Medical image segmentation method based on adaptive grouping strategy and Kolmogorov-Arnold Networks

Abstract

The invention relates to a medical image segmentation method based on an adaptive grouping strategy and Kolmogorov-Arnold Networks, and belongs to the field of medical image segmentation. According to the method, a medical image segmentation framework, namely, Kolmogorov-Arnold Network (KAN) with adaptive Group Stratification and Contextual Transformation based on Unet configuration (KAC-Unet), is provided, and the medical image segmentation framework, namely, KAC-Unet, is used for segmenting the medical image, and is used for segmenting the medical image based on the KAC-Unet, the KAC-Unet, the KAC-Unet, the KAC-Unet, the KAC-Unet, the KAC-Unet, the KAC-Unet, the KAC-Unet, the KAC-Unet, the KAC-Unet, the KAC-Unet and the KAC-Unet. Specifically, firstly, an adaptive group strategy (AGS) is put forward, and the problem of information flow interruption is relieved by effectively keeping the consistency of the number of packet channels; then, a Shift Tokenized Kolmogorov Arnold network module (STKB) is provided, and a feature extraction module aims at enhancing the capability of capturing complex features in the medical image through flexible nonlinear transformation and shift operation.

Description

Technical Field

[0001] The invention relates to a medical image segmentation method based on an adaptive grouping strategy and Kolmogorov-Arnold Networks. Background Art

[0002] Medical image segmentation is of great significance for modern clinical diagnosis and therapeutic intervention. The accuracy and efficiency of segmentation greatly improve the quality of clinical decision-making and significantly optimize patient treatment outcomes. This technology enables clinicians to more accurately identify and locate lesion areas, promoting the development of highly personalized medical image segmentation for patients. Traditional medical image segmentation techniques include threshold segmentation, region growing algorithms, and edge detection algorithms. Although these techniques are simple and efficient, they are limited to manually performing specific tasks. Fortunately, with the rapid development of artificial intelligence, deep learning-based algorithms have made significant progress in medical image segmentation. This technology can segment anatomical structures or lesion areas, greatly enhancing medical technology and measurement accuracy. For example, Unet's unique encoder-decoder architecture has shown excellent performance in medical image segmentation. TransUnet combines Transformer with Unet and applies it to the field of medical image segmentation. It demonstrates the ability of domain transfer and provides valuable insights for subsequent research. nnUnet performs well in complex medical image segmentation tasks, especially multi-organ segmentation, through its innovative multi-scale feature fusion strategy. UNeXt adopts a grouping extraction strategy for high-dimensional feature extraction, which enhances the model's ability to recognize subtle textures in medical images. However, in existing image processing research, most methods use fixed-size and non-learnable convolution kernels for feature extraction. Fixed convolution limits the model's ability to represent complex medical images to a certain extent. In addition, although multi-layer feedforward neural networks learn feature representations of data by stacking simple linear transformations and nonlinear activation functions, this method still lacks the ability to generalize to different data features, resulting in limited adaptability in some medical image domains.

[0003] In summary, the challenges of current deep learning-based medical image segmentation methods are: (1) how to dynamically adopt image feature extraction strategies based on image features of different learning depths. (2) using learnable activation functions to learn more detailed features of medical images. Therefore, it is necessary to design a medical image segmentation model based on a dynamic grouping strategy and a nonlinear deep learning segmentation model with good generalization performance.

[0004] The performance of existing methods for medical image segmentation needs to be further improved. Traditional convolutional neural networks (CNNs) usually use fixed activation functions, linear layers, and channel grouping strategies for feature extraction. When applied to complex medical image textures, excessive stacking of linear layers limits the model's ability to capture fine-grained features. Fixed grouping methods usually lead to an imbalanced feature distribution between groups, resulting in the loss of key information. Summary of the invention

[0005] The purpose of the present invention is to overcome the defects of the prior art, provide a medical image segmentation method based on an adaptive grouping strategy and Kolmogorov-Arnold Networks, and propose a medical image segmentation framework, namely, Kolmogorov-Arnold Network (KAN) with the Adaptive Group Strategy and Contextual Transformerbased on Unet structure (KAC-Unet); specifically, firstly, an adaptive group strategy (AGS) is proposed, which alleviates the problem of information flow interruption by effectively maintaining the consistency of the number of grouping channels; then, a ShiftTokenized KolmogorovArnold network module (STKB) is proposed, and this feature extraction module aims to enhance the ability to capture complex features in medical images through flexible nonlinear transformations and shift operations.

[0006] To achieve the above-mentioned purpose, the technical scheme of the present invention is: a medical image segmentation method based on adaptive grouping strategy and Kolmogorov-Arnold Networks, and a medical image segmentation framework KAC-Unet is proposed; specifically, an adaptive group strategy AGS is proposed to alleviate the problem of information flow interruption by effectively maintaining the consistency of the number of grouping channels; and a ShiftTokenized Kolmogorov-Arnold network module STKB is proposed to enhance the ability to capture complex features in medical images through flexible nonlinear transformations and shift operations.

[0007] In one embodiment of the present invention, the method is specifically implemented as follows:

[0008] First, in the convolution stage, nonlinear convolution is used to extract low-dimensional features in the image;

[0009] Subsequently, a context-shifted attention mechanism is introduced in each skip connection module to strengthen feature interrelationships and integrate contextual information;

[0010] Next, the features are refined into high-dimensional information through a feature adaptive grouping displacement module based on AGS and STKB. The feature adaptive grouping displacement module enhances the focus on local information and extracts the nonlinear relationship of features in complex medical images through a series of nonlinear transformations and shift operations.

[0011] Finally, the decoder generates the segmentation result of the image.

[0012] In one embodiment of the present invention, AGS dynamically determines the number of groups according to the input feature dimension, ensures that the number of channels in each group is the same as the number of group channels, and realizes uniform channel allocation, which is specifically implemented as follows:

[0013]

[0014] x1,x2,x3,…,x N =AGroup(X) (2)

[0015] Where D represents the number of channels in the input feature map, C represents the number of channels in each group, N represents the number of groups, and x i Represents the feature information in each group, AGroup(X) represents adaptive grouping, and X represents the input feature.

[0016] In one embodiment of the present invention, after AGS groups the features, STKB performs displacement according to the axis of the channel of the features, which is specifically implemented as follows:

[0017] x s1 ,x s2 ,x s3 ,…,x sN =Shift(x1,x2,x3,…,x N ),

[0018] X shift =CAT(x s1 ,x s2 ,x s3 ,…,x sN ) (3)

[0019] Among them, x si Represents x i The result after shifting, X shift is x si The result after splicing, CAT(·) indicates the splicing operation, and Shift indicates the shift operation.

[0020] In one embodiment of the present invention, the specific implementation method of the decoder generating the segmentation result of the picture is as follows:

[0021] After the input feature X is processed by the feature adaptive grouping and shifting module, it is tokenized to generate tokens; then, the tokens are passed to the KAN network for further feature extraction and dimensionality reduction; finally, the reduced dimensionality features are passed to the deep convolutional layer, which is defined as follows:

[0022]

[0023] X Wshift =Shift W (X); T w =Tokenize(X Wshift ) (5)

[0024] Y=f(DWConv(KAN(T w ))) (6)

[0025] in Represents the learnable activation function of the i-th layer in the KAN network, with dimension n in ×n out ; T w represents the feature obtained by horizontal displacement projection, DWConv(·) represents the depthwise separable convolution, Tokenize represents the tokenization operation, and Shift W represents lateral displacement;

[0026] In order to enable the model to fully capture the spatial features in medical images, feature Y is first vertically displaced and tokenized, and then the tokenized results are input into the KAN network and deep convolution network and normalized. Next, the ReLU activation function is used to enhance the expressiveness of the features. On this basis, the features extracted in the width and height dimensions are input into the KAN network and deep convolution operation again to further extract and integrate the features. Finally, the initial features are combined with the processed features through the residual connection technology, as shown below:

[0027] Y Hshift =Shift H (Y); T H =Tokenize(Y Hshift ) (7)

[0028] Y1=Relu(BN(DWConv(KAN(T H )))) (8)

[0029] Y=T+DWConv(KAN(Y1)) (9)

[0030] Among them, BN(·) represents normalization, T represents the initial feature, Shift H Represents a vertical displacement operation.

[0031] In one embodiment of the present invention, the skip connection module extracts static and dynamic context information of the input features through a combination of 3×3 convolution and 1×1 convolution, and then fuses this information to output a feature representation containing rich context to enhance the performance of the visual recognition task, as shown below:

[0032]

[0033] W atten =Conv2d(Relu(BN(CAT(K,Q)))) (11)

[0034] Out = K + Sostmax (W atten )⊙V (12)

[0035]

[0036] Among them, X represents the output of the encoder at a certain layer, Q, K and V represent dynamic image representation, and W atten represents the dynamic attention weight, ⊙ represents the matrix multiplication operation, Softmax(·) converts the input score into the classification probability, and represents the output of the encoder and decoder at layer l-1, represents the concatenated result, BN(·) represents normalization, and Conv2d represents two-dimensional convolution.

[0037] In one embodiment of the present invention, the method uses a combination of binary cross entropy (BCE) and Dice loss to train the model.

[0038] In one embodiment of the present invention, in the predicted value The loss between y and the target value y is as follows:

[0039]

[0040] Among them, α and β represent the weights of the loss.

[0041] The present invention also provides a medical image segmentation system based on an adaptive grouping strategy and Kolmogorov-Arnold Networks, comprising a memory, a processor, and computer program instructions stored in the memory and capable of being executed by the processor. When the processor executes the computer program instructions, the method steps described above can be implemented.

[0042] The present invention also provides a computer-readable storage medium, on which computer program instructions that can be executed by a processor are stored. When the processor executes the computer program instructions, the method steps described above can be implemented.

[0043] Compared with the prior art, the present invention has the following beneficial effects: In order to improve the segmentation accuracy of medical images, an innovative segmentation method is proposed in the present invention, which aims to learn more detailed features of medical images through more nonlinear learnable functions and flexible adaptive adjustment mechanisms. In the early stage of the image processing process, the present invention first extracts low-dimensional features in the image through nonlinear convolution technology. Next, in order to strengthen the relationship between these features and integrate contextual information, the present invention introduces the context conversion attention mechanism in each jump connection layer. Subsequently, the feature extraction module based on the adaptive grouping displacement strategy proposed in the present invention is used to further refine these features into high-dimensional information. This module particularly strengthens the focus on local information and captures the complex nonlinear relationship of features in medical images through nonlinear transformation and displacement operations. Finally, the decoder integrates these high-dimensional features to generate accurate segmentation results of the image, completing the complete process from preliminary feature extraction to final segmentation output. The results are compared with several state-of-the-art methods on a real prostate MRI medical image dataset, proving the effectiveness of the method proposed in the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] Figure 1 This is a structural diagram of the method of the present invention.

[0045] Figure 2 The visualization results of the method proposed in this invention on three medical datasets. DETAILED DESCRIPTION

[0046] The technical solution of the present invention is described in detail below in conjunction with the accompanying drawings.

[0047] It should be noted that the following detailed descriptions are exemplary and are intended to provide further explanation of the present application. Unless otherwise specified, all technical and scientific terms used herein have the same meanings as those commonly understood by those skilled in the art to which the present application belongs.

[0048] It should be noted that the terms used herein are only for describing specific embodiments and are not intended to limit the exemplary embodiments according to the present application. As used herein, unless the context clearly indicates otherwise, the singular form is also intended to include the plural form. In addition, it should be understood that when the terms "comprise" and / or "include" are used in this specification, it indicates the presence of features, steps, operations, devices, components and / or combinations thereof.

[0049] The present invention proposes a medical image segmentation method based on an adaptive grouping strategy and Kolmogorov-Arnold Networks, and proposes a medical image segmentation framework KAC-Unet; specifically, an adaptive group strategy AGS is proposed to alleviate the problem of information flow interruption by effectively maintaining the consistency of the number of grouping channels; a ShiftTokenized Kolmogorov-Arnold network module STKB is also proposed to enhance the ability to capture complex features in medical images through flexible nonlinear transformations and shift operations.

[0050] 1. Method structure

[0051] like Figure 1 As shown, this embodiment provides a medical image segmentation method based on an adaptive grouping strategy and Kolmogorov-Arnold Networks, which is summarized as follows Figure 1 As shown. The method proposes a medical image segmentation framework KAC-Unet. First, in the convolution stage, the present invention uses nonlinear convolution to extract low-dimensional features in the image. Subsequently, the present invention introduces a context conversion attention mechanism in each jump connection layer to strengthen the relationship between features and integrate contextual information. Next, the features are refined into high-dimensional information through the feature extraction module based on the adaptive grouping shift strategy proposed in the present invention. This module enhances the attention to local information and extracts the nonlinear relationship of features in complex medical images through a series of nonlinear transformations and shift operations. Finally, the decoder generates the segmentation result of the image.

[0052] 1.1 Feature Adaptive Grouping Shift Module

[0053] The feature channel strategy aims to enhance the performance and stability of the model by dynamically adjusting the number of groups. Specifically, AGS dynamically determines the number of groups based on the input feature dimension, ensuring that the number of channels in each group is the same as the number of group channels, achieving a more uniform channel distribution. This flexible grouping strategy effectively alleviates the problem of insufficient or redundant feature extraction caused by unbalanced channel distribution within the group. It improves the accuracy and robustness of the model in medical image segmentation tasks. The fixed and regular channel division method makes the structure of the model clearer and easier to understand, helps to explain and analyze the working principle of the model, and provides a clear direction for further improvement and optimization. As shown below:

[0054]

[0055] x1,x2,x3,…,x N =AGroup(X) (2)

[0056] Where D represents the number of channels in the input feature map, C represents the number of channels in each group, N represents the number of groups, and x i represents the feature information in each group, and AGroup(X) represents adaptive grouping.

[0057] After grouping the features, the present invention performs displacement according to the axis of the feature channel. This helps the convolution module to focus on more local information in the subsequent feature extraction process, as shown below:

[0058] x s1 ,x s2 ,x s3 ,…,x sN =Shift(x1,x2,x3,…,x N ),

[0059] X shift =CAT(x s1 ,x s2 ,x s3 ,…,x sN ) (3)

[0060] Among them, x si Represents x i The result after shifting, X shift is x si The result after splicing, CAT(·) indicates the splicing operation, and Shift indicates the shift operation.

[0061] 1.2 Tokenized Nonlinear Feature Extraction Module

[0062] The present invention performs adaptive grouping and shifting operations on the input feature X, and tokenizes it to generate tokens. Then, the present invention passes these tokens to the KAN network (across width) for further feature extraction and dimensionality reduction. The KAN network has stronger feature capture and expression capabilities. It introduces a complex transformation mechanism based on B-splines, which can handle more subtle and complex image features than MLP. Finally, these features are passed to the deep convolutional layer, which further enhances the model's perception of local features and provides a richer and more accurate feature representation for the final image segmentation task, which is defined as follows:

[0063]

[0064] X Wshift =Shift W (X); T w =Tokenize(X Wshift )(5)

[0065] Y=f(DWConv(KAN(T w))) (6)

[0066] in Represents the learnable activation function of the i-th layer in the KAN network, with dimension n in ×n out Shift W represents the lateral displacement, T w It represents the feature obtained by horizontal displacement projection, DWConv(·) represents the depthwise separable convolution, and Tokenize represents the tokenization operation.

[0067] In order to enable the model to fully capture the spatial features in medical images, we first vertically shift and tokenize feature Y, and then input these tokenized results into the KAN network and deep convolution network and normalize them. Next, we use the ReLU activation function to enhance the expressiveness of the features. On this basis, we input the features extracted in the width and height dimensions into the KAN network and deep convolution operations again to further extract and integrate the features. Finally, through the residual connection, we combine the initial features with the processed features to enhance the segmentation accuracy of the model. As shown below:

[0068] Y Hshift =Shift H (Y); T H =Tokenize(Y Hshift )(7)

[0069] Y1=Relu(BN(DWConv(KAN(T H )))) (8)

[0070] Y=T+DWConv(KAN(Y1)) (9)

[0071] Among them, BN(·) represents normalization, T represents the initial feature, Shift H Indicates vertical displacement.

[0072] 1.3 Skip Connection Module

[0073] In the skip connection, the present invention extracts static and dynamic context information of the input features through a combination of 3×3 convolution and 1×1 convolution. Then, this information is fused to output a feature representation containing rich context to enhance the performance of visual recognition tasks. As shown below:

[0074]

[0075] W atten =Conv2d(Relu(BN(CAT(K,Q)))) (11)

[0076] Out = K + Sostmax (W atten )⊙V (12)

[0077] Where X represents the output of the encoder at a certain layer, Q, K and V represent dynamic image representations, and W atten represents the dynamic attention weight, ⊙ represents the matrix multiplication operation, Softmax(·) converts the input score into a classification probability, BN(·) represents normalization, and Conv2d represents a two-dimensional convolution.

[0078]

[0079] in and Represents the output of the encoder and decoder at layer l-1. Indicates the result after splicing.

[0080] 1.4 Loss Function

[0081] The present invention uses a combination of binary cross entropy (BCE) and Dice loss to train the model. The loss between y and the target value y is as follows:

[0082]

[0083] Where α and β represent the weights of the loss.

[0084] 2. Experiment

[0085] In order to ensure the applicability of the model, the present invention conducts experiments on three public datasets, including the Prostate dataset, the BUSI dataset and the ISIC dataset.

[0086] The Prostate dataset is a merger of three different prostate T2-weighted MRI datasets: NCIISBI 2013, I2CVB, and PROMISE 2012, including a total of 761 medical images for prostate segmentation. The size of the original images ranges between 369×369 and 389×389 pixels. The present invention standardizes all images to a uniform size of 256×256 pixels.

[0087] The BUSI dataset includes a wide range of ultrasound images, including normal, benign, and malignant breast cancer cases, and their corresponding segmentation maps. We selected 647 medical images, including both benign and malignant breast tumors. To ensure consistency, all images were resized from 557×473 pixels (approximately) to 256×256 pixels.

[0088] The ISIC 2018 dataset is a comprehensive collection of dermoscopic images, released by the International Skin Imaging Collaboration. The dataset includes 2,549 high-resolution medical images, including benign and malignant skin lesions. The original image resolution ranges from 640×480 to 3008×2000 pixels. The present invention standardizes the resolution of all images to a consistent 512×512 pixels.

[0089] All methods are implemented using Python 3.9 and PyTorch 2.0.1. Training and testing are performed on a 12-core machine equipped with an Intel Xeon Silver 4214R 2.4GHz CPU (128GB RAM) and an NVIDIA GeForce RTX 3090 GPU with 24GB VRAM. During the training phase, the present invention uses the Adam optimizer with a learning rate of 1e-3 and a momentum of 0.9. The present invention also uses a cosine annealing learning rate scheduler with a minimum learning rate of 1e-5. The batch size is set to 8. The present invention trains KAC-Unet for a total of 400 rounds. The dataset is randomly divided into training and test sets 3 times in a ratio of 8:2, and the results are averaged and the variance is reported.

[0090] In order to ensure a comprehensive evaluation, the present invention uses the average dice coefficient (Dice) and the average intersection over union (IoU) as evaluation indicators. The IoU indicator mainly evaluates the overlap between the predicted segmentation and the true segmentation label, and is defined as follows:

[0091]

[0092] TP is a true positive, FP is a false positive, and FN is a false negative.

[0093] The Dice coefficient, also known as the F1 score, is a metric used to measure the similarity between two sets in binary classification problems. Compared with IoU, Dice is more sensitive to small areas. Its definition is as follows:

[0094]

[0095] 2.1 Qualitative comparison

[0096] In order to qualitatively compare the lesion segmentation effects, the segmentation results were visualized. Figure 2 The results of 6 medical image segmentation on three datasets, BUSI, ISIC, and Prostate, are shown. These localization results are almost similar to the baseline.

[0097] 2.2 Quantitative comparison

[0098] Table 1 Quantitative comparison of the proposed method with the most advanced algorithm on real data sets

[0099]

[0100] Table 1 shows the quantitative experimental results of KAC-Unet on three datasets, and the best results are in bold. The experimental results show that KAC-Unet outperforms all other algorithms in these medical image segmentation datasets. In BUSI, the IoU and Dice indicators of KAC-Unet reached 69.29% and 77.57%, respectively. Compared with the second and third methods U-KAN and UNeXt, the IoU indicator of KAC-Unet is 3.11% and 1.2% higher, and the Dice indicator is 2.63% and 0.92% higher, respectively. The reason is that the present invention adopts flexible and variable feature extraction strategies and pays more attention to extracting detailed information of the image. On the ISIC dataset, the present invention also obtains the highest IoU value of 82.34% and Dice value of 89.40%. Similarly, on the Prostate dataset, the present invention also achieves the best indicators. VM-MaMba and EGE-Unet tend to prioritize global features at the expense of local pathological details, which limits their segmentation capabilities, while the present invention solves the above problems through a multi-stage, flexible feature extraction strategy.

[0101] The present invention also provides a medical image segmentation system based on an adaptive grouping strategy and Kolmogorov-Arnold Networks, comprising a memory, a processor, and computer program instructions stored in the memory and capable of being executed by the processor. When the processor executes the computer program instructions, the method steps described above can be implemented.

[0102] The present invention also provides a computer-readable storage medium, on which computer program instructions that can be executed by a processor are stored. When the processor executes the computer program instructions, the method steps described above can be implemented.

[0103] Those skilled in the art will appreciate that the embodiments of the present application may be provided as methods, systems, or computer program products. Therefore, the present application may adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment in combination with software and hardware. Moreover, the present application may adopt the form of a computer program product implemented in one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) that include computer-usable program code.

[0104] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0105] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0106] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 A step that specifies a function in one or more boxes.

[0107] The above is only a preferred embodiment of the present invention, and does not limit the present invention in other forms. Any technician familiar with the profession may use the above disclosed technical content to change or modify it into an equivalent embodiment with equivalent changes. However, any simple modification, equivalent change and modification made to the above embodiment according to the technical essence of the present invention without departing from the technical solution of the present invention still belongs to the protection scope of the technical solution of the present invention.

Claims

1. A medical image segmentation method based on adaptive grouping strategy and Kolmogorov-Arnold Networks, characterized in that: A medical image segmentation framework KAC-Unet is proposed. Specifically, an adaptive group strategy AGS is proposed to alleviate the problem of information flow interruption by effectively maintaining the consistency of the number of grouping channels. A Shift Tokenized Kolmogorov Arnold network module STKB is also proposed to enhance the ability to capture complex features in medical images through flexible nonlinear transformations and shift operations.

2. The medical image segmentation method based on adaptive grouping strategy and Kolmogorov-Arnold Networks according to claim 1, characterized in that: The method is specifically implemented as follows: First, in the convolution stage, nonlinear convolution is used to extract low-dimensional features in the image; Subsequently, a context-shifted attention mechanism is introduced in each skip connection module to strengthen feature interrelationships and integrate contextual information; Next, the features are refined into high-dimensional information through a feature adaptive grouping displacement module based on AGS and STKB. The feature adaptive grouping displacement module enhances the focus on local information and extracts the nonlinear relationship of features in complex medical images through a series of nonlinear transformations and shift operations. Finally, the decoder generates the segmentation result of the image.

3. A medical image segmentation method based on adaptive grouping strategy and Kolmogorov-Arnold Networks according to claim 1 or 2, characterized in that: AGS dynamically determines the number of groups based on the input feature dimension, ensuring that the number of channels in each group is the same as the number of group channels, achieving uniform channel distribution. The specific implementation is as follows: x1,x2,x3,…,x N =AGroup (X) (2) Where D represents the number of channels in the input feature map, C represents the number of channels in each group, N represents the number of groups, and x i Represents the feature information in each group, AGroup(X) represents adaptive grouping, and X represents the input feature.

4. The medical image segmentation method based on adaptive grouping strategy and Kolmogorov-Arnold Networks according to claim 3, characterized in that: After AGS groups the features, STKB shifts the axes of the feature channels. The specific implementation is as follows: x s1 ,x s2 ,x s3 ,…,x sN =Shift(x1,x2,x3,…,x N ), X shift =CAT(x s1 ,x s2 ,x s3 ,…,x sN ) (3) Among them, x si Represents x i The result after shifting, X shift is x si The result after splicing, CAT(·) indicates the splicing operation, and Shift indicates the shift operation.

5. The medical image segmentation method based on adaptive grouping strategy and Kolmogorov-Arnold Networks according to claim 4, characterized in that: The specific implementation method of the decoder to generate the segmentation result of the picture is as follows: After the input feature X is processed by the feature adaptive grouping and shifting module, it is tokenized to generate tokens; then, the tokens are passed to the KAN network for further feature extraction and dimensionality reduction; finally, the reduced dimensionality features are passed to the deep convolutional layer, which is defined as follows: X Wshift =Shift W (X);T w =Tokenize(X Wshift )(5) Y=f(DWConv(KAN ( T w ))) (6) in Represents the learnable activation function of the i-th layer in the KAN network, with dimension n in ×n out ; T w represents the feature obtained by horizontal displacement projection, Tokenize represents the tokenization operation, DWConv(·) represents the depthwise separable convolution, Shift W represents lateral displacement; In order to enable the model to fully capture the spatial features in medical images, feature Y is first vertically displaced and tokenized, and then the tokenized results are input into the KAN network and deep convolution network and normalized. Next, the ReLU activation function is used to enhance the expressiveness of the features. On this basis, the features extracted in the width and height dimensions are input into the KAN network and deep convolution operation again to further extract and integrate the features. Finally, the initial features are combined with the processed features through the residual connection technology, as shown below: Y Hshift =Shift H (Y);T H =Tokenize(Y Hshift ) (7) Y1=Relu)BN)DWConv ) WHERE(T H )))) (8) Y=T+DWConv(KAN(Y1)) (9) Among them, BN(·) represents normalization, T represents the initial feature, Shift H Represents a vertical displacement operation.

6. According to the method for medical image segmentation based on adaptive grouping strategy and Kolmogorov-Arnold Networks in claim 2, the skip connection module extracts static and dynamic context information of the input features through a combination of 3×3 convolution and 1×1 convolution, and then fuses the information to output a feature representation containing rich context to enhance the performance of the visual recognition task, as shown below: W atten =Conv2d(Relu(BN(CAT(K,Q)))) (11) Out=K+Sostmax(W atten )⊙V (12) in, X represents the output of the encoder at a certain layer, Q, K and V represent dynamic image representation, and W atten represents the dynamic attention weight, ⊙ represents the matrix multiplication operation, Softmax(·) converts the input score into the classification probability, and represents the output of the encoder and decoder at layer l-1, It represents the result after concatenation, BN represents normalization, and Conv2d represents a two-dimensional convolution operation.

7. The medical image segmentation method based on adaptive grouping strategy and Kolmogorov-Arnold Networks according to claim 1, characterized in that: This method uses a combination of binary cross entropy (BCE) and Dice loss to train the model.

8. The medical image segmentation method based on adaptive grouping strategy and Kolmogorov-Arnold Networks according to claim 7, characterized in that: In the predicted value The loss between y and the target value y is as follows: Among them, α and β represent the weights of the loss.

9. A medical image segmentation system based on adaptive grouping strategy and Kolmogorov-Arnold Networks, characterized in that: The method comprises a memory, a processor and computer program instructions stored in the memory and capable of being executed by the processor. When the processor executes the computer program instructions, the method steps as claimed in any one of claims 1 to 8 can be implemented.

10. A computer-readable storage medium having stored thereon computer program instructions that can be executed by a processor, and when the processor executes the computer program instructions, the method steps according to any one of claims 1 to 8 can be implemented.

Images