Colorectal tumor image region segmentation method and system based on AI

Abstract

The invention belongs to the field of image region segmentation, and particularly relates to an AI-based colorectal tumor image region segmentation method and system, and the method comprises the following steps: S1, obtaining original 3D colorectal tumor image data; according to the method, the high-quality 3D colorectal tumor image data is processed through the surface parameterization technology, the obtained 2D feature map is more suitable for processing the hybrid FMU-attention network model, and the processing efficiency and the processing speed of the hybrid FMU-attention network model are greatly improved; in addition, an AI framework is formed by jointly fusing a hybrid FMU-attention network model, a Vision Transformers module and a DeepLabV3 + model, the AI framework enables the boundary to be more accurate and the multi-scale performance to be more excellent, and it is ensured that the tumor can be effectively recognized and segmented regardless of the size.

Description

Technical Field

[0001] This invention relates to the field of image region segmentation technology, and in particular to an AI-based method and system for colorectal tumor image region segmentation. Background Technology

[0002] Colorectal cancer, also known as colon and rectal cancer, is a common malignant tumor of the digestive tract that occurs in the colon or rectum. Clinically, patients mainly present with blood in the stool, changes in stool consistency, and changes in bowel habits. In advanced stages, it can lead to intestinal obstruction, abdominal masses, and cachexia, endangering life. Colorectal cancer usually begins as a benign tumor, initially mostly appearing as polyps, which develop into cancer over time. According to data released by the National Cancer Center of China in February 2022, colorectal cancer ranks second in incidence, and its incidence and mortality rates are on the rise. When segmenting colorectal tumor images, convolutional neural network (CNN) technology is commonly used. CNN is a deep learning neural network specifically designed to process data with specific grid-like topological structures (most typically images), laying a solid foundation for the automatic segmentation of colorectal tumors. However, this CNN technology still has the following shortcomings when segmenting colorectal tumor images: (1) The core of convolutional neural network technology is local convolution operation. Each neuron only processes local information within its receptive field. Although the receptive field can be enlarged by stacking multiple layers of convolution, it is still very difficult and inefficient to model the relationship between two distant regions in an image, which makes CNNs have obvious shortcomings in long-distance modeling. (2) Convolutional neural network technology is greatly limited in its ability to capture remote context information; (3) Convolutional neural network technology cannot effectively identify tumors with blurred boundaries; To address the shortcomings of convolutional neural network technology in segmenting colorectal tumor images, we propose an AI-based method and system for colorectal tumor image region segmentation. Summary of the Invention

[0003] This invention proposes an AI-based method and system for colorectal tumor image region segmentation, which overcomes the shortcomings of convolutional neural network technology in segmenting colorectal tumor images.

[0004] To achieve the above objectives, the present invention adopts the following technical solution: An AI-based method for colorectal tumor image region segmentation includes the following steps: S1: Acquire raw 3D colorectal tumor image data; S2: The raw 3D colorectal tumor image data first removes invalid data and selects raw 3D colorectal tumor image data with clear images, good contrast, good colon filling and cleanliness. The raw 3D colorectal tumor image data is then preprocessed to obtain high-quality 3D colorectal tumor image data. S3: High-quality 3D colorectal tumor image data are processed using surface parameterization technology to obtain multiple standard 2D feature maps; S4: The 2D feature map in S3 is input into the trained hybrid FMU-attention network model. This model consists of a shallow hybrid FMU-attention processing layer, a medium hybrid FMU-attention processing layer, and a deep hybrid FMU-attention processing layer. The shallow hybrid FMU-attention processing layer processes the input 2D feature map to obtain a low-level feature map, which is then input into the medium hybrid FMU-attention processing layer for further processing to obtain a medium-level feature map. This medium-level feature map is then input into the deep hybrid FMU-attention processing layer for further processing to obtain a high-level feature map, which is then output by the hybrid FMU-attention network model. S5: The Vision Transformers module performs secondary processing on the high-level feature maps output by the hybrid FMU-attention network model, outputting a refined feature sequence rich in global relationships. This sequence is then reshaped back into the feature map format to obtain a refined feature map. S6: The DeepLabV3+ model receives the refined feature map output by the Vision Transformers module and the low-level feature map provided by the shallow hybrid FMU-attention processing layer. It then uses ASPP to fuse multi-scale context and combines it with high-resolution details through the decoder to finally output a segmentation map with clear boundaries and accurate classification.

[0005] Preferably, in step S1, the original 3D colorectal tumor image data includes the original 3D colorectal tumor image and its corresponding image label.

[0006] Preferably, in step S2, the original 3D colorectal tumor image data with clear images, good contrast, and good colon filling and cleanliness undergoes image preprocessing, which specifically includes the following steps: S201: First, the input raw 3D colorectal tumor image data is subjected to Z-Score normalization. The calculation formula for Z-Score normalization is: ,in This is the pixel value matrix of the original 3D colorectal tumor image. The mean pixel data of the entire original 3D colorectal tumor image. The standard deviation of the pixel data of the entire original 3D colorectal tumor image is given, and it is consistent with... The calculation range is consistent, resulting in the original image after Z-Score normalization. Original image obey" For 0 and The original image has a distribution of "1". It is a three-dimensional matrix of H×W×D, where H, W, and D are the original image, respectively. The number of voxels in the height, width, and depth directions; S202: Original Image The target image is adjusted to have equal voxel spacing in the X, Y and Z dimensions by physical coordinate alignment and interpolation calculation. S203: Divide the target image obtained in S202 into several small, non-overlapping blocks. For each block, calculate the grayscale histogram, crop the histogram according to the set cropping limit, and equalize the cropped histogram. Finally, use bilinear interpolation to combine the blocks to obtain the enhanced new image. S204: Gaussian blur is used to extract the low-frequency background from the new image obtained in S203, while suppressing high-frequency edges and noise. The calculation formula for Gaussian blur processing is as follows: ,in It is a two-dimensional matrix of the new image. It is a Gaussian convolution kernel. This is the output image after Gaussian blurring. S205: Subtract the low-frequency background from the new image obtained from S203 to separate the high-frequency components representing edges, textures, and noise. The formula for calculating the separation of high-frequency components is as follows: ,in It is a two-dimensional matrix of the new image. This is the output image after Gaussian blurring. The high-frequency components are separated and output as the detail layer; S206: The separated high-frequency components are amplified and then added back to the new image obtained in S203 to obtain high-quality 3D colorectal tumor image data. The calculation formula used is as follows: ,in It is a two-dimensional matrix of the new image. This is the output image of the separated high-frequency components. k is the enhancement coefficient, also known as the gain or intensity factor. k > 0, and it is usually between 0.5 and 2. The larger the k, the stronger the edge enhancement effect. This provides high-quality 3D colorectal tumor image data.

[0007] Preferably, the specific steps of S202 are as follows: S2021: The target size after resampling is calculated based on the principle of physical length conservation. The calculation formula is as follows: , , ,in , and These represent the number of voxels in the height, width, and depth of the target image, respectively. This is the rounding function. , ,and The original images are shown in the X, Y, and Z directions, respectively. voxel spacing, Usually, the original image is taken. Minimum spacing of voxels, where H, W, and D are the original image values. The number of voxels in the height, width, and depth directions; S2022: Mapping physical coordinates to voxel coordinates: Original image The physical coordinates corresponding to the voxel coordinates (i, j, k) are: (x, y, z) = , where (i∈[0, H−1], j∈[0, W−1], k∈[0, D−1]); Voxel coordinates of the target image in the resampled image The corresponding physical coordinates are: ,in( ∈[0, -1], ∈[0, -1], ∈[0, -1]); S023: For each voxel coordinate of the target image From the original image using trilinear interpolation Obtain the corresponding physical coordinates The pixel value.

[0008] Preferably, in step S3, the high-quality 3D colorectal tumor image data is processed using surface parameterization technology, and the specific process includes the following steps: S301: High-quality 3D colorectal tumor image data is used to extract triangular meshes from colorectal tumor images using the Marching Cubes algorithm. The triangular mesh is a 3D surface mesh composed of vertices V, edges E, and faces F, denoted as M=(V, E, F). S302: The 3D surface mesh in S301 is mapped one-to-one to a 2D planar mesh through geometric surface parameterization. The geometric surface parameterization adopts the wrap angle parameter method, which first uses the Cauchy-Riemann equation for mapping. The formula of the Cauchy-Riemann equation is: , where (u, v) are the coordinates of the mapped 2D plane, and (x, y) are the local coordinates of the 3D surface; Then calculate the conformally parameterized angle distortion energy function. The formula is as follows: ,in Let dS be the area integral of the 3D surface M, dS be the area element of the 3D surface, ∇u be the gradient of the u-component of the coordinates in the 2D plane, and ∇v be the gradient of the v-component of the coordinates in the 2D plane. S303: After the mapping in S302 is completed, a 2D image is created. The width of the 2D image is W, the height is H, and each pixel q in the 2D image is... , Each pixel q corresponds to a point p in a 2D planar grid, where the coordinates of point p in the 2D planar grid are (u, v). S304: Traverse every triangular facet of the 2D planar mesh. The centroid coordinate method is used to determine whether point p is inside the current triangle T. If p is inside the current triangle T, the next step is executed. If p is not inside the current triangle T, it is marked as background and the background value is assigned to 0. Then the next triangle T is checked. S305: Once p is determined to be on the triangular facet The centroid coordinates relative to the triangular facet are calculated using the centroid coordinate formula. The formula for the centroid coordinates is: ,and ; S306: Find the 3D facet F corresponding to the triangular facet T of the 2D planar mesh. , , Then, the same set of barycentric coordinates calculated on the 2D planar mesh are used. To the vertices of the 3D surface , , eigenvalues , , Linear interpolation is performed, and the linear interpolation formula is: This ultimately yields multiple standard 2D feature maps.

[0009] Preferably, in step S4, the shallow, medium, and deep hybrid FMU-attention processing layers in the hybrid FMU-attention network model transmit data sequentially from front to back. The feature map input to the shallow hybrid FMU-attention processing layer is the 2D feature map input to the hybrid FMU-attention network model, and the high-level feature map output by the deep hybrid FMU-attention processing layer is the high-level feature map output by the hybrid FMU-attention network model. Each of the shallow, medium, and deep hybrid FMU-attention processing layers consists of an FMU-attention block and a downsampling layer connected to it. The specific steps for processing using FMU-attention blocks and downsampling layers are as follows: S401: The shape of the input feature map is... Where B is the number of feature maps input at once, C is the number of channels in the feature map, H is the number of vertical pixels in the feature map, and W is the number of horizontal pixels in the feature map. Then, layer normalization is performed on the input feature maps. Specifically, all channels C and all spatial channels (H, W) of each sample are normalized. The normalization formula is as follows: ,in It is the mean of all features for each sample. It is the standard deviation of all features for each sample. These are learnable affine transformation parameters. It is a very small number used for numerical stability, and X is the input feature map. , This is the feature map after normalization. S402: The normalized feature map Xnorm is copied twice, one copy is input into the FMU path, and the other copy is input into the spatial attention path; S403: The FMU path first processes the input feature map Perform a Fourier transform; the Fourier transform formula is as follows: , It is a frequency domain feature in complex form, which can be decomposed into real parts. and the virtual part ; Then, regarding the real part... and the virtual part The formulas used when applying multilayer perceptrons are as follows: , The two MLPs in this process share a structure but do not share parameters; R and I are identifiers of frequency domain feature components. Finally, the processed real part is... and the virtual part The process of combining the elements and then transforming them back into the spatial domain uses the following formula: , Where i is the imaginary unit, , This is the output of the FMU path; S404: Spatial attention path uses a 1×1 convolution to integrate the feature map The multi-channel features are compressed into a single-channel feature map, and then a spatial attention map is generated by passing it through an activation function. The formula for the activation function is: ,in It is a 1×1 convolution with 1 output channel, which performs a weighted summation of information from all channels to obtain a spatial saliency map. , ,in addition It's the Sigmoid function, which compresses the value to... Within the interval, the spatial attention map is finally obtained. And output it; S405: Output of the FMU path Spatial attention map output by spatial attention path The formula for element-wise multiplication is: ,in Representing element-wise multiplication, spatial attention graph The shape is The characteristics after fusion are obtained. ; S406: Features after fusion and the normalized feature map Perform residual concatenation to obtain the output feature map Y, Y= ; S407: The downsampling layer uses a convolutional layer with a stride of 2 and a kernel size of 3×3 to compress and abstract the feature map Y output by S406, and finally outputs a feature map with a reduced spatial size, increased number of channels, and expanded receptive field.

[0010] Preferably, in step S5, the Vision Transformers module performs secondary processing on the high-level feature map output by the hybrid FMU-attention network model. The specific steps are as follows: S501: The size of the input high-level feature map is ,in H represents the number of channels in the high-level feature map output by the hybrid FMU-attention network model, H represents the height of the feature map input to the shallow hybrid FMU-attention processing layer, and W represents the width of the feature map input to the shallow hybrid FMU-attention processing layer. The Vision Transformers module reshapes the input high-level feature map into a series of feature tokens. The formula used is: , Where N is the number of tokens, i.e., the total number of pixels in the feature map, and each token is a... A 3D vector represents the features of a spatial location; S502: A series of feature tokens The linear projection layer maps the dimensions required by the VisionTransformers module, using the following formula: , Where N is the number of tokens, It is the projection weight matrix. For a series of feature tokens, , where is the bias term, and D is the hidden dimension of the Transformer model. The mapped sequence; S503: Mapped sequence First, positional encoding is added to obtain the sequence. , , ,in This is a positional encoding matrix, where each position has a unique D-dimensional vector, and N is the number of tokens. The sequence... The input is then fed into a Transformer encoder, which consists of a multi-head self-attention layer and a feedforward network. The Transformer encoder outputs a refined feature sequence rich in global relationships. , Where N is the number of tokens, D is the Transformer hidden dimension, and L is the number of Transformer encoder layers, this feature sequence can be reconstructed back into a feature map format to obtain a refined feature map, and the size of this refined feature map is [size missing]. ,in These represent the number of channels, height, and width of the refined feature map, respectively.

[0011] Preferably, in step S6, the specific steps are as follows: S601: Refined feature maps of the Vision Transformers module output by the ASPP module in the DeepLabV3+ model. Parallel multi-scale convolution and pooling are performed to fuse multi-scale context and obtain rich semantic features. ,in , These represent the number of channels, height, and width of the refined feature map, respectively. Number of output channels ; S602: The decoder first enriches semantic features Upsampling by a factor of 4 yields the upsampled ASPP features. Its formula is: , The number of output channels ; S603: Decoder for low-level features Perform a 1×1 convolution to reduce the number of channels and obtain the projected low-level features. ; The convolution formula is: , in , , Low-level features The number of channels, Low-level features The number of channels, H is the height of the feature map input to the shallow hybrid FMU-attention processing layer, and W is the width of the feature map input to the shallow hybrid FMU-attention processing layer; S604: Upsample the ASPP features Low-level features after projection By splicing the features together, we can obtain the spliced ​​features. , , ,in , Low-level features The number of channels, Then, a 3×3 convolution is used to process the concatenated features. Perform smooth fusion to obtain the fused feature map. Its smoothing formula is: , Its output channel number is set to ,in (•) represents a two-dimensional convolution function using a 3×3 kernel. (•) represents the modified linear activation function; S605: Merged feature map Upsampled by a factor of 2, the feature map resolution H×W is restored to that of the input to the shallow hybrid FMU-attention processing layer, resulting in the output feature map. , Feature map Then, a 1×1 prediction convolutional layer is used for processing, and the final output is a segmentation map with clear boundaries and accurate classification.

[0012] This invention also proposes an AI-based colorectal tumor image region segmentation system, including a colorectal tumor image data acquisition module, a colorectal tumor image data processing module, an image data geometric surface parameterization processing module, a hybrid FMU-attention network model, a Vision Transformers module, a DeepLabV3+ model, and an output module; The colorectal tumor image data acquisition module is used to acquire raw 3D colorectal tumor image data and input it into the colorectal tumor image data processing module; The colorectal tumor image data processing module is used to remove invalid data from the original 3D colorectal tumor image data and select the original 3D colorectal tumor image data with clear images, good contrast, good colon filling and cleanliness. The colorectal tumor image data processing module also performs image preprocessing on the selected original 3D colorectal tumor image data and outputs high-quality 3D colorectal tumor image data to the image data geometric surface parameterization processing module. The image data geometric surface parameterization processing module is used to process high-quality 3D colorectal tumor image data through surface parameterization technology to obtain multiple standard 2D feature maps, which are then output to the hybrid FMU-attention network model. The hybrid FMU-attention network model consists of shallow hybrid FMU-attention processing layers, medium hybrid FMU-attention processing layers, and deep hybrid FMU-attention processing layers. The shallow hybrid FMU-attention processing layer processes the input 2D feature map to obtain a low-level feature map, which is then fed into the medium hybrid FMU-attention processing layer for further processing to obtain a medium-level feature map. This medium-level feature map is then fed into the deep hybrid FMU-attention processing layer for further processing to obtain a high-level feature map. The hybrid FMU-attention network model outputs the high-level feature map to the Vision Transformers module. The Vision Transformers module is used to perform secondary processing on the high-level feature maps output by the hybrid FMU-attention network model to obtain refined feature maps, and then output the refined feature maps to the DeepLabV3+ model. The DeepLabV3+ model receives the refined feature map output by the Vision Transformers module and the low-level feature map provided by the shallow hybrid FMU-attention processing layer. It then uses ASPP to fuse multi-scale context and combines it with high-resolution details through a decoder to obtain a segmentation map with clear boundaries and accurate classification. The DeepLabV3+ model outputs the segmentation map with clear boundaries and accurate classification through the output module.

[0013] Compared with the prior art, the beneficial effects of the present invention are: 1. This invention processes high-quality 3D colorectal tumor image data using surface parameterization technology to obtain multiple standard 2D feature maps. These 2D feature maps are more suitable for processing by the hybrid FMU-attention network model, which greatly improves the processing efficiency and speed of the hybrid FMU-attention network model. 2. This invention employs a hybrid FMU-attention network model composed of shallow, medium, and deep hybrid FMU-attention processing layers to process standard 2D feature maps. The spatial attention path of the hybrid FMU-attention network model dynamically adjusts the weights of the feature maps in the spatial dimension, generating higher attention weights at the tumor-normal tissue boundary, dynamically optimizing the allocation of computational resources, and establishing long-distance dependencies of spatial context. Combined with the FMU path of the hybrid FMU-attention network model, this achieves rapid and efficient capture of global context and periodic texture features. 3. The Vision Transformers module in this invention performs secondary processing on the high-level feature maps output by the hybrid FMU-attention network model, and establishes deeper and more accurate global relationships between pixels through its own attention mechanism. 4. The DeepLabV3+ model in this invention is used to receive the refined feature map output by the Vision Transformers module and the low-level feature map provided by the shallow hybrid FMU-attention processing layer. It uses ASPP to fuse multi-scale context and then combines it with high-resolution details through the decoder. Thus, it can capture context information and finely recover boundaries at multiple scales, making the boundaries more accurate and the multi-scale performance more outstanding, ensuring that tumors of all sizes can be effectively identified and segmented. This invention processes high-quality 3D colorectal tumor image data using surface parameterization technology. The resulting 2D feature maps are more suitable for processing by a hybrid FMU-attention network model, greatly improving the processing efficiency and speed of the hybrid FMU-attention network model. In addition, a hybrid FMU-attention network model, Vision Transformers module, and DeepLabV3+ model are integrated into an AI architecture. This AI architecture makes the boundaries more accurate and the multi-scale performance more outstanding, ensuring that tumors of all sizes can be effectively identified and segmented. Attached Figure Description

[0014] Figure 1 This is a flowchart of an AI-based image region segmentation method for colorectal tumors proposed in this invention; Figure 2 This is a flowchart illustrating the process of FMU-attention blocks and downsampling layers working together in an AI-based colorectal tumor image region segmentation method proposed in this invention. Figure 3 This is a connection block diagram of an AI-based colorectal tumor image region segmentation system proposed in this invention. Detailed Implementation

[0015] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0016] Example 1 Reference Figure 1-2 This embodiment proposes an AI-based method for colorectal tumor image region segmentation, including the following steps: S1: Obtain raw 3D colorectal tumor image data, which includes raw 3D colorectal tumor images and their corresponding image labels; S2: The raw 3D colorectal tumor image data first removes invalid data, and selects raw 3D colorectal tumor image data with clear images, good contrast, good colon filling and cleanliness. The raw 3D colorectal tumor image data with clear images, good contrast, good colon filling and cleanliness undergoes image preprocessing, which specifically includes the following steps: S201: First, the input raw 3D colorectal tumor image data is subjected to Z-Score normalization. The calculation formula for Z-Score normalization is: ,in This is the pixel value matrix of the original 3D colorectal tumor image. The mean pixel data of the entire original 3D colorectal tumor image. The standard deviation of the pixel data of the entire original 3D colorectal tumor image is given, and it is consistent with... The calculation range is consistent, resulting in the original image after Z-Score normalization. Original image obey" For 0 and The original image has a distribution of "1". It is a three-dimensional matrix of H×W×D, where H, W, and D are the original image, respectively. The number of voxels in the height, width, and depth directions; S202: Original Image The target image is adjusted to have equal voxel spacing in the X, Y and Z dimensions by physical coordinate alignment and interpolation calculation. S203: Divide the target image obtained in S202 into several small, non-overlapping blocks. For each block, calculate the grayscale histogram, crop the histogram according to the set cropping limit, and equalize the cropped histogram. Finally, use bilinear interpolation to combine the blocks to obtain the enhanced new image. S204: Gaussian blur is used to extract the low-frequency background from the new image obtained in S203, while suppressing high-frequency edges and noise. The calculation formula for Gaussian blur processing is as follows: ,in It is a two-dimensional matrix of the new image. It is a Gaussian convolution kernel. This is the output image after Gaussian blurring. S205: Subtract the low-frequency background from the new image obtained from S203 to separate the high-frequency components representing edges, textures, and noise. The formula for calculating the separation of high-frequency components is as follows: ,in It is a two-dimensional matrix of the new image. This is the output image after Gaussian blurring. The high-frequency components are separated and output as the detail layer; S206: The separated high-frequency components are amplified and then added back to the new image obtained in S203 to obtain high-quality 3D colorectal tumor image data. The calculation formula used is as follows: ,in It is a two-dimensional matrix of the new image. This is the output image of the separated high-frequency components. k is the enhancement coefficient, also known as the gain or intensity factor. k > 0, and it is usually between 0.5 and 2. The larger the k, the stronger the edge enhancement effect. High-quality 3D colorectal tumor image data; S3: High-quality 3D colorectal tumor image data are processed using surface parametric techniques. The specific process includes the following steps: S301: High-quality 3D colorectal tumor image data is used to extract triangular meshes from colorectal tumor images using the Marching Cubes algorithm. The triangular mesh is a 3D surface mesh composed of vertices V, edges E, and faces F, denoted as M=(V, E, F). S302: The 3D surface mesh in S301 is mapped one-to-one to a 2D planar mesh through geometric surface parameterization. The geometric surface parameterization adopts the wrap angle parameter method, which first uses the Cauchy-Riemann equation for mapping. The formula of the Cauchy-Riemann equation is: , where (u, v) are the coordinates of the mapped 2D plane, and (x, y) are the local coordinates of the 3D surface; Then calculate the conformally parameterized angle distortion energy function. The formula is as follows: ,in Let dS be the area integral of the 3D surface M, dS be the area element of the 3D surface, ∇u be the gradient of the u-component of the coordinates in the 2D plane, and ∇v be the gradient of the v-component of the coordinates in the 2D plane. S303: After the mapping in S302 is completed, a 2D image is created. The width of the 2D image is W, the height is H, and each pixel q in the 2D image is... , Each pixel q corresponds to a point p in a 2D planar grid, where the coordinates of point p in the 2D planar grid are (u, v). S304: Traverse every triangular facet of the 2D planar mesh. The centroid coordinate method is used to determine whether point p is inside the current triangle T. If p is inside the current triangle T, the next step is executed. If p is not inside the current triangle T, it is marked as background and the background value is assigned to 0. Then the next triangle T is checked. S305: Once p is determined to be on the triangular facet The centroid coordinates relative to the triangular facet are calculated using the centroid coordinate formula. The formula for the centroid coordinates is: ,and ; S306: Find the 3D facet F corresponding to the triangular facet T of the 2D planar mesh. , , Then, the same set of barycentric coordinates calculated on the 2D planar mesh are used. To the vertices of the 3D surface , , eigenvalues , , Linear interpolation is performed, and the linear interpolation formula is: This process ultimately yields multiple standard 2D feature maps, which are more suitable for processing by the hybrid FMU-attention network model, greatly improving the processing efficiency and speed of the hybrid FMU-attention network model. S4: The 2D feature map in S306 is input into the trained hybrid FMU-attention network model. This model consists of a shallow hybrid FMU-attention processing layer, a medium hybrid FMU-attention processing layer, and a deep hybrid FMU-attention processing layer. The shallow hybrid FMU-attention processing layer processes the input 2D feature map to obtain a low-level feature map, which is then input into the medium hybrid FMU-attention processing layer for further processing to obtain a medium-level feature map. This medium-level feature map is then input into the deep hybrid FMU-attention processing layer for further processing to obtain a high-level feature map, which is then output by the hybrid FMU-attention network model. S5: The Vision Transformers module performs secondary processing on the high-level feature maps output by the hybrid FMU-attention network model, outputting a refined feature sequence rich in global relationships. This sequence is then reshaped back into the feature map format to obtain a refined feature map. S6: The DeepLabV3+ model receives the refined feature map output by the Vision Transformers module and the low-level feature map provided by the shallow hybrid FMU-attention processing layer, and uses ASPP to fuse multi-scale context. Then, it combines high-resolution details through the decoder to finally output a segmentation map with clear boundaries and accurate classification. This embodiment processes high-quality 3D colorectal tumor image data using surface parameterization technology. The resulting 2D feature map is more suitable for processing by the hybrid FMU-attention network model, greatly improving the processing efficiency and speed of the hybrid FMU-attention network model. In addition, the hybrid FMU-attention network model, the Vision Transformers module, and the DeepLabV3+ model are integrated into an AI architecture. This AI architecture makes the boundaries more accurate and the multi-scale performance more outstanding, ensuring that tumors of all sizes can be effectively identified and segmented.

[0017] Specifically, the steps of S202 in this embodiment are as follows: S2021: The target size after resampling is calculated based on the principle of physical length conservation. The calculation formula is as follows: , , ,in , and These represent the number of voxels in the height, width, and depth of the target image, respectively. This is the rounding function. , ,and The original images are shown in the X, Y, and Z directions, respectively. voxel spacing, Usually, the original image is taken. Minimum spacing of voxels, where H, W, and D are the original image values. The number of voxels in the height, width, and depth directions; S2022: Mapping physical coordinates to voxel coordinates: Original image The physical coordinates corresponding to the voxel coordinates (i, j, k) are: (x, y, z) = , where (i∈[0, H−1], j∈[0, W−1], k∈[0, D−1]); Voxel coordinates of the target image in the resampled image The corresponding physical coordinates are: ,in( ∈[0, -1], ∈[0, -1], ∈[0, -1]); S023: For each voxel coordinate of the target image From the original image using trilinear interpolation Obtain the corresponding physical coordinates The pixel values ​​are used to obtain the target image.

[0018] In S4 of this embodiment, the shallow, medium, and deep hybrid FMU-attention processing layers in the hybrid FMU-attention network model transmit data sequentially from front to back. The feature map input to the shallow hybrid FMU-attention processing layer is the 2D feature map input to the hybrid FMU-attention network model, and the high-level feature map output by the deep hybrid FMU-attention processing layer is the high-level feature map output by the hybrid FMU-attention network model. Each of the shallow, medium, and deep hybrid FMU-attention processing layers consists of an FMU-attention block and a downsampling layer connected to it. The specific steps for processing using FMU-attention blocks and downsampling layers are as follows: S401: The shape of the input feature map is... Where B is the number of feature maps input at once, C is the number of channels in the feature map, H is the number of vertical pixels in the feature map, and W is the number of horizontal pixels in the feature map. Then, layer normalization is performed on the input feature maps. Specifically, all channels C and all spatial channels (H, W) of each sample are normalized. The normalization formula is as follows: ,in It is the mean of all features for each sample. It is the standard deviation of all features for each sample. These are learnable affine transformation parameters. It is a very small number used for numerical stability, and X is the input feature map. , This is the feature map after normalization. S402: Feature map after normalization Make two copies, one input into the FMU path and the other input into the spatial attention path; S403: The FMU path first processes the input feature map Perform a Fourier transform; the Fourier transform formula is as follows: , It is a frequency domain feature in complex form, which can be decomposed into real parts. and the virtual part ; Then, regarding the real part... and the virtual part The formulas used when applying multilayer perceptrons are as follows: , The two MLPs in this process share a structure but do not share parameters; R and I are identifiers of frequency domain feature components. Finally, the processed real part is... and the virtual part The process of combining the elements and then transforming them back into the spatial domain uses the following formula: , Where i is the imaginary unit, , This is the output of the FMU path; S404: Spatial attention path uses a 1×1 convolution to integrate the feature map The multi-channel features are compressed into a single-channel feature map, and then a spatial attention map is generated by passing it through an activation function. The formula for the activation function is: ,in It is a 1×1 convolution with 1 output channel, which performs a weighted summation of information from all channels to obtain a spatial saliency map. , ,in addition It's the Sigmoid function, which compresses the value to... Within the interval, the spatial attention map is finally obtained. And output it; S405: Output of the FMU path Spatial attention map output by spatial attention path The formula for element-wise multiplication is: ,in Representing element-wise multiplication, spatial attention graph The shape is The characteristics after fusion are obtained. ; S406: Features after fusion and the normalized feature map Perform residual concatenation to obtain the output feature map Y, Y= ; S407: The downsampling layer uses a convolutional layer with a stride of 2 and a kernel size of 3×3 to compress and abstract the feature map Y output by S406, and finally outputs a feature map with a reduced spatial size, increased number of channels, and expanded receptive field. The spatial attention path of the hybrid FMU-attention network model in this embodiment dynamically adjusts the weights of the feature map in the spatial dimension, generating higher attention weights at the tumor-normal tissue boundary, dynamically optimizing the allocation of computational resources, and establishing long-distance dependencies of spatial context. Combined with the FMU path of the hybrid FMU-attention network model, it achieves fast and efficient capture of global context and periodic texture features.

[0019] In S5 of this embodiment, the Vision Transformers module performs secondary processing on the high-level feature map output by the hybrid FMU-attention network model. The specific steps are as follows: S501: The size of the input high-level feature map is ,in H represents the number of channels in the high-level feature map output by the hybrid FMU-attention network model, H represents the height of the feature map input to the shallow hybrid FMU-attention processing layer, and W represents the width of the feature map input to the shallow hybrid FMU-attention processing layer. The Vision Transformers module reshapes the input high-level feature map into a series of feature tokens. The formula used is: , Where N is the number of tokens, i.e., the total number of pixels in the feature map, and each token is a... A 3D vector represents the features of a spatial location; S502: A series of feature tokens The linear projection layer maps the dimensions required by the VisionTransformers module, using the following formula: , Where N is the number of tokens, It is the projection weight matrix. For a series of feature tokens, , where is the bias term, and D is the hidden dimension of the Transformer model. The mapped sequence; S503: Mapped sequence First, positional encoding is added to obtain the sequence. , , ,in This is a positional encoding matrix, where each position has a unique D-dimensional vector, and N is the number of tokens. The sequence... The input is then fed into a Transformer encoder, which consists of a multi-head self-attention layer and a feedforward network. The Transformer encoder outputs a refined feature sequence rich in global relationships. , Where N is the number of tokens, D is the Transformer hidden dimension, and L is the number of Transformer encoder layers, this feature sequence can be reconstructed back into a feature map format to obtain a refined feature map, and the size of this refined feature map is [size missing]. ,in These represent the number of channels, height, and width of the refined feature map, respectively. The Vision Transformers module in this embodiment performs secondary processing on the high-level feature maps output by the hybrid FMU-attention network model, and establishes deeper and more accurate global relationships between pixels through its own attention mechanism.

[0020] The specific steps of S6 in this embodiment are as follows: S601: Refined feature maps of the Vision Transformers module output by the ASPP module in the DeepLabV3+ model. Parallel multi-scale convolution and pooling are performed to fuse multi-scale context and obtain rich semantic features. ,in , These represent the number of channels, height, and width of the refined feature map, respectively. Number of output channels ; S602: The decoder first enriches semantic features Upsampling by a factor of 4 yields the upsampled ASPP features. Its formula is: , The number of output channels ; S603: Decoder for low-level features Perform a 1×1 convolution to reduce the number of channels and obtain the projected low-level features. ; The convolution formula is: , in , , Low-level features The number of channels, Low-level features The number of channels, H is the height of the feature map input to the shallow hybrid FMU-attention processing layer, and W is the width of the feature map input to the shallow hybrid FMU-attention processing layer; S604: Upsample the ASPP features Low-level features after projection By splicing the features together, we can obtain the spliced ​​features. , , ,in , Low-level features The number of channels, Then, a 3×3 convolution is used to process the concatenated features. Perform smooth fusion to obtain the fused feature map. Its smoothing formula is: , Its output channel number is set to ,in (•) represents a two-dimensional convolution function using a 3×3 kernel. (•) represents the modified linear activation function; S605: Merged feature map Upsampled by a factor of 2, the feature map resolution H×W is restored to that of the input to the shallow hybrid FMU-attention processing layer, resulting in the output feature map. , Feature map Then, a 1×1 prediction convolutional layer is used for processing, and the final output is a segmentation map with clear boundaries and accurate classification. The DeepLabV3+ model in this embodiment can capture contextual information and finely restore boundaries at multiple scales, resulting in more accurate boundaries and superior multi-scale performance, ensuring that tumors of all sizes can be effectively identified and segmented.

[0021] Example 2 Based on the same inventive concept as the AI-based colorectal tumor image region segmentation method in the foregoing embodiments, and referring to... Figure 3 This embodiment proposes an AI-based colorectal tumor image region segmentation system, including a colorectal tumor image data acquisition module, a colorectal tumor image data processing module, an image data geometric surface parameterization processing module, a hybrid FMU-attention network model, a Vision Transformers module, a DeepLabV3+ model, and an output module; The colorectal tumor image data acquisition module is used to acquire raw 3D colorectal tumor image data and input it into the colorectal tumor image data processing module; The colorectal tumor image data processing module is used to remove invalid data from the original 3D colorectal tumor image data and select the original 3D colorectal tumor image data with clear images, good contrast, good colon filling and cleanliness. The colorectal tumor image data processing module also performs image preprocessing on the selected original 3D colorectal tumor image data and outputs high-quality 3D colorectal tumor image data to the image data geometric surface parameterization processing module. The image data geometric surface parameterization processing module is used to process high-quality 3D colorectal tumor image data through surface parameterization technology to obtain multiple standard 2D feature maps, which are then output to the hybrid FMU-attention network model. The hybrid FMU-attention network model consists of shallow hybrid FMU-attention processing layers, medium hybrid FMU-attention processing layers, and deep hybrid FMU-attention processing layers. The shallow hybrid FMU-attention processing layer processes the input 2D feature map to obtain a low-level feature map, which is then fed into the medium hybrid FMU-attention processing layer for further processing to obtain a medium-level feature map. This medium-level feature map is then fed into the deep hybrid FMU-attention processing layer for further processing to obtain a high-level feature map. The hybrid FMU-attention network model outputs the high-level feature map to the Vision Transformers module. The Vision Transformers module is used to perform secondary processing on the high-level feature maps output by the hybrid FMU-attention network model to obtain refined feature maps, and then output the refined feature maps to the DeepLabV3+ model. The DeepLabV3+ model receives the refined feature map output by the Vision Transformers module and the low-level feature map provided by the shallow hybrid FMU-attention processing layer. It then uses ASPP to fuse multi-scale context and combines it with high-resolution details through a decoder to obtain a segmentation map with clear boundaries and accurate classification. The DeepLabV3+ model outputs the segmentation map with clear boundaries and accurate classification through the output module.

[0022] The AI-based colorectal tumor image region segmentation system provided in this embodiment of the invention can execute the AI-based colorectal tumor image region segmentation method provided in any embodiment of the invention, and has the corresponding functional modules and beneficial effects of the execution method.

[0023] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. An AI-based method for segmenting colorectal tumor image regions, characterized in that, Includes the following steps: S1: Acquire raw 3D colorectal tumor image data; S2: The raw 3D colorectal tumor image data first removes invalid data and selects raw 3D colorectal tumor image data with clear images, good contrast, good colon filling and cleanliness. The raw 3D colorectal tumor image data is then preprocessed to obtain high-quality 3D colorectal tumor image data. S3: High-quality 3D colorectal tumor image data are processed using surface parameterization technology to obtain multiple standard 2D feature maps; S4: The 2D feature map in S3 is input into the trained hybrid FMU-attention network model. This model consists of a shallow hybrid FMU-attention processing layer, a medium hybrid FMU-attention processing layer, and a deep hybrid FMU-attention processing layer. The shallow hybrid FMU-attention processing layer processes the input 2D feature map to obtain a low-level feature map, which is then input into the medium hybrid FMU-attention processing layer for further processing to obtain a medium-level feature map. This medium-level feature map is then input into the deep hybrid FMU-attention processing layer for further processing to obtain a high-level feature map, which is then output by the hybrid FMU-attention network model. S5: The Vision Transformers module performs secondary processing on the high-level feature maps output by the hybrid FMU-attention network model, outputting a refined feature sequence rich in global relationships. This sequence is then reshaped back into the feature map format to obtain a refined feature map. S6: The DeepLabV3+ model receives the refined feature map output by the Vision Transformers module and the low-level feature map provided by the shallow hybrid FMU-attention processing layer. It then uses ASPP to fuse multi-scale context and combines it with high-resolution details through the decoder to finally output a segmentation map with clear boundaries and accurate classification.

2. The AI-based method for colorectal tumor image region segmentation according to claim 1, characterized in that, In S1, the original 3D colorectal tumor image data includes the original 3D colorectal tumor image and its corresponding image label.

3. The AI-based method for colorectal tumor image region segmentation according to claim 1, characterized in that, In step S2, the raw 3D colorectal tumor image data with clear images, good contrast, and good colon filling and cleanliness undergoes image preprocessing, which specifically includes the following steps: S201: First, the input raw 3D colorectal tumor image data is subjected to Z-Score normalization. The calculation formula for Z-Score normalization is: ,in This is the pixel value matrix of the original 3D colorectal tumor image. The mean pixel data of the entire original 3D colorectal tumor image. The standard deviation of the pixel data of the entire original 3D colorectal tumor image is given, and it is consistent with... The calculation range is consistent, resulting in the original image after Z-Score normalization. Original image obey" 0 and The original image has a distribution of "1". It is a three-dimensional matrix of H×W×D, where H, W, and D are the original image, respectively. The number of voxels in the height, width, and depth directions; S202: Original Image The target image is adjusted to have equal voxel spacing in the X, Y and Z dimensions by physical coordinate alignment and interpolation calculation. S203: Divide the target image obtained in S202 into several small, non-overlapping blocks. For each block, calculate the grayscale histogram, crop the histogram according to the set cropping limit, and equalize the cropped histogram. Finally, use bilinear interpolation to combine the blocks to obtain the enhanced new image. S204: Gaussian blur is used to extract the low-frequency background from the new image obtained in S203, while suppressing high-frequency edges and noise. The calculation formula for Gaussian blur processing is as follows: ,in It is a two-dimensional matrix of the new image. It is a Gaussian convolution kernel. This is the output image after Gaussian blurring. S205: Subtract the low-frequency background from the new image obtained from S203 to separate the high-frequency components representing edges, textures, and noise. The formula for calculating the separation of high-frequency components is as follows: ,in It is a two-dimensional matrix of the new image. This is the output image after Gaussian blurring. The high-frequency components are separated and output as the detail layer; S206: The separated high-frequency components are amplified and then added back to the new image obtained in S203 to obtain high-quality 3D colorectal tumor image data. The calculation formula used is as follows: ,in It is a two-dimensional matrix of the new image. This is the output image of the separated high-frequency components. k is the enhancement coefficient, also known as the gain or intensity factor. k > 0, and it is usually between 0.5 and 2. The larger the k, the stronger the edge enhancement effect. This provides high-quality 3D colorectal tumor image data.

4. The AI-based method for colorectal tumor image region segmentation according to claim 3, characterized in that, The specific steps of S202 are as follows: S2021: The target size after resampling is calculated based on the principle of physical length conservation. The calculation formula is as follows: ， ， ,in , and These represent the number of voxels in the height, width, and depth of the target image, respectively. This is the rounding function. , ,and The original images are shown in the X, Y, and Z directions, respectively. voxel spacing, Usually, the original image is taken. Minimum spacing of voxels, where H, W, and D are the original image values. The number of voxels in the height, width, and depth directions; S2022: Mapping physical coordinates to voxel coordinates: Original image The physical coordinates corresponding to the voxel coordinates (i, j, k) are: (x, y, z) = , where (i∈[0, H−1], j∈[0, W−1], k∈[0, D−1]); Voxel coordinates of the target image in the resampled image The corresponding physical coordinates are: ,in( ∈[0, -1], ∈[0, -1], ∈[0, -1]); S023: For each voxel coordinate of the target image From the original image using trilinear interpolation Obtain the corresponding physical coordinates The pixel value.

5. The AI-based method for colorectal tumor image region segmentation according to claim 1, characterized in that, In step S3, the high-quality 3D colorectal tumor image data is processed using surface parameterization technology. The specific process includes the following steps: S301: High-quality 3D colorectal tumor image data is processed using the Marching Cubes algorithm to extract triangular meshes from the colorectal tumor images. These triangular meshes are 3D surface meshes composed of vertices V, edges E, and faces F, denoted as... ; S302: The 3D surface mesh in S301 is mapped one-to-one to a 2D planar mesh through geometric surface parameterization. The geometric surface parameterization adopts the wrap angle parameter method, which first uses the Cauchy-Riemann equation for mapping. The formula of the Cauchy-Riemann equation is: , where (u, v) are the coordinates of the mapped 2D plane, and (x, y) are the local coordinates of the 3D surface; Then calculate the conformally parameterized angle distortion energy function. The formula is as follows: ,in Let dS be the area integral of the 3D surface M, dS be the area element of the 3D surface, ∇u be the gradient of the u-component of the coordinates in the 2D plane, and ∇v be the gradient of the v-component of the coordinates in the 2D plane. S303: After the mapping in S302 is completed, a 2D image is created. The width of the 2D image is W, the height is H, and each pixel q in the 2D image is... , Each pixel q corresponds to a point p in a 2D planar grid, where the coordinates of point p in the 2D planar grid are (u, v). S304: Traverse every triangular facet of the 2D planar mesh. The centroid coordinate method is used to determine whether point p is inside the current triangle T. If p is inside the current triangle T, the next step is executed. If p is not inside the current triangle T, it is marked as background and the background value is assigned to 0. Then the next triangle T is checked. S305: Once p is determined to be on the triangular facet The centroid coordinates relative to the triangular facet are calculated using the centroid coordinate formula. The formula for the centroid coordinates is: ,and ; S306: Find the 3D facet F corresponding to the triangular facet T of the 2D planar mesh. , , Then, the same set of barycentric coordinates calculated on the 2D planar mesh are used. To the vertices of the 3D surface , , eigenvalues , , Linear interpolation is performed, and the linear interpolation formula is: This ultimately yields multiple standard 2D feature maps.

6. The AI-based method for colorectal tumor image region segmentation according to claim 1, characterized in that, In S4, the shallow, medium, and deep hybrid FMU attention processing layers in the hybrid FMU attention network model transmit data sequentially from front to back. The feature map input to the shallow hybrid FMU attention processing layer is the 2D feature map input to the hybrid FMU attention network model, and the high-level feature map output by the deep hybrid FMU attention processing layer is the high-level feature map output by the hybrid FMU attention network model. Each of the shallow, medium, and deep hybrid FMU attention processing layers consists of an FMU attention block and a downsampling layer connected to it. The specific steps for processing using FMU-attention blocks and downsampling layers are as follows: S401: The shape of the input feature map is... Where B is the number of feature maps input at once, C is the number of channels in the feature map, H is the number of vertical pixels in the feature map, and W is the number of horizontal pixels in the feature map. Then, layer normalization is performed on the input feature maps. Specifically, all channels C and all spatial channels (H, W) of each sample are normalized. The normalization formula is as follows: ,in It is the mean of all features for each sample. It is the standard deviation of all features for each sample. These are learnable affine transformation parameters. It is a very small number used for numerical stability, and X is the input feature map. , This is the feature map after normalization. S402: Feature map after normalization Make two copies, one input into the FMU path and the other input into the spatial attention path; S403: The FMU path first processes the input feature map Perform a Fourier transform; the Fourier transform formula is as follows: , It is a frequency domain feature in complex form, which can be decomposed into real parts. and the virtual part ; Then the real part and the virtual part The formulas used when applying multilayer perceptrons are as follows: ， The two MLPs in this process share a structure but do not share parameters; R and I are identifiers of frequency domain feature components. Finally, the processed real part is... and the virtual part The process of combining the elements and then transforming them back into the spatial domain uses the following formula: ， Where i is the imaginary unit, , This is the output of the FMU path; S404: Spatial attention path uses a 1×1 convolution to integrate the feature map The multi-channel features are compressed into a single-channel feature map, and then a spatial attention map is generated by passing it through an activation function. The formula for the activation function is: ,in It is a 1×1 convolution with 1 output channel, which performs a weighted summation of information from all channels to obtain a spatial saliency map. , ,in addition It's the Sigmoid function, which compresses the value to... Within the interval, the spatial attention map is finally obtained. And output it; S405: Output of the FMU path Spatial attention map output by spatial attention path The formula for element-wise multiplication is: ,in Representing element-wise multiplication, spatial attention graph The shape is The characteristics after fusion are obtained. ; S406: Features after fusion and the normalized feature map Perform residual concatenation to obtain the output feature map Y, Y= ; S407: The downsampling layer uses a convolutional layer with a stride of 2 and a kernel size of 3×3 to compress and abstract the feature map Y output by S406, and finally outputs a feature map with a reduced spatial size, increased number of channels, and expanded receptive field.

7. The AI-based method for colorectal tumor image region segmentation according to claim 1, characterized in that, In S5, the Vision Transformers module performs secondary processing on the high-level feature maps output by the hybrid FMU-attention network model. The specific steps are as follows: S501: The size of the input high-level feature map is ,in H represents the number of channels in the high-level feature map output by the hybrid FMU-attention network model, H represents the height of the feature map input to the shallow hybrid FMU-attention processing layer, and W represents the width of the feature map input to the shallow hybrid FMU-attention processing layer. The Vision Transformers module reshapes the input high-level feature map into a series of feature tokens. The formula used is: , Where N is the number of tokens, i.e., the total number of pixels in the feature map, and each token is a... A 3D vector represents the features of a spatial location; S5 02: A series of feature tokens The linear projection layer maps the dimensions required by the Vision Transformers module using the following formula: , Where N is the number of tokens, It is the projection weight matrix. For a series of feature tokens, , where is the bias term, and D is the hidden dimension of the Transformer model. The mapped sequence; S503: Mapped sequence First, positional encoding is added to obtain the sequence. , , ,in This is a positional encoding matrix, where each position has a unique D-dimensional vector, and N is the number of tokens. The sequence... The input is then fed into a Transformer encoder, which consists of a multi-head self-attention layer and a feedforward network. The Transformer encoder outputs a refined feature sequence rich in global relationships. , Where N is the number of tokens, D is the Transformer hidden dimension, and L is the number of Transformer encoder layers, this feature sequence can be reconstructed back into a feature map format to obtain a refined feature map, and the size of this refined feature map is [size missing]. ,in These represent the number of channels, height, and width of the refined feature map, respectively.

8. The AI-based method for colorectal tumor image region segmentation according to claim 7, characterized in that, The specific steps in S6 are as follows: S601: Refined feature maps of the Vision Transformers module output by the ASPP module in the DeepLabV3+ model. Parallel multi-scale convolution and pooling are performed to fuse multi-scale context and obtain rich semantic features. ,in , These represent the number of channels, height, and width of the refined feature map, respectively. Number of output channels ; S602: The decoder first enriches semantic features Upsampling by a factor of 4 yields the upsampled ASPP features. Its formula is: , The number of output channels ; S603: Decoder for low-level features Perform a 1×1 convolution to reduce the number of channels and obtain the projected low-level features. ; The convolution formula is: , in , , Low-level features The number of channels, Low-level features The number of channels, H is the height of the feature map input to the shallow hybrid FMU-attention processing layer, and W is the width of the feature map input to the shallow hybrid FMU-attention processing layer; S604: Upsample the ASPP features Low-level features after projection By splicing the features together, we can obtain the spliced ​​features. , ， ,in , Low-level features The number of channels, Then, a 3×3 convolution is used to process the concatenated features. Perform smooth fusion to obtain the fused feature map. Its smoothing formula is: ， Its output channel number is set to ,in (•) represents a two-dimensional convolution function using a 3×3 kernel. (•) represents the modified linear activation function; S605: Merge the feature map Upsampled by a factor of 2, the feature map resolution H×W is restored to that of the input to the shallow hybrid FMU-attention processing layer, resulting in the output feature map. , Feature map Then, a 1×1 prediction convolutional layer is used for processing, and the final output is a segmentation map with clear boundaries and accurate classification.

9. An AI-based colorectal tumor image region segmentation system, used to implement the method described in any one of claims 1-8, characterized in that, It includes a colorectal tumor image data acquisition module, a colorectal tumor image data processing module, an image data geometric surface parameterization processing module, a hybrid FMU-attention network model, a Vision Transformers module, a DeepLabV3+ model, and an output module; The colorectal tumor image data acquisition module is used to acquire raw 3D colorectal tumor image data and input it into the colorectal tumor image data processing module; The colorectal tumor image data processing module is used to remove invalid data from the original 3D colorectal tumor image data and select the original 3D colorectal tumor image data with clear images, good contrast, good colon filling and cleanliness. The colorectal tumor image data processing module also performs image preprocessing on the selected original 3D colorectal tumor image data and outputs high-quality 3D colorectal tumor image data to the image data geometric surface parameterization processing module. The image data geometric surface parameterization processing module is used to process high-quality 3D colorectal tumor image data through surface parameterization technology to obtain multiple standard 2D feature maps, which are then output to the hybrid FMU-attention network model. The hybrid FMU-attention network model consists of shallow hybrid FMU-attention processing layers, medium hybrid FMU-attention processing layers, and deep hybrid FMU-attention processing layers. The shallow hybrid FMU-attention processing layer processes the input 2D feature map to obtain a low-level feature map, which is then fed into the medium hybrid FMU-attention processing layer for further processing to obtain a medium-level feature map. This medium-level feature map is then fed into the deep hybrid FMU-attention processing layer for further processing to obtain a high-level feature map. The hybrid FMU-attention network model outputs the high-level feature map to the Vision Transformers module. The Vision Transformers module is used to perform secondary processing on the high-level feature maps output by the hybrid FMU-attention network model to obtain refined feature maps, and then output the refined feature maps to the DeepLabV3+ model. The DeepLabV3+ model receives the refined feature map output by the Vision Transformers module and the low-level feature map provided by the shallow hybrid FMU-attention processing layer. It then uses ASPP to fuse multi-scale context and combines it with high-resolution details through a decoder to obtain a segmentation map with clear boundaries and accurate classification. The DeepLabV3+ model outputs the segmentation map with clear boundaries and accurate classification through the output module.

Images