A colorectal cancer pathological image semantic segmentation method and system

Abstract

The application discloses a kind of colorectal cancer pathological image semantic segmentation method and system, it is related to image segmentation computer vision technical field, method includes: by image acquisition equipment to target pathological area is carried out image acquisition, based on convolution kernel is carried out grid division to the image dataset of target pathological area, according to the feature comparison of the image data training set of introduced N pseudo time sequence conversion window, memory sliding window module is embedded into colon cancer image semantic segmentation network, according to layered pseudo time sequence semantic feature to image data training set and image data test set are carried out boundary remediation correction, update to colorectal cancer pathological image semantic segmentation network, by the weight model of colorectal cancer lesion segmentation to the image dataset in correction image data test set is segmented, it has solved current visual segmentation algorithm under the technical problem of low segmentation precision and segmentation difficulty in colorectal cancer pathological image, can realize accurate identification to colorectal cancer lesion.

Description

Technical Field

[0001] This invention relates to the field of image segmentation and computer vision technology, specifically to a semantic segmentation method and system for colorectal cancer pathological images. Background Technology

[0002] Medical image segmentation is a fundamental problem in computer-aided diagnostic systems and intelligent analysis of colorectal cancer histopathology. Colorectal cancer accounts for approximately 8% of all cancer diagnoses and cancer-related deaths worldwide each year. By dividing image data using convolutional kernels and then further segmenting the image data to obtain a sufficient quantity and quality of segmentation results, reliable evidence can be provided for the clinical diagnosis, timely treatment, and pathological research of colorectal cancer. Therefore, semantic segmentation of colorectal cancer pathological images can help doctors make more accurate and faster diagnoses, thereby reducing workload, diagnostic time, and the error rate due to subjective factors.

[0003] Current visual segmentation algorithms suffer from high segmentation difficulty and low segmentation accuracy in colorectal cancer pathological images, making them unsuitable for accurate identification of colorectal cancer lesions. Summary of the Invention

[0004] This application provides a semantic segmentation method and system for colorectal cancer pathological images, which solves the technical problems of high segmentation difficulty and low segmentation accuracy of current visual segmentation algorithms in colorectal cancer pathological images, and can be used to achieve accurate identification of colorectal cancer lesions.

[0005] In view of the above problems, this application provides a semantic segmentation method and system for colorectal cancer pathological images.

[0006] Firstly, this application provides a semantic segmentation method for colorectal cancer pathological images. The method includes: acquiring images of a target pathological region using an image acquisition device to obtain an image dataset of the target pathological region; dividing the image dataset of the target pathological region into a grid based on convolutional kernels to determine an image data training set and an image data test set; performing feature comparison on the image data training set according to N pseudo-temporal transformation windows to determine hierarchical pseudo-temporal semantic features, where N is a positive integer greater than or equal to 3; embedding a memory sliding window module into a colorectal cancer image semantic segmentation network, and determining hierarchical pseudo-temporal semantic features based on the hierarchical transformation windows. Pseudo-temporal semantic features are used to perform boundary compensation and correction on the image data training set and image data test set to obtain corrected image data training set and corrected image data test set; pseudo-temporal hierarchical semantic information is encoded using a shared weighted memory convolution operator to update the semantic segmentation network for colorectal cancer pathological images; based on the updated semantic segmentation network for colorectal cancer pathological images, a weight model for segmenting colorectal cancer lesions is trained according to the corrected image data training set, and the image dataset containing lesions within the target pathological region in the corrected image data test set is segmented using the weight model for segmenting colorectal cancer lesions.

[0007] Secondly, this application provides a semantic segmentation system for colorectal cancer pathological images. The system includes: an image acquisition module for acquiring images of the target pathological region using an image acquisition device to obtain an image dataset of the target pathological region; a grid partitioning module for partitioning the image dataset of the target pathological region into a grid based on a convolutional kernel to determine the image data training set and the image data test set; a feature comparison module for performing feature comparison on the image data training set according to the introduced N pseudo-temporal transformation windows to determine hierarchical pseudo-temporal semantic features, where N is a positive integer greater than or equal to 3; and a compensation and correction module for embedding a memory sliding window module into the colorectal cancer image. In the semantic segmentation network, boundary compensation and correction are performed on the image data training set and image data test set based on hierarchical pseudo-temporal semantic features to obtain corrected image data training set and corrected image data test set; Network update module: The network update module is used to encode pseudo-temporal hierarchical semantic information using shared weight memory convolution operators to update the semantic segmentation network for colorectal cancer pathological images; Model training module: The model training module is used to train the weight model for colorectal cancer lesion segmentation based on the updated semantic segmentation network for colorectal cancer pathological images and the corrected image data training set, and then segment the image dataset containing lesions within the target pathological region in the corrected image data test set using the weight model for colorectal cancer lesion segmentation.

[0008] One or more technical solutions provided in this application have at least the following technical effects or advantages: This application provides a semantic segmentation method for colorectal cancer pathological images. The method involves acquiring images of the target pathological region using an image acquisition device, dividing the image dataset of the target pathological region into a grid based on convolutional kernels, comparing features of the image data training set according to N pseudo-temporal transformation windows, embedding a memory sliding window module into the colorectal cancer image semantic segmentation network, correcting the boundaries of the image data training set and the image data test set based on hierarchical pseudo-temporal semantic features, updating the colorectal cancer pathological image semantic segmentation network, and segmenting the image dataset in the corrected image data test set using a weight model for colorectal cancer lesion segmentation. This method solves the technical problems of high segmentation difficulty and low segmentation accuracy of current visual segmentation algorithms in colorectal cancer pathological images, and can be used to achieve accurate identification of colorectal cancer lesions. Attached Figure Description

[0009] Figure 1 This application provides a schematic flowchart of a semantic segmentation method for colorectal cancer pathological images; Figure 2This application provides a schematic diagram of the process for determining the image data training set and image data test set in a semantic segmentation method for colorectal cancer pathological images; Figure 3 This application provides a schematic diagram of the structure of a semantic segmentation system for colorectal cancer pathological images.

[0010] Figure labeling: a) Image acquisition module, b) Grid partitioning module, c) Feature comparison module, d) Compensation and correction module, e) Network update module, f) Model training module. Detailed Implementation

[0011] This application provides a semantic segmentation method for colorectal cancer pathological images. The method involves acquiring images of the target pathological region using an image acquisition device to obtain an image dataset of the target pathological region. A convolutional kernel then divides the image dataset into a grid, determining the training and test sets. Next, N pseudo-temporal transformation windows are introduced to perform feature comparison on the training set, yielding hierarchical pseudo-temporal semantic features. A memory sliding window module is embedded into the colorectal cancer image semantic segmentation network. Boundary correction and filling are performed on the training and test sets based on the hierarchical pseudo-temporal semantic features. A shared-weight memory convolution operator is then used to encode the pseudo-temporal hierarchical semantic information, updating the colorectal cancer pathological image semantic segmentation network. Subsequently, a weighted model for colorectal cancer lesion segmentation is trained using the corrected training set. Finally, the weighted model for colorectal cancer lesion segmentation is used to segment the image dataset containing lesions within the target pathological region of the corrected test set. This method solves the technical problems of high segmentation difficulty and low accuracy in current visual segmentation algorithms for colorectal cancer pathological images, and can be used to achieve accurate identification of colorectal cancer lesions.

[0012] Example 1: As Figure 1 As shown, this application provides a semantic segmentation method for colorectal cancer pathological images, applied to an intelligent management system. This intelligent management system is communicatively connected to an image acquisition device. The method includes: Step S100: Acquire images of the target pathological area using an image acquisition device to obtain an image dataset of the target pathological area; Specifically, the image acquisition device is an electronic colonoscope. The electronic colonoscope is small, slender, and flexible, with a diameter of about one centimeter. It can be inserted into the rectum through the anus to clearly display images of lesions inside the colon and large intestine on a computer screen for observation. The lens can acquire images from multiple angles and directions. Images of the target pathological area are acquired using this device, and the acquired images are then processed and collected to obtain an image dataset of the target pathological area. This image dataset provides the data foundation for subsequent grid division.

[0013] Step S200: Divide the image dataset of the target pathological region into grids based on the convolution kernel to determine the image data training set and the image data test set; Specifically, a convolution kernel refers to a function that, when processing a target image, takes a weighted average of pixels in a small region of the input image and outputs each corresponding pixel in the output image. The weights are defined by a function called the convolution kernel.

[0014] Mesh partitioning involves dividing image data into many small units for analysis. In this case, a convolutional kernel is used to partition the image dataset of the target pathological region into a mesh. One-quarter of the mesh is used as a test set to evaluate network performance, while the remaining three-quarters are used as a training set to train the network model. This convolutional kernel processing makes the analysis of local features clearer and more accurate.

[0015] Step S300: Perform feature comparison on the image data training set according to the introduced N pseudo-temporal transformation windows to determine the hierarchical pseudo-temporal semantic features, where N is a positive integer greater than or equal to 3; Specifically, the pseudo-temporal transformation window is a non-temporal transformation window. It allows for continuous convolution operations on the image training set without affecting other image data. First, the non-temporal image is divided into several equal parts using the pseudo-temporal transformation window. Then, continuous convolution operations are performed on each of these parts. After each operation, the parts are restored, resulting in hierarchical pseudo-temporal semantic features. There are N pseudo-temporal transformation windows, where N is a positive integer greater than or equal to 3. A larger N indicates more pseudo-temporal semantic features. Determining the hierarchical pseudo-temporal semantic features allows for continuous optimization of the network's performance.

[0016] Step S400: Embed the memory sliding window module into the colorectal cancer image semantic segmentation network, and perform boundary compensation correction on the image data training set and the image data test set according to the hierarchical pseudo-temporal semantic features to obtain the corrected image data training set and the corrected image data test set; Specifically, the memory sliding window module consists of shared-weight memory convolution and a sliding window mechanism. The sliding window mechanism only performs boundary correction on portions within the sliding window, and the window slides along the boundary, ensuring complete processing of boundary information without affecting other data and reducing the possibility of omissions. Shared-weight memory convolution refers to parameters with similar features sharing a single weight parameter, reducing the number of parameters and improving learning efficiency. After using the sliding window to extract a certain number of small-sized images from each image, the shared-weight memory convolution extracts feature space features, compensating for the loss of boundary information caused by the equal division process. Adding the memory sliding window module further enhances the network's segmentation capabilities.

[0017] Step S500: Use the shared weight memory convolution operator to encode pseudo-temporal hierarchical semantic information to update the semantic segmentation network of the colorectal cancer pathological image; Specifically, the shared-weighted memory convolution operator is an operator of shared-weighted memory convolution. Using the shared-weighted memory convolution operator instead of ordinary convolution as the encoder to encode pseudo-temporal hierarchical semantic information layer by layer can enhance the consistency of multi-scale spatial features and improve the effectiveness and breadth of feature acquisition during network encoding. By encoding pseudo-temporal hierarchical semantic information using the shared-weighted memory convolution operator, a semantic segmentation network for colorectal cancer pathological images is updated. During decoding, layer-by-layer upsampling and skip connections with encoded features are used to fuse low-level cell shape, color, and size features with high-level semantic information layer by layer, improving the network's ability to identify and segment targets, thus constructing a semantic segmentation network suitable for colorectal cancer pathological images. After spatial feature enhancement of the semantic features in the pseudo-temporal transformation window using shared-weighted memory convolution, the overall performance of the network is further optimized.

[0018] Step S600: Based on the updated semantic segmentation network for colorectal cancer pathological images, a weight model for segmenting colorectal cancer lesions is trained according to the training set of the corrected image data. The weight model for segmenting colorectal cancer lesions is then used to segment the image dataset containing lesions within the target pathological region in the test set of the corrected image data.

[0019] Specifically, a pseudo-temporal hierarchical memory semantic segmentation network for colorectal cancer pathological images is trained on a corrected image data training set. This generates a weighted model suitable for segmenting two semantic regions: colorectal cancer cells and background. The effectiveness of this weighted model is then validated on a test set. The network's ability to segment colorectal cancer lesions is comprehensively evaluated using four metrics. The weighted model for colorectal cancer lesion segmentation is then used to segment the image dataset containing lesions within the target pathological region of the corrected image data test set. The weighted model for colorectal cancer lesion segmentation demonstrates better segmentation of the image dataset containing lesions within the target pathological region of the corrected image data test set.

[0020] Furthermore, this application provides a semantic segmentation method for colorectal cancer pathological images. This method acquires images of the target pathological region using an image acquisition device, divides the image dataset of the target pathological region into a grid based on convolutional kernels, performs feature comparison on the image data training set according to N pseudo-temporal transformation windows, embeds a memory sliding window module into the colorectal cancer image semantic segmentation network, performs boundary compensation and correction on the image data training set and image data test set based on hierarchical pseudo-temporal semantic features, updates the colorectal cancer pathological image semantic segmentation network, and segments the image dataset in the corrected image data test set using a weight model for colorectal cancer lesion segmentation. This method solves the technical problems of high segmentation difficulty and low segmentation accuracy of current visual segmentation algorithms in colorectal cancer pathological images, and can be used to achieve accurate identification of colorectal cancer lesions.

[0021] Furthermore, step S100 of this application also includes: Step S110: Determine the pixel range based on the image acquisition boundary of the image acquisition device; Step S120: Within the pixel range, the lesion area is marked to obtain multiple pathological area data; Step S130: Integrate the images corresponding to the multiple pathological region data to obtain the image dataset of the target pathological region; Specifically, the image boundary data of the image acquisition device is defined as the pixel range of the acquired image. Generally, the pixel range of images acquired by image acquisition devices is from 25199×11044 to 41832×30822. The lesion area is annotated using annotation software such as Photoshop. During the annotation process, relevant knowledge of colorectal cancer cell pathology must be incorporated, and the annotation must be carried out under the strict guidance and supervision of relevant pathology experts. The target pathological region is mainly divided into two semantic segmentation regions: colorectal cancer cells and background. The resulting annotation is the image dataset of the target pathological region. This image dataset provides the data foundation for subsequent grid division.

[0022] Furthermore, such as Figure 2 As shown, step S200 of this application further includes: Step S210: Based on the convolution kernel, the image data in the image dataset of the target pathological region are sequentially divided into grids to obtain the divided image information; Step S220: Traverse and recognize the segmented image information to obtain grid image recognition information; Step S230: Based on the grid image recognition information, determine the image data training set and the image data test set for the target case region.

[0023] Specifically, the image data in the target pathological region image dataset is sequentially divided into grids, resulting in numerous small units. This division of image information is then used to evaluate the feature matching degree based on the value of the convolution kernel at local feature locations. If the evaluation is unsatisfactory, the corresponding image data is re-divided and evaluated again. All divided image information is then recognized, yielding the resulting grid image recognition information, with a one-to-one correspondence between the divided image information and the grid image recognition information. Based on the characteristics exhibited by the grid image recognition information, the image data in the image dataset is divided into a training set and a test set, with the training set showing a higher feature matching degree.

[0024] Furthermore, step S300 of this application also includes: Step S310: The image data training set is segmented through each of the N pseudo-temporal transformation windows to determine the image slice data; Step S320: Perform continuous convolution operation on the image slice data to obtain convolutional image slices; Step S330: The convolutional image slices are restored, and the restored image data is compared with the image data training set to determine the hierarchical pseudo-temporal semantic features; Specifically, N pseudo-temporal transformation windows, namely CW1, CW2, ..., CWn, divide the non-temporal image into different equal parts. CW1 remains unchanged, CW2 is divided into 4 equal parts, CW3 is divided into 16 equal parts, and CWn is divided into 4 equal parts. n -1The image is divided into equal parts, and the resulting slices are the image slice data. First, the small slices in each CW are flattened. After flattening, three consecutive convolutions are performed. After each convolution operation, the small slices are restored and stitched back to the normal image size. The restored image data is then compared with the image data training set to obtain three layers of semantic features under three pseudo-temporal conditions. This transforms non-temporal images into hierarchical pseudo-temporal semantic features. The more pseudo-temporal transformation windows (i.e., the larger the value of N), the better the network performance can be optimized. Determining these hierarchical pseudo-temporal semantic features enables the establishment of global and local information interaction, expanding the semantic horizon.

[0025] Furthermore, step S440 of this application also includes: Step S410: Use the memory sliding window to slide through the image dataset to obtain M images, where M is a positive integer greater than or equal to 4; Step S420: Then, slide the M images through the memory sliding window to obtain 2M images, wherein the size of the 2M images is smaller than the size of the M images; Step S430: Extract spatial features from the 2M images based on shared weighted memory convolution to obtain spatial features; Step S440: Perform feature fusion between the hierarchical pseudo-temporal semantic features and the spatial features, and perform image boundary correction on the image data training set and the image data test set according to the feature fusion result, and determine the corrected image data training set and the updated image data test set; Specifically, a memory sliding window dataset is used for image segmentation. For each image, M images are continuously slid along its edges, where M is a positive integer not less than 4. A larger value of M improves the compensation of boundary information, but M is limited by the network's computational power. These M images are then further segmented using the memory sliding window, resulting in a total of 2M images. These 2M images are a refinement of the original M images, so their size is smaller than the original M images. Feature matching is performed using shared-weight memory convolution, and the spatial features of the 2M images are extracted. The hierarchical pseudo-temporal semantic features are then fused with the spatial features to obtain the feature fusion result. This fusion result compensates for the loss of boundary information caused during the segmentation process, resulting in a corrected image training set and an updated image test set. The N pseudo-temporal transformation windows provide hierarchical semantic features, enabling the network to capture both global and local semantic information. Further embedding of the memory sliding window module can compensate for the loss of boundary information and some spatial feature information caused during image segmentation. Memory sliders can improve the ability to capture and identify lesions, enhance the overall performance of the network, and thus improve the accuracy and efficiency of semantic segmentation of colorectal cancer pathological images.

[0026] Furthermore, step S600 of this application also includes: Step S610: Construct a weighted model for segmenting the colorectal cancer lesions based on a BP neural network; Step S620: The weighted model for colorectal cancer lesion segmentation includes a lesion proportion layer, a lesion weight configuration layer, a lesion precision control layer, and a lesion recall probability layer; Step S630: Perform data annotation on the corrected image data training set to obtain a first constructed dataset, wherein the first constructed dataset includes a first training set and a first validation set; Step S640: Supervised training and validation of the weight model for colorectal cancer lesion segmentation are performed using the first training set and the first validation set until the weight model for colorectal cancer lesion segmentation converges or the accuracy reaches the preset requirements.

[0027] Specifically, a backpropagation (BP) neural network is a multi-layered feedforward neural network, characterized by forward signal propagation and backward error propagation. The BP neural network uses current data to find the weight relationship between the input and output, and then uses this weight relationship for simulation. For example, using temperature, humidity, and air pressure as inputs and weather conditions as outputs, the BP neural network is trained using the input-output relationship of the current data. Then, by inputting today's temperature, humidity, and air pressure data into this BP neural network, the output, i.e., today's weather conditions, can be obtained. Similarly, a weight model for colorectal cancer lesion segmentation can also be obtained using a BP neural network.

[0028] The weighted model for colorectal cancer lesion segmentation includes a lesion proportion layer, a lesion weight configuration layer, a lesion precision control layer, and a lesion recall probability layer. The lesion proportion layer represents the proportion of the colorectal cancer lesion size to the entire image. The lesion weight configuration layer indicates the severity of the lesion; colorectal cancer presents differently in early, middle, and late stages. The lesion precision control layer represents the clarity of the lesion image boundaries captured by the image acquisition device. Blurred edges and overexposure often reduce clarity around lesions, affecting subsequent analysis. The lesion recall probability layer is the ratio of precisely segmented images of colorectal lesions to the sum of precisely segmented and inaccurately segmented images, representing the lesion segmentation capability.

[0029] The corrected image data training set is labeled one by one, and the labeled result is the first constructed dataset. The first constructed dataset includes a first training set and a first validation set. The weight model for colorectal cancer lesion segmentation is trained and validated using the first training set and the first validation set. First, the weight model for colorectal cancer lesion segmentation is trained using the training set to obtain results, and then the results are validated using the validation set. This process continues until the weight model for colorectal cancer lesion segmentation converges or its accuracy reaches a preset requirement. Supervised training and validation of the weight model for colorectal cancer lesion segmentation can improve its accuracy.

[0030] Furthermore, step S600 of this application also includes: Step S650: The target pathological region in the corrected image data test set is processed by the lesion proportion layer in the weight model of the colorectal cancer lesion segmentation to determine the average crossover ratio of the lesions; Step S660: The target pathological region in the corrected image data test set is processed by the lesion weight configuration layer in the weight model of colorectal cancer lesion segmentation to determine the lesion weight coefficient; Step S670: The target pathological region in the corrected image data test set is processed by the lesion accuracy control layer in the weight model of the colorectal cancer lesion segmentation to determine the lesion location accuracy; Step S680: The target pathological region in the test set of the corrected image data is processed by the lesion recall probability layer in the weight model of the colorectal cancer lesion segmentation to determine the lesion recall rate; Step S690: Segment the image dataset based on the average crossover ratio (CLORD) of the lesions in the target pathological region, the lesion weight coefficient, the lesion location accuracy, and the lesion recall rate.

[0031] Specifically, the area of ​​the target pathological region in the corrected image data test set is calculated based on the lesion proportion layer in the weight model for colorectal cancer lesion segmentation. Then, the area of ​​the non-pathological region is calculated, and their intersection and union are compared to obtain the average intersection-union ratio (IU / U) of the lesions. The target pathological region in the corrected image data test set corresponding to the lesion weight configuration layer in the weight model for colorectal cancer lesion segmentation is multiplied by its corresponding weight ratio to obtain its lesion weight coefficient. Similarly, the target pathological region in the corrected image data test set corresponding to the lesion precision control layer in the weight model for colorectal cancer lesion segmentation is found, and multiplied by its corresponding weight ratio to obtain its lesion location precision. The lesion recall rate is obtained by multiplying the target pathological region in the corrected image data test set corresponding to the lesion recall probability layer in the weight model for colorectal cancer lesion segmentation by its corresponding weight ratio. The image dataset is then segmented based on the average IU / U, lesion weight coefficient, lesion location precision, and lesion recall rate of the lesions in the target pathological region.

[0032] Example 2: Based on the same inventive concept as the semantic segmentation method for colorectal cancer pathological images in the foregoing examples, such as... Figure 3 As shown, this application provides a semantic segmentation system for colorectal cancer pathological images, the system comprising: Image acquisition module a: The image acquisition module is used to acquire images of the target pathological area through an image acquisition device to obtain an image dataset of the target pathological area; Grid partitioning module b: This module is used to partition the image dataset of the target pathological region into a grid based on the convolution kernel, thereby determining the image data training set and the image data test set. Feature comparison module c: The feature comparison module is used to perform feature comparison on the image data training set according to the introduced N pseudo-temporal transformation windows to determine the hierarchical pseudo-temporal semantic features, where N is a positive integer greater than or equal to 3; Compensation and Correction Module d: The compensation and correction module is used to embed the memory sliding window module into the colorectal cancer image semantic segmentation network, and to perform boundary compensation and correction on the image data training set and the image data test set according to the hierarchical pseudo-temporal semantic features, so as to obtain the corrected image data training set and the corrected image data test set; Network update module e: The network update module is used to update the semantic segmentation network of the colorectal cancer pathological image by encoding pseudo-temporal hierarchical semantic information using a shared weight memory convolution operator; Model training module f: The model training module is used to train a weight model for colorectal cancer lesion segmentation based on the updated colorectal cancer pathological image semantic segmentation network and the corrected image data training set, and to segment the image dataset containing lesions in the target pathological region in the corrected image data test set through the weight model for colorectal cancer lesion segmentation.

[0033] Furthermore, the system also includes: Range determination module: The range determination module is used to determine the pixel range based on the image acquisition boundary of the image acquisition device; Lesion region annotation module: The lesion region annotation module is used to annotate the lesion region within the pixel range and obtain multiple pathological region data; Image integration module: The image integration module is used to integrate the images corresponding to the multiple pathological region data to obtain the image dataset of the target pathological region; Furthermore, the system also includes: Image data grid partitioning module: The image data grid partitioning module is used to partition the image data in the image dataset of the target pathological region into grids based on the convolution kernel, and obtain partitioned image information; Grid image recognition module: The grid image recognition module is used to traverse and recognize the divided image information to obtain grid image recognition information; Region data determination module: The region data determination module is used to determine the image data training set and the image data test set of the target case region based on the grid image recognition information.

[0034] Furthermore, the system also includes: Slice Data Module: The slice data module is used to segment the image data training set through each of the N pseudo-temporal transformation windows to determine the image slice data; Continuous convolution module: The continuous convolution module is used to perform continuous convolution operations on the image slice data to obtain convolutional image slices; Feature comparison module: The feature comparison module is used to restore the convolutional image slices, compare the restored image data with the image data training set, and determine the hierarchical pseudo-temporal semantic features.

[0035] Furthermore, the system also includes: Sliding window module: The sliding window module is used to slide the image dataset through the memory sliding window to obtain M images, where M is a positive integer greater than or equal to 4; Secondary sliding window capture module: The secondary sliding window capture module is used to perform sliding window capture on the M images again through the memory sliding window to obtain 2M images, wherein the size of the 2M images is smaller than the size of the M images; Spatial feature extraction module: The spatial feature extraction module is used to extract spatial features from the 2M images based on shared weighted memory convolution to obtain spatial features; The image data test set update module is used to fuse the hierarchical pseudo-temporal semantic features with the spatial features, and perform image boundary correction on the image data training set and the image data test set according to the feature fusion result, and determine the corrected image data training set and the updated image data test set. Furthermore, the system also includes: Weight model construction module: The weight model construction module is used to construct a weight model for colorectal cancer lesion segmentation based on a BP neural network; Weighted model layering module: The weighted model layering module is used for the weighted model of colorectal cancer lesion segmentation, which includes a lesion proportion layer, a lesion weight configuration layer, a lesion precision control layer, and a lesion recall probability layer; Corrected image data annotation module: The corrected image data annotation module is used to annotate the corrected image data training set to obtain a first constructed dataset, wherein the first constructed dataset includes a first training set and a first validation set; Supervised training and verification module: The supervised training and verification module is used to supervise the training and verification of the weight model for colorectal cancer lesion segmentation using the first training set and the first verification set until the weight model for colorectal cancer lesion segmentation converges or the accuracy reaches the preset requirements.

[0036] Furthermore, the system also includes: Average Cross-Union Ratio (CUI) Determination Module: The average cross-union ratio determination module is used to process the target pathological region in the corrected image data test set through the lesion proportion layer in the weight model of the colorectal cancer lesion segmentation to determine the average cross-union ratio of the lesions; Weight coefficient determination module: The weight coefficient determination module is used to process the target pathological region in the corrected image data test set through the lesion weight configuration layer in the weight model of colorectal cancer lesion segmentation, and determine the lesion weight coefficient; Location accuracy determination module: The location accuracy determination module is used to process the target pathological region in the corrected image data test set through the lesion accuracy control layer in the weight model of colorectal cancer lesion segmentation to determine the lesion location accuracy; Lesion recall module: The lesion recall module is used to process the target pathological region in the test set of the corrected image data through the lesion recall probability layer in the weight model of colorectal cancer lesion segmentation to determine the lesion recall rate; Image dataset segmentation module: The image dataset segmentation module is used to segment the image dataset based on the average crossover ratio of the lesions in the target pathological region, the weight coefficient of the lesions, the location accuracy of the lesions, and the recall rate of the lesions.

[0037] Through the foregoing detailed description of a semantic segmentation method for colorectal cancer pathological images, those skilled in the art can clearly understand the semantic segmentation method and system for colorectal cancer pathological images in this embodiment. As for the device disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and relevant parts can be referred to the description in the method section.

[0038] The above description of the disclosed embodiments enables those skilled in the art to make or use this application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of this application. Therefore, this application is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A semantic segmentation method for colorectal cancer pathological images, characterized in that, The method includes: Images of the target pathological area are acquired using an image acquisition device to obtain an image dataset of the target pathological area; Based on the convolution kernel, the image dataset of the target pathological region is divided into grids to determine the image data training set and the image data test set. Based on the N pseudo-temporal transformation windows introduced, feature comparison is performed on the image data training set to determine the hierarchical pseudo-temporal semantic features, where N is a positive integer greater than or equal to 3; The memory sliding window module is embedded into the semantic segmentation network of colorectal cancer pathological images. Based on the hierarchical pseudo-temporal semantic features, boundary compensation and correction are performed on the image data training set and the image data test set to obtain the corrected image data training set and the corrected image data test set. The semantic segmentation network for colorectal cancer pathological images is updated by encoding pseudo-temporal hierarchical semantic information using a shared weighted memory convolution operator. Based on the updated semantic segmentation network for colorectal cancer pathological images, a weight model for segmenting colorectal cancer lesions is trained according to the training set of the corrected image data. The weight model for segmenting colorectal cancer lesions is then used to segment the image dataset containing lesions within the target pathological region in the test set of the corrected image data.

2. The method as described in claim 1, characterized in that, The method for obtaining the image dataset of the target pathological region further includes: The pixel range is determined based on the image acquisition boundary of the image acquisition device. Within the specified pixel range, the lesion area is marked to obtain multiple pathological area data; The images corresponding to the multiple pathological region data are integrated to obtain the image dataset of the target pathological region.

3. The method as described in claim 1, characterized in that, The method for determining the image data training set and the image data test set further includes: Based on the convolution kernel, the image data in the image dataset of the target pathological region are sequentially divided into grids to obtain the divided image information; The segmented image information is traversed and identified to obtain grid image recognition information; Based on the grid image recognition information, the image data training set and the image data test set for the target pathological region are determined.

4. The method as described in claim 1, characterized in that, The method for determining the hierarchical pseudo-temporal semantic features further includes: The image data training set is segmented by using each of the N pseudo-temporal transformation windows to determine the image slice data; Perform continuous convolution operations on the image slice data to obtain convolutional image slices; The convolutional image slices are restored, and the restored image data is compared with the image data training set to determine the hierarchical pseudo-temporal semantic features.

5. The method as described in claim 1, characterized in that, The method for obtaining the corrected image data training set and the updated image data test set further includes: The image dataset is slide-windowed through the memory window to obtain M images, where M is a positive integer greater than or equal to 4; Then, the M images are slid-down through the memory sliding window to obtain 2M images, wherein the size of the 2M images is smaller than the size of the M images; Spatial features are extracted from the 2M images based on shared-weighted memory convolution to obtain spatial features; The hierarchical pseudo-temporal semantic features are fused with the spatial features. Based on the feature fusion result, the image data training set and the image data test set are respectively corrected to compensate for image boundaries, and the corrected image data training set and the updated image data test set are determined.

6. The method as described in claim 1, characterized in that, The method also includes: A weighted model for segmenting colorectal cancer lesions is constructed based on a backpropagation neural network. The weighted model for colorectal cancer lesion segmentation includes a lesion proportion layer, a lesion weight configuration layer, a lesion precision control layer, and a lesion recall probability layer. The corrected image data training set is labeled to obtain a first constructed dataset, wherein the first constructed dataset includes a first training set and a first validation set; The weighted model for colorectal cancer lesion segmentation is trained and validated using the first training set and the first validation set until the weighted model for colorectal cancer lesion segmentation converges or the accuracy reaches the preset requirements.

7. The method as described in claim 6, characterized in that, The method also includes: The target pathological region in the corrected image data test set is processed by the lesion proportion layer in the weighted model of colorectal cancer lesion segmentation to determine the average crossover ratio of lesions; The target pathological region in the corrected image data test set is processed by the lesion weight configuration layer in the weight model for colorectal cancer lesion segmentation to determine the lesion weight coefficient. The target pathological region in the corrected image data test set is processed by the lesion accuracy control layer in the weight model of colorectal cancer lesion segmentation to determine the lesion location accuracy; The lesion recall rate is determined by processing the target pathological region in the corrected image data test set through the lesion recall probability layer in the weighted model of colorectal cancer lesion segmentation. The image dataset is segmented based on the average crossover ratio (CLORD) of the lesions in the target pathological region, the lesion weight coefficient, the lesion location accuracy, and the lesion recall rate.

8. A semantic segmentation system for pathological images of colorectal cancer, characterized in that, The system includes: Image acquisition module: The image acquisition module is used to acquire images of the target pathological area through an image acquisition device to obtain an image dataset of the target pathological area; Grid partitioning module: The grid partitioning module performs grid partitioning on the image dataset of the target pathological region based on the convolution kernel to determine the image data training set and the image data test set; Feature comparison module: The feature comparison module is used to perform feature comparison on the image data training set according to the introduced N pseudo-temporal transformation windows to determine the hierarchical pseudo-temporal semantic features, where N is a positive integer greater than or equal to 3; The compensation and correction module is used to embed the memory sliding window module into the semantic segmentation network of colorectal cancer pathological images, and to perform boundary compensation and correction on the image data training set and the image data test set according to the hierarchical pseudo-temporal semantic features to obtain the corrected image data training set and the corrected image data test set. Network update module: The network update module is used to update the semantic segmentation network of the colorectal cancer pathological image by encoding pseudo-temporal hierarchical semantic information using a shared weight memory convolution operator; Model training module: The model training module is used to train a weight model for colorectal cancer lesion segmentation based on the updated colorectal cancer pathological image semantic segmentation network and the training set of the corrected image data. The weight model for colorectal cancer lesion segmentation is used to segment the image dataset containing lesions in the target pathological region of the test set of the corrected image data.

Images