Nonlinear medical image segmentation method

Abstract

The invention aims to provide a non-linear medical image segmentation method, which comprises the following steps of: A, constructing a neural network which comprises a front enhancement network, a coding network and a decoding network; the front enhanced network comprises an elliptic convolution module ELPCB and a mixed pooling module MPB; the coding network and the decoding network are both constructed based on an FKC-VSS module; b, an original image is firstly input into a front enhancement network for processing, in the front enhancement network, an input result is divided into four paths, and the first path is processed by an elliptic convolution module ELPCB; the second path is subjected to convolution processing; the third path is subjected to convolution processing; the fourth path is processed by a mixed pooling module MPB; the four paths of processing results are spliced through a Concat function, and then are sequentially processed through a Shuffle function and convolution processing to obtain a front enhanced junction; and C, after the coding network codes the pre-enhancement result, the pre-enhancement result is input into a decoding network to be processed, and then a final output result is obtained. According to the method, noise can be suppressed, brightness can be balanced, complex and changeable shape information can be extracted, and pixel-level segmentation prediction can be generated.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of computer image processing, in particular to a nonlinear medical image segmentation method. BACKGROUND

[0002] In the medical image segmentation task, effectively suppressing noise and balancing brightness is an essential part of achieving accurate medical image segmentation. To address this issue, current methods mainly use multi-scale feature aggregation strategies, combining shallow information with deep semantic information, using deep semantic information to suppress noise and balance brightness at the shallow level, and have achieved good results. In addition, some advanced convolutional neural network architectures have emerged, which further improve the accuracy and stability of medical image segmentation by improving network structure and optimizing loss function. Experimental results show that this method has achieved significant segmentation results in multiple medical segmentation scenarios, providing strong technical support for automated diagnosis and treatment.

[0003] Although the current end-to-end medical image segmentation method has achieved certain results, the existing method generally uses multi-scale feature aggregation to deal with noise interference and brightness imbalance, but in the process of fusing multi-scale features, noise may be introduced into deep semantic features, resulting in poor segmentation results. SUMMARY

[0004] The present application aims to provide a nonlinear medical image segmentation method that can effectively suppress noise, balance brightness, and extract complex and variable shape information to generate pixel-level segmentation predictions.

[0005] The technical solution of the present application is as follows: The nonlinear medical image segmentation method comprises the following steps: A. Constructing a neural network, the neural network comprising a pre-enhancement network, an encoding network, and a decoding network; The pre-enhancement network comprises an elliptical convolution module ELPCB and a hybrid pooling module MPB; the encoding network and the decoding network are both constructed based on the FKC-VSS module; B. The original image is first input into the pre-enhancement network for processing, and the processing process is as follows: In the pre-enhancement network, the input result is divided into four paths, the first path is processed by the elliptical convolution module ELPCB to obtain the first path processing result; the second path is processed by convolution to obtain the second path processing result; the third path is processed by convolution to obtain the third path processing result; the fourth path is processed by the hybrid pooling module MPB to obtain the fourth path processing result; the four path processing results are concatenated by the Concat function, then processed by the Shuffle function and convolution to obtain the pre-enhancement result, which is input into the b encoding network; C, the encoding network encodes the pre-enhanced result, and inputs the decoding network to obtain the final output result after processing by the decoding network.

[0006] The processing process in the elliptical convolution module ELPCB is as follows: The input result is divided into three paths, the first path is processed by X_ELPC elliptical convolution, the obtained X_ELPC elliptical convolution result is multiplied by the scaling factor X_Scale to obtain the first path result; the second path is processed by convolution and Scale function in turn to obtain the second path result; the third path is processed by Y_ELPC elliptical convolution, the obtained Y_ELPC elliptical convolution result is multiplied by the scaling factor Y_Scale to obtain the third path result; The first path result is subtracted from the second path result, and the obtained result is multiplied by the input result to obtain the first multiplication result; the third path result is subtracted from the second path result, and the obtained result is multiplied by the input result to obtain the second multiplication result; the first multiplication result and the second multiplication result are added, multiplied by the coefficient O_Scale, and then processed by convolution to obtain the output result.

[0007] The X_ELPC elliptical convolution and the Y_ELPC elliptical convolution are constructed as follows: Based on the standard elliptical equation d = ((x / a)^2) + ((y / b)^2), the major axis and the minor axis of the X_ELPC elliptical convolution and the Y_ELPC elliptical convolution are set respectively; wherein the major axis of the X_ELPC elliptical convolution is located on the X axis, and the minor axis is located on the Y axis; the major axis of the Y_ELPC elliptical convolution is located on the Y axis, and the minor axis is located on the X axis; The processing process of the X_ELPC elliptical convolution and the Y_ELPC elliptical convolution is as follows: The coordinates (x, y) of each element of the input result are input into the standard elliptical equation for calculation, if d is less than or equal to 1, it is an element in the elliptical receptive field, if d is greater than 1, it is not an element in the elliptical receptive field, and is set to 0.

[0008] The processing process of the mixed pooling module MPB is as follows: The input result is processed by convolution to obtain a convolution result, the convolution result is divided into three paths, the first path is processed by maximum pooling, the second path is processed by average pooling, and the third path is processed by minimum pooling, the three pooling results are added, and then processed by two CBR modules, up-sampling and convolution to obtain the output result.

[0009] The processing process in the encoding network is as follows: The input result is first up-sampled by the Patch Embedding module, and then sequentially processed by the first FKVSL module, the Patch Merging module, the second FKVSL module, the Patch Merging module, the third FKVSL module, the Patch Merging module, and the fourth FKVSL module; wherein the first FKVSL module processing result, the second FKVSL module processing result, the third FKVSL module processing result, and the fourth FKVSL module processing result are respectively input into the decoding network; The processing process in the Patch Merging module is as follows: The input result is first processed by the Linear function to double the channel number, and then processed by the LayerNorm function to obtain the output result.

[0010] The processing process in the decoding network is as follows: The fourth FKVSL module processing result is input into the fifth FKVSL module for processing to obtain the fifth FKVSL module processing result; the fifth FKVSL module processing result is added to the third FKVSL module processing result, and then sequentially processed by the Patch Expanding module and the sixth FKVSL module to obtain the sixth FKVSL module processing result; the sixth FKVSL module processing result is added to the second FKVSL module processing result, and then sequentially processed by the Patch Expanding module and the seventh FKVSL module to obtain the seventh FKVSL module processing result; the seventh FKVSL module processing result is added to the first FKVSL module processing result, and then sequentially processed by the Patch Expanding module and the eighth FKVSL module to obtain the eighth FKVSL module processing result; the eighth FKVSL module processing result is processed by the Final Project module to obtain the output result. The processing process in the Patch Expanding module is as follows: The input result is first processed by the Linear function to halve the channel number, and then processed by the LayerNorm function to obtain the output result. The processing process in the Final Project module is as follows: the input result is first processed by the Linear function to quarter the channel number, and then processed by the LayerNorm function to obtain the output result.

[0011] Each FKVSL module has the same structure, and the processing process is as follows: The input result is sequentially processed by three FKC-VSS modules, and the obtained result is processed by the Sample function to obtain the output result. The processing process in the FKC-VSS module is as follows: After the input result is processed by the Chunk function, it is divided into two paths, one path is processed by the FKC module, and the other path is processed by the VSS module, the processing results of the two paths are spliced by the Concat function, then processed by the Shuffle function, the obtained result is added to the input result, and the output result is obtained.

[0012] The processing process in the FKC module is as follows: After the input result is processed by the Chunk function, it is divided into four paths, the first path is processed by the FastKAN convolution of the 3*3 convolution kernel and the FastKAN convolution of the 1*1 convolution kernel in turn, to obtain the first path result; the second path is processed by the FastKAN convolution of the 3*3 convolution kernel and the FastKAN convolution of the 1*1 convolution kernel in turn, to obtain the second path result; the third path is processed by the FastKAN convolution of the 3*3 convolution kernel and the FastKAN convolution of the 1*1 convolution kernel in turn, to obtain the third path result; the fourth path is processed by the FastKAN convolution of the 3*3 convolution kernel and the FastKAN convolution of the 1*1 convolution kernel in turn, to obtain the fourth path result; The four path results are spliced by the Concat function, then processed by the Shuffle function, to obtain the output result.

[0013] The processing process in the VSS module is as follows: After the input result is processed by the LN function, the obtained result is divided into two paths, the first path is processed by the Linear function and the SiLU function in turn, to obtain the first path result; the second path is processed by the Linear function, the depth separable convolution, the SiLU function, the SS2D function and the LN function in turn, to obtain the second path result; the first path result and the second path result are multiplied, then processed by the Linear function, and the obtained result is added to the input result, to obtain the output result.

[0014] The beneficial effects of the present application are as follows: The present application proposes a novel general segmentation model based on FastKAN and SSM, the pre-enhancement network of which can effectively suppress noise and balance brightness in the shallow layer, meanwhile, the long-distance dependence modeling capability of SSM is utilized to capture global information, combined with the nonlinear modeling capability of FastKAN, so that the model can understand the continuously changing shape information in the axial slice to enhance the boundary recognition capability.

[0015] In order to more accurately perform pixel-level segmentation, the present application designs a pre-enhancement network, which extracts shallow features of the output image, captures texture, edge, local and structural feature information, to suppress noise and balance brightness, and provides a feature map with a large amount of detail information for the following encoding network.

[0016] In the encoder-decoder network designed in this invention, the core module of FKC-VSS is used to extract feature information from complex medical data. Specifically, FastKAN convolutional dimensionality reduction mapping and nonlinear modeling are employed to enable the model to understand complex and varied shape information, extracting local detail information and optimizing boundary regions. The global modeling capability of Visual SSM (VSSM) is used to capture contextual information, allowing the model to focus on lesion areas while understanding structural information.

[0017] The method proposed in this invention improves segmentation performance by employing hybrid pooling, elliptical receptive fields, and function approximation to achieve noise suppression, dimensionality reduction mapping, brightness balancing, and nonlinear approximation. Unlike typical medical image segmentation methods, this invention proposes a nonlinear medical image segmentation method with brightness balancing and noise suppression capabilities. This method effectively suppresses noise, balances brightness, and extracts complex and variable shape information, generating pixel-level segmentation predictions. Experiments demonstrate that this structure effectively improves network performance. Attached Figure Description

[0018] Figure 1 A schematic diagram of the neural network structure in Example 1; Figure 2 This is a schematic diagram of the receptive field of X_ELPC elliptical convolution and Y_ELPC elliptical convolution in Example 1; Figure 3 This is a schematic diagram of the pre-reinforcement network in Example 1; Figure 4 This is a schematic diagram of the FKVSL module in Example 1; Figure 5 A comparison diagram showing the effects of the segmentation method provided in Example 1 and the segmentation method in Reference 1. Detailed Implementation

[0019] The present invention will now be described in detail with reference to the accompanying drawings and embodiments. Example 1

[0020] A nonlinear medical image segmentation method includes the following steps: A. Construct a neural network, such as Figure 1 As shown, the neural network includes a pre-enhancement network, an encoding network, and a decoding network; The aforementioned pre-enhancement network includes an elliptic convolution module (ELPCB) and a hybrid pooling module (MPB); both the encoding and decoding networks are constructed based on the FKC-VSS module. B. The original image is first input into the pre-enhancement network for processing, and the processing procedure is as follows: like Figure 3As shown, in the pre-enhancement network, the input results are divided into four paths. The first path is processed by the elliptic convolution module ELPCB to obtain the first processing result; the second path is processed by convolution to obtain the second processing result; the third path is processed by convolution to obtain the third processing result; and the fourth path is processed by the hybrid pooling module MPB to obtain the fourth processing result. The four processing results are concatenated by the Concat function and then sequentially processed by the Shuffle function and convolution to obtain the pre-enhancement result. The pre-enhancement result is then input into the b-encoding network. C. After the encoding network encodes the pre-enhancement result, it is input into the decoding network, which processes it to obtain the final output result.

[0021] like Figure 3 As shown, the processing procedure in the elliptic convolution module ELPCB is as follows: The input results are divided into three paths. The first path is processed by X_ELPC elliptical convolution. The X_ELPC elliptical convolution result is multiplied by the scaling factor X_Scale to obtain the first path result. The second path is processed by convolution and the scaling function in sequence to obtain the second path result. The third path is processed by Y_ELPC elliptical convolution. The Y_ELPC elliptical convolution result is multiplied by the scaling factor Y_Scale to obtain the third path result. Subtracting the first result from the second result, multiplying the result by the input result to obtain the first multiplication result; subtracting the third result from the second result, multiplying the result by the input result to obtain the second multiplication result; adding the first multiplication result and the second multiplication result, multiplying by the coefficient O_Scale, and then performing convolution to obtain the output result.

[0022] The X_ELPC elliptical convolution and Y_ELPC elliptical convolution are constructed as follows: Based on the standard elliptic equation d=((x / a)^2) +((y / b)^2), the major and minor axes of X_ELPC elliptic convolution and Y_ELPC elliptic convolution are set respectively; where the major axis of X_ELPC elliptic convolution is located on the X-axis and the minor axis is located on the Y-axis; the major axis of Y_ELPC elliptic convolution is located on the Y-axis and the minor axis is located on the X-axis. The processing steps for X_ELPC elliptical convolution and Y_ELPC elliptical convolution are as follows: Input the coordinates (x, y) of each element in the input result into the standard ellipse equation for calculation. If d is less than or equal to 1, it is an element within the ellipse receptive field; if d is greater than 1, it is not an element within the ellipse receptive field and is set to 0.

[0023] The effective receptive fields of X_ELPC elliptic convolution and Y_ELPC elliptic convolution are as follows: Figure 2As shown, the left image represents the effective receptive field of the X_ELPC elliptical convolution, and the right image represents the effective receptive field of the Y_ELPC elliptical convolution. The elliptical regions are the effective receptive fields, while other locations are ineffective receptive fields with a weight of 0. The entire convolution kernel size is 21px, with the major axis of the ellipse being 10px and the minor axis being 7px.

[0024] like Figure 3 As shown, the MPB (Multi-Pool Pooling) module's processing procedure is as follows: The input result is convolved to obtain the convolution result, which is divided into three paths. The first path is processed by max pooling, the second path is processed by average pooling, and the third path is processed by min pooling. The three pooling results are added together and then processed by two CBR modules, upsampling, and convolution to obtain the output result.

[0025] The processing procedure in the aforementioned coding network is as follows: The input result is first upsampled by the Patch Embedding module, and then processed sequentially by the first FKVSL module, the PatchMerging module, the second FKVSL module, the Patch Merging module, the third FKVSL module, the Patch Merging module, and the fourth FKVSL module. The processing results of the first FKVSL module, the second FKVSL module, the third FKVSL module, and the fourth FKVSL module are then input into the decoding network. The processing procedure in the Patch Merging module is as follows: The input result is first processed by the Linear function to double the number of channels, and then processed by the LayerNorm function to obtain the output result.

[0026] The processing procedure in the decoding network is as follows: The result processed by the fourth FKVSL module is input into the fifth FKVSL module for further processing, resulting in the fifth FKVSL module's result. This result is then added to the result processed by the third FKVSL module, and subsequently processed by the Patch Expanding module and the sixth FKVSL module, resulting in the sixth FKVSL module's result. This result is then added to the result processed by the second FKVSL module, and subsequently processed by the Patch Expanding module and the seventh FKVSL module, resulting in the seventh FKVSL module's result. This result is then added to the result processed by the first FKVSL module, and subsequently processed by the Patch Expanding module and the eighth FKVSL module, resulting in the eighth FKVSL module's result. Finally, this eighth FKVSL module's result is processed by the Final Project module to obtain the output result. The processing procedure in the Patch Expanding module is as follows: The input result is first processed by the Linear function to reduce the number of channels to half of the original number, and then processed by the LayerNorm function to obtain the output result. The processing procedure in the Final Project module is as follows: the input result is first processed by the Linear function to reduce the number of channels to one-quarter of the original number, and then processed by the LayerNorm function to obtain the output result.

[0027] like Figure 4 As shown, all FKVSL modules have the same structure, and their processing procedure is as follows: The input results are processed sequentially through three FKC-VSS modules, and the results are then processed by the Sample function to obtain the output results. The processing procedure in the FKC-VSS module is as follows: The input result is processed by the Chunk function and then split into two paths. One path is processed by the FKC module, and the other path is processed by the VSS module. The two processing results are concatenated by the Concat function, then processed by the Shuffle function. The result is added to the input result to obtain the output result.

[0028] The processing procedure in the FKC module is as follows: The input result is processed by the Chunk function and divided into four paths. The first path is processed by FastKAN convolution with a 3*3 kernel and FastKAN convolution with a 1*1 kernel, resulting in the first path result. The second path is processed by FastKAN convolution with a 3*3 kernel and FastKAN convolution with a 1*1 kernel, resulting in the second path result. The third path is processed by FastKAN convolution with a 3*3 kernel and FastKAN convolution with a 1*1 kernel, resulting in the third path result. The fourth path is processed by FastKAN convolution with a 3*3 kernel and FastKAN convolution with a 1*1 kernel, resulting in the fourth path result. The four results are concatenated using the Concat function, then processed by the Shuffle function to obtain the output result.

[0029] The processing procedure in the VSS module is as follows: After the input result is processed by the LN function, the result is divided into two paths. The first path is processed by the Linear function and the SiLU function in sequence to obtain the first path result. The second path is processed by the Linear function, the depthwise separable convolution, the SiLU function, the SS2D function, and the LN function in sequence to obtain the second path result. The first path result and the second path result are multiplied, processed by the Linear function, and the result is added to the input result to obtain the output result. Example 2

[0030] For the quantitative performance evaluation of the final segmentation map, we adopted the same performance measurement standards as in Reference 1, namely, Dice similarity coefficient (Dice), Jaccard index (JI), accuracy (ACC), sensitivity (Sen), and specificity (Spe) as evaluation indicators for the skin lesion segmentation model. The specific evaluation is shown in the following formula.

[0031] (1) (2) (3) (4) Wherein, TP (true positive) represents the number of correctly segmented skin lesion pixels, TN (true negative) represents the number of correctly segmented background pixels, FP (false positive) represents the incorrectly predicted background pixels as skin lesion pixels, and FN (false negative) represents the incorrectly predicted skin lesion pixels as background pixels.

[0032] Document 1: Li C, Liu The parameters used in Reference 1, as in its original text, are guaranteed to be optimal for the model. Since our structure must include a decoding network... Figure 5 The table shows the ground truth segmentation maps and the segmentation maps detected by the method in Reference 1 for medical images selected from the ISIC2017 dermoscopy dataset and the ClinicDB colorectal polyp dataset. Table 1 shows a quantitative comparison of the ground truth segmentation maps and the segmentation maps detected by the method in Reference 1 for medical images selected from the ISIC2017 dermoscopy dataset and the ClinicDB colorectal polyp dataset. The optimal segmentation detected by the method in Example 1 is also shown.

[0033] Based on the experimental results, the segmentation method in Example 1 is superior to the segmentation method in Reference 1.

[0034] Table 1 Quantitative data from ISIC17 Method Dice mIoU Sen Spe U-KAN 87.98 78.55 86.41 97.98 ANNet (ours) 88.79 79.84 88.63 97.78 Table 2 Quantitative data from ClinicDB Method Dice mIoU Sen Spe U-KAN 91.80 84.84 92.22 99.22 ANNet (ours) 92.47 86.00 93.76 99.20

Claims

1. A nonlinear medical image segmentation method, characterized in that, Includes the following steps: A. Construct a neural network, which includes a pre-amplification network, an encoding network, and a decoding network; The aforementioned pre-enhancement network includes an elliptic convolution module (ELPCB) and a hybrid pooling module (MPB); both the encoding and decoding networks are constructed based on the FKC-VSS module. B. The original image is first input into the pre-enhancement network for processing, and the processing procedure is as follows: In the pre-enhancement network, the input results are divided into four paths. The first path is processed by the elliptic convolution module ELPCB to obtain the first path processing result; the second path is processed by convolution to obtain the second path processing result; and the third path is processed by convolution to obtain the third path processing result. The fourth path is processed by the hybrid pooling module MPB to obtain the fourth path processing result; The four processing results are concatenated by the Concat function, and then sequentially processed by the Shuffle function and convolution to obtain the pre-enhancement result, which is then input into the b-encoding network. C. After the encoding network encodes the pre-enhancement result, it is input into the decoding network, which processes it to obtain the final output result.

2. The nonlinear medical image segmentation method as described in claim 1, characterized in that: The processing procedure in the elliptic convolution module ELPCB is as follows: The input results are divided into three paths. The first path is processed by X_ELPC elliptical convolution. The X_ELPC elliptical convolution result is multiplied by the scaling factor X_Scale to obtain the first path result. The second path is processed by convolution and the scaling function in sequence to obtain the second path result. The third path is processed by Y_ELPC elliptical convolution. The Y_ELPC elliptical convolution result is multiplied by the scaling factor Y_Scale to obtain the third path result. Subtracting the first result from the second result, multiplying the result by the input result to obtain the first multiplication result; subtracting the third result from the second result, multiplying the result by the input result to obtain the second multiplication result; adding the first multiplication result and the second multiplication result, multiplying by the coefficient O_Scale, and then performing convolution to obtain the output result.

3. The nonlinear medical image segmentation method as described in claim 2, characterized in that: The X_ELPC elliptical convolution and Y_ELPC elliptical convolution are constructed as follows: Based on the standard elliptic equation d=((x / a)^2) +((y / b)^2), the major and minor axes of X_ELPC elliptic convolution and Y_ELPC elliptic convolution are set respectively; where the major axis of X_ELPC elliptic convolution is located on the X-axis and the minor axis is located on the Y-axis; the major axis of Y_ELPC elliptic convolution is located on the Y-axis and the minor axis is located on the X-axis. The processing steps for X_ELPC elliptical convolution and Y_ELPC elliptical convolution are as follows: Input the coordinates (x, y) of each element in the input result into the standard ellipse equation for calculation. If d is less than or equal to 1, it is an element within the ellipse receptive field; if d is greater than 1, it is not an element within the ellipse receptive field and is set to 0.

4. The nonlinear medical image segmentation method as described in claim 1, characterized in that: The hybrid pooling module (MPB) processes the following steps: The input result is convolved to obtain the convolution result, which is divided into three paths. The first path is processed by max pooling, the second path is processed by average pooling, and the third path is processed by min pooling. The three pooling results are added together and then processed by two CBR modules, upsampling, and convolution to obtain the output result.

5. The nonlinear medical image segmentation method as described in claim 1, characterized in that: The processing procedure in the aforementioned coding network is as follows: The input result is first upsampled by the Patch Embedding module, and then processed sequentially by the first FKVSL module, the PatchMerging module, the second FKVSL module, the Patch Merging module, the third FKVSL module, the Patch Merging module, and the fourth FKVSL module. The processing results of the first FKVSL module, the second FKVSL module, the third FKVSL module, and the fourth FKVSL module are then input into the decoding network. The processing procedure in the Patch Merging module is as follows: The input result is first processed by the Linear function to double the number of channels, and then processed by the LayerNorm function to obtain the output result.

6. The nonlinear medical image segmentation method as described in claim 4, characterized in that: The processing procedure in the decoding network is as follows: The result processed by the fourth FKVSL module is input into the fifth FKVSL module for further processing, resulting in the fifth FKVSL module's result. This result is then added to the result processed by the third FKVSL module, and subsequently processed by the Patch Expanding module and the sixth FKVSL module, resulting in the sixth FKVSL module's result. This result is then added to the result processed by the second FKVSL module, and subsequently processed by the Patch Expanding module and the seventh FKVSL module, resulting in the seventh FKVSL module's result. This result is then added to the result processed by the first FKVSL module, and subsequently processed by the Patch Expanding module and the eighth FKVSL module, resulting in the eighth FKVSL module's result. Finally, this eighth FKVSL module's result is processed by the Final Project module to obtain the output result. The processing procedure in the Patch Expanding module is as follows: The input result is first processed by the Linear function to reduce the number of channels to half of the original number, and then processed by the LayerNorm function to obtain the output result. The processing procedure in the Final Project module is as follows: the input result is first processed by the Linear function to reduce the number of channels to one-quarter of the original number, and then processed by the LayerNorm function to obtain the output result.

7. The nonlinear medical image segmentation method as described in claim 4 or 5, characterized in that: All FKVSL modules have the same structure, and their processing procedure is as follows: The input results are processed sequentially through three FKC-VSS modules, and the results are then processed by the Sample function to obtain the output results. The processing procedure in the FKC-VSS module is as follows: The input result is processed by the Chunk function and then split into two paths. One path is processed by the FKC module, and the other path is processed by the VSS module. The two processing results are concatenated by the Concat function, then processed by the Shuffle function. The result is added to the input result to obtain the output result.

8. The nonlinear medical image segmentation method as described in claim 6, characterized in that: The processing procedure in the FKC module is as follows: The input result is processed by the Chunk function and divided into four paths. The first path is processed by FastKAN convolution with a 3*3 kernel and FastKAN convolution with a 1*1 kernel, resulting in the first path result. The second path is processed by FastKAN convolution with a 3*3 kernel and FastKAN convolution with a 1*1 kernel, resulting in the second path result. The third path is processed by FastKAN convolution with a 3*3 kernel and FastKAN convolution with a 1*1 kernel, resulting in the third path result. The fourth path is processed by FastKAN convolution with a 3*3 kernel and FastKAN convolution with a 1*1 kernel, resulting in the fourth path result. The four results are concatenated using the Concat function, then processed by the Shuffle function to obtain the output result.

9. The nonlinear medical image segmentation method as described in claim 6, characterized in that: The processing procedure in the VSS module is as follows: After the input result is processed by the LN function, the result is divided into two paths. The first path is processed by the Linear function and the SiLU function in sequence to obtain the first path result. The second path is processed by the Linear function, the depthwise separable convolution, the SiLU function, the SS2D function, and the LN function in sequence to obtain the second path result. The first path result and the second path result are multiplied, processed by the Linear function, and the result is added to the input result to obtain the output result.

Images