Semi-supervised medical image segmentation method based on high-value region adaptive learning

Abstract

The invention discloses a semi-supervised medical image segmentation method based on high-value region adaptive learning, and the method comprises the steps: constructing a semi-supervised medical image segmentation model based on high-value region adaptive learning, and the model comprises two parallel networks which have the same structure and independently update parameters; the two parallel networks output an original prediction result and a disturbed prediction result, and different value regions, namely a reliable and unstable region and an unreliable and unstable region, are screened out based on confidence distribution of the disturbed prediction result; training a reliable and unstable area through a confidence-guided cross prototype consistency learning module, and training an unreliable and unstable area through a dynamic teacher competition teaching module; and training the model by using the training set, selecting an optimal weight based on the verification set, inputting the test set into the model with the optimal weight, and outputting a result. Through the differential learning strategy, the method is accurately focused on the high-value region of the medical image, and the segmentation effect of the model in the key region of the image is remarkably improved.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of semi-supervised learning and image segmentation, and particularly relates to a semi-supervised medical image segmentation method based on high-value region adaptive learning. BACKGROUND

[0002] Medical image segmentation, as a key technology in intelligent medical diagnosis systems, its core task is to accurately identify and segment various important anatomical structures (such as heart, liver and other organ tissues) and pathological features (such as tumor lesions, vascular plaques and other abnormal areas) through intelligent analysis of CT, MRI, ultrasound and other medical image data. This technology can provide objective and quantitative imaging basis for clinical diagnosis and treatment, and plays an irreplaceable role in medical scenarios such as early disease screening (such as lung nodule detection), precise surgical planning (such as tumor boundary definition), treatment effect evaluation (such as lesion volume change monitoring). In recent years, with the rapid development of deep learning technology, the full convolutional network architecture represented by U-Net and V-Net, and the latest introduction of Vision Transformer model have made breakthrough progress in medical image segmentation. However, this kind of method usually needs a large number of professional annotated medical image samples to train a model with good generalization ability. However, medical image annotation work faces special difficulties: first, the professionalism of medical images is extremely strong, which requires annotators to have professional medical knowledge background and rich clinical experience. Secondly, due to the problems of patient privacy protection and the barriers of data sharing between different medical institutions, the publicly available high-quality medical image dataset is very limited.

[0003] Based on the above problems, semi-supervised learning technology shows unique application value. This method can use a small amount of labeled data and a large amount of unlabeled data for model training, and through the design of effective regularization constraints and pseudo-label generation mechanism, it can fully exploit the rich information contained in unlabeled data, thereby significantly reducing the dependence of the model on labeled data. In the field of medical image segmentation, semi-supervised learning not only can alleviate the problem of insufficient annotation resources, but also can improve the recognition ability of the model to complex anatomical structures and pathological changes by using a large amount of unlabeled data naturally generated in clinical practice, and provides a new technical path for developing more intelligent and reliable medical image analysis systems.

[0004] Consistency regularization and pseudo-label methods have become the mainstream research direction of semi-supervised learning. Firstly, the consistency regularization method is based on the smoothing assumption, that is, the input model data should remain similar under different levels of perturbation (data, model and feature level perturbation). For example, Bai et al. mix labeled data and unlabeled data in both directions to generate new samples. Chen et al. learn independently by initializing two networks with different parameters, and promote the same structure model to reach agreement on prediction. Ouali et al. add noise, random occlusion and other operations in the feature space to make the main decoder and auxiliary decoder produce consistent prediction results in the feature space under different perturbations. Secondly, the pseudo-label method is based on the clustering assumption, and its core idea is to first train the model using labeled data, then generate pseudo-labels for unlabeled data using the trained model to expand the dataset, and then retrain the model using the newly expanded dataset. For example, Feng et al. identify errors in pseudo-labels by comparing the prediction differences of different models, and then train each other. Sohn et al. encourage the model to keep the prediction results consistent on the strong augmented data and the pseudo-labels generated by the weak augmented data. Although the above methods improve the model performance in different ways (such as data diversification or improving the quality of pseudo-labels), there are various types of tissues in medical images, and the anatomical structure of different tissues is complex, resulting in blurred boundaries between the segmentation target and other tissues, especially in the edge region. The existing methods ignore the prediction complexity difference of different regions in the image, and also fail to effectively identify and learn more valuable regions, and fail to classify different regions in the image in detail, ignore the exploration of more valuable regions, and lack of differential learning strategies for different value regions, thus limiting the learning ability of the model in the key segmentation region of the tissue edge, hindering the further improvement of the segmentation performance of the model.

[0005] Obviously, the mainstream semi-supervised learning method faces two main challenges: first, it ignores the prediction complexity difference of different regions in the image, and also fails to effectively identify and learn more valuable regions; second, it fails to design a targeted training strategy according to the characteristics of the data. SUMMARY

[0006] In order to overcome the shortcomings of the prior art, the purpose of the present application is to provide a semi-supervised medical image segmentation method based on high-value region adaptive learning, which solves the problem of limiting model performance due to the lack of image region value evaluation and differential learning strategy in the existing semi-supervised learning method.

[0007] The present application is realized by the following technical solutions:

[0008] A semi-supervised medical image segmentation method based on high-value region adaptive learning, comprising the following steps,

[0009] S1, pre-process the data set, and divide the pre-processed data set into a training set, a validation set and a test set;

[0010] S2, constructing a semi-supervised medical image segmentation model based on high-value region adaptive learning, the semi-supervised medical image segmentation model comprising two parallel networks with the same structure and independently updating parameters, by applying random disturbance in the feature space, the two parallel networks output original prediction results and disturbance prediction results, based on the confidence distribution of the disturbance prediction results, screening out different value regions: reliable unstable regions and unreliable stable regions; training the reliable unstable regions through a confidence-guided cross-prototype consistency learning module, and training the unreliable stable regions through a dynamic teacher competition teaching module;

[0011] S3, using random initialization parameters, training the semi-supervised medical image segmentation model using the training set, selecting the optimal weight based on the validation set, and obtaining a semi-supervised medical image segmentation model with the optimal weight;

[0012] S4, inputting the test set into the semi-supervised medical image segmentation model with the optimal weight, and outputting the medical image segmentation result.

[0013] Further, the data set pre-processing in step S1 specifically comprises uniformly cutting the image size, normalizing the image to standardize the image distribution, and rotating and flipping the image to improve the diversity of the image, and the training set comprises labeled data and unlabeled data.

[0014] Further, the two parallel networks with the same structure and independently updating parameters in step S2 are U-Net or V-Net, denoted as f A (·) and f B (·) respectively.

[0015] Further, the screening of different value regions in step S2 specifically comprises:

[0016] S2.11, inputting the training set into the encoders of f A (·) and f B (·), the encoders output original features F1, and the original features F1 output by the encoders and the disturbance features F2 of the original features F1 are input into the corresponding decoders and to obtain the final prediction results:

[0017]

[0018] wherein, F a1 represents the original features output by the encoder of f A (·), represents the original features output by the encoder of fA The original prediction result of (·), F a2 represents f A (·) The original feature F output by the encoder a1 The disturbance characteristics of represents f A (·) perturbation prediction results, F b1 represents f B (·) The original features output by the encoder, represents f B The original prediction result of (·), F b2 represents f B (·) The original feature F output by the encoder b1 The disturbance characteristics of represents f B The perturbation prediction results of (·) are: and Common composition The decoder output includes the prediction results of labeled data and unlabeled data, that is,

[0019] S2.12, generate confidence maps of two parallel networks based on the perturbation prediction results obtained in step S2.11:

[0020]

[0021] Among them, con a2 Indicated by f A The confidence map generated by the perturbation prediction results of (·) is b2 Indicated by f B Confidence map generated by the perturbation prediction results of (·);

[0022] S2.13, based on the confidence map obtained in step S2.12, perform pixel-by-pixel averaging on the two confidence maps to obtain a comprehensive confidence map, and sort all pixels in the comprehensive confidence map in ascending order according to their confidence values:

[0023]

[0024] S2.14, select the K pixels with the lowest confidence after sorting, and calculate the mean of the K pixels as the adaptive threshold:

[0025]

[0026] Where K = β·(H·W·D), where H, W, and D refer to the height, width, and depth of the image, respectively. β is a hyperparameter with a value between (0, 1) that controls the number of pixels to be filtered. γ represents the confidence threshold.

[0027] S2.15, according to the confidence threshold obtained in step S2.14, the predicted pixels of the two networks are interactively classified, when the confidence of the two pixels is higher than the confidence threshold at the same time, a reliable stable area Ω1 is formed; when the confidence of the two pixels is lower than the confidence threshold at the same time, an unreliable stable area Ω3 is formed; when the confidence of the two pixels is one higher than the confidence threshold and the other lower than the confidence threshold, a reliable unstable area Ω2 is formed, specifically:

[0028] Ω1={con a2 ≥γ&con b2 <γ} (6)

[0029] Ω2={(con a2 >γ&con b2 <γ)|(con a2 <γ&con b2 >γ)} (7)

[0030] Ω3={con a2 <γ&con b2 <γ} (8).

[0031] Further, the step S2 of the cross prototype consistency learning module guided by confidence training reliable unstable area is specifically:

[0032] S2.21, the features of the decoder and the second last layer are up-sampled using trilinear interpolation, so that the features match the size of the pseudo label, the corresponding class features are screened out by using the pseudo label, and the class feature prototypes of the unlabeled data are generated by calculating the mean value of different class features, specifically:

[0033]

[0034] Where, B U represents the batch size of the unlabeled data, (h, w, z) represents the spatial coordinates, represents the original features of the decoder the second last layer, represents the indicator function, represents the pseudo label generated by , and represents the original features of the decoder the second last layer, represents the pseudo label generated by , and and represent the c-th class feature prototype from the decoder and respectively;

[0035] S2.22, the class feature prototype obtained according to step S2.21, the prediction probability of the model is approximated by calculating the cosine similarity of the perturbed feature and the class feature prototype:

[0036]

[0037] wherein, F d represents the feature of the current sample, p c represents the class feature prototype, represents the decoder the perturbed feature of the second-to-last layer, represents the decoder the perturbed feature of the second-to-last layer, sim a represents the cosine similarity of the perturbed feature and the class feature prototype of f A (·), sim b represents the cosine similarity of the perturbed feature and the class feature prototype of f B (·);

[0038] S2.23, pixel screening is performed on the confidence map of each network perturbation prediction to generate a position mask to determine the position of the reliable unstable region Ω2, specifically:

[0039] Mask a = con a2 <γ&con b2 >γ, Mask b = con a2 >γ&con b2 <γ (11)

[0040] wherein, Mask a represents the position mask of the reliable unstable region of f A (·), Mask b represents the position mask of the reliable unstable region of f B (·);

[0041] S2.24, the cross prototype consistency loss is calculated using the position mask obtained in step S2.23, specifically:

[0042]

[0043] wherein, ⊙ represents pixel-by-pixel multiplication operation; represents the cross prototype consistency loss of f A (·), represents the cross prototype consistency loss of f B (·);

[0044] S2.25, calculate the cosine similarity loss of the feature prototype and the class feature prototype:

[0045]

[0046] wherein, denotes the cosine similarity loss of f A (·), denotes the cosine similarity loss of f B (·);

[0047] S2.26, the total loss of the cross prototype consistency learning module training the reliable unstable area includes the cross prototype consistency loss and the cosine similarity loss of the feature prototype and the class feature prototype, specifically:

[0048]

[0049] wherein, denotes the total loss of f A (·) in the reliable unstable area training, denotes the total loss of f B (·) in the reliable unstable area training.

[0050] Further, the step S2 of training the unreliable stable area by the dynamic teacher competition teaching module is specifically:

[0051] S2.31, generate the confidence map of the original prediction result of f A (·) and f B (·):

[0052]

[0053] wherein, con a1 denotes the confidence map generated by the original prediction result of f A (·), and con b1 denotes the confidence map generated by the original prediction result of f B (·);

[0054] S2.32, compare the generated confidence map, select the prediction probability corresponding to the pixel with higher confidence as the new teacher prediction, and perform probability sharpening to convert it into a soft pseudo label, specifically:

[0055]

[0056] wherein, T is a hyperparameter for controlling the output distribution of the model, denotes the newly generated teacher prediction, denotes the soft pseudo label generated by ;

[0057] S2.33, pixel screening is performed on the confidence map of each network perturbation prediction to generate a position mask to determine the location of the unreliable stable region Ω3, specifically:

[0058] Mask = con a2 <γ & con b2 <γ (19)

[0059] wherein, Mask represents the position mask of the unreliable stable region of f A (·) and f B (·);

[0060] S2.34, the prediction of the perturbation feature is supervised using the soft pseudo label generated in step S2.32, specifically:

[0061]

[0062] wherein, represents the consistency loss of f A (·) trained in the unreliable stable region, represents the consistency loss of f B (·) trained in the unreliable stable region;

[0063] S2.35, the uncertainty is quantified by calculating the information entropy and added as a regularization term to the loss function:

[0064]

[0065] wherein, H a1 represents the information entropy of the original prediction of f A (·) in the unreliable stable region, H a2 represents the information entropy of the perturbation prediction of f A (·) in the unreliable stable region, H b1 represents the information entropy of the original prediction of f B (·) in the unreliable stable region, H b2 represents the information entropy of the perturbation prediction of f B (·) in the unreliable stable region;

[0066] S2.36, the total loss of the unreliable stable region trained by the dynamic teacher competition teaching module includes the mean square error loss and the information entropy, specifically:

[0067]

[0068] wherein, represents the total loss of f A (·) trained in the unreliable stable region, represents the total loss of fB (·) the total loss of the model trained in the unreliable stable region.

[0069] Further, the total loss function of the semi-supervised medical image segmentation model in the step S3 is set as:

[0070] For the labeled data, the combination of cross-entropy loss and Dice loss is used as the loss function of supervised learning:

[0071]

[0072] Wherein, y represents the label, the supervised loss of f A (·), the supervised loss of f B (·);

[0073] For unlabeled data, cross-entropy loss and cosine similarity loss are introduced through CG-CPCL module, and DTCT module is used to combine MSE loss and information entropy, and finally an unsupervised loss function is formed:

[0074]

[0075] Wherein, the unsupervised loss of f A (·), the unsupervised loss of f B (·);

[0076] The total loss function is defined as:

[0077]

[0078] Wherein, L a the total loss of f A (·), L b the total loss of f B (·), and λ is a Gaussian warm-up function changing with time, used to balance the supervised loss and the unsupervised loss.

[0079] Further, the model in the step S3 is trained using SGD optimizer, the weight decay factor is set to 10 -4 , the momentum is set to 0.9, and the initial learning rate is set to 10 -2 .

[0080] Compared with the prior art, the present application has the following advantages:

[0081] (1) The semi-supervised medical image segmentation model based on high-value region adaptive learning is constructed, the high-value region is divided into two categories of reliable unstable region and unreliable stable region, different learning strategies are designed according to the characteristics of the reliable unstable region and the unreliable stable region, the confidence guided cross prototype consistency learning module is used to approximate the probability distribution of the voxels in each category by calculating the cosine similarity between the features and the class feature prototype, and the global constraint is introduced, so that the excessive dependence of the model on the local features can be effectively avoided; the dynamic teacher competitive teaching module selects the pixels with higher confidence from the two prediction results without feature disturbance, generates a new teacher prediction, and converts it into a more reliable soft pseudo label by using a probability sharpening function, so that the negative influence of noise and uncertainty on model training can be effectively relieved, the competitive learning between models is promoted, and the adaptability of the model to complex regions is significantly enhanced.

[0082] (2) The confidence guided cross prototype consistency learning module solves the problem that the model cannot accurately identify the key region features due to unstable prediction in the region. Unlike the mainstream method of directly processing the model decision space, the confidence guided cross prototype consistency learning module constructs the class prototype and optimizes the consistency between the features and the prototype, and introduces the confidence constraint, realizes the learning from low confidence prediction to high confidence prediction, can effectively guide the model to learn more reliable feature representation in the reliable unstable region, and enhances the identification ability of the model in the key region.

[0083] (3) The dynamic teacher competitive teaching module is proposed to solve the problem that the accumulation of false signals caused by unreliable prediction in the region leads to prediction deviation. Unlike the mainstream method of using a single model to generate a comprehensive reliable prediction as a pseudo label, the dynamic teacher competitive teaching selects the more reliable pixels in the model prediction without disturbance as the teacher prediction through pixel-level reliability evaluation, and generates a new prediction as a soft pseudo label in a competitive manner by using the prediction difference between the two networks, so as to supervise the model prediction with disturbance and provide more valuable learning signals for the model, thereby effectively reducing the prediction deviation of the model.

[0084] In summary, the present application is widely applied in the fields of medical image analysis, lesion detection, surgical navigation and the like, not only lays a solid foundation for precision medicine and intelligent diagnosis, but also provides a new research idea and technical means for the cross field of artificial intelligence and medicine. BRIEF DESCRIPTION OF DRAWINGS

[0085] Figure 1 The flow principle block diagram of the present application is shown in the figure;

[0086] Figure 2 The framework diagram of the DTCT module proposed by the present application is shown in the figure;

[0087] Figure 3 The visualization comparison chart of the segmentation results of the present application and the existing mainstream method on the ACDC dataset and the AbdomenCT-1K dataset. DETAILED DESCRIPTION

[0088] The present application will be further described in detail below in conjunction with specific examples, which are an explanation of the present application rather than a limitation.

[0089] As shown in the drawings, Figure 1 The present application proposes a semi-supervised medical image segmentation method based on adaptive learning of high-value regions, constructs a semi-supervised medical image segmentation model based on adaptive learning of high-value regions (ALHVR), which is composed of two segmentation networks with the same structure and independent parameter updating. By applying random perturbation in the feature space, each network outputs the original prediction and the perturbed prediction. Based on the confidence distribution of the perturbed prediction results, different value regions are selected, and differentiated training strategies are designed for the characteristics of these regions to further improve the segmentation performance of the model. Specifically, a confidence-guided cross-prototype consistency learning (CG-CPCL) module and a dynamic teacher competition teaching (DTCT) module are proposed, which are used to train reliable unstable regions and unreliable stable regions, respectively. Compared with existing mainstream semi-supervised learning methods, the present application not only deeply explores the regions with more learning value in semi-supervised medical image segmentation, but also divides the high-value regions into two categories: reliable unstable regions and unreliable stable regions. In addition, different learning strategies are designed for the characteristics of these high-value regions to further improve the segmentation performance of the model in key regions. The present application has significant application value, especially in the field of medical image segmentation, and shows a wide application prospect.

[0090] A semi-supervised medical image segmentation method based on adaptive learning of high-value regions, the specific steps are as follows:

[0091] (1) Data set source and preprocessing operation:

[0092] The ACDC dataset consists of cardiac dynamic magnetic resonance imaging scans of 100 patients and their corresponding labels, which are divided into three classes: right ventricle, left ventricle, and myocardium. 70 cases are used for training, 10 cases are used for validation, and 20 cases are used for testing. In the training process, the original 3D scan data is converted into 2D slices, and the size of all slices is adjusted to 256*256 pixels as input. In the present application, 10% and 20% of the training dataset are used as labeled data, and the rest of the data are used as unlabeled data.

[0093] The AbdomenCT-1K dataset contains more than 1000 (1K) CT scan images from 12 medical centers, which are divided into four classes including liver, kidney, spleen, and pancreas. The resolution of each CT scan is 512*512 pixels, and the pixel size and slice thickness (between 1.25 and 5 mm) vary. 30 cases are selected for training, and 11 cases are selected for testing. In the training process, the original 3D scan data is converted into 2D slices, and the size of all slices is adjusted to 256*256 pixels as input. In the present application, 5% and 10% of the training dataset are used as labeled data, and the rest of the data are used as unlabeled data.

[0094] The Brats dataset consists of preoperative magnetic resonance imaging scans of 335 patients with glioma, including 259 high-grade glioma (HGG) and 76 low-grade glioma (LGG). Each patient's magnetic resonance imaging scan includes four modes: T1, T1Gd, T2, and T2-FLAIR, and is labeled by expert pixel-by-pixel inspection. 250 cases are used for training, 25 cases are used for validation, and 60 cases are used for testing. In the training process, the data is randomly cropped into small blocks of 96*96*96 voxels as input; in the testing phase, a sliding window of the same size is used to extract features with a fixed stride of 64*64*64. In the present application, 10% of the training data are used as labeled data, and the rest of the data are used as unlabeled data.

[0095] (2) The overall architecture of the ALHVR model proposed in the present application: ALHVR is composed of two parallel networks f A (·) and f B (·) with the same structure and independent parameter updates, the training set is input into the encoders of f A (·) and f B (·), the encoder outputs the original feature F1, and the original feature F1 and the perturbation feature F2 of the original feature F1 are input into the corresponding decoders and to generate prediction outputs of different perspectives. The model is trained in a traditional supervised manner on the labeled dataset D LThe two networks are trained on the labeled dataset D, while consistency between different view predictions is enforced on the unlabeled dataset D U The two networks are trained on the labeled dataset D, while consistency between different view predictions is enforced on the unlabeled dataset D

[0096] τ←random(0,1) (1)

[0097]

[0098] The original features and perturbed features of the two networks are input into the corresponding decoders and to obtain the final prediction output:

[0099]

[0100] where F a1 represents the original feature output by the f A (·) encoder, represents the original prediction result of f A (·), F a2 represents the perturbed feature of the original feature F A output by the f a1 (·) encoder, represents the perturbed prediction result of f A (·), F b1 represents the original feature output by the f B (·) encoder, represents the original prediction result of f B (·), F b2 represents the perturbed feature of the original feature F B output by the f b1 (·) encoder, represents the perturbed prediction result of f B (·), and together constitute The decoder output includes the prediction results of the labeled data and the unlabeled data, i.e. (3) Different value region classification method: In order to solve the limitation of single model in evaluating the difficulty of pixel prediction, the present application uses a pixel classification method based on double network cognitive interaction. By fully utilizing the consistency and difference of the prediction results of the double networks, an adaptive threshold is used to realize fine evaluation of the difficulty of pixel prediction. Specifically, the prediction confidence map of the two networks is used to generate an adaptive threshold, and the prediction result here uses the prediction with feature perturbation.

[0101] (a) First, two network predicted confidence maps are generated:

[0102]

[0103] where con a2 represents the confidence map generated by the perturbation prediction result of f A (·), con b2 represents the confidence map generated by the perturbation prediction result of f B (·);

[0104] (b) Then, the two confidence maps are averaged pixel by pixel to obtain a comprehensive confidence map. Then, all pixels in the comprehensive confidence map are sorted in ascending order according to the confidence values:

[0105]

[0106] (c) Finally, the lowest K pixels in the sorted confidence are selected, and the mean value thereof is calculated as an adaptive threshold:

[0107]

[0108] where K = β(H·W·D), H, W, and D represent the height, width, and depth of the image, respectively. β is a hyperparameter with a value between (0, 1) for controlling the proportion of the number of selected pixels, and γ represents the confidence threshold.

[0109] (e) According to the confidence threshold, the predicted pixels of the two networks are interactively classified. By comparing the relationship between the corresponding pixels in the two confidence maps and the threshold, the pixels can be divided into three categories: when the confidence of both pixels is higher or lower than the threshold, the reliable stable region Ω1 and the unreliable stable region Ω3 are formed, respectively; when the confidence of the two pixels is one higher than the threshold and the other lower than the threshold, then the reliable unstable region Ω2 is formed. These regions can be defined as:

[0110] Ω1 = {con a2 ≥ γ & con b2 ≥ γ} (8)

[0111] Ω2 = {(con a2 > γ & con b2 < γ) | (con a2 < γ & con b2 > γ)} (9)

[0112] Ω3 = {con a2 < γ & con b2 < γ} (10)

[0113] Since the reliable stable region is an easy-to-learn region for both networks to reach consensus, the reliable unstable region is a region where both networks have differences, and the unreliable stable region is a difficult-to-learn region for both networks to reach consensus. Therefore, compared with the reliable stable region, focusing on the reliable unstable region and the unreliable stable region can better improve the learning ability of the model and has higher learning value.

[0114] (4) The training process of the CG-CPCL module designed by the application:

[0115]

[0116] The above describes the algorithm pseudo code of the CG-CPCL module, and the CG-CPCL approximates the probability distribution of the voxels in each class by calculating the cosine similarity between the perturbation features and the class feature prototypes. In this way, by introducing global constraints, the model can effectively avoid excessive dependence on local features. In addition, in order to ensure that the high-confidence prediction of one network in the dual network can guide the low-confidence prediction of the other network to learn, the application adopts a confidence condition guiding mechanism to realize cross-consistency learning, thereby effectively improving the stability of the model to the prediction divergence region.

[0117] (a) First, the class feature prototype of the unlabeled data is generated, and the decoder and The features of the second-to-last layer are upsampled using trilinear interpolation to match the size of the pseudo label, and the pseudo label is used to filter the features of the corresponding class, and the class feature prototype of the unlabeled data is generated by calculating the mean of the features of different classes. Specifically:

[0118]

[0119] Where B U represents the batch size of the unlabeled data, (h, w, z) represents the spatial coordinates, represents the original features of the decoder of the second-to-last layer, represents the indicator function, represents the pseudo label generated by , and represents the original features of the decoder of the second-to-last layer, represents the pseudo label generated by , and and respectively represent the c-th class feature prototype from the decoder and .

[0120] (b) After obtaining the class feature prototype, the cosine similarity between the perturbed feature and the class feature prototype is calculated to approximate the prediction probability of the model:

[0121]

[0122] wherein, F d represents the feature of the current sample, p c represents the class feature prototype, represents the decoder the perturbed feature of the second-to-last layer, represents the decoder the perturbed feature of the second-to-last layer, sim a represents the perturbed feature of f A (·) and the cosine similarity between the class feature prototype, sim b represents the perturbed feature of f B (·) and the cosine similarity between the class feature prototype;

[0123] (c) In order to ensure that the low-confidence prediction in the Ω2 region is only learned from reliable targets, a pixel screening is required for the confidence map of each network prediction, so as to generate a location mask containing the interaction information of the two networks, and use it as a guide signal for model learning:

[0124] Mask a = con a2 <γ&con b2 >γ, Mask b = con a2 >γ&con b2 <γ (13)

[0125] wherein, Mask a represents the location mask of the reliable unstable region of f A (·), Mask b represents the location mask of the reliable unstable region of f B (·);

[0126] (d) Then, the cross-prototype consistency loss can be calculated using the location mask, and the formula is as follows:

[0127]

[0128] wherein, ⊙ represents the pixel-by-pixel multiplication operation; represents the cross-prototype consistency loss of f A (·), represents the cross-prototype consistency loss of f B (·);

[0129] In addition, in order to further enhance the model's ability to learn category features, the feature-prototype cosine similarity loss is also calculated:

[0130]

[0131] in, represents f A cosine similarity loss of (·), represents f B Cosine similarity loss of (·);

[0132] (e) Finally, the total loss for each model training reliable unstable region is composed of the cross-prototype consistency loss and the feature-prototype cosine similarity loss:

[0133]

[0134] in, represents f A (·) total loss for training in reliably unstable regions, represents f B (·) Total loss for training in reliably unstable regions.

[0135] (5) Training process of the DTCT module designed by the present invention: Figure 2 This paper describes the process of the DTCT module. DTCT generates new teacher predictions by selecting the pixels with higher confidence from the two unperturbed predictions. Using a probabilistic sharpening function, these new teacher predictions are converted into more reliable soft pseudo labels, effectively mitigating the negative impact of noise and uncertainty on model training. Ultimately, the newly generated reliable pseudo labels are used to supervise the perturbed predictions, providing a more accurate supervisory signal for perturbed predictions. In this way, the DTCT module not only promotes competitive learning between models but also significantly enhances the model's adaptability to complex regions.

[0136] (a) To generate new teacher predictions, we perform pixel-by-pixel confidence comparison based on the prediction confidence map without feature perturbation, retaining pixels with higher confidence as new predictions while discarding pixels with lower confidence. First, we generate f A (·) and f B (·) Confidence map of the original prediction result:

[0137]

[0138] Among them, con a1 Indicated by f A The confidence map generated by the original prediction result of (·), conn b1 Indicated by f Bthe confidence map generated by the original prediction result of

[0139] (b) Then, by comparing the generated confidence map, the prediction probability corresponding to the pixel with higher confidence is selected as the new teacher prediction, and it is converted into a soft pseudo label by probability sharpening. This process can be defined as:

[0140]

[0141] where T is a hyperparameter that controls the output distribution of the model, represents the newly generated teacher prediction, represents the soft pseudo label generated by

[0142] (c) For each network perturbation prediction confidence map, pixel screening is performed to generate a position mask to determine the location of the unreliable stable area Ω3, specifically:

[0143] Mask = con a2 <γ&con b2 <γ (21)

[0144] where Mask represents the position mask of the unreliable stable area of f A (·) and f b (·);

[0145] (d) Then, the newly generated soft pseudo label is used to supervise the perturbation version prediction, and the formula is as follows:

[0146]

[0147] where represents the consistency loss of f A (·) trained in the unreliable stable area, represents the consistency loss of f B (·) trained in the unreliable stable area;

[0148] (e) In order to enhance the model's uncertainty perception ability in the unreliable stable area, guide the model to focus on the area with high uncertainty, and effectively reduce the uncertainty in these areas, the uncertainty is quantified by calculating the information entropy, and it is added to the loss function as a regularization term:

[0149]

[0150] where H a1 represents the information entropy of the original prediction of f A (·) in the unreliable stable area, H a2 represents the information entropy of the new prediction of f A ​The information entropy H of the disturbance prediction of (·) on the unreliable stable region b1 represents f B The information entropy H of the original prediction of (·) on the unreliable stable region b2 represents f B The information entropy H of the disturbance prediction of (·) on the unreliable stable region

[0151] (f) Finally, the total loss of each model training the unreliable stable region is composed of the mean square error loss and the information entropy:

[0152]

[0153] wherein, represents f A The total loss of (·) trained on the unreliable stable region, represents f B The total loss of (·) trained on the unreliable stable region.

[0154] (6) The setting of the total loss function in the application: for the labeled data, the combination of cross entropy loss and Dice loss is used as the loss function of supervised learning:

[0155]

[0156] wherein, y represents a label, represents f A The supervised loss of (·), represents f B The supervised loss of (·);

[0157] For unlabeled data, cross entropy loss and cosine similarity loss are introduced through the CG-CPCL module, and DTCT module is used to combine MSE loss and information entropy, and finally an unsupervised loss function is formed:

[0158]

[0159] wherein, represents f A The unsupervised loss of (·), represents f B The unsupervised loss of (·);

[0160] The total loss of each network is defined as:

[0161]

[0162] wherein, L a represents f A The total loss of (·), L b represents f Bthe total loss of (·), λ is a Gaussian warm-up function that changes over time, used to balance the supervised loss and unsupervised loss.

[0163] The effects of the present application can be further illustrated by the following experiments.

[0164] All experiments of the present application are implemented using Pytorch1.7 framework and NVIDIA GeForce RTX 3090 GPU, the model is trained using SGD optimizer, the weight decay factor is set to 10 -4 , the momentum is set to 0.9, and the initial learning rate is set to 10 -2 . In the segmentation task of ACDC dataset and AbdomenCT-1K dataset, the backbone network uses U-Net, the batch size is set to 16, each batch contains 8 labeled data and 8 unlabeled data, and the number of iterations in the training stage is set to 30000. In the segmentation task of Brats dataset, the backbone network uses V-Net. The batch size is set to 4 respectively, each batch contains 2 labeled data and 2 unlabeled data respectively, and the number of iterations in the training stage is set to 15000. In the total loss function of the network, the calculation formula of the Gaussian warm-up function is: where t represents the current iteration number, t max represents the maximum iteration number. In the probability sharpening function, the hyperparameter T is usually set to 0.1 according to experience. When calculating the confidence threshold, the hyperparameter β is set to 0.8, 0.7 and 0.7 in ACDC, AbdomenCT-1K and Brats three datasets respectively.

[0165] Firstly, the performance of the present application is evaluated on the ACDC dataset, and the comparison methods include MT, ICT, CCT, DTC, CPS, MC-Net, URPC, SS-Net, MCF, AC-MT, UG-MCL, EVIL, AAU and MLRP. In addition, the indicators obtained by training U-Net using 10%, 20% and 100% of the labeled data are used as the baseline and upper limit of the reference performance. From Table 1, it can be observed that the present application has significantly improved in the four evaluation indicators compared with the baseline model. Specifically, when training using 10% of the labeled data, Dice is improved from 81.59% to 90.56%; when training using 20% of the labeled data, Dice is improved from 84.83% to 91.09%, close to the upper limit of the full supervision learning 91.10% (training U-Net using 100% of the labeled data). Compared with the best method, ALHVR still maintains the best performance in two training settings. When training using 10% of the labeled data, Dice is improved by 3.31% compared with the suboptimal method (EVIL). When the labeled data increases to 20%, Dice is improved by 2.69% compared with the suboptimal method UG-MCL.

[0166] Table 1 Comparison with state-of-the-art segmentation methods on ACDC dataset, where bold font indicates the best indicator results

[0167]

[0168]

[0169] To further verify the generalization ability of the present application on different datasets, performance evaluation was also conducted on the five-classification dataset AbdomenCT-1K. The comparative methods used were the same as those of the ACDC dataset. As shown in Table 2, when 5% of the labeled data was used for training, the Dice of ALHVR increased by 17.36% compared with the baseline model; when the amount of labeled data was increased to 20%, the Dice increased by 10.02%. Compared with the state-of-the-art methods, the present application still remained optimal in the segmentation task of the four organs. When 5% and 10% of the labeled data was used for training, the Dice of ALHVR increased by 3.78% and 1.13% compared with the suboptimal methods (MC-Net and MLRP) corresponding thereto, respectively. In addition, from the attached figures, it can be seen that the present application can more accurately segment the branch or edge region of the four organs. Figure 3 It can be seen that the present application can more accurately segment the branch or edge region of the four organs.

[0170] Table 2 Comparison with state-of-the-art segmentation methods on AbdomenCT-1K dataset, where bold font indicates the best indicator results

[0171]

[0172]

[0173] In addition to the experiment on the organ segmentation task, the present application also performed performance evaluation on the brain tumor dataset Brats. The comparative methods used were the same as those of the ACDC dataset. As shown in Table 3, when only 10% of the labeled data was used for training, the Dice of ALHVR increased by 5.91% compared with the baseline model. In addition, compared with the state-of-the-art methods, ALHVR still occupied an advantage, and the Dice increased by 1.18% compared with the suboptimal method (MLRP).

[0174] Table 3 Comparison with state-of-the-art segmentation methods on Brats dataset, where bold font indicates the best indicator results

[0175]

Claims

1. A semi-supervised medical image segmentation method based on adaptive learning of high-value regions, characterized in that: The following steps are included: S1, preprocess the data set and divide the preprocessed data set into training set, validation set and test set; S2. Constructing a semi-supervised medical image segmentation model based on adaptive learning of high-value regions. The semi-supervised medical image segmentation model includes two parallel networks with identical structures and independently updated parameters. By applying random perturbations in the feature space, the two parallel networks output original prediction results and perturbed prediction results. Based on the confidence distribution of the perturbation prediction results, different value regions are screened out, including reliable unstable regions and unreliable stable regions. The reliable unstable regions are trained through a confidence-guided cross-prototype consistency learning module, and the unreliable stable regions are trained through a dynamic teacher competition teaching module. S3, using random initialization parameters, training the semi-supervised medical image segmentation model using the training set, selecting the optimal weight based on the validation set, and obtaining a semi-supervised medical image segmentation model with the optimal weight; S4, input the test set into the semi-supervised medical image segmentation model with optimal weights, and output the medical image segmentation results.

2. The semi-supervised medical image segmentation method based on high-value region adaptive learning according to claim 1, characterized in that: The data set preprocessing in step S1 is specifically to normalize the image distribution by randomly cropping the image to a uniform size, and to improve the diversity of the image by rotating and flipping. The training set includes labeled data and unlabeled data.

3. The semi-supervised medical image segmentation method based on high-value region adaptive learning according to claim 1, characterized in that: The two parallel networks with the same structure and independent parameter update in step S2 are U-Net or V-Net, respectively represented by f A (·) and f B (·).

4. The semi-supervised medical image segmentation method based on high-value region adaptive learning according to claim 3, characterized in that: The different value areas screened out in step S2 are specifically: S2.11, input the training set into f A (·) and f B In the encoder of (·), the encoder outputs the original feature F1, and the original feature F1 output by the encoder and the perturbation feature F2 of the original feature F1 are input to the corresponding decoder and , we get the final prediction result: Among them, F a1 represents f A (·) The original features output by the encoder, represents f A The original prediction result of (·), F a2 represents f A (·) The original feature F output by the encoder a1 The disturbance characteristics of represents f A (·) perturbation prediction results, F b1 represents f B (·) The original features output by the encoder, represents f b The original prediction result of (·), F b2 represents f B (·) The original feature F output by the encoder b1 The disturbance characteristics of represents f B The perturbation prediction results of (·) are: and Common composition The decoder output includes the prediction results of labeled data and unlabeled data, that is, S2.12, generate confidence maps of two parallel networks based on the perturbation prediction results obtained in step S2.11: Among them, con a2 Indicated by f A The confidence map generated by the perturbation prediction results of (·) is b2 Indicated by f B Confidence map generated by the perturbation prediction results of (·); S2.13, based on the confidence map obtained in step S2.12, perform pixel-by-pixel averaging on the two confidence maps to obtain a comprehensive confidence map, and sort all pixels in the comprehensive confidence map in ascending order according to their confidence values: S2.14, select the K pixels with the lowest confidence after sorting, and calculate the mean of the K pixels as the adaptive threshold: Where K = β·(H·W·D), where H, W, and D refer to the height, width, and depth of the image, respectively. β is a hyperparameter with a value between (0, 1) that controls the number of pixels to be filtered. γ represents the confidence threshold. S2.15, interactively classify the predicted pixels of the two networks according to the confidence threshold obtained in step S2.

14. When the confidence of the two pixels is higher than the confidence threshold at the same time, a reliable stable region Ω1 is formed; when the confidence of the two pixels is lower than the confidence threshold at the same time, an unreliable stable region Ω3 is formed; when the confidence of one pixel is higher than the confidence threshold and the other is lower than the confidence threshold, a reliable unstable region Ω2 is formed. Specifically: Ω1={with a2 ≥γ&con b2 ≥γ} (6) Ω2={(con a2 >c&con b2 <c)|(con a2 <c&con b2 >c)} (7) Ω3={with a2 <γ&con b2 <γ} (8)。 5. The semi-supervised medical image segmentation method based on high-value region adaptive learning according to claim 1, characterized in that: The specific training of the reliable unstable region by the confidence-guided cross-prototype consistency learning module in step S2 is: S2.21, for decoder and The features of the penultimate layer are upsampled using trilinear interpolation to match the size of the features with the pseudo-labels. The pseudo-labels are used to filter out the features of the corresponding categories, and the class feature prototypes of the unlabeled data are generated by calculating the mean of the features of different categories. Specifically: Among them, B U represents the batch size of unlabeled data, (h,w,z) represents the spatial coordinates, Representation decoder The original features of the penultimate layer, represents the indicator function, Indicated by The pseudo labels generated, Representation decoder The original features of the penultimate layer, Indicated by The pseudo labels generated, and Respectively represent the decoder and The c-th characteristic prototype; S2.22, based on the class feature prototype obtained in step S2.21, approximate the model's prediction probability by calculating the cosine similarity between the perturbation feature and the class feature prototype: in, F d Represents the characteristics of the current sample, p c Represents the class feature prototype, Representation decoder The perturbation features of the penultimate layer, Decoder The perturbation features of the penultimate layer, sim a represents f A The cosine similarity between the perturbation feature and the prototype of the class feature of (·) is sim b represents f B The cosine similarity between the perturbation feature and the prototype of the class feature (·); S2.23, perform pixel screening on the confidence map of each network perturbation prediction and generate a position mask to determine the location of the reliable unstable region Ω2, specifically: Mask a =with a2 <γ&con b2 >γ,Mask b =with a2 >γ&con b2 <γ (11) Among them, Mask a represents f A (·) is the location mask of the reliable unstable area, Mask b represents f B (·) is the position mask of the reliable unstable regions; S2.24, using the position mask obtained in step S2.23, calculate the cross-prototype consistency loss, specifically: Among them, ⊙ represents the pixel-by-pixel product operation; represents f A (·) cross-prototype consistency loss, represents f B (·) cross-prototype consistency loss; S2.25, calculate the cosine similarity loss of features and class feature prototypes: in, represents f A cosine similarity loss of (·), represents f B Cosine similarity loss of (·); S2.26, the total loss of the cross-prototype consistency learning module for training reliable unstable regions includes the cross-prototype consistency loss and the cosine similarity loss of feature and class feature prototypes, specifically: in, represents f A (·) total loss for training in reliably unstable regions, represents f B (·) Total loss for training in reliably unstable regions.

6. The semi-supervised medical image segmentation method based on high-value region adaptive learning according to claim 1, characterized in that: The specific steps of training the unreliable stable region by the dynamic teacher competition teaching module in step S2 are as follows: S2.31, generate f A (·) and f B (·) Confidence map of the original prediction result: Among them, con a1 Indicated by f A The confidence map generated by the original prediction result of (·) is b1 Indicated by f B Confidence map generated by the original prediction results of (·); S2.32, compare the generated confidence maps, select the predicted probability corresponding to the pixel with higher confidence as the new teacher prediction, and perform probability sharpening to convert it into a soft pseudo label, specifically: Among them, T is a hyperparameter that controls the distribution of model output, represents the newly generated teacher prediction, Indicated by Generated soft fake labels; S2.33, perform pixel screening on the confidence map of each network perturbation prediction and generate a position mask to determine the location of the unreliable stable region Ω3, specifically: Mask=con a2 <γ&con b2 <γ (19) Among them, Mask represents f A (·) and f B (·) is the position mask of the unreliable stable regions; S2.34, using the soft pseudo labels generated in step S2.32 to supervise the prediction of the perturbation features, specifically: in, represents f A (·) consistency loss for training in unreliable stable regions, represents f B (·) consistency loss for training in unreliable stable regions; S2.35, quantify uncertainty by calculating information entropy and add it to the loss function as a regularization term: in, H a1 represents f A The information entropy of the original prediction of (·) on the unreliable stable region, H a2 represents f A The information entropy of the perturbation prediction of (·) on the unreliable stable region, H b1 represents f B The information entropy of the original prediction of (·) on the unreliable stable region, H b2 represents f B The information entropy of the perturbation prediction of (·) on the unreliable stable region; S2.36, the total loss of the dynamic teacher competition teaching module training unreliable stable region includes mean square error loss and information entropy, specifically: in, represents f A (·) the total loss of training in the unreliable stable region, represents f B (·) Total loss for training in the unreliable stable region.

7. The semi-supervised medical image segmentation method based on high-value region adaptive learning according to claim 1, characterized in that: The total loss function for training the semi-supervised medical image segmentation model using the training set in step S3 is set to: For labeled data, a combination of cross entropy loss and Dice loss is used as the loss function for supervised learning: Among them, y represents the label, represents f A (·) supervision loss, represents f B (·) loss of supervision; For unlabeled data, the cross entropy loss and cosine similarity loss are introduced through the CG-CPCL module, and the DTCT module is used to combine the MSE loss and information entropy to finally form an unsupervised loss function: in, represents f A The unsupervised loss of (·), represents f B Unsupervised loss of (·); The total loss function is defined as: Among them, L a represents f A (·) total loss, L b represents f B (·) is the total loss, and λ is a time-varying Gaussian warm-up function used to balance the supervised loss and unsupervised loss.

8. The semi-supervised medical image segmentation method based on high-value region adaptive learning according to claim 1, characterized in that: In step S3, the model is trained using the SGD optimizer, and the weight decay factor is set to 10 -4 , the momentum is set to 0.9, and the initial learning rate is set to 10 -2 .

Images