Medical image segmentation method and device, computer equipment and storage medium

Abstract

The invention relates to a medical image segmentation method and device, computer equipment and a storage medium. The method comprises the following steps: introducing a KAN module on the basis of a teacher-student model of an interpolation consistency training framework, and constructing a medical image segmentation network in the teacher-student model, wherein the medical image segmentation network comprises a cascade convolution encoder, a cascade KAN encoder, a cascade KAN decoder and a cascade convolution decoder. According to the invention, nonlinear mapping and sparse representation are introduced by using the KAN network to extract more discriminative features, and the teacher-student model under an interpolation consistency training framework is used to effectively improve the complementarity of marked and unmarked data, so that the learning ability and generalization ability of the model are enhanced, the marking cost is reduced, and the marking efficiency is improved. A good segmentation result can be obtained only through a small amount of mark data, the number of parameters of the model can be greatly reduced, the test time can be greatly shortened, and the medical image segmentation efficiency is improved while the segmentation effect of the medical image segmentation network is improved.

Description

Technical Field

[0001] The present application relates to the field of image processing technology, and in particular to a medical image segmentation method, apparatus, computer equipment, and storage medium. Background Art

[0002] Medical image segmentation is a crucial component of medical image analysis, significantly impacting a range of clinical applications, including organ segmentation, tumor identification, and personalized treatment planning. It is crucial for early disease diagnosis, disease monitoring, and treatment efficacy assessment. Accurate segmentation of medical images enables detailed quantification of anatomical structures, pathological regions, and lesion boundaries, providing physicians with accurate auxiliary diagnostic information and playing a key role in clinical decision-making. In recent years, deep learning methods have significantly advanced medical image segmentation. However, traditional fully supervised learning methods exhibit significant limitations in real-world clinical settings with limited labeled data, making them difficult to meet the demands of diverse clinical scenarios. Semi-supervised learning (SSL) methods have emerged as an effective approach to address this problem. SSL leverages limited labeled data and abundant unlabeled data, reducing labeling costs while enhancing model generalization and robustness. In recent years, various SSL frameworks have emerged, such as the mean teacher model (MT), uncertainty-aware mean teacher (UAMT), and entropy minimization (EN). These frameworks employ different strategies to effectively utilize unlabeled data to enhance model performance.

[0003] In addition, many studies have expanded upon these frameworks and derived a variety of high-performance semi-supervised medical image segmentation models. For example, an adversarial self-assembly network effectively utilizes unlabeled data by combining adversarial learning and self-assembly mechanisms, so that ASE-Net improves segmentation performance by generating pseudo-labels and utilizing adversarial learning mechanisms to enhance the robustness of the network. A SSL method based on a graph convolutional network (GCN) captures spatial relationships and contextual information in medical images by constructing a graph structure, thereby improving segmentation accuracy. A SSL method based on a graph convolutional network (GCN) captures spatial relationships and contextual information in medical images by constructing a graph structure. Although these methods have made significant progress in SSL, they still have difficulty coping with the inherent complexity and variability of various medical images, and suffer from problems of poor results and low efficiency. Summary of the Invention

[0004] The first object of the present invention is to overcome the problems existing in the above-mentioned prior art and to provide a medical image segmentation method, apparatus, computer equipment and storage medium that can improve the image segmentation effect of medical images with complex anatomical structures.

[0005] To achieve the above object, the present invention adopts the following technical solutions:

[0006] A medical image segmentation method, comprising:

[0007] Constructing a medical image segmentation network; the medical image segmentation network includes a student model and a teacher model, wherein the student model and the teacher model respectively include a cascaded convolutional encoder, a cascaded KAN encoder, a cascaded KAN decoder, and a cascaded convolutional decoder; the cascaded convolutional encoder, the cascaded KAN encoder, the cascaded KAN decoder, and the cascaded convolutional decoder are sequentially cascaded;

[0008] The input data is processed by the cascaded convolution encoder to obtain a multi-level convolution down-sampling feature map, the last level of convolution down-sampling feature map is processed by the cascaded KAN encoder to obtain a multi-level KAN down-sampling feature map, the last level of KAN down-sampling feature map is processed by the cascaded KAN decoder to obtain a multi-level KAN up-sampling feature map, and the last level of KAN up-sampling feature map is processed by the cascaded convolution decoder to obtain a multi-level convolution up-sampling feature map; wherein, the multi-level convolution down-sampling feature map, the multi-level KAN down-sampling feature map, the multi-level KAN up-sampling feature map and the multi-level convolution up-sampling feature map are skip-connected according to the size information of the feature map;

[0009] Activate and output the last level of convolution upsampling feature map to obtain a prediction map; the prediction map includes the prediction map output by the student model and the prediction map output by the teacher model;

[0010] Obtaining labeled samples and unlabeled samples of a medical image, and inputting the labeled samples and the unlabeled samples into the medical image segmentation network;

[0011] Based on an interpolation consistency training algorithm, the medical image segmentation network is trained according to the labeled samples, the unlabeled samples, the pre-constructed loss function, and the prediction graph to obtain a trained medical image segmentation network;

[0012] Medical image segmentation is performed using the trained medical image segmentation network.

[0013] As a preferred embodiment, the cascaded KAN encoder includes at least two KAN encoders; wherein the KAN encoder includes: a feature reshaping module, a word segmentation module, a KAN module and a deep convolution module.

[0014] As a preferred embodiment, it further includes: the cascaded KAN decoder includes at least two KAN decoders; wherein the KAN decoder includes: a feature reconstruction module, a word segmentation module, a KAN module and a deep convolution module.

[0015] As a preferred embodiment, it also includes: based on the interpolation consistency training algorithm, the medical image segmentation network is trained according to the labeled samples, the unlabeled samples, the pre-constructed loss function and the prediction graph to obtain a trained medical image segmentation network; the pre-constructed loss function includes a supervision loss function and a consistency loss function.

[0016] As a preferred embodiment, the method further includes: based on a multi-task learning mechanism, the student model and the teacher model respectively and simultaneously predict the segmentation mask and the signed distance map;

[0017] Based on the interpolation consistency training algorithm, the medical image segmentation network is trained according to the labeled samples, the unlabeled samples, the pre-constructed loss function, the segmentation mask prediction map and the signed distance prediction map output by the student model and the teacher model to obtain a trained medical image segmentation network.

[0018] As a preferred embodiment, the supervision loss function includes a segmentation mask supervision loss function and a signed distance graph supervision loss function; the consistency loss function includes a segmentation mask consistency loss function and a signed distance graph supervision loss function. Figure 1 Consistency loss function.

[0019] As a preferred implementation, it further includes: the KAN module includes multiple KAN layers.

[0020] A second object of the present invention is to provide a medical image segmentation device, comprising:

[0021] A network construction module is used to construct a medical image segmentation network; the medical image segmentation network includes a student model and a teacher model, the student model and the teacher model respectively include a cascaded convolutional encoder, a cascaded KAN encoder, a cascaded KAN decoder and a cascaded convolutional decoder; the cascaded convolutional encoder, the cascaded KAN encoder, the cascaded KAN decoder and the cascaded convolutional decoder are cascaded in sequence;

[0022] A network processing module, configured to process the input data through the cascaded convolution encoder to obtain a multi-level convolution downsampling feature map, process the last level of convolution downsampling feature map through the cascaded KAN encoder to obtain a multi-level KAN downsampling feature map, process the last level of KAN downsampling feature map through the cascaded KAN decoder to obtain a multi-level KAN upsampling feature map, and process the last level of KAN upsampling feature map through the cascaded convolution decoder to obtain a multi-level convolution upsampling feature map; wherein, the multi-level convolution downsampling feature map, the multi-level KAN downsampling feature map, the multi-level KAN upsampling feature map and the multi-level convolution upsampling feature map are jump-connected according to the size information of the feature map; the last level of convolution upsampling feature map is activated and output to obtain a prediction map; the prediction map includes the prediction map output by the student model and the prediction map output by the teacher model;

[0023] a data input module, configured to obtain labeled samples and unlabeled samples of medical images, and input the labeled samples and the unlabeled samples into the medical image segmentation network;

[0024] a training module, configured to train the medical image segmentation network based on an interpolation consistency training algorithm according to the labeled samples, the unlabeled samples, the pre-constructed loss function, and the prediction graph to obtain a trained medical image segmentation network;

[0025] A using module is used to perform medical image segmentation using the trained medical image segmentation network.

[0026] A third object of the present invention is to provide a computer device comprising a memory and a processor, wherein the memory stores a computer program, and when the processor executes the computer program, the following steps are implemented:

[0027] Constructing a medical image segmentation network; the medical image segmentation network includes a student model and a teacher model, wherein the student model and the teacher model respectively include a cascaded convolutional encoder, a cascaded KAN encoder, a cascaded KAN decoder, and a cascaded convolutional decoder; the cascaded convolutional encoder, the cascaded KAN encoder, the cascaded KAN decoder, and the cascaded convolutional decoder are sequentially cascaded;

[0028] The input data is processed by the cascaded convolution encoder to obtain a multi-level convolution down-sampling feature map, the last level of convolution down-sampling feature map is processed by the cascaded KAN encoder to obtain a multi-level KAN down-sampling feature map, the last level of KAN down-sampling feature map is processed by the cascaded KAN decoder to obtain a multi-level KAN up-sampling feature map, and the last level of KAN up-sampling feature map is processed by the cascaded convolution decoder to obtain a multi-level convolution up-sampling feature map; wherein, the multi-level convolution down-sampling feature map, the multi-level KAN down-sampling feature map, the multi-level KAN up-sampling feature map and the multi-level convolution up-sampling feature map are skip-connected according to the size information of the feature map;

[0029] Activate and output the last level of convolution upsampling feature map to obtain a prediction map; the prediction map includes the prediction map output by the student model and the prediction map output by the teacher model;

[0030] Obtaining labeled samples and unlabeled samples of a medical image, and inputting the labeled samples and the unlabeled samples into the medical image segmentation network;

[0031] Based on an interpolation consistency training algorithm, the medical image segmentation network is trained according to the labeled samples, the unlabeled samples, the pre-constructed loss function, and the prediction graph to obtain a trained medical image segmentation network;

[0032] Medical image segmentation is performed using the trained medical image segmentation network.

[0033] A fourth object of the present invention is to provide a computer-readable storage medium having a computer program stored thereon, wherein when the computer program is executed by a processor, the following steps are implemented:

[0034] Constructing a medical image segmentation network; the medical image segmentation network includes a student model and a teacher model, wherein the student model and the teacher model respectively include a cascaded convolutional encoder, a cascaded KAN encoder, a cascaded KAN decoder, and a cascaded convolutional decoder; the cascaded convolutional encoder, the cascaded KAN encoder, the cascaded KAN decoder, and the cascaded convolutional decoder are sequentially cascaded;

[0035] The input data is processed by the cascaded convolution encoder to obtain a multi-level convolution down-sampling feature map, the last level of convolution down-sampling feature map is processed by the cascaded KAN encoder to obtain a multi-level KAN down-sampling feature map, the last level of KAN down-sampling feature map is processed by the cascaded KAN decoder to obtain a multi-level KAN up-sampling feature map, and the last level of KAN up-sampling feature map is processed by the cascaded convolution decoder to obtain a multi-level convolution up-sampling feature map; wherein, the multi-level convolution down-sampling feature map, the multi-level KAN down-sampling feature map, the multi-level KAN up-sampling feature map and the multi-level convolution up-sampling feature map are skip-connected according to the size information of the feature map;

[0036] Activate and output the last level of convolution upsampling feature map to obtain a prediction map; the prediction map includes the prediction map output by the student model and the prediction map output by the teacher model;

[0037] Obtaining labeled samples and unlabeled samples of a medical image, and inputting the labeled samples and the unlabeled samples into the medical image segmentation network;

[0038] Based on an interpolation consistency training algorithm, the medical image segmentation network is trained according to the labeled samples, the unlabeled samples, the pre-constructed loss function, and the prediction graph to obtain a trained medical image segmentation network;

[0039] Medical image segmentation is performed using the trained medical image segmentation network.

[0040] The medical image segmentation method, apparatus, computer device, and storage medium described above introduce a KAN module into a teacher-student model within an interpolation consistency training framework, thereby constructing a medical image segmentation network comprising a cascaded convolutional encoder, a cascaded KAN encoder, a cascaded KAN decoder, and a cascaded convolutional decoder within the teacher-student model. In the teacher-student model, the cascaded convolutional encoder processes input data to obtain a multi-level convolutional downsampling feature map, the cascaded KAN encoder processes the last level of convolutional downsampling feature map to obtain a multi-level KAN downsampling feature map, the cascaded KAN decoder processes the last level of KAN downsampling feature map to obtain a multi-level KAN upsampling feature map, and the cascaded convolutional decoder processes the last level of KAN upsampling feature map to obtain a multi-level convolutional upsampling feature map. The multi-level convolutional downsampling feature map, the multi-level KAN downsampling feature map, the multi-level KAN upsampling feature map, and the multi-level convolutional upsampling feature map are skip-connected based on feature map size information. The last level of convolutional upsampling feature map is activated and output to obtain a prediction map. Then, based on the interpolation consistency training algorithm, the network is trained according to labeled samples, unlabeled samples, a pre-built loss function, and a prediction graph to obtain a trained medical image segmentation network for medical image segmentation. The present invention utilizes the KAN network to introduce nonlinear mapping and sparse representation to extract more discriminative features. The teacher-student model under the interpolation consistency training framework is used to effectively improve the complementarity of labeled and unlabeled data, thereby enhancing the learning and generalization capabilities of the model, reducing labeling costs, and obtaining good segmentation results with only a small amount of labeled data. At the same time, it can greatly reduce the number of model parameters and testing time, thereby improving the segmentation effect of the medical image segmentation network while improving the efficiency of medical image segmentation. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] Figure 1 FIG1 is a flow chart of a medical image segmentation method in one embodiment.

[0042] Figure 2 2 is a structural diagram of the Tok-KAN module in one embodiment.

[0043] Figure 3 FIG. 4 is a structural diagram of a KAN module in one embodiment.

[0044] Figure 4 2 is a framework diagram of a medical image segmentation method in one embodiment.

[0045] Figure 5 A visual comparison diagram of various methods in one embodiment.

[0046] Figure 6 This is a diagram showing the visual evaluation of ablation study results using validation images in one embodiment.

[0047] Figure 7 FIG. 4 is a structural block diagram of a medical image segmentation device in one embodiment.

[0048] Figure 8 FIG. 1 is a diagram showing the internal structure of a computer device in one embodiment. DETAILED DESCRIPTION

[0049] In order to make the purpose, technical solutions and advantages of this application more clear, the following further describes this application in detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain this application and are not intended to limit this application.

[0050] The medical image segmentation method provided in the present application can be applied in the following application environment. A medical image segmentation method is executed by a terminal, and a KAN module is introduced on the basis of the teacher-student model of the interpolation consistency training framework to construct a medical image segmentation network including a cascaded convolution encoder, a cascaded KAN encoder, a cascaded KAN decoder and a cascaded convolution decoder in the teacher-student model. The input data is processed by the cascaded convolution encoder to obtain a multi-level convolution downsampling feature map, the last level of convolution downsampling feature map is processed by the cascaded KAN encoder to obtain a multi-level KAN downsampling feature map, the last level of KAN downsampling feature map is processed by the cascaded KAN decoder to obtain a multi-level KAN upsampling feature map, and the last level of KAN upsampling feature map is processed by the cascaded convolution decoder to obtain a multi-level convolution upsampling feature map; wherein, the multi-level convolution downsampling feature map, the multi-level KAN downsampling feature map, the multi-level KAN upsampling feature map and the multi-level convolution upsampling feature map are jump-connected according to the size information of the feature map; the last level of convolution upsampling feature map is activated and output to obtain a prediction map. Then, based on the interpolation consistency training algorithm, the network is trained according to the labeled samples, unlabeled samples, a pre-built loss function, and the prediction graph to obtain a trained medical image segmentation network for medical image segmentation. The terminal can be, but is not limited to, various personal computers, laptops, smartphones, tablets, and portable wearable devices.

[0051] In one embodiment, Figure 1 As shown, a medical image segmentation method is provided, comprising the following steps:

[0052] Step 102: construct a medical image segmentation network.

[0053] The medical segmentation network constructed in this application introduces the Kolmogorov-Arnold network (KAN) into the teacher-student model within the interpolation consistency training (ICT) framework.

[0054] Specifically, the student and teacher models are improved upon U-Net, with multiple Tok-KAN modules added between the traditional U-Net encoder and decoder, forming a cascaded KAN encoder and decoder. The U-Net network was first used to solve medical image segmentation problems and is a classic network in image generation models. The network structure consists of an encoder and a decoder, with the encoder using downsampling and the decoder using upsampling. Jump connections are used between the downsampled and upsampled feature maps, allowing for more efficient utilization of the feature maps extracted at each layer. KAN is a novel neural network architecture that improves network performance and interpretability by replacing weight parameters with learnable single-variable functions. It has the potential to become a key driver of the development of deep learning models.

[0055] In this embodiment, the student model and the teacher model respectively include a cascaded convolutional encoder, a cascaded KAN encoder, a cascaded KAN decoder and a cascaded convolutional decoder; the cascaded convolutional encoder, the cascaded KAN encoder, the cascaded KAN decoder and the cascaded convolutional decoder are cascaded in sequence.

[0056] The Tok-KAN module structure is as follows Figure 2 As shown, the system includes a feature reshaping module 202, a tokenization module 204, a KAN module 206, and a deep convolution module 208. The feature reshaping module 202 is used to reshape the input features into a series of flattened two-dimensional blocks. The tokenization module 204 is used to obtain token information. In this embodiment, the tokenization module 204 uses a convolutional layer with a kernel size of 3 to perform linear projection on the flattened two-dimensional blocks. This convolutional layer effectively encodes positional information as token information, and its performance can exceed that of standard positional encoding techniques. After obtaining the token information, the KAN module 206 processes the token information, and then passes it through the deep convolution module 208. After convolution, the output is activated through batch normalization (BN) and the ReLU activation function.

[0057] The Tok-KAN module can be defined as:

[0058]

[0059] in is the output feature map of the nth layer.

[0060] The KAN module in the Tok-KAN module includes multiple KAN layers, the structure is as follows Figure 3 As shown. Assuming that the KAN module includes L layers of KAN layers, the L-layer KAN module can be characterized as a nested set of multiple KAN layers:

[0061]

[0062] in, Table ith layer of the KAN module.

[0063] In this embodiment, .

[0064] Step 104: Process the input data through a cascaded convolutional encoder to obtain a multi-level convolution downsampling feature map, process the last level of convolution downsampling feature map through a cascaded KAN encoder to obtain a multi-level KAN downsampling feature map, process the last level of KAN downsampling feature map through a cascaded KAN decoder to obtain a multi-level KAN upsampling feature map, and process the last level of KAN upsampling feature map through a cascaded convolutional decoder to obtain a multi-level convolution upsampling feature map.

[0065] Among them, the multi-level convolution downsampling feature map, the multi-level KAN downsampling feature map, the multi-level KAN upsampling feature map and the multi-level convolution upsampling feature map are jump-connected according to the size information of the feature map.

[0066] The framework of this application method is as follows Figure 4 As shown in FIG, the constructed medical segmentation network is in the interpolation consistency training (ICT) framework. The student model and the teacher model respectively include a cascaded convolutional encoder 402, a cascaded KAN encoder 404, a cascaded KAN decoder 406 and a cascaded convolutional decoder 408. The student model It is trained not only on the labeled dataset but also on interpolated samples of unlabeled data. Mainly generate pseudo labels from unlabeled data to guide The learning process. and teacher model Maintaining consistency on interpolated samples effectively utilizes the complementarity of labeled and unlabeled data, thereby extracting more reliable feature information from unlabeled data and enhancing the learning and generalization capabilities of the model.

[0067] Furthermore, by implementing skip connections between the cascaded KAN encoder 404 and the cascaded KAN decoder 406 based on the size of the feature maps, the network can reinject high-level semantic information extracted during the encoder phase into the decoder phase, helping the network better restore detailed information. Furthermore, skip connections can help the network better address the vanishing gradient problem, mitigating the gradual decrease in gradients as they propagate deep into the network, thereby improving network training efficiency and convergence speed.

[0068] and Inserting a KAN network between the traditional U-Net network encoder and decoder can extract more discriminative features by introducing nonlinear mapping and sparse representation. This significantly improves the model's ability to extract features, enabling it to better capture complex local and global information, thereby generating more reliable pseudo-labels and improving the quality of the learning signal provided to the student model. As a result, the student model can effectively learn from high-quality pseudo-labels and gradually optimize its segmentation performance, achieving better segmentation results in the SSL environment. In addition, the introduction of the KAN network can significantly reduce the number of model parameters and shorten model testing time, thereby improving the model's image segmentation efficiency.

[0069] Step 106: Activate and output the last level of convolution up-sampled feature map to obtain a prediction map.

[0070] The prediction graph includes the prediction graph output by the student model and the prediction graph output by the teacher model.

[0071] Step 108: Obtain labeled samples and unlabeled samples of the medical image, and input the labeled samples and unlabeled samples into a medical image segmentation network.

[0072] Step 110 , based on the interpolation consistency training algorithm, the medical image segmentation network is trained according to the labeled samples, the unlabeled samples, the pre-built loss function and the prediction graph to obtain a trained medical image segmentation network.

[0073] Step 112: perform medical image segmentation using the trained medical image segmentation network.

[0074] In the above-mentioned medical image segmentation method, a KAN module is introduced into the teacher-student model of the interpolation consistency training framework to construct a medical image segmentation network comprising a cascaded convolutional encoder, a cascaded KAN encoder, a cascaded KAN decoder, and a cascaded convolutional decoder. In the teacher-student model, the cascaded convolutional encoder processes the input data to obtain a multi-level convolutional downsampling feature map, the cascaded KAN encoder processes the last level of convolutional downsampling feature map to obtain a multi-level KAN downsampling feature map, the cascaded KAN decoder processes the last level of KAN downsampling feature map to obtain a multi-level KAN upsampling feature map, and the cascaded convolutional decoder processes the last level of KAN upsampling feature map to obtain a multi-level convolutional upsampling feature map. Skip connections are established between the multi-level convolutional downsampling feature map, the multi-level KAN downsampling feature map, the multi-level KAN upsampling feature map, and the multi-level convolutional upsampling feature map based on the feature map size information. The last level of convolutional upsampling feature map is activated and output to obtain a prediction map. Then, based on the interpolation consistency training algorithm, the network is trained according to labeled samples, unlabeled samples, a pre-built loss function, and a prediction graph to obtain a trained medical image segmentation network for medical image segmentation. The present invention utilizes the KAN network to introduce nonlinear mapping and sparse representation to extract more discriminative features. The teacher-student model under the interpolation consistency training framework is used to effectively improve the complementarity of labeled and unlabeled data, thereby enhancing the learning and generalization capabilities of the model, reducing labeling costs, and obtaining good segmentation results with only a small amount of labeled data. At the same time, it can greatly reduce the number of model parameters and testing time, thereby improving the segmentation effect of the medical image segmentation network while improving the efficiency of medical image segmentation.

[0075] In one embodiment, the method further includes: based on an interpolation consistency training algorithm, training a medical image segmentation network according to labeled samples, unlabeled samples, a pre-constructed loss function, and a prediction graph to obtain a trained medical image segmentation network; the pre-constructed loss function includes a supervision loss function and a consistency loss function.

[0076] In one embodiment, it also includes: based on a multi-task learning mechanism, the student model and the teacher model simultaneously predict the segmentation mask and the signed distance map respectively; based on the interpolation consistency training algorithm, according to the labeled samples, unlabeled samples, and a pre-constructed loss function, the segmentation mask prediction map and the signed distance prediction map output by the student model and the teacher model are used to train the medical image segmentation network to obtain a trained medical image segmentation network.

[0077] Experiments have shown that after the introduction of the KAN network, the segmentation results in some scenarios still differ significantly from the true values. To address this technical issue, this embodiment proposes further improvements by introducing a multi-task learning mechanism that enables the model to simultaneously predict the segmentation mask (SM) and the signed distance map (SDM). The SDM provides distance information from each pixel to the nearest boundary, providing critical spatial context, enabling the model to more accurately understand the structural relationship within the image, thereby further improving segmentation accuracy. The SDM is generated through a 3D convolution block and then activated by a tanh function. The multi-task strategy enhances the learning ability of the model by promoting information sharing between the segmentation task and the auxiliary tasks, thereby further improving its generalization and robustness.

[0078] use Represent a set of N labeled data, let Represents a set of M unlabeled data, where .here, Represents input data, represents the segmentation mask, Denotes the ground truth SDM derived from y. Given an unlabeled sample , The model's prediction confidence map is defined as follows:

[0079]

[0080]

[0081] in, represents the mixing operation applied to the unlabeled data and is defined as:

[0082]

[0083] in, is the sampling coefficient of the Beta distribution. In addition, for the teacher model , the pseudo label is given by the following formula:

[0084]

[0085]

[0086] in, yes The moving average of .

[0087] In one embodiment, the supervision loss function includes a segmentation mask supervision loss function and a signed distance graph supervision loss function; the consistency loss function includes a segmentation mask consistency loss function and a signed distance Figure 1 Consistency loss function.

[0088] In this embodiment, the loss function consists of two main parts: supervision loss and consistency loss Given that this method adopts multi-task training to predict SM and SDM from the input image, the supervision loss and consistency loss are further divided into two sub-parts. Specifically, the supervision loss is composed of the supervision loss of SM and SDM supervision loss Consistency loss includes SM consistency loss and SDM consistency loss Cross Entropy (CE) Loss and dice losses The supervised loss for calculating the segmentation mask is and consistency loss Then use the mean square error (MSE) loss Therefore, the supervision loss and consistency loss as follows:

[0089]

[0090]

[0091]

[0092]

[0093] in and The definition is as follows:

[0094]

[0095]

[0096] The total loss is calculated as the supervision loss and consistency loss The weighted sum of:

[0097]

[0098] Weighting Factor The formula is: , Indicates the number of iterations of the current training, which is used to dynamically adjust with the training process .

[0099] It should be understood that although Figure 1The steps in the flowchart are shown in sequence as indicated by the arrows, but these steps are not necessarily executed in the order indicated by the arrows. Unless otherwise specified in this document, there is no strict order restriction for the execution of these steps, and these steps can be executed in other orders. In addition, Figure 1 At least part of the steps may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily executed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be executed in turn or alternately with other steps or at least part of the sub-steps or stages of other steps.

[0100] In a specific embodiment, the cardiac segmentation dataset in the 2017 Automatic Cardiac Diagnosis Challenge (ACDC) is used to evaluate the segmentation performance of the medical image segmentation model proposed in this application. The dataset contains 200 annotated cardiac magnetic resonance (MR) images from 100 patients, covering five different pathological conditions: healthy individuals, dilated cardiomyopathy, previous myocardial infarction, right ventricular abnormalities, and hypertrophic cardiomyopathy. Segmentation masks of the myocardium (My), right ventricle (RV), and left ventricle (LV) are provided. Images of 70 patients are randomly assigned for training, while images of the remaining 30 patients are designated for validation. In order to prevent overfitting and improve the generalization ability of the model, data augmentation techniques such as random flipping, random cropping, and random rotation in the range of -20 to 20 degrees are applied.

[0101] In addition, the model's performance was quantitatively evaluated using three commonly used evaluation metrics: 95% Hausdorff distance (HD95), Jaccard similarity coefficient (JSC), and Dice similarity coefficient (DSC). These metrics provide a comprehensive analysis of the model's segmentation accuracy, thus helping to fully assess its effectiveness.

[0102] To validate the proposed method's efficiency in leveraging both labeled and unlabeled data, we conducted a detailed comparison with six state-of-the-art SSL methods, including CPS, ICT, MT, UAMT, ASE-Ne, and MS2MPS. To ensure a fair comparison, all methods used the same network backbone as their foundation and did not apply any post-processing strategies. Furthermore, all methods maintained consistent batch size, maximum number of iterations, and optimizer settings.

[0103] Table 1 shows a quantitative comparison of the segmentation performance of all methods using labeled samples from 3 and 14 patients. Experimental results demonstrate that our method consistently achieves superior segmentation performance under both conditions. Compared to the best-performing method, our method achieved an average improvement of 8.45% in DSC, 10.57% in JSC, and a 30.87% reduction in HD95 when using 3 labeled patients. When using 14 labeled patients, our method achieved an improvement of 1.26% in DSC and 1.27% in JSC, while simultaneously reducing HD95 by 15.1%. These findings demonstrate the significant advantages of our method in leveraging both labeled and unlabeled data, effectively improving segmentation accuracy. This capability not only addresses the challenges posed by limited labeled samples but also provides strong evidence for the future adoption of SSL methods in clinical practice.

[0104] Figure 5 Qualitative results for all methods are presented, with the first two rows showing the results when tested using 3 labeled patients, and the last two rows showing the results when tested using 14 labeled patients. Visual analysis shows that all compared methods exhibit both over-segmentation and under-segmentation issues, particularly in the first case, where the RV organ is significantly over-segmented. In the second case, all methods, except the interpolation consistency training (ICT) method, produce segmentation results that are closely consistent with the ground truth. For the visualization results of the 14 cases, the ASE-Net and MS2MPS methods achieve the best segmentation performance among all compared methods, although there is still a certain gap between them and the true annotations. Overall, the segmentation results of the proposed method are closest to the true annotations in all visualization cases, further validating the effectiveness and superiority of the method in medical image segmentation tasks.

[0105] Table 1 Quantitative comparison of this method with other methods. This table shows the mean and standard error (in brackets)

[0106]

[0107] In another specific embodiment, an ablation experiment was conducted based on three labeled patient data sets to verify the impact of the KAN network and multi-task learning mechanism on the performance of the proposed method. First, the UK-ICT model was constructed by introducing the Tok-KAN module into the teacher-student model of the ICT framework. The introduction of the KAN module significantly improved the model's feature extraction capabilities, enabling it to better capture complex local and global information. This improvement led to improved segmentation accuracy. Second, based on this, the UKMT-ICT model was further integrated with a multi-task strategy. This multi-task strategy enhances the model's learning capabilities by promoting information sharing between the segmentation task and auxiliary tasks, further improving its generalization and robustness.

[0108] Table 2. Results of ablation studies using three labeled patients

[0109]

[0110] Table 2 presents quantitative comparison results from ablation experiments, demonstrating that the combination of the KAN network and the multi-task learning mechanism significantly improves model performance, particularly in scenarios with limited labeled data. Compared to the ICT method, the UKMT-ICT method achieves an average improvement of 20.56% in DSC, an average improvement of 30.39% in JSC, and an average reduction of 66.68% in HD95. Furthermore, Table 3 lists the single-case test time and model parameter size for the ICT, UK-ICT, and UKMT-ICT models. The results show that the introduction of the KAN module reduces model parameters by 80.74% and test time by 18.28%. After the introduction of the multi-task learning mechanism, the model parameters and test time remain largely unchanged compared to before the introduction.

[0111] Table 3. Number of parameters and computational complexity of each module of the proposed method.

[0112]

[0113] Figure 6Qualitative comparison results of ablation experiments are presented. Before the introduction of the KAN network and multi-task learning mechanism, the model exhibited obvious over-segmentation and under-segmentation problems in the segmentation task. For example, in the first and second test patients, the left ventricle was severely over-segmented, while the right ventricle was under-segmented. In addition, over-segmentation of the right ventricle was also observed in the third and fourth test patients. After the introduction of KAN, the over-segmentation problem of the left ventricle was significantly improved, but the segmentation results were still significantly different from the true values. With the further introduction of the multi-task strategy, the segmentation performance of the model was greatly improved. Although the right ventricle was still under-segmented in the second test patient, the segmentation results produced by UKMT-ICT were generally very close to the true values. This shows that KAN and multi-task learning mechanisms are crucial to improving the segmentation performance of the model, enabling it to produce excellent results even in more complex situations.

[0114] Table 4. Performance of our method on a validation set with different numbers of labeled and unlabeled images.

[0115]

[0116] In addition, this example also evaluated the effectiveness of the UKMT-ICT model at different ratios of labeled to unlabeled samples: 3%, 5%, 12.5%, 25%, 50%, 75%, and 100%, with 100% representing fully supervised segmentation. Table 4 summarizes the segmentation results of the UKMT-ICT model at different ratios of labeled to unlabeled data. Clearly, the segmentation results using 75% labeled data are very close to those obtained under full supervision (100% labeled data), demonstrating that the model maintains high segmentation accuracy even with reduced labeled data. The segmentation performance of the SSL model gradually improves with increasing numbers of labeled training images. Notably, the shift from 3% to 5% labeled data leads to a significant improvement in model performance, demonstrating that even with limited labeled data, effective utilization of unlabeled data can significantly improve segmentation results. When the labeled data ratio exceeds 25%, model performance saturates, and further improvement is relatively slow. Overall, the UKMT-ICT model demonstrates good robustness across different ratios of labeled data, achieving satisfactory segmentation results with as little as 3% labeled data. This verifies the effectiveness and advantages of the model in a semi-supervised learning environment, especially when labeled data is scarce, it can effectively utilize unlabeled data for auxiliary training, thereby improving segmentation performance.

[0117] In one embodiment, Figure 7As shown, a medical image segmentation device is provided, comprising: a network construction module 702, a network processing module 704, a data input module 706, a training module 708 and a use module 710, wherein:

[0118] A network construction module 702 is used to construct a medical image segmentation network; the medical image segmentation network includes a student model and a teacher model, wherein the student model and the teacher model respectively include a cascaded convolutional encoder, a cascaded KAN encoder, a cascaded KAN decoder, and a cascaded convolutional decoder; the cascaded convolutional encoder, the cascaded KAN encoder, the cascaded KAN decoder, and the cascaded convolutional decoder are cascaded in sequence;

[0119] The network processing module 704 is configured to process the input data through a cascaded convolutional encoder to obtain a multi-level convolution downsampling feature map, process the last level of convolution downsampling feature map through a cascaded KAN encoder to obtain a multi-level KAN downsampling feature map, process the last level of KAN downsampling feature map through a cascaded KAN decoder to obtain a multi-level KAN upsampling feature map, and process the last level of KAN upsampling feature map through a cascaded convolutional decoder to obtain a multi-level convolution upsampling feature map; wherein the multi-level convolution downsampling feature map, the multi-level KAN downsampling feature map, the multi-level KAN upsampling feature map, and the multi-level convolution upsampling feature map are jump-connected based on the size information of the feature maps; the last level of convolution upsampling feature map is activated and output to obtain a prediction map; the prediction map includes the prediction map output by the student model and the prediction map output by the teacher model;

[0120] A data input module 706 is used to obtain labeled samples and unlabeled samples of medical images and input the labeled samples and unlabeled samples into a medical image segmentation network;

[0121] A training module 708 is configured to train the medical image segmentation network based on an interpolation consistency training algorithm according to labeled samples, unlabeled samples, a pre-built loss function, and a prediction graph to obtain a trained medical image segmentation network;

[0122] The module 710 is used to perform medical image segmentation using a trained medical image segmentation network.

[0123] The training module 708 is also used to train the medical image segmentation network based on the interpolation consistency training algorithm according to the labeled samples, the unlabeled samples, the pre-constructed loss function and the prediction graph to obtain a trained medical image segmentation network; the pre-constructed loss function includes a supervision loss function and a consistency loss function.

[0124] The training module 708 is also used to simultaneously predict the segmentation mask and the signed distance map by the student model and the teacher model based on a multi-task learning mechanism; based on the interpolation consistency training algorithm, the medical image segmentation network is trained according to the labeled samples, the unlabeled samples, and the pre-constructed loss function, and the segmentation mask prediction map and the signed distance prediction map output by the student model and the teacher model to obtain a trained medical image segmentation network.

[0125] The specific limitations of medical image segmentation can be found in the limitations of the medical image segmentation method described above and will not be further elaborated here. Each module in the aforementioned medical image segmentation apparatus may be implemented in whole or in part via software, hardware, or a combination thereof. Each of these modules may be embedded in or independent of a processor within a computer device in hardware form, or may be stored in a computer device memory in software form, allowing the processor to call and execute the corresponding operations of each module.

[0126] In one embodiment, a computer device is provided. The computer device may be a terminal, and its internal structure diagram may be as follows: Figure 8 As shown. The computer device includes a processor, a memory, a network interface, a display screen and an input device connected via a system bus. The processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The network interface of the computer device is used to communicate with an external terminal via a network connection. When the computer program is executed by the processor, a medical image segmentation method is implemented. The display screen of the computer device can be a liquid crystal display screen or an electronic ink display screen, and the input device of the computer device can be a touch layer covering the display screen, or a button, trackball or touchpad provided on the computer device housing, or an external keyboard, touchpad or mouse, etc.

[0127] Those skilled in the art will understand that Figure 8 The structure shown in the figure is only a block diagram of a part of the structure related to the solution of the present application, and does not constitute a limitation on the computer device to which the solution of the present application is applied. The specific computer device may include more or fewer components than shown in the figure, or combine certain components, or have a different component arrangement.

[0128] In one embodiment, a computer device is provided, including a memory and a processor. The memory stores a computer program, and the processor implements the steps in the above method embodiment when executing the computer program.

[0129] In one embodiment, a computer-readable storage medium is provided, on which a computer program is stored. When the computer program is executed by a processor, the steps in the above method embodiment are implemented.

[0130] Those skilled in the art will understand that all or part of the processes in the above-mentioned embodiments can be implemented by instructing the relevant hardware through a computer program. The computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, storage, database or other media used in the embodiments provided in this application may include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in many forms such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), Synchronous Link DRAM (SLDRAM), Rambus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0131] The technical features of the above embodiments can be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0132] The above-described embodiments merely represent several implementation methods of the present application. While the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the present invention. It should be noted that a person skilled in the art could make various modifications and improvements without departing from the spirit of the present application, all of which fall within the scope of protection of the present application. Therefore, the scope of protection of the present patent application shall be determined by the appended claims.

Claims

1. A medical image segmentation method, characterized in that: The method comprises: Constructing a medical image segmentation network; the medical image segmentation network includes a student model and a teacher model, wherein the student model and the teacher model respectively include a cascaded convolutional encoder, a cascaded KAN encoder, a cascaded KAN decoder and a cascaded convolutional decoder that are cascaded in sequence; The input data is processed by the cascaded convolution encoder to obtain a multi-level convolution down-sampling feature map, the last level of convolution down-sampling feature map is processed by the cascaded KAN encoder to obtain a multi-level KAN down-sampling feature map, the last level of KAN down-sampling feature map is processed by the cascaded KAN decoder to obtain a multi-level KAN up-sampling feature map, and the last level of KAN up-sampling feature map is processed by the cascaded convolution decoder to obtain a multi-level convolution up-sampling feature map; wherein, the multi-level convolution down-sampling feature map, the multi-level KAN down-sampling feature map, the multi-level KAN up-sampling feature map and the multi-level convolution up-sampling feature map are skip-connected according to the size information of the feature map; The last level of convolution upsamples the feature map activation output to obtain the prediction maps output by the student model and the teacher model respectively; Obtaining labeled samples and unlabeled samples of medical images and inputting them into the medical image segmentation network; Based on the interpolation consistency training algorithm, the medical image segmentation network is trained according to the labeled samples, the unlabeled samples, the pre-constructed loss function and the prediction graph to obtain a trained medical image segmentation network for medical image segmentation.

2. The method according to claim 1, characterized in that The cascade KAN encoder includes at least two KAN encoders; wherein the KAN encoder includes: a feature reshaping module, a word segmentation module, a KAN module and a deep convolution module.

3. The method according to claim 1, characterized in that The cascade KAN decoder includes at least two KAN decoders; wherein the KAN decoder includes: a feature reshaping module, a word segmentation module, a KAN module and a deep convolution module.

4. The method according to claim 1, wherein The interpolation consistency training algorithm is based on training the medical image segmentation network according to the labeled samples, the unlabeled samples, the pre-constructed loss function and the prediction graph to obtain a trained medical image segmentation network, including: Based on the interpolation consistency training algorithm, the medical image segmentation network is trained according to the labeled samples, the unlabeled samples, the pre-constructed loss function and the prediction graph to obtain a trained medical image segmentation network; the pre-constructed loss function includes a supervision loss function and a consistency loss function.

5. The method according to claim 4, characterized in that The interpolation consistency training algorithm is based on training the medical image segmentation network according to the labeled samples, the unlabeled samples, the pre-constructed loss function and the prediction graph to obtain a trained medical image segmentation network, including: Based on a multi-task learning mechanism, the student model and the teacher model simultaneously predict the segmentation mask and the signed distance map respectively; Based on the interpolation consistency training algorithm, the medical image segmentation network is trained according to the labeled samples, the unlabeled samples, the pre-constructed loss function, the segmentation mask prediction map and the signed distance prediction map output by the student model and the teacher model to obtain a trained medical image segmentation network.

6. The method according to claim 5, characterized in that The supervised loss function includes a segmentation mask supervised loss function and a signed distance map supervised loss function; the consistency loss function includes a segmentation mask consistency loss function and a signed distance map consistency loss function.

7. The method according to any one of claims 2 to 6, characterized in that The KAN module includes multiple KAN layers.

8. A medical image segmentation device, characterized in that: The device comprises: A network construction module for constructing a medical image segmentation network; the medical image segmentation network includes a student model and a teacher model, wherein the student model and the teacher model respectively include a cascaded convolutional encoder, a cascaded KAN encoder, a cascaded KAN decoder and a cascaded convolutional decoder that are cascaded in sequence; A network processing module, configured to process the input data through the cascaded convolution encoder to obtain a multi-level convolution downsampling feature map, process the last level of convolution downsampling feature map through the cascaded KAN encoder to obtain a multi-level KAN downsampling feature map, process the last level of KAN downsampling feature map through the cascaded KAN decoder to obtain a multi-level KAN upsampling feature map, and process the last level of KAN upsampling feature map through the cascaded convolution decoder to obtain a multi-level convolution upsampling feature map; wherein, the multi-level convolution downsampling feature map, the multi-level KAN downsampling feature map, the multi-level KAN upsampling feature map and the multi-level convolution upsampling feature map are jump-connected according to the size information of the feature map; the last level of convolution upsampling feature map is activated and output to obtain the prediction maps output by the student model and the teacher model respectively; A data input module, configured to obtain labeled samples and unlabeled samples of medical images and input them into the medical image segmentation network; a training module, configured to train the medical image segmentation network based on an interpolation consistency training algorithm according to the labeled samples, the unlabeled samples, the pre-constructed loss function, and the prediction graph to obtain a trained medical image segmentation network; A using module is used to perform medical image segmentation using the trained medical image segmentation network.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the method according to any one of claims 1 to 7 are implemented.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 7 are implemented.

Images