Frequency-based adaptive interactive learning semi-supervised medical segmentation method and system

Abstract

The application discloses a frequency-based adaptive interactive learning semi-supervised medical segmentation method and system, and belongs to the technical field of medical image segmentation. The method comprises the following steps: acquiring an image to be segmented and inputting the image to a pre-trained semi-supervised medical segmentation model for processing to obtain an image segmentation result; the semi-supervised medical segmentation model comprises a high-frequency subnetwork and a low-frequency subnetwork, and the training process of the semi-supervised medical segmentation model comprises an internal learning process and an external learning process. The specific training steps are as follows: a high-frequency view and a low-frequency view are generated through a frequency double-view flow architecture; in the internal learning process, the high-frequency view and the low-frequency view are taken as inputs to construct a complementary view, and the corresponding subnetwork state is updated through cross-pseudo-label supervision and mutual uncertainty distance supervision, and the corresponding subnetwork state is taken as the initial state of the external learning process; in the external learning process, the corresponding subnetwork state is updated through cross-supervision loss and adaptive self-correction loss, and the corresponding subnetwork state is taken as the initial state of the internal learning process in the next cycle.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of medical image segmentation, and particularly relates to a frequency-based adaptive interactive learning semi-supervised medical segmentation method and system. BACKGROUND

[0002] The statements herein are provided only to enhance understanding of the present application and are not necessarily intended to constitute the prior art.

[0003] Medical image segmentation plays a crucial role in modern computer-aided diagnosis systems, aiding in the accurate identification of tissues and organs in imaging modalities such as magnetic resonance imaging (MRI) and computed tomography (CT). Most current state-of-the-art segmentation methods rely on large-scale datasets with complete annotations, but obtaining such pixel-level labels for medical images is both laborious and time-consuming, requiring the expertise of trained radiologists. Semi-supervised medical image segmentation (SSMIS) aims to train segmentation models by combining a small set of labeled data with a large amount of unlabeled data, with the core idea being to effectively utilize the information in unlabeled data. Pseudo-labeling provides a direct and effective strategy to achieve this goal.

[0004] In the prior art, there are many studies that adopt self-training pseudo-labeling methods, generating pseudo-labels from weakly perturbed inputs based on a single model and using them to supervise the predictions of strongly perturbed inputs, thereby maintaining consistency between different transformations of medical images, as shown in Figure 1 (a).

[0005] Currently, the concept of co-training has been integrated into deep learning, and some studies combine co-training architectures with pseudo-labeling to utilize more diverse feature representations. As shown in Figure 1 (b), co-training trains multiple subnetworks on different views of the same instance, allowing each subnetwork to capture complementary information. Instead of relying on self-generated pseudo-labels, it employs cross-pseudo-labeling, where each subnetwork generates pseudo-labels to supervise the predictions of other subnetworks, forming a cross-supervision mechanism. Compared to self-training methods, this collaborative pseudo-labeling strategy allows subnetworks to infer the same instance from different perspectives, thereby improving the diversity and robustness of generated pseudo-labels.

[0006] However, while pseudo-labels can enhance the ability to learn from unlabelled data, they also pose a significant challenge: when the generated pseudo-labels are inaccurate, the model can reinforce these incorrect predictions during training, which can lead to error accumulation and reduce model stability. This is known as confirmation bias, which remains a persistent problem for pseudo-label-based methods. To address this issue, existing methods introduce confidence-based strategies to improve pseudo-label quality by iteratively retaining high-confidence pseudo-labels and filtering out low-confidence pseudo-labels, which aim to improve the reliability of pseudo-labels across consecutive training steps. However, since incorrect pseudo-labels can also exhibit high confidence, this assumption may not always hold true, and therefore, without a feedback mechanism to improve or correct these errors, the effectiveness of such methods remains limited.

[0007] In summary, the inventors have found that existing semi-supervised medical image segmentation methods suffer from at least the following shortcomings:

[0008] (1) Pseudo-label-based methods can suffer from confirmation bias, where incorrect or inaccurate pseudo-labels are reinforced during training, leading to error accumulation and reduced model stability, which in turn results in decreased model performance;

[0009] (2) Confidence-based strategies can suffer from unreliable pseudo-label confidence, making it unreliable to improve pseudo-label quality by filtering high-confidence pseudo-labels, which limits the effectiveness of such methods. SUMMARY

[0010] The present application aims to overcome the shortcomings of the prior art and provide a frequency-based adaptive interactive learning semi-supervised medical segmentation method and system that aims to improve pseudo-label reliability and self-correction ability, address the limitations of pseudo-label-based methods, and mitigate confirmation errors through the establishment of an adaptive feedback loop, thereby improving segmentation model stability and ultimately improving medical image segmentation accuracy.

[0011] To achieve the above-mentioned purpose, the present application is realized by the following technical solutions:

[0012] In one aspect, the technical solutions of the present application provide a frequency-based adaptive interactive learning semi-supervised medical segmentation method, comprising:

[0013] obtaining an image to be segmented;

[0014] inputting the image to be segmented into a pre-trained semi-supervised medical segmentation model for processing to obtain an image segmentation result;

[0015] The semi-supervised medical segmentation model includes a high-frequency subnetwork and a low-frequency subnetwork, and the training process includes an internal learning process and an external learning process, and the specific training steps are as follows:

[0016] The high-frequency view and the low-frequency view are generated through a frequency dual-view flow architecture.

[0017] In the internal learning process, the complementary view is constructed by taking the high-frequency view and the low-frequency view as input, and the corresponding subnetwork state is updated through cross-pseudo-label supervision and mutual uncertainty distance supervision, and is taken as the initial state of the external learning process.

[0018] In the external learning process, the corresponding subnetwork state is updated through cross-supervision loss and adaptive self-correction loss, and is taken as the initial state of the internal learning process of the next cycle.

[0019] In at least one embodiment, the high-frequency view and the low-frequency view are generated through a frequency dual-view flow architecture, and the specific process includes:

[0020] The high-frequency components and the low-frequency components of the labeled images and the unlabeled images are extracted based on a Gaussian filter respectively.

[0021] The enhanced high-frequency view and the enhanced low-frequency view are constructed based on a set of enhancement factors respectively.

[0022] The local features are extracted from the enhanced low-frequency view by using a high-frequency subnetwork based on CNN, and the long-distance dependency in the enhanced high-frequency view is captured by using a low-frequency subnetwork based on Transformer.

[0023] In at least one embodiment, the cross-pseudo-label supervision includes the following specific process:

[0024] The weak enhancement variant and the strong enhancement variant of the high-frequency view and the low-frequency view are created respectively by using a random data enhancement strategy.

[0025] The high-frequency one-hot pseudo-label and the low-frequency one-hot pseudo-label are generated respectively based on the logit output of the strong enhancement variant.

[0026] The obtained pseudo-labels are used to supervise the subnetworks through cross-pseudo-label supervision, and the cross-pseudo-label supervision loss is the sum of the cross-entropy loss and the Dice loss of each subnetwork.

[0027] In at least one embodiment, the mutual uncertainty distance supervision optimizes the prediction of the unlabeled samples by using uncertainty information, including class entropy minimization and entropy distance minimization; and the loss function of the mutual uncertainty distance supervision is a weighted combination of the class entropy minimization loss and the entropy distance minimization loss.

[0028] In at least one embodiment, in the internal learning process, updating the state of the corresponding subnetwork is specifically:

[0029] The inner loop loss of the high-frequency subnetwork and the low-frequency subnetwork is calculated respectively;

[0030] Based on the inner loop loss, the state parameters of the high-frequency subnetwork and the low-frequency subnetwork are updated by gradient descent; the updated high-frequency subnetwork state parameters and low-frequency subnetwork state parameters are used as the initial state parameters of the high-frequency subnetwork and the low-frequency subnetwork in the external learning process respectively.

[0031] In at least one embodiment, the inner loop loss of the high-frequency subnetwork is the sum of the cross-pseudo label supervision loss and the mutual uncertainty distance supervision loss of the high-frequency subnetwork; the inner loop loss of the low-frequency subnetwork is the sum of the cross-pseudo label supervision loss and the mutual uncertainty distance supervision loss of the low-frequency subnetwork.

[0032] In at least one embodiment, in the external learning process, the pseudo label confidence is dynamically adjusted through a performance feedback coefficient, the adaptive self-correction loss of the corresponding subnetwork is calculated based on the performance feedback coefficient, and the learning strategy of the subnetwork is adjusted through the adaptive self-correction loss.

[0033] In at least one embodiment, in the external learning process, updating the state of the corresponding subnetwork is specifically:

[0034] The outer loop loss of the high-frequency subnetwork and the low-frequency subnetwork is calculated respectively;

[0035] Based on the outer loop loss, the state parameters of the high-frequency subnetwork and the low-frequency subnetwork are updated by gradient descent; the updated high-frequency subnetwork state parameters and low-frequency subnetwork state parameters are used as the initial state parameters of the high-frequency subnetwork and the low-frequency subnetwork in the internal learning process of the next cycle respectively.

[0036] In at least one embodiment, the outer loop loss of the high-frequency subnetwork is the sum of the cross-supervision loss and the adaptive self-correction loss of the high-frequency subnetwork; the inner loop loss of the low-frequency subnetwork is the sum of the cross-supervision loss and the mutual uncertainty distance supervision loss of the low-frequency subnetwork.

[0037] In another aspect, the technical scheme of the present application also provides a frequency-based adaptive interactive learning semi-supervised medical segmentation system, comprising:

[0038] An acquisition module configured to acquire a to-be-segmented image;

[0039] An image segmentation module configured to input the to-be-segmented image into a pre-trained semi-supervised medical segmentation model for processing to obtain an image segmentation result;

[0040] The semi-supervised medical segmentation model includes a high-frequency subnetwork and a low-frequency subnetwork, and the training process includes an internal learning process and an external learning process, and the specific training steps are as follows:

[0041] The high-frequency view and the low-frequency view are generated through the frequency dual-view flow architecture.

[0042] In the internal learning process, the complementary view is constructed by taking the high-frequency view and the low-frequency view as input, and the corresponding subnetwork state is updated through cross-pseudo-label supervision and mutual uncertainty distance supervision, and is taken as the initial state of the external learning process.

[0043] In the external learning process, the corresponding subnetwork state is updated through cross-supervision loss and adaptive self-correction loss, and is taken as the initial state of the internal learning process of the next cycle.

[0044] The beneficial effects of the technical scheme of the present application are as follows:

[0045] The frequency dual-view flow architecture is constructed by integrating the frequency characteristics of the image and the frequency trend of the network, the complementary features of the high-frequency view and the low-frequency view are used to improve the model's ability to capture details and structures, the mutual uncertainty distance (MUD) loss is introduced in the internal learning process to reduce the uncertainty gap between labeled and unlabeled samples, effectively improving the reliability of the pseudo-label of the unlabeled medical image, so that the subnetwork can generate high-quality pseudo-labels, the adaptive self-correction (ASC) loss is set in the external learning process to introduce a feedback mechanism, dynamically adjust the subnetwork to improve the model's sensitivity to pseudo-label quality, effectively improve the training stability, solve the confirmation bias and unreliable pseudo-label confidence in the prior art, improve the stability of the segmentation model, and further improve the medical image segmentation accuracy. BRIEF DESCRIPTION OF DRAWINGS

[0046] The drawings accompanying the specification of the present application form part of the present application and serve to provide a further understanding of the present application, the exemplary embodiments of the present application and their descriptions serve to explain the present application and do not constitute an improper limitation on the present application.

[0047] Figure 1 is a comparison diagram of the prior art and the method proposed in embodiment 1 of the present application, wherein (a) is a segmentation diagram based on the self-training pseudo-labeling method of the prior art, (b) is a segmentation diagram based on the cooperative training architecture and pseudo-labeling combination method in the prior art, and (c) is a segmentation diagram of the method proposed in embodiment 1 of the present application;

[0048] Figure 2 is a schematic diagram of the frequency-based adaptive interactive learning framework of the semi-supervised medical segmentation model proposed in embodiment 1 of the present application;

[0049] Figure 3 is the enhancement frequency view diagram of the existing pseudo-label generation strategy and the pseudo-label generation strategy of the semi-supervised medical segmentation model proposed in Embodiment 1 of the present application in three groups of samples; wherein (a) is an original image, (b) is a low-frequency component diagram extracted by Gaussian filtering, (c) is an enhanced low-frequency image, (d) is a high-frequency component diagram extracted by Gaussian filtering, and (e) is an enhanced high-frequency image;

[0050] Figure 4 is the calculation and comparison result of the average confidence of each class in the labeled medical images and unlabeled medical images in the ACDC dataset mentioned in Embodiment 1 of the present application;

[0051] Figure 5 is the result diagram of five representative cases of the frequency-based adaptive interactive learning semi-supervised medical segmentation method disclosed in Embodiment 1 of the present application and the segmentation of the competitive method;

[0052] Figure 6 is the visualization result diagram of the frequency-based adaptive interactive learning semi-supervised medical segmentation method disclosed in Embodiment 1 of the present application compared with other SOTA methods. DETAILED DESCRIPTION

[0053] It should be pointed out that the following detailed description is exemplary and is intended to provide further explanation of the present application. Unless otherwise specified, all technical and scientific terms used in the present application have the same meaning as generally understood by those skilled in the art to which the present application belongs.

[0054] As introduced in the background, the purpose of the present application is to overcome the shortcomings of the prior art, and to provide a frequency-based adaptive interactive learning semi-supervised medical segmentation method and system, which aims to improve the reliability of pseudo-labels and self-correction ability, solve the limitations of pseudo-label-based methods and reduce confirmation errors by establishing an adaptive feedback loop, thereby improving the stability of the segmentation model and further improving the segmentation accuracy of medical images.

[0055] Embodiment 1

[0056] In a typical embodiment of the present application, the present embodiment discloses a frequency-based adaptive interactive learning semi-supervised medical segmentation method, comprising the following steps:

[0057] S1, obtaining an image to be segmented;

[0058] S2, inputting the image to be segmented into a pre-trained semi-supervised medical segmentation model for processing to obtain an image segmentation result.

[0059] Wherein, the semi-supervised medical segmentation model includes a high-frequency subnetwork and a low-frequency subnetwork, and the training process thereof includes an internal learning process and an external learning process, and the specific training steps are as follows:

[0060] A high-frequency view and a low-frequency view are generated through a frequency dual-view stream architecture.

[0061] In the internal learning process, complementary views are constructed by taking the high-frequency view and the low-frequency view as inputs, and the corresponding subnetwork states are updated through cross-pseudo-label supervision and mutual uncertainty distance (MUD) supervision, and are taken as the initial states of the external learning process;

[0062] In the external learning process, the corresponding subnetwork states are updated through cross-supervision loss and adaptive self-correction loss, and are taken as the initial states of the internal learning process of the next cycle.

[0063] The above frequency-based adaptive interactive learning semi-supervised medical segmentation method will be described in detail below in combination with the accompanying drawings and specific embodiments.

[0064] S1, an image to be segmented is acquired.

[0065] In this step, the image to be segmented can be a cardiac image, a prostate image, a laparoscopic kidney image, or other medical images, which are not limited herein. In this embodiment, the left ventricle, right ventricle, myocardium, and other cardiac images in the ACDC dataset can be used as training data, the prostate images in the Promise12 dataset can be used as training data, and the laparoscopic kidney images in the KiTS19 dataset can be used as training data, which are not limited herein.

[0066] S2, the image to be segmented is input to a pre-trained semi-supervised medical segmentation model for processing to obtain an image segmentation result.

[0067] The semi-supervised medical segmentation model in this embodiment is constructed based on a frequency-based adaptive interactive learning (FAM) framework. The FAM framework is a double-layer learning framework composed of an internal learning cycle and an external learning cycle, which iteratively refines pseudo-labels and enhances model learning. In this embodiment, the FAM is established based on a frequency dual-view stream (FDvS) cooperative training framework, which uses a special subnetwork design to extract high-frequency and low-frequency representations from input data. The frequency dual-view stream (FDvS) architecture utilizes the frequency characteristics of the network and input data to construct multiple complementary views, further improving the quality of pseudo-labels.

[0068] Specifically, the training process of the semi-supervised medical segmentation model based on the frequency-based adaptive interactive learning (FAM) framework is as follows:

[0069] S201. Determination of dataset

[0070] In a semi-supervised medical segmentation task, given a medical image dataset D , which consists of a labeled sample set N with samples and an unlabeled sample set M with samples, and , respectively represent the medical image and its corresponding segmentation label of the i th labeled sample, and represents the medical image of the i th unlabeled sample. The goal of semi-supervised medical segmentation is to leverage the unlabeled set to improve performance beyond what can be achieved using only the labeled set .

[0071] As shown in Figure 2 , the frequency-based adaptive mutual learning (FAM) framework consists of two main processes: an internal learning process and an external learning process, and both the internal learning process and the external learning process include a high-frequency subnetwork and a low-frequency subnetwork. In this embodiment, the high-frequency subnetwork is a CNN structure, and the low-frequency subnetwork is a Transformer structure.

[0072] S202. Frequency dual-view stream architecture

[0073] Co-training is a process in which multiple subnetworks learn from different but complementary views, aiming to leverage the comprehensive information of input data to improve performance. However, a key challenge of the co-training framework is to prevent the subnetworks from collapsing into similar representations. To solve this problem, this embodiment introduces a frequency-based co-training framework that integrates the frequency characteristics of images and networks while preventing collapse through input and subnetwork diversity.

[0074] For image diversity, this embodiment introduces enhanced frequency view processing to extract information from the frequency domain. High-frequency components and low-frequency components in the frequency spectrum capture different types of information, complementing each other to form a comprehensive representation. As shown in Figure 3 , as can be clearly seen from columns (b) and (d), high-frequency components are generally related to rapid intensity changes, emphasizing fine details and local information, which is crucial for depicting edges and boundaries. While low-frequency components encode smoother changes, they can capture the overall structure and global features of an image. Therefore, this embodiment constructs a pair of enhanced frequency views, the specific process being as follows:

[0075] First, based on a Gaussian filter, the high-frequency components and low-frequency components of the image are extracted:

[0076] , ;

[0077] wherein, is the extracted high-frequency component, is the extracted low-frequency component, G represents a Gaussian filter, denotes the original image, α is a Gaussian convolution kernel.

[0078] Then, based on the set enhancement factor, an enhanced high-frequency view and an enhanced low-frequency view are respectively constructed:

[0079] , ;

[0080] wherein, k is the enhancement factor, is the enhanced high-frequency view, is the enhanced low-frequency view.

[0081] For network diversity, the embodiment adopts a dual-stream architecture to exploit the frequency trends of different neural networks. Existing research has revealed the relationship between frequency characteristics and network operations: models based on the Transformer structure have a global attention mechanism, with low-pass filtering characteristics; on the contrary, models based on the CNN structure emphasize local information through their focused receptive fields, usually exhibiting high-pass filtering characteristics. Based on these characteristics, the embodiment extracts local features from the enhanced low-frequency view using a high-frequency subnetwork based on the CNN structure ( ), and captures long-distance dependencies in the enhanced high-frequency view using a low-frequency subnetwork based on the Transformer ( ). Due to the differences in structure and frequency trends, the frequency dual-view stream architecture naturally prevents the network from learning similar representations.

[0082] S203. Internal learning process

[0083] In the internal learning process, the parameters of the two subnetworks are and , and first, the high-frequency view and the low-frequency view of the labeled image and the unlabeled image are taken as input to build complementary views for the frequency dual-view stream architecture, and the prediction results of the unlabeled image obtained from one view are used as pseudo labels for the other view.

[0084] In the embodiment, the internal learning framework establishes cross-pseudo label supervision and mutual uncertainty distance (MUD) supervision between the frequency-specific subnetworks, and updates the state parameters of the corresponding subnetworks through cross-pseudo label supervision loss and mutual uncertainty distance (MUD) supervision loss.

[0085] where the specific process of cross-pseudo-label supervision is as follows:

[0086] Given unlabeled medical images, first generate high-frequency views and low-frequency views through high-pass and low-pass filtering operations, for each unlabeled image and low-frequency view , use random data augmentation strategies to create weak augmented variants and strong augmented variants .

[0087] Then the subnetwork produces logit output:

[0088] , ;

[0089] where and represent the initial parameters of t training iterations in the internal learning process. High-frequency one-hot pseudo labels and low-frequency one-hot pseudo labels are generated by performing argmax operations on the outputs and respectively.

[0090] The obtained one-hot pseudo labels are used to cross-pseudo-label supervise the subnetwork, and the loss function combines the cross-entropy loss and Dice loss of each subnetwork:

[0091] , ;

[0092] where represents the cross-pseudo-label supervision loss of the high-frequency subnetwork, represents the cross-pseudo-label supervision loss of the low-frequency subnetwork. For ease of subsequent understanding, the cross-pseudo-label supervision loss is denoted as .

[0093] Similarly, for labeled samples , similarly obtain and through weak augmentation.

[0094] Figure 4The calculation and comparison results of the average confidence of each class on the labeled medical images and the unlabeled medical images in the ACDC dataset are given. It can be observed that the average confidence of the background (BG) class in the unlabeled medical images is approximately equal to the average confidence of the background (BG) class in the labeled medical images. However, compared with the labeled medical images, the predicted average confidence of the left ventricle (LV), the right ventricle (RV) and the myocardium (MYO) in the unlabeled medical images is significantly lower. This phenomenon is attributed to the lack of explicit supervision signal in the unlabeled medical images, which leads to greater uncertainty of the model in segmenting anatomical structures. In order to alleviate this problem, the embodiment introduces mutual uncertainty distance (MUD) supervision in the internal learning process, which uses uncertainty information to optimize the model's prediction of unlabeled samples. This module encourages the model to generate more confident and reliable predictions, thereby bridging the confidence gap between labeled and unlabeled medical images.

[0095] Specifically, the mutual uncertainty distance (MUD) supervision consists of two parts, namely class entropy minimization (CEM) and entropy distance minimization (EDM), which are described in detail below taking the high-frequency view as an example.

[0096] Class entropy minimization (CEM) aims to minimize the difference between the class-specific prediction confidence in the output of labeled samples and unlabeled samples. Specifically, for a multi-class segmentation task, the entropy of the logit output of each class is first calculated to evaluate the respective uncertainty, denoted as:

[0097]

[0098] and represent the entropy values of the th class in the high-frequency domain of the unlabeled sample and the labeled sample, respectively, and are the corresponding weighted softmax outputs.

[0099] To enhance the robustness of the model to uncertain regions, the Euclidean distance between these entropy values is minimized, denoted as:

[0100]

[0101] denotes the L2 norm. Minimizing this loss encourages the model to align the confidence distribution between the unlabeled samples and the labeled samples for each class, which helps to reduce the impact of uncertainty in ambiguous regions, thereby achieving more consistent and robust predictions.

[0102] ​​​​​Entropy distance minimization (EDM) aims to calculate the distance between the confidence entropy of each class in the labeled samples and the confidence entropy of each class in the unlabeled samples, and minimize the difference between these respective distances to make the uncertainty distribution consistent between the labeled samples and the unlabeled samples. Specifically, first define the uncertainty distance as follows:

[0103] , ;

[0104] wherein, and measure the pairwise Euclidean distances between the entropy values of the classes and in the labeled sample set and the unlabeled sample set, respectively, which capture the structural relationship between the classes according to the uncertainty pattern.

[0105] Then, the entropy distance minimization loss is formulated to enforce the consistency between the intra-class entropy structure of the labeled samples and the unlabeled samples, represented as:

[0106] ;

[0107] wherein, C is the total number of classes. By minimizing the entropy distance minimization loss , the relative uncertainty distribution between the labeled set and the unlabeled set can be adjusted to determine the generalization of the model's confidence pattern consistent on the two datasets.

[0108] Further, the mutual uncertainty distance (MUD) supervision loss is defined as a weighted combination of the class entropy minimization (CEM) loss and the entropy distance minimization (EDM) loss, represented as:

[0109] ;

[0110] wherein, and are predefined weights. Then, the mutual uncertainty distance (MUD) supervision loss is incorporated into the inner loop loss of the high-frequency subnetwork and the low-frequency subnetwork.

[0111] Specifically, the inner loop loss of the high-frequency subnetwork is the sum of the cross-pseudo label supervision loss of the high-frequency subnetwork and the mutual uncertainty distance (MUD) supervision loss, and the inner loop loss of the low-frequency subnetwork is the sum of the cross-pseudo label supervision loss of the low-frequency subnetwork and the mutual uncertainty distance (MUD) supervision loss, specifically represented as:

[0112] , .

[0113] Based on the inner loop loss, the state parameters of the high-frequency subnetwork and the low-frequency subnetwork are updated by gradient descent, which is specifically represented as:

[0114] , ;

[0115] In the formula, and are the updated high-frequency subnetwork state parameters and low-frequency subnetwork state parameters, respectively, which will be used as the initial state parameters of the high-frequency subnetwork and the low-frequency subnetwork in the external learning process, represents the learning rate, and are the inner loop losses of the high-frequency subnetwork ( ) and the low-frequency subnetwork ( ), respectively.

[0116] The embodiment reduces the uncertainty gap between labeled samples and unlabeled samples by introducing mutual uncertainty distance (MUD) supervision in the internal learning process, improves the reliability of pseudo labels and reduces the accumulation of errors, thereby ensuring the generation of high-quality pseudo labels.

[0117] S204. External learning process

[0118] On the basis of improving the internal learning process, the embodiment proposes a novel external learning mechanism, which enables the subnetwork to adaptively correct the residual confirmation bias through performance-driven self-optimization, and enables each subnetwork to dynamically adjust its pseudo label utilization rate according to the cross-view supervision effectiveness by setting a double feedback system in the external learning process.

[0119] First, let , , and represent the logit output from the internal learning process, and the updated output logit is calculated as follows:

[0120] , .

[0121] In order to measure the reliability of the cross-view pseudo label, the embodiment introduces a performance feedback coefficient to quantify the optimization improvement of the labeled sample and dynamically adjust the pseudo label confidence, which is specifically represented as:

[0122] ,

[0123] ;

[0124] In the formula, reflects the pseudo label The quality of guidance, Reflect pseudo labels The performance feedback coefficient is a positive value, indicating useful pseudo-label knowledge, and its magnitude indicates confidence. A negative performance feedback coefficient indicates misleading supervision, and its absolute value quantifies the severity of the error.

[0125] Then, the sub-network calculates the adaptive self-correction (ASC) loss based on the performance feedback coefficient and adjusts its learning strategy through the adaptive self-correction (ASC) loss. The adaptive self-correction (ASC) loss of the sub-network is specifically expressed as:

[0126] , .

[0127] The above formula automatically amplifies credible pseudo-labels while suppressing unreliable pseudo-labels.

[0128] For labeled samples, cross-supervision loss is also combined to utilize certain supervisory signals from labeled samples, specifically expressed as:

[0129] , .

[0130] In this embodiment, the outer loop loss of the high-frequency sub-network is the sum of the cross-supervision loss and the adaptive self-correction loss of the high-frequency sub-network, and the inner loop loss of the low-frequency sub-network is the sum of the cross-supervision loss and the mutual uncertainty distance (MUD) supervision loss of the low-frequency sub-network, which can be specifically expressed as:

[0131] , ;

[0132] Based on the outer loop loss, the state parameters of the high-frequency sub-network and the low-frequency sub-network are updated accordingly by gradient descent, which is expressed as:

[0133] , ;

[0134] In the formula, the updated high-frequency sub-network state parameters are and low-frequency sub-network state parameters They are used as the initial state parameters of the high-frequency sub-network and the low-frequency sub-network in the internal learning process of the next cycle, respectively. In this way, the training process of the two-layer optimization framework is completed and the training of the next cycle is entered. When the number of training times reaches the set epoch value, the training is stopped and the image to be segmented obtained in S1 is input into the trained semi-supervised medical segmentation model for processing to obtain the image segmentation result.

[0135] The embodiment introduces a feedback mechanism to dynamically adjust the sub-networks by introducing an adaptive self-correction (ASC) loss in the internal learning process, so that the model can continuously adapt through self-correction according to the performance feedback of the labeled samples.

[0136] In order to verify the effectiveness of the above-mentioned frequency-based adaptive interactive learning semi-supervised medical segmentation method, the embodiment performs comparative experiments on two popular data sets, the ACDC data set and the PROMISE12 data set. Among them, the ACDC data set contains 100 short-axis cine MRI images and has expert annotations of three categories: left ventricle (LV), right ventricle (RV), and myocardium (MYO). The data set is divided into 70 images for training, 10 images for verification, and 20 images for testing. The PROMISE12 data set includes 50 prostate MRI scans collected from multiple medical centers, with different scanning protocols. The data set is divided into 35 images for training and 15 images for testing.

[0137] In order to quantitatively evaluate the performance of the proposed method, the embodiment uses the following indicators to evaluate the overlap and boundary accuracy: (1) Jaccard index, which calculates the union and intersection (IoU) of the prediction and ground truth, providing a measure of overlap between the two groups; (2) Dice Similarity Coefficient (DSC), which quantifies the overlap between the prediction and the ground truth, but is particularly sensitive to small targets; (3) 95th percentile Hausdorff distance (95HD), which measures the maximum distance between the predicted boundary and the ground truth; (4) Average Surface Distance (ASD), which calculates the average distance between corresponding points on the predicted surface and the ground truth.

[0138] The frequency-based adaptive interactive learning semi-supervised medical segmentation method proposed in the embodiment is implemented using PyTorch and runs on an NVIDIA Tesla V100 GPU with 32GB of memory. The HF network is a convolution-based UNet model, while the LF network is a transformer-based Swin-UNet model. The MRI scans are preprocessed into 2D slices, resized to 256x256 for the ACDC data set and 224x224 for the Promise data set, consistent with related work. The batch size is set to 16, consisting of 8 labeled samples and 8 unlabeled samples. During internal learning and external learning, stochastic gradient descent (SGD) is used as the optimizer for the HF and LF models. The initial learning rate is , the momentum is 0.9, and the weight decay is In addition, in ASC, pseudo-labels are generated from weakly augmented data to supervise the prediction of strongly augmented data. Weak augmentation includes random rotation, flipping, and color jitter; strong augmentation further combines a grid masking strategy, in which certain parts of the image are masked with a structured grid pattern.

[0139] To evaluate the effectiveness of the proposed method, it is compared with two supervised methods: U-Net and Swin-unet used as backbone models in the architecture, and eight semi-supervised methods: DTC, URPC, MC-Net, SS-net, SCP-Net, PLGCL, Cross-Teaching, BCP, ABD on the ACDC dataset. For a fair comparison, 5% and 10% of medical images in the entire ACDC dataset are selected as labeled samples for the experiment. Table 1 shows the quantitative evaluation as follows:

[0140] Table 1 Quantitative evaluation on the ACDC dataset

[0141]

[0142] As can be seen from Table 1, the proposed method produces a DSC of 90.83%, a Jaccard index of 83.57%, a 95HD of 1.12% and an ASD of 0.27% under 10% labeled samples, which improves the DSC by 11.42% on U-Net and 46.51% on Swin-UNet compared with the supervised baseline trained with 10% labeled samples, which shows that the proposed method achieves superior segmentation performance compared with fully supervised models under the same ratio of labeled samples. Further comparison with semi-supervised methods shows that the proposed method improves the DSC and Jaccard by 1.02% and 1.62% respectively under 10% labeled samples, which proves its superiority in segmentation quality. In addition, the proposed method improves the 95HD by 0.34 and the ASD by 0.22, which shows more accurate boundary delineation. Furthermore, under only 5% labeled samples, the proposed method always outperforms other methods in all indicators. Figure 5 Five representative cases of segmentation by the proposed method and competing methods are qualitatively compared, including Swin-UNet, MC-Net, SS-Net, Cross-Teaching, BCP and ABD, and the cases are arranged in order of segmentation difficulty from (a) to (e). From Figure 5 As can be seen, the proposed method has more superior segmentation performance in handling large and small structures and multi-class and single-class segmentation tasks in cardiac imaging, and can always achieve more accurate boundary delineation and better preservation of segmented structures compared with other methods.

[0143] The embodiment also extends the method to the PROMISE12 dataset to evaluate its generalization ability. Following the experimental protocol adopted by the state-of-the-art methods (CCT, URPC, SS-Net, SLC-Net, SCP-Net, ABD) on the PROMISE12 dataset, 20% of the images are used as labeled samples, and the remaining 80% are regarded as unlabeled samples, and the performance is evaluated using DSC and ASD indicators. As shown in Table 2, the method achieves a DSC of 82.14% and an ASD of 1.30%. Compared with the fully supervised method, the method achieves an improvement of 15.13% in DSC and a reduction of 12.57% in ASD. In addition, the FAM of the method outperforms other state-of-the-art semi-supervised methods in all evaluation indicators, establishing a new state-of-the-art level for the prostate segmentation task.

[0144] Table 2 Comparison of generalization ability on PROMISE12 dataset

[0145]

[0146] Figure 6 Visualization results compared with other SOTA methods (Swin-UNet, BCP, ABD) are provided, which include four examples, where the segmentation difficulty gradually increases from (a) to (d) as the target size decreases. From Figure 6 It can be seen that in the relatively easy cases (a) and (b), most methods achieve high segmentation accuracy, and the main difference appears in the boundary accuracy. However, as the difficulty increases, the accuracy of other methods gradually decreases, especially in (d), they show obvious errors in segmenting small targets. In contrast, even in challenging cases, the method can always maintain high segmentation quality and well preserve the boundaries.

[0147] Embodiment 2

[0148] In a typical embodiment of the present application, the embodiment discloses a frequency-based adaptive interactive learning semi-supervised medical segmentation system, comprising:

[0149] The acquisition module is configured to acquire an image to be segmented;

[0150] The image segmentation module is configured to input the image to be segmented into a pre-trained semi-supervised medical segmentation model for processing to obtain an image segmentation result;

[0151] The semi-supervised medical segmentation model includes a high-frequency subnetwork and a low-frequency subnetwork, and its training process includes an internal learning process and an external learning process, and the specific training steps are as follows:

[0152] generate a high-frequency view and a low-frequency view through a frequency dual-view stream architecture;

[0153] In the internal learning process, the complementary views are constructed with the high-frequency view and the low-frequency view as inputs, and the corresponding subnetwork states are updated through cross-pseudo-label supervision and mutual uncertainty distance (MUD) supervision, and are used as the initial states of the external learning process;

[0154] In the external learning process, the corresponding subnetwork states are updated through cross-supervision loss and adaptive self-correction loss, and are used as the initial states of the internal learning process in the next cycle.

[0155] The above is only the preferred embodiment of the present application, and is not used to limit the present application. For those skilled in the art, the present application can have various modifications and changes. Any modification, equivalent replacement, improvement, etc. within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims

1. A frequency-based adaptive interactive learning semi-supervised medical segmentation method, characterized in that, The method comprises the following steps: obtaining an image to be segmented; inputting the image to be segmented into a pre-trained semi-supervised medical segmentation model for processing to obtain an image segmentation result; wherein the semi-supervised medical segmentation model comprises a high-frequency subnetwork and a low-frequency subnetwork, and the training process of the semi-supervised medical segmentation model comprises an internal learning process and an external learning process, and the specific training steps are as follows: generating a high-frequency view and a low-frequency view through a frequency dual-view stream architecture; in the internal learning process, a complementary view is constructed by taking the high-frequency view and the low-frequency view as inputs, and the corresponding subnetwork state is updated through cross-pseudo-label supervision and mutual uncertainty distance supervision, and the initial state of the external learning process is taken as the initial state of the internal learning process of the next cycle; in the external learning process, the corresponding subnetwork state is updated through cross-supervision loss and adaptive self-correction loss, and the initial state of the internal learning process of the next cycle is taken as the initial state of the high-frequency subnetwork and the low-frequency subnetwork of the external learning process; specifically, the mutual uncertainty distance supervision optimizes the prediction of unlabeled samples by using uncertainty information, including class entropy minimization and entropy distance minimization; the loss function of the mutual uncertainty distance supervision is a weighted combination of the class entropy minimization loss and the entropy distance minimization loss; the class entropy minimization loss is represented as: ； wherein represents the entropy value of the class of the unmarked sample; represents the L2 norm;​ the entropy distance minimization loss is represented as: ； Where, and Represents the classes in the labeled sample set and the unlabeled sample set respectively and The pairwise Euclidean distance between the entropy values ​​of ; C is the total number of categories; the adaptive self-correction loss is specifically represented as: ， ； In the formula, represents the high-frequency subnetwork adaptive self-correction loss; reflects the guidance quality of the pseudo label ; represents the low-frequency subnetwork adaptive self-correction loss; reflects the guidance quality of the pseudo label ; and represent the logit output generated by the subnetwork; represents the high-frequency one-hot pseudo label; represents the low-frequency one-hot pseudo label; and are the updated high-frequency subnetwork state parameters and low-frequency subnetwork state parameters, respectively.

2. The frequency-based adaptive interactive learning semi-supervised medical segmentation method of claim 1, wherein, the high-frequency view and the low-frequency view are generated through the frequency dual-view stream architecture, and the specific process comprises: extracting high-frequency components and low-frequency components of labeled images and unlabeled images based on Gaussian filters respectively; constructing enhanced high-frequency views and enhanced low-frequency views based on a set of enhancement factors respectively; extracting local features from the enhanced low-frequency views by using a high-frequency subnetwork based on CNN, and capturing long-distance dependency in the enhanced high-frequency views by using a low-frequency subnetwork based on Transformer.

3. The frequency-based adaptive interactive learning semi-supervised medical segmentation method of claim 1, wherein, the specific process of the cross-pseudo-label supervision comprises: creating weak enhancement variants and strong enhancement variants of the high-frequency view and the low-frequency view respectively using a random data enhancement strategy; generating high-frequency one-hot pseudo-labels and low-frequency one-hot pseudo-labels based on the logit outputs of the strong enhancement variants respectively; using the obtained pseudo-labels to perform cross-pseudo-label supervision on the subnetworks, and the cross-pseudo-label supervision loss is the sum of the cross-entropy loss and the Dice loss of each subnetwork.

4. The frequency-based adaptive interactive learning semi-supervised medical segmentation method of claim 1, wherein, in the internal learning process, the corresponding subnetwork state is updated as follows: calculating the inner loop loss of the high-frequency subnetwork and the low-frequency subnetwork respectively; updating the state parameters of the high-frequency subnetwork and the low-frequency subnetwork by gradient descent based on the inner loop loss; the updated state parameters of the high-frequency subnetwork and the low-frequency subnetwork are taken as the initial state parameters of the high-frequency subnetwork and the low-frequency subnetwork of the external learning process respectively.

5. The frequency-based adaptive interactive learning semi-supervised medical segmentation method of claim 4, wherein, the inner loop loss of the high-frequency subnetwork is the sum of the cross-pseudo-label supervision loss and the mutual uncertainty distance supervision loss of the high-frequency subnetwork; the inner loop loss of the low-frequency subnetwork is the sum of the cross-pseudo-label supervision loss and the mutual uncertainty distance supervision loss of the low-frequency subnetwork.

6. The frequency-based adaptive interactive learning semi-supervised medical segmentation method of claim 1, wherein, in the external learning process, the performance feedback coefficient is used to dynamically adjust the pseudo-label confidence, the adaptive self-correction loss of the corresponding subnetwork is calculated based on the performance feedback coefficient, and the learning strategy of the subnetwork is adjusted through the adaptive self-correction loss.

7. The frequency-based adaptive interactive learning semi-supervised medical segmentation method of claim 1, wherein, In the external learning process, updating the state of the corresponding subnetwork is specifically: The outer loop loss of the high-frequency subnetwork and the low-frequency subnetwork is calculated respectively. Based on the outer loop loss, the state parameters of the high-frequency subnetwork and the low-frequency subnetwork are updated through gradient descent; the updated high-frequency subnetwork state parameters and low-frequency subnetwork state parameters are used as the initial state parameters of the high-frequency subnetwork and the low-frequency subnetwork in the next cycle of the internal learning process.

8. The frequency-based adaptive interactive learning semi-supervised medical segmentation method of claim 7, wherein, The outer loop loss of the high-frequency subnetwork is the sum of the cross-supervision loss and the adaptive self-correction loss of the high-frequency subnetwork; the inner loop loss of the low-frequency subnetwork is the sum of the cross-supervision loss and the mutual uncertainty distance supervision loss of the low-frequency subnetwork.

9. A frequency-based adaptive interactive learning semi-supervised medical segmentation system, characterized in that, It comprises: An acquisition module configured to acquire an image to be segmented; An image segmentation module configured to input the image to be segmented into a pre-trained semi-supervised medical segmentation model for processing to obtain an image segmentation result; The semi-supervised medical segmentation model comprises a high-frequency subnetwork and a low-frequency subnetwork, and its training process comprises an internal learning process and an external learning process, and the specific training steps are as follows: Generate a high-frequency view and a low-frequency view through a frequency dual-view flow architecture; In the internal learning process, complementary views are constructed by taking the high-frequency view and the low-frequency view as input, and the state of the corresponding subnetwork is updated through cross-pseudo label supervision and mutual uncertainty distance supervision, and is used as the initial state of the external learning process; In the external learning process, the state of the corresponding subnetwork is updated through cross-supervision loss and adaptive self-correction loss, and is used as the initial state of the internal learning process in the next cycle; Specifically, the mutual uncertainty distance supervision optimizes the prediction of unlabeled samples by using uncertainty information optimization model, including class entropy minimization and entropy distance minimization; the loss function of the mutual uncertainty distance supervision is a weighted combination of the class entropy minimization loss and the entropy distance minimization loss; The class entropy minimization loss is represented as: ； wherein represents the entropy value of the class of the unmarked sample; represents the L2 norm;​ The entropy distance minimization loss is represented as: ； Where, and Represents the classes in the labeled sample set and the unlabeled sample set respectively and The pairwise Euclidean distance between the entropy values ​​of ; C is the total number of categories; The adaptive self-correction loss is specifically represented as: ， ； In the formula, represents the high-frequency subnetwork adaptive self-correction loss; reflects the guidance quality of the pseudo label ; represents the low-frequency subnetwork adaptive self-correction loss; reflects the guidance quality of the pseudo label ; and represent the logit output generated by the subnetwork; represents the high-frequency one-hot pseudo label; represents the low-frequency one-hot pseudo label; and are the updated high-frequency subnetwork state parameters and low-frequency subnetwork state parameters, respectively.

Images