from three medical image segmentation models as the student model, and the other two are selected as teacher models, including: For i For a training round, the number i The model with the remainder after division by 3 is used as the student model, and the other two models are used as the teacher models.
3. The medical image segmentation method based on multi-teacher network and pseudo-label comparison generation according to claim 1, characterized in that: The supervised loss and unsupervised loss are calculated, and the weighted sum is used to obtain the overall loss, including: Calculate the supervised loss based on the student model’s predictions and annotations for the labeled data; Calculate the unsupervised loss based on the prediction results of the student model on the unlabeled data, the unlabeled data pseudo labels and the unlabeled data confidence matrix obtained by comparative analysis of the teacher model; The supervised loss and the unsupervised loss are weighted together to get the overall loss.
4. The medical image segmentation method based on multi-teacher network and pseudo-label comparison generation according to claim 3 is characterized in that: The supervised loss includes BCE loss and IoU loss.
Citation Information
Patent Citations
RGB image semi-supervised target detection method based on double-pseudo-label optimization learning
CN115393687A
3D coronary artery image segmentation method based on semi-supervised consistency learning
CN115908463A
Semi-supervised image segmentation method based on uncertain false label correction
CN116206106A
Semi-supervised image semantic segmentation method based on similarity weighted multi-teacher network model
CN116433909A
Federal unified segmentation method and system for partial category labeling
CN116994061A
