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

Abstract

The invention discloses a frequency-based self-adaptive interactive learning semi-supervised medical segmentation method and system, and belongs to the technical field of medical image segmentation, and the method comprises the steps: obtaining a to-be-segmented image, inputting the to-be-segmented image into a pre-trained semi-supervised medical segmentation model for processing, and obtaining an image segmentation result; the semi-supervised medical segmentation model comprises a high-frequency sub-network and a low-frequency sub-network, 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 double-view stream architecture; in the internal learning process, a high-frequency view and a low-frequency view are used as input to construct a complementary view, and a corresponding sub-network state is updated through cross pseudo tag supervision and mutual uncertainty distance supervision and is used as an initial state of the external learning process; in the external learning process, the corresponding sub-network state is updated through the cross supervision loss and the self-adaptive self-correction loss, and the state is used as the initial state of the internal learning process of the next cycle.

Description

Technical Field

[0001] The present invention belongs to the technical field of medical image segmentation, and in particular relates to a frequency-based adaptive interactive learning semi-supervised medical segmentation method and system. Background Art

[0002] The statements herein merely provide background information related to the present invention and do not necessarily constitute prior art.

[0003] Medical image segmentation plays a crucial role in modern computer-aided diagnosis systems, facilitating the precise 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, fully annotated datasets, but obtaining such pixel-level labels for medical images is 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. Its core idea is to effectively leverage the information in the unlabeled data. Pseudo-labeling provides a straightforward and effective strategy to achieve this goal.

[0004] In the existing technology, many studies have adopted self-training pseudo-labeling methods to generate pseudo labels from weakly perturbed inputs based on a single model and use them to supervise the prediction of strongly perturbed inputs, thereby maintaining consistency between different transformations of medical images, such as Figure 1 (a) shown.

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

[0006] However, while pseudo-labeling can enhance the ability to learn from unlabeled data, it also poses a significant challenge: when the generated pseudo-labels are inaccurate, the model may reinforce these incorrect predictions during training, which may lead to error accumulation and reduce model stability. This is called confirmation bias, which remains an ongoing problem for pseudo-label-based methods. To address this issue, existing methods have introduced confidence-based strategies to improve pseudo-label quality. By iteratively retaining high-confidence pseudo-labels and filtering out low-confidence pseudo-labels, these methods aim to improve the reliability of pseudo-labels in consecutive training steps. However, since incorrect pseudo-labels may also exhibit high confidence, this assumption may not always hold. Therefore, if there is no feedback mechanism to improve or correct these errors, the effectiveness of this approach remains limited.

[0007] In summary, the inventors have found that existing semi-supervised medical image segmentation methods have at least the following deficiencies: (1) Pseudo-label-based methods may suffer from confirmation bias. Wrong or inaccurate pseudo-labels are reinforced during training, which leads to error accumulation and reduced model stability, thus causing model performance degradation. (2) Confidence-based strategies may have the problem of unreliable pseudo-label confidence, which makes it not always feasible to improve the quality of pseudo-labels by screening high-confidence pseudo-labels, resulting in limited effectiveness of such methods. Summary of the Invention

[0008] The purpose of the present invention is to overcome the deficiencies in the above-mentioned prior art and to provide a frequency-based adaptive interactive learning semi-supervised medical segmentation method and system, aiming to improve the reliability and self-correction ability of pseudo-labels, 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.

[0009] In order to achieve the above object, the present invention is implemented through the following technical solutions: On the one hand, the technical solution of the present invention provides a frequency-based adaptive interactive learning semi-supervised medical segmentation method, comprising: Obtain the image to be segmented; Input the image to be segmented into a pre-trained semi-supervised medical segmentation model for processing to obtain the image segmentation result; The semi-supervised medical segmentation model includes a high-frequency sub-network and a low-frequency sub-network. The training process includes an internal learning process and an external learning process. The specific training steps are as follows: Generate high-frequency views and low-frequency views through a frequency dual-view stream architecture; In the internal learning process, high-frequency views and low-frequency views are used as input to construct complementary views, and the corresponding sub-network states are updated through cross-pseudo-label supervision and mutual uncertainty distance supervision, which serve as the initial state of the external learning process; During the external learning process, the corresponding sub-network state is updated through cross-supervision loss and adaptive self-correction loss, and is used as the initial state of the internal learning process of the next cycle.

[0010] In at least one embodiment, a high-frequency view and a low-frequency view are generated using a frequency dual-view stream architecture, and the specific process includes: Based on Gaussian filter, high-frequency components and low-frequency components of labeled images and unlabeled images are extracted respectively; An enhanced high-frequency view and an enhanced low-frequency view are constructed based on a set enhancement factor; A CNN-based high-frequency sub-network is used to extract local features from the enhanced low-frequency views, and a Transformer-based low-frequency sub-network is used to capture long-range dependencies in the enhanced high-frequency views.

[0011] In at least one embodiment, the cross-pseudo-label supervision process includes: Use random data augmentation strategies to create weakly augmented and strongly augmented variants of high-frequency and low-frequency views, respectively. Generate high-frequency one-hot pseudo labels and low-frequency one-hot pseudo labels based on the logit output of the strong enhancement variant; The obtained pseudo labels are used to perform cross-pseudo-label supervision on the sub-network, and the cross-pseudo-label supervision loss is the sum of the cross entropy loss and the Dice loss of each sub-network.

[0012] In at least one embodiment, the mutual uncertainty distance supervision utilizes uncertainty information to optimize the model's prediction of unlabeled samples, including category entropy minimization and entropy distance minimization; the loss function of the mutual uncertainty distance supervision is a weighted combination of category entropy minimization loss and entropy distance minimization loss.

[0013] In at least one embodiment, during the internal learning process, updating the corresponding sub-network state is specifically as follows: Calculate the inner loop loss of the high-frequency sub-network and the low-frequency sub-network respectively; Based on the inner loop loss, the state parameters of the high-frequency sub-network and the low-frequency sub-network are updated accordingly by gradient descent; the updated state parameters of the high-frequency sub-network and the low-frequency sub-network are used as the initial state parameters of the high-frequency sub-network and the low-frequency sub-network in the external learning process respectively.

[0014] In at least one embodiment, the inner loop loss of the high-frequency sub-network is the sum of the cross-pseudo-label supervision loss and the mutual uncertainty distance supervision loss of the high-frequency sub-network; the inner loop loss of the low-frequency sub-network is the sum of the cross-pseudo-label supervision loss and the mutual uncertainty distance supervision loss of the low-frequency sub-network.

[0015] In at least one embodiment, during the external learning process, the pseudo-label confidence is dynamically adjusted by the performance feedback coefficient, and the adaptive self-correction loss of the corresponding sub-network is calculated based on the performance feedback coefficient, and the sub-network adjusts its learning strategy through the adaptive self-correction loss.

[0016] In at least one embodiment, during the external learning process, updating the corresponding sub-network state is specifically as follows: Calculate the outer loop loss of the high-frequency sub-network and the low-frequency sub-network respectively; 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; the updated state parameters of the high-frequency sub-network and the low-frequency sub-network 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.

[0017] In at least one 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; the inner loop loss of the low-frequency sub-network is the sum of the cross-supervision loss and the mutual uncertainty distance supervision loss of the low-frequency sub-network.

[0018] On the other hand, the technical solution of the present invention also provides a frequency-based adaptive interactive learning semi-supervised medical segmentation system, comprising: The acquisition module is configured to: acquire the image to be segmented; 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 the image segmentation result; The semi-supervised medical segmentation model includes a high-frequency sub-network and a low-frequency sub-network. The training process includes an internal learning process and an external learning process. The specific training steps are as follows: Generate high-frequency views and low-frequency views through a frequency dual-view stream architecture; In the internal learning process, high-frequency views and low-frequency views are used as input to construct complementary views, and the corresponding sub-network states are updated through cross-pseudo-label supervision and mutual uncertainty distance supervision, which serve as the initial state of the external learning process; During the external learning process, the corresponding sub-network state is updated through cross-supervision loss and adaptive self-correction loss, and is used as the initial state of the internal learning process of the next cycle.

[0019] The beneficial effects of the technical solution of the present invention are as follows: The present invention constructs a frequency dual-view stream architecture by integrating the frequency characteristics of the image and the frequency trend of the network, and improves the model's ability to capture details and structures through the complementary characteristics of high-frequency and low-frequency views; by introducing the mutual uncertainty distance (MUD) loss in the internal learning process, the uncertainty gap between labeled and unlabeled samples is reduced, and the reliability of the pseudo-labels of unlabeled medical images is effectively improved, so that the sub-network can generate high-quality pseudo-labels; by setting the adaptive self-correction (ASC) loss in the external learning process to introduce a feedback mechanism, the sub-network is dynamically adjusted to make the model more sensitive to the quality of pseudo-labels, effectively improve the training stability, solve the problems of possible confirmation bias and unreliable pseudo-label confidence in the existing technology, improve the stability of the segmentation model, and thus improve the accuracy of medical image segmentation. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] The accompanying drawings, which constitute a part of the present invention, are used to provide a further understanding of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute improper limitations on the present invention.

[0021] Figure 1 Schematic diagrams comparing the prior art and the method proposed in Example 1 of the present invention, wherein (a) is a schematic diagram of segmentation based on the prior art self-training pseudo-labeling method, (b) is a schematic diagram of segmentation based on the prior art method combining collaborative training architecture and pseudo-labeling, and (c) is a schematic diagram of segmentation based on the method proposed in Example 1 of the present invention; Figure 2 is a schematic diagram of a frequency-based adaptive interactive learning framework for a semi-supervised medical segmentation model proposed in Example 1 of the present invention; Figure 3 Schematic diagrams of enhanced frequency views of three groups of samples using an existing pseudo-label generation strategy and a pseudo-label generation strategy for a semi-supervised medical segmentation model proposed in Example 1 of the present invention; wherein (a) is the original image, (b) is a schematic diagram of the low-frequency component extracted by Gaussian filtering, (c) is the enhanced low-frequency image, (d) is a schematic diagram of the high-frequency component extracted by Gaussian filtering, and (e) is the enhanced high-frequency image; Figure 4 The calculation and comparison results of the average confidence of each category in the labeled medical images and the unlabeled medical images in the ACDC dataset mentioned in Example 1 of the present invention are as follows; Figure 5 Schematic diagram of five representative case results of the frequency-based adaptive interactive learning semi-supervised medical segmentation method and the competitive method segmentation disclosed in Example 1 of the present invention; Figure 6 This is a schematic diagram of the visualization results of the frequency-based adaptive interactive learning semi-supervised medical segmentation method disclosed in Example 1 of the present invention compared with other SOTA methods. DETAILED DESCRIPTION

[0022] It should be noted that the following detailed description is illustrative and is intended to provide further explanation of the present invention. Unless otherwise specified, all technical and scientific terms used in the present invention have the same meaning as commonly understood by those skilled in the art to which the present invention belongs.

[0023] As introduced in the background technology, the purpose of the present invention is to overcome the shortcomings of the above-mentioned prior art and provide a frequency-based adaptive interactive learning semi-supervised medical segmentation method and system, aiming to improve the reliability and self-correction ability of pseudo-labels, 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.

[0024] Example 1 In a typical embodiment of the present invention, this embodiment discloses a frequency-based adaptive interactive learning semi-supervised medical segmentation method, comprising the following steps: S1, obtain the image to be segmented; S2. Input the image to be segmented into a pre-trained semi-supervised medical segmentation model for processing to obtain the image segmentation result.

[0025] Among them, the semi-supervised medical segmentation model includes a high-frequency sub-network and a low-frequency sub-network. Its training process includes an internal learning process and an external learning process. The specific training steps are as follows: Generate high-frequency views and low-frequency views through a frequency dual-view stream architecture; In the internal learning process, high-frequency and low-frequency views are used as input to construct complementary views, and the corresponding sub-network states are updated through cross-pseudo-label supervision and mutual uncertainty distance (MUD) supervision, which serves as the initial state of the external learning process; During the external learning process, the corresponding sub-network state is updated through cross-supervision loss and adaptive self-correction loss, and is used as the initial state of the internal learning process of the next cycle.

[0026] The above-mentioned frequency-based adaptive interactive learning semi-supervised medical segmentation method will be described in detail below with reference to the accompanying drawings and specific implementation methods.

[0027] S1. Obtain the image to be segmented.

[0028] In this step, the image to be segmented can be a cardiac image, or a medical image such as a prostate image or a laparoscopic kidney image, which is not limited here. In this embodiment, cardiac images such as the left ventricle, right ventricle, and myocardium in the ACDC dataset can be used as training data, or prostate images in the Promise12 dataset can be used as training data, or laparoscopic kidney images in the KiTS19 dataset can be used as training data, which is not limited here.

[0029] S2. Input the image to be segmented into the pre-trained semi-supervised medical segmentation model for processing to obtain the image segmentation result.

[0030] The semi-supervised medical segmentation model in this embodiment is constructed using a frequency-based adaptive interaction learning (FAM) framework. This FAM architecture is a two-layer learning framework, consisting of an internal learning loop and an external learning loop, which refines the pseudo labels and enhances model learning in an iterative manner. In this embodiment, the FAM is established based on the frequency dual-view stream (FDvS) co-training framework, which uses a dedicated sub-network design to extract high-frequency and low-frequency representations from the input data. The frequency dual-view stream (FDvS) architecture utilizes the frequency characteristics of the network and the input data to construct multiple complementary views, further improving the quality of the pseudo labels.

[0031] Specifically, the training process of the semi-supervised medical segmentation model constructed based on the frequency-based adaptive interaction learning (FAM) framework is as follows: S201. Determination of the dataset In the semi-supervised medical segmentation task, a medical image dataset is given D , which is composed of a labeled sample set with N samples and an unlabeled sample set with M samples , and N << M. Among them, and respectively represent the medical image of the i th labeled sample and its corresponding segmentation label, and represents the medical image of the i th unlabeled sample. The goal of semi-supervised medical segmentation is to utilize the unlabeled set to improve performance beyond what can be achieved by only using the labeled set .

[0032] For example, as Figure 2As shown, the frequency-based adaptive interactive 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 sub-network and a low-frequency sub-network. In this embodiment, the high-frequency sub-network is a CNN structure and the low-frequency sub-network is a Transformer structure.

[0033] S202. Frequency Dual-View Stream Architecture Co-training enables multiple sub-networks to learn from different but complementary perspectives, aiming to leverage the comprehensive information of the input data to improve performance. However, a key challenge in co-training frameworks is preventing the sub-networks from collapsing into similar representations. To address this issue, this paper introduces a frequency-based co-training framework that integrates the frequency characteristics of images and networks while preventing collapse due to input and sub-network diversity.

[0034] For image diversity, this embodiment introduces enhanced frequency view processing to extract information from the frequency domain. The high-frequency components and low-frequency components in the spectrum capture different types of information and complement each other to form a comprehensive representation. Figure 3 As shown in columns (b) and (d), it is clear that high-frequency components are typically associated with rapid intensity changes, emphasizing fine details and local information, which is crucial for depicting edges and boundaries. Low-frequency components, on the other hand, encode smoother changes and are able to capture the overall structure and global features of the image. Therefore, this embodiment constructs a pair of enhanced frequency views. The specific process is as follows: First, the high-frequency and low-frequency components of the image are extracted based on the Gaussian filter: , ; Where, is the extracted high frequency component, is the extracted low-frequency component, G represents Gaussian filtering, Indicates the original image. α is the Gaussian convolution kernel.

[0035] Then, the enhanced high-frequency view and enhanced low-frequency view are constructed based on the set enhancement factor: , ; Where, k is the enhancement factor, For an enhanced high-frequency view, for an enhanced low-frequency view.

[0036] In view of network diversity, this 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: Transformer-based models have a global attention mechanism and exhibit low-pass filtering characteristics; in contrast, CNN-based models emphasize local information through their focused receptive fields and generally exhibit high-pass filtering characteristics. Based on these characteristics, this embodiment utilizes a high-frequency sub-network based on the CNN structure ( ) extracts local features from the enhanced low-frequency view and utilizes the low-frequency sub-network based on Transformer ( ) captures long-range dependencies in enhanced high-frequency views. Due to the differences in structure and frequency trends, the frequency dual-view stream architecture naturally prevents the network from learning similar representations.

[0037] S203. Internal learning process During the internal learning process, the parameters of the two sub-networks are and , first mark the image and unlabeled images The high-frequency view and low-frequency view of are taken as input to build complementary views for the frequency dual-view flow architecture, and the prediction results of the unlabeled image obtained from one view are used as the pseudo-label of the other view.

[0038] In this embodiment, the internal learning framework establishes cross-pseudo-label supervision and mutual uncertainty distance (MUD) supervision between frequency-specific sub-networks, and updates the state parameters of the corresponding sub-networks through cross-pseudo-label supervision loss and mutual uncertainty distance (MUD) supervision loss.

[0039] Among them, the specific process of cross-pseudo-label supervision is as follows: Given an unlabeled medical image, we first generate high-frequency and low-frequency views through high-pass and low-pass filtering operations. For each high-frequency view of an unlabeled image, and low frequency view , using a random data augmentation strategy to create weakly augmented variants ( ) and the strongly enhanced variant ( ).

[0040] The subnetwork then produces the logit output: , ; Where, and Represents the initial parameters of t training iterations in the internal learning process. and Perform argmax operation to generate high-frequency one-hot pseudo labels and low-frequency unique hot pseudo labels .

[0041] The obtained unique pseudo labels are used to perform cross-pseudo label supervision on the sub-network, and its loss function combines the cross entropy loss of each sub-network and Dice loss : , ; in, represents the cross pseudo-label supervision loss of the high-frequency sub-network, Represents the cross pseudo-label supervision loss of the low-frequency sub-network. For the convenience of subsequent understanding, the cross pseudo-label supervision loss is expressed as .

[0042] Similarly, for the labeled samples , similar to that obtained by weak enhancement and .

[0043] Figure 4 The calculation and comparison of the average confidence for each class in labeled and unlabeled medical images from the ACDC dataset are presented. It can be observed that the average confidence of the background (BG) class in the unlabeled medical images is approximately equal to that in the labeled medical images. However, the average confidence of the predictions for the left ventricle (LV), right ventricle (RV), and myocardium (MYO) in the unlabeled medical images is significantly lower than that in the labeled medical images. This phenomenon is attributed to the lack of clear supervisory signal in the unlabeled medical images, which causes the model to exhibit greater uncertainty when segmenting anatomical structures. To alleviate this problem, this embodiment introduces mutual uncertainty distance (MUD) supervision into the internal learning process, leveraging uncertainty information to optimize the model's predictions for unlabeled samples. This module encourages the model to generate more confident and reliable predictions, thereby narrowing the confidence gap between labeled and unlabeled medical images.

[0044] Specifically, mutual uncertainty distance (MUD) supervision consists of two parts: category entropy minimization (CEM) and entropy distance minimization (EDM). The following takes high-frequency views as an example to explain them in detail.

[0045] 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 multi-class segmentation tasks, the entropy of the logit output of each class is first calculated to evaluate the degree of uncertainty of each class, expressed as: , ; Where, and Represent the high frequency domain of unlabeled samples and labeled samples respectively. The entropy of the class, and is the corresponding weighted softmax output.

[0046] In order to enhance the robustness of the model to uncertain areas, the Euclidean distance between these entropy values is minimized, which is expressed as: ; Where, Denotes the L2 norm. Minimizing this loss encourages the model to align the confidence distributions between unlabeled and labeled samples for each class. This alignment helps mitigate the effects of uncertainty in ambiguous regions, leading to more consistent and robust predictions.

[0047] Entropy distance minimization (EDM) aims to calculate the distance between the confidence entropy of each class in the labeled sample and the confidence entropy of each class in the unlabeled sample, and minimize the difference between these corresponding distances to make the uncertainty distribution between the labeled and unlabeled samples consistent. Specifically, the uncertainty distance is first defined as follows: , ; Where, and Measure the class of labeled sample set and unlabeled sample set separately and These distances capture the structural relationships between categories based on uncertainty patterns.

[0048] Then, formulate the entropy distance minimization loss To enforce the consistency between the intra-class entropy structures of labeled and unlabeled samples, it is expressed as: ; Where C is the total number of categories. By minimizing the loss of entropy distance Minimizing this,adjusts the relative uncertainty distribution between the labeled and unlabeled,sets, thus determining that the confidence pattern of the model,generalizes consistently on both datasets.

[0049] Furthermore, the mutual uncertainty distance (MUD) supervision loss is defined as a weighted combination of the category entropy minimization (CEM) loss and the entropy distance minimization (EDM) loss, expressed as: ; Where, and are predefined weights. The mutual uncertainty distance (MUD) supervision loss is then incorporated into the inner loop loss of the high-frequency sub-network and the low-frequency sub-network.

[0050] Specifically, the inner loop loss of the high-frequency sub-network is the sum of the cross-pseudo-label supervision loss and the mutual uncertainty distance (MUD) supervision loss of the high-frequency sub-network, and the inner loop loss of the low-frequency sub-network is the sum of the cross-pseudo-label supervision loss and the mutual uncertainty distance (MUD) supervision loss of the low-frequency sub-network, which can be expressed as: , .

[0051] Based on the inner loop loss, the state parameters of the high-frequency sub-network and the low-frequency sub-network are updated accordingly through gradient descent, which is specifically expressed as: , ; Where, and are the updated high-frequency sub-network state parameters and low-frequency sub-network state parameters, which will be used as the initial state parameters of the high-frequency sub-network and the low-frequency sub-network in the external learning process. represents the learning rate, and They are high-frequency sub-networks ( ) and the low-frequency sub-network ( ) of the inner loop loss.

[0052] This embodiment reduces the uncertainty gap between labeled and unlabeled samples by introducing mutual uncertainty distance (MUD) supervision in the internal learning process, improves pseudo-label reliability and alleviates error accumulation, thereby ensuring high-quality pseudo-label generation.

[0053] S204. External learning process Based on the improvement of the internal learning process, this embodiment proposes a novel external learning mechanism, which enables the sub-network to adaptively self-correct the residual confirmation bias through performance-driven self-optimization. By setting up a dual feedback system in the external learning process, each sub-network can dynamically adjust its pseudo-label utilization according to the effectiveness of cross-view supervision.

[0054] First, 、 、 and Represents the logit output from the internal learning process. The output logit after the internal learning state is updated is calculated as follows: , .

[0055] In order to measure the reliability of cross-view pseudo labels, this embodiment introduces a performance feedback coefficient To quantify the optimization improvement of labeled samples, dynamically adjust the pseudo-label confidence, which is specifically expressed as: , ; Where, Reflect pseudo labels 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.

[0056] 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: , .

[0057] The above formula automatically amplifies credible pseudo-labels while suppressing unreliable pseudo-labels.

[0058] For labeled samples, cross-supervision loss is also combined to utilize certain supervisory signals from labeled samples, specifically expressed as: , .

[0059] 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: , ; 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: , ; 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.

[0060] This embodiment dynamically adjusts the sub-network by introducing an adaptive self-correction (ASC) loss in the internal learning process to introduce a feedback mechanism, so that the model can continuously adapt through self-correction based on the performance feedback of labeled samples.

[0061] To verify the effectiveness of the frequency-based adaptive interactive learning semi-supervised medical segmentation method described above, this example conducted comparative experiments on two popular datasets: the ACDC dataset and the PROMISE12 dataset. The ACDC dataset contains 100 short-axis cine MRI images with expert annotations for three categories: left ventricle (LV), right ventricle (RV), and myocardium (MYO). The dataset is divided into 70 images for training, 10 images for validation, and 20 images for testing. The PROMISE12 dataset includes 50 prostate MRI scans collected from multiple medical centers with different scanning protocols. The dataset is divided into 35 images for training and 15 images for testing.

[0062] To quantitatively evaluate the performance of the proposed method, this embodiment utilizes the following metrics to assess overlap and boundary accuracy: (1) Jaccard index, which calculates the intersection over union (IoU) of the prediction and the ground truth, providing a measure of the 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 objects; (3) 95th percentile Hausdorff distance (95HD), which measures the maximum distance between the predicted boundary and the ground truth; and (4) average surface distance (ASD), which calculates the average distance between corresponding points on the predicted surface and the ground truth.

[0063] The frequency-based adaptive interactive learning semi-supervised medical segmentation method proposed in this example is implemented using PyTorch and run 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. MRI scans are preprocessed into 2D slices, resized to 256×256 for the ACDC dataset and 224×224 for the Promise dataset, consistent with related work. The batch size is set to 16, consisting of 8 labeled samples and 8 unlabeled samples. During internal and external learning, stochastic gradient descent (SGD) is used as the optimizer for the HF and LF models. The initial learning rate is , momentum is 0.9, and 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 jittering; strong augmentation further incorporates a grid masking strategy, in which certain parts of the image are masked in a structured grid pattern.

[0064] To evaluate the effectiveness of our method, we compared it with two supervised methods on the ACDC dataset: U-Net and Swin-unet, which are used as backbone models in the architecture, and eight semi-supervised methods: DTC, URPC, MC-Net, SS-net, SCP-Net, PLGCL, Cross-Teaching, BCP, and ABD. For fair comparison, 5% and 10% of the medical images in the entire ACDC dataset were selected as labeled samples for the experiments. Table 1 shows the quantitative evaluation, as follows: Table 1 Quantitative evaluation on the ACDC dataset

[0065] As shown in Table 1, our method achieves a DSC of 90.83%, a Jaccard index of 83.57%, a 95HD of 1.12%, and an ASD of 0.27% with 10% labeled samples. Compared to the supervised baseline trained with 10% labeled samples, our method improves DSC by 11.42% on U-Net and 46.51% on Swin-UNet, demonstrating that, with the same labeled sample ratio, our method achieves superior segmentation performance compared to fully supervised models. Further comparisons with semi-supervised methods show that, with 10% labeled samples, our method improves DSC and Jaccard by 1.02% and 1.62%, respectively, demonstrating its superiority in segmentation quality. Furthermore, our method improves 95HD by 0.34% and ASD by 0.22, indicating more accurate boundary delineation. Furthermore, with only 5% labeled samples, our method consistently outperforms other methods on all metrics. Figure 5 Qualitative comparisons are made between the proposed method and competing methods on five representative cases of segmentation, including Swin-UNet, MC-Net, SS-Net, Cross-Teaching, BCP, and ABD. The cases are arranged in order of segmentation difficulty from (a) to (e). Figure 5 It can be seen that our method has superior segmentation performance in handling large and small structures in cardiac imaging, as well as multi-class and single-class segmentation tasks. Compared with other methods, our method can always achieve more accurate boundary delineation and better segmentation structure preservation.

[0066] This example also extends this method to the PROMISE12 dataset to evaluate its generalization ability. Following the experimental protocol adopted by state-of-the-art methods (CCT, URPC, SS-Net, SLC-Net, SCP-Net, ABD) on the PROMISE12 dataset, 20% of the images were used as labeled samples, the remaining 80% were considered unlabeled samples, and performance was evaluated using the DSC and ASD metrics. As shown in Table 2, this method achieved a DSC of 82.14% and an ASD of 1.30%. Compared with the fully supervised method, this method achieved a 15.13% improvement in DSC and a 12.57% reduction in ASD. Furthermore, our FAM outperformed other state-of-the-art semi-supervised methods on all evaluation metrics, establishing a new state-of-the-art for the prostate segmentation task.

[0067] Table 2 Comparison of generalization ability on PROMISE12 dataset

[0068] Figure 6A visualization of the results compared to other SOTA methods (Swin-UNet, BCP, ABD) is provided. The figure includes four examples where the segmentation difficulty increases from (a) to (d) as the object size decreases. Figure 6 As can be seen, in relatively easy cases (a) and (b), most methods achieve high segmentation accuracy, with the main differences appearing in boundary accuracy. However, as the difficulty increases, the accuracy of other methods gradually decreases, especially in (d), where they show significant errors in segmenting small objects. In contrast, even in challenging cases, our method can consistently maintain high segmentation quality and well preserve boundaries.

[0069] Example 2 In a typical embodiment of the present invention, this embodiment discloses a frequency-based adaptive interactive learning semi-supervised medical segmentation system, comprising: The acquisition module is configured to: acquire the image to be segmented; 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 the image segmentation result; Among them, the semi-supervised medical segmentation model includes a high-frequency sub-network and a low-frequency sub-network. Its training process includes an internal learning process and an external learning process. The specific training steps are as follows: Generate high-frequency views and low-frequency views through a frequency dual-view stream architecture; In the internal learning process, high-frequency and low-frequency views are used as input to construct complementary views, and the corresponding sub-network states are updated through cross-pseudo-label supervision and mutual uncertainty distance (MUD) supervision, which serves as the initial state of the external learning process; During the external learning process, the corresponding sub-network state is updated through cross-supervision loss and adaptive self-correction loss, and is used as the initial state of the internal learning process of the next cycle.

[0070] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Those skilled in the art will readily appreciate that various modifications and variations of the present invention are possible. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the present invention shall be included within the scope of protection of the present invention.

Claims

1. Frequency-based adaptive interactive learning semi-supervised medical segmentation method, characterized by: include: Obtain the image to be segmented; Input the image to be segmented into a pre-trained semi-supervised medical segmentation model for processing to obtain the image segmentation result; The semi-supervised medical segmentation model includes a high-frequency sub-network and a low-frequency sub-network. The training process includes an internal learning process and an external learning process. The specific training steps are as follows: Generate high-frequency views and low-frequency views through a frequency dual-view stream architecture; In the internal learning process, high-frequency views and low-frequency views are used as input to construct complementary views, and the corresponding sub-network states are updated through cross-pseudo-label supervision and mutual uncertainty distance supervision, which serve as the initial state of the external learning process; During the external learning process, the corresponding sub-network state is updated through cross-supervision loss and adaptive self-correction loss, and is used as the initial state of the internal learning process of the next cycle.

2. The frequency-based adaptive interactive learning semi-supervised medical segmentation method according to claim 1, characterized in that: The high-frequency view and low-frequency view are generated through the frequency dual-view flow architecture. The specific process includes: Based on Gaussian filter, high-frequency components and low-frequency components of labeled images and unlabeled images are extracted respectively; An enhanced high-frequency view and an enhanced low-frequency view are constructed based on a set enhancement factor; A CNN-based high-frequency sub-network is used to extract local features from the enhanced low-frequency views, and a Transformer-based low-frequency sub-network is used to capture long-range dependencies in the enhanced high-frequency views.

3. The frequency-based adaptive interactive learning semi-supervised medical segmentation method according to claim 1, characterized in that: The specific process of cross-pseudo-label supervision includes: Use random data augmentation strategies to create weakly augmented and strongly augmented variants of high-frequency and low-frequency views, respectively. Generate high-frequency one-hot pseudo labels and low-frequency one-hot pseudo labels based on the logit output of the strong enhancement variant; The obtained pseudo labels are used to perform cross-pseudo-label supervision on the sub-network, and the cross-pseudo-label supervision loss is the sum of the cross entropy loss and the Dice loss of each sub-network.

4. The frequency-based adaptive interactive learning semi-supervised medical segmentation method according to claim 1, characterized in that: The mutual uncertainty distance supervision utilizes uncertainty information to optimize the model's prediction of unlabeled samples, including minimizing category entropy and entropy distance; the loss function of the mutual uncertainty distance supervision is a weighted combination of category entropy minimization loss and entropy distance minimization loss.

5. The frequency-based adaptive interactive learning semi-supervised medical segmentation method according to claim 1, characterized in that: During the internal learning process, the corresponding sub-network status is updated as follows: Calculate the inner loop loss of the high-frequency sub-network and the low-frequency sub-network respectively; Based on the inner loop loss, the state parameters of the high-frequency sub-network and the low-frequency sub-network are updated accordingly by gradient descent; the updated state parameters of the high-frequency sub-network and the low-frequency sub-network are used as the initial state parameters of the high-frequency sub-network and the low-frequency sub-network in the external learning process respectively.

6. The frequency-based adaptive interactive learning semi-supervised medical segmentation method according to claim 5, characterized in that: The inner loop loss of the high-frequency sub-network is the sum of the cross-pseudo-label supervision loss and the mutual uncertainty distance supervision loss of the high-frequency sub-network; the inner loop loss of the low-frequency sub-network is the sum of the cross-pseudo-label supervision loss and the mutual uncertainty distance supervision loss of the low-frequency sub-network.

7. The frequency-based adaptive interactive learning semi-supervised medical segmentation method according to claim 1, characterized in that: During the external learning process, the pseudo-label confidence is dynamically adjusted through the performance feedback coefficient, and the adaptive self-correction loss of the corresponding sub-network is calculated based on the performance feedback coefficient. The sub-network adjusts its learning strategy through the adaptive self-correction loss.

8. The frequency-based adaptive interactive learning semi-supervised medical segmentation method according to claim 1, characterized in that: During the external learning process, the corresponding sub-network status is updated as follows: Calculate the outer loop loss of the high-frequency sub-network and the low-frequency sub-network respectively; 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; the updated state parameters of the high-frequency sub-network and the low-frequency sub-network 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.

9. The frequency-based adaptive interactive learning semi-supervised medical segmentation method according to claim 8, characterized in that: 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; the inner loop loss of the low-frequency sub-network is the sum of the cross-supervision loss and the mutual uncertainty distance supervision loss of the low-frequency sub-network.

10. Frequency-based adaptive interactive learning semi-supervised medical segmentation system, characterized by: include: The acquisition module is configured to: acquire the image to be segmented; 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 the image segmentation result; The semi-supervised medical segmentation model includes a high-frequency sub-network and a low-frequency sub-network. The training process includes an internal learning process and an external learning process. The specific training steps are as follows: Generate high-frequency views and low-frequency views through a frequency dual-view stream architecture; In the internal learning process, high-frequency views and low-frequency views are used as input to construct complementary views, and the corresponding sub-network states are updated through cross-pseudo-label supervision and mutual uncertainty distance supervision, which serve as the initial state of the external learning process; During the external learning process, the corresponding sub-network state is updated through cross-supervision loss and adaptive self-correction loss, and is used as the initial state of the internal learning process of the next cycle.

Images