A liver image recognition system and method based on a graph neural network

Abstract

This invention discloses a liver image recognition system and method based on graph neural networks, belonging to the field of medical image processing and intelligent recognition technology. It includes a liver image receiving module for receiving liver image data and performing slice processing, grayscale standardization, and spatial location information association processing on the liver image data to generate standardized liver image data. In this invention, by performing slice processing, grayscale standardization, and spatial location information association processing on the liver image data, a unified correspondence is formed between multiple liver image slices in three dimensions: slice number, grayscale expression, and spatial location. Results of liver parenchyma regions, candidate abnormal regions, and boundary transition regions are extracted. The adjacency relationships, grayscale variation relationships, and boundary continuity relationships of adjacent regions within the same slice are uniformly organized, thereby improving the data orderliness, regional discrimination ability, and completeness of the expression of relationships within a single slice of multi-slice liver images.

Description

Technical Field

[0001] This invention relates to the field of medical image processing and intelligent recognition technology, and in particular to a liver image recognition system and method based on graph neural networks. Background Technology

[0002] Liver image recognition is a crucial component of medical image analysis, widely applied in scenarios such as liver region extraction, abnormal region identification, and image-assisted analysis. Existing technologies typically rely on liver image data to perform grayscale processing, region segmentation, category discrimination, and result output on the image content of each slice, obtaining information on the distribution of liver parenchyma regions, abnormal related regions, and their boundaries. In multi-slice liver image processing, related technologies also combine and organize the recognition results of each slice by considering the positional correspondence between slices, changes in regional contours, and continuous changes in the image, thereby improving the overall recognition effect. With the increase in the amount of liver image data and the increasing requirements for recognition accuracy, how to simultaneously consider the expression of regional content, boundary connectivity, and cross-slice continuity within a multi-slice scope has gradually become a key focus in liver image recognition. This has led to increased attention being paid to the processing capabilities of existing technologies in complex boundary regions and continuous slice scenarios.

[0003] Current liver image recognition methods, when processing multi-slice liver images, typically focus on region segmentation or category determination within a single slice, lacking a unified approach to slice numbering, spatial location, and grayscale representation. This makes it difficult to establish stable correspondences between different slices. Even when liver parenchyma regions and candidate abnormal regions can be extracted, it is difficult to create clear boundary transitions at their junctions, leading to unstable attribution and discontinuous connectivity between adjacent slices. Furthermore, existing technologies lack a unified organization for relationships between adjacent regions within a slice and for cross-slice correspondences, making it difficult to form a complete representation of liver structure. If the initial discrimination results are output directly without incorporating boundary consistency results into connectivity correction and re-propagation, inconsistencies between region categories and boundary connectivity can easily arise, ultimately affecting the stability and cross-slice consistency of liver recognition results. Summary of the Invention

[0004] The purpose of this invention is to address the shortcomings of existing technologies by proposing a liver image recognition system and method based on graph neural networks.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: a liver image recognition system based on graph neural networks, comprising: a liver image receiving module, used to receive liver image data, perform slice processing, grayscale standardization processing, and spatial location information association processing on the liver image data to generate standardized liver image data; a region unit generation module, used to perform liver parenchyma region extraction processing, candidate abnormal region extraction processing, and boundary transition region extraction processing on the standardized liver image data to generate a set of liver region units; and an intra-slice relationship construction module, used to perform regional adjacency relationship analysis processing, grayscale change relationship analysis processing, and boundary continuity relationship analysis processing on the set of liver region units, and perform boundary coupling adjustment processing on the analysis results according to boundary coupling adjustment parameters to generate intra-slice structural relationship data; and cross-slice... The slice relationship construction module performs position mapping, contour continuation analysis, and region correspondence processing on liver region units in adjacent slices, and performs continuous compensation adjustment processing on the analysis results based on continuous compensation adjustment parameters to generate cross-slice continuous relationship data. The liver structure semantic graph generation module performs node organization processing and edge connection organization processing on the set of liver region units, intra-slice structural relationship data, and cross-slice continuous relationship data to generate liver structure semantic graph data. The graph propagation recognition module performs liver structure semantic graph boundary write-back convergence recognition algorithm processing on the liver structure semantic graph data to generate region category discrimination results and boundary consistency results. The recognition result output module performs boundary correction processing and recognition result generation processing on the region category discrimination results and boundary consistency results to generate liver recognition results.

[0006] As a further description of the above technical solution:

[0007] The liver image receiving module performs slice numbering and slice location matching processing on the liver image data according to the acquisition sequence to form slice matching results. Based on the slice matching results, grayscale range unification and grayscale distribution correction processing are performed on the image grayscale of each slice to form grayscale standardization results. Based on the slice matching results, the spatial location information corresponding to each slice is read, and the spatial location information is associated with the slice number to form spatial location information association results. The grayscale standardization results and spatial location information association results are integrated according to the slice number to generate standardized liver image data.

[0008] As a further description of the above technical solution:

[0009] The region unit generation module performs liver main body range identification processing based on the grayscale distribution and spatial location of each slice in the standardized liver image data to form the liver main body range result. Based on the liver main body range result, it performs region connectivity processing and region contour extraction processing on the standardized liver image data to form the initial region division result. Based on the initial region division result, it performs liver parenchyma region extraction processing on the region located in the middle of the liver main body and with continuous grayscale distribution to form the liver parenchyma region result. Based on the initial region division result, it performs candidate abnormal region extraction processing on the region located in the middle of the liver main body and with grayscale distribution deviating from the liver parenchyma region result to form the candidate abnormal region result. Based on the liver parenchyma region result and the candidate abnormal region result, it performs edge transition identification processing and boundary range extraction processing on the boundary position between the two to form the boundary transition region result. The liver parenchyma region result, the candidate abnormal region result, and the boundary transition region result are processed according to the slice number and spatial location to generate a set of liver region units.

[0010] As a further description of the above technical solution:

[0011] The intra-slice relation construction module performs intra-slice aggregation processing on the liver parenchyma region results, candidate abnormal region results, and boundary transition region results in the liver region unit set according to the slice number, forming intra-slice region aggregation results;

[0012] Based on the region aggregation results within the slice, region contact recognition and region spacing determination are performed on adjacent regions in the same slice to form region adjacency relationship analysis results;

[0013] Based on the intra-slice region aggregation results, grayscale reading processing of boundary positions and grayscale change trend processing are performed on adjacent regions in the same slice to form grayscale change relationship analysis results; based on the intra-slice region aggregation results, boundary edge extraction processing and edge extension direction correspondence processing are performed on adjacent regions in the same slice to form boundary continuity relationship analysis results; based on the region adjacency relationship analysis results, grayscale change relationship analysis results, and boundary continuity relationship analysis results, boundary coupling adjustment processing is performed according to boundary coupling adjustment parameters to form intra-slice relationship adjustment results; based on the intra-slice relationship adjustment results, connection relationship organization processing is performed on adjacent regions in the same slice to generate intra-slice structural relationship data.

[0014] As a further description of the above technical solution:

[0015] The cross-slice relationship construction module performs slice correspondence processing on adjacent slices in the liver region unit set based on slice number, forming cross-slice region processing results. Based on the cross-slice region processing results, it performs region position correspondence analysis on the liver parenchyma region results, candidate abnormal region results, and boundary transition region results in adjacent slices, forming position mapping processing results. Based on the cross-slice region processing results, it performs region contour extension direction reading processing and region contour continuous change processing on corresponding regions in adjacent slices, forming contour continuity analysis results. Based on the cross-slice region processing results, it performs region overlap range comparison processing and region change continuity judgment processing on corresponding regions in adjacent slices, forming region correspondence association results. Based on the position mapping processing results, contour continuity analysis results, and region correspondence association results, it performs continuous compensation adjustment processing according to continuous compensation adjustment parameters, forming cross-slice relationship adjustment results. Based on the cross-slice relationship adjustment results, it performs continuous connection relationship tissue processing on corresponding regions in adjacent slices, generating cross-slice continuous relationship data.

[0016] As a further description of the above technical solution:

[0017] The liver structure semantic graph generation module performs node aggregation processing on the liver parenchyma region results, candidate abnormal region results, and boundary transition region results in the liver region unit set based on slice number and spatial location, forming a graph node organization result. Based on the graph node organization result, it performs node identifier writing processing and node position mapping processing on the liver parenchyma region results, candidate abnormal region results, and boundary transition region results, forming a node organization result. Based on the intra-slice structural relationship data, it performs intra-slice connection mapping processing and intra-slice connection merging processing on adjacent regions in the same slice, forming an intra-slice edge organization result. Based on the cross-slice continuous relationship data, it performs cross-slice connection mapping processing and cross-slice connection merging processing on corresponding regions in adjacent slices, forming a cross-slice edge organization result. Based on the node organization result, intra-slice edge organization result, and cross-slice edge organization result, it performs node edge mapping integration processing, forming a graph structure integration result. Based on the graph structure integration result, it performs graph data organization processing to generate liver structure semantic graph data.

[0018] As a further description of the above technical solution:

[0019] The graph propagation recognition module processes the liver structure semantic graph data using a boundary write-back convergence recognition algorithm. Based on slice numbers and spatial locations, it performs propagation preparation processing on the graph nodes and connections in the liver structure semantic graph data, generating a graph propagation preparation result. Based on the graph propagation preparation result, it performs graph neural network propagation processing, generating node propagation update results. Based on the node propagation update results, it performs region category matching processing on the liver parenchyma region results, candidate abnormal region results, and boundary transition region results, generating initial region category discrimination results. Based on the node propagation update results, it further compares the boundary transition region results with adjacent liver parenchyma region results and candidate abnormal region results. The connection relationships between the results of the normal regions are processed by boundary correspondence sorting to form the initial boundary consistency results. Based on the initial region classification results and the initial boundary consistency results, the connection relationships corresponding to the boundary transition region results are processed by writeback correction according to the boundary writeback adjustment parameters, and the writeback correction results are written into the liver structure semantic map data to form the writeback corrected liver structure semantic map data. Based on the writeback corrected liver structure semantic map data, graph neural network propagation processing is performed again, and the region category correspondence sorting and boundary correspondence sorting processing are performed on the propagated graph nodes to generate region category classification results and boundary consistency results.

[0020] As a further description of the above technical solution:

[0021] The recognition result output module performs region attribution reading processing on the liver parenchyma region results, candidate abnormal region results, and boundary transition region results in the region category discrimination results based on the slice number and spatial location, forming a region attribution consolidation result; based on the region attribution consolidation result and boundary consistency result, it performs boundary attribution determination processing on the boundary correspondence between the boundary transition region results and adjacent liver parenchyma region results and candidate abnormal region results, forming a boundary correction result; based on the boundary correction result, it performs boundary merging processing on the boundary transition region results, and performs attribution update processing on the boundary transition region results to the liver parenchyma region results or candidate abnormal region results respectively, forming a region correction result; based on the region correction result and region category discrimination result, it performs intra-slice recognition region integration processing and cross-slice recognition region correspondence processing, forming a recognition region consolidation result; based on the recognition region consolidation result, it performs recognition result generation processing to generate a liver recognition result.

[0022] As a further description of the above technical solution:

[0023] The liver image receiving module receives liver image data, generates standardized liver image data, and outputs the standardized liver image data to the region unit generation module. The region unit generation module receives the standardized liver image data, generates a set of liver region units, and outputs the set of liver region units to the intra-slice relationship construction module, the cross-slice relationship construction module, and the liver structure semantic map generation module, respectively. The intra-slice relationship construction module receives the set of liver region units, generates intra-slice structural relationship data, and outputs the intra-slice structural relationship data to the liver structure semantic map generation module. The cross-slice relationship construction module receives the set of liver region units, generates cross-slice continuous relationship data, and outputs the cross-slice continuous relationship data to the liver structure semantic map generation module. The liver structure semantic map generation module receives the set of liver region units, intra-slice structural relationship data, and cross-slice continuous relationship data, generates liver structure semantic map data, and outputs the liver structure semantic map data to the graph propagation recognition module. The graph propagation recognition module receives the liver structure semantic map data, generates region category discrimination results and boundary consistency results, and outputs the region category discrimination results and boundary consistency results to the recognition result output module. The recognition result output module receives the region category discrimination results and boundary consistency results and generates the liver recognition result.

[0024] As a further description of the above technical solution:

[0025] A liver image recognition method based on graph neural networks, the steps of which are as follows:

[0026] The system receives liver imaging data and performs slice numbering and slice location matching processing on the liver imaging data according to the acquisition order. Based on the slice matching results, it performs grayscale range unification processing and grayscale distribution correction processing on the image grayscale of each slice. It also performs association writing processing on the spatial location information corresponding to each slice and the slice number. Finally, it performs corresponding integration processing on the grayscale standardization results and spatial location information association results according to the slice number to generate standardized liver imaging data.

[0027] Based on the grayscale distribution and spatial location of each slice in the standardized liver image data, the main body of the liver is identified. Based on the results of the main body of the liver, the standardized liver image data is processed by region connectivity sorting and region contour extraction. Based on the initial region division results, liver parenchyma region extraction, candidate abnormal region extraction, edge transition identification and boundary range extraction are performed. The liver parenchyma region results, candidate abnormal region results and boundary transition region results are processed according to the slice number and spatial location to generate a set of liver region units.

[0028] Based on the slice number, the results of the liver parenchyma region, candidate abnormal region, and boundary transition region in the liver region unit set are processed by the same slice aggregation. Based on the region aggregation results within the slice, adjacent regions in the same slice are processed by region contact recognition, region spacing determination, gray-scale reading of boundary position, gray-scale change trend sorting, boundary edge extraction, and edge extension direction correspondence processing. Based on the boundary coupling adjustment parameters, the results of region adjacency relationship analysis, gray-scale change relationship analysis, and boundary continuity relationship analysis are processed by boundary coupling adjustment to generate structural relationship data within the slice.

[0029] Based on the slice number, adjacent slices in the liver region unit set are processed to perform front-to-back slice correspondence processing. Based on the cross-slice region processing results, corresponding regions in adjacent slices are processed to perform region position correspondence analysis, region contour extension direction reading, region contour continuous change processing, region overlap range comparison, and region change continuous determination processing. Based on the continuous compensation adjustment parameters, the position mapping processing results, contour continuation analysis results, and region correspondence association results are processed to perform continuous compensation adjustment processing to generate cross-slice continuous relationship data.

[0030] Based on slice number and spatial location, node aggregation, node identifier writing, and node position correspondence processing are performed on the liver parenchyma region results, candidate abnormal region results, and boundary transition region results in the liver region unit set. Based on the intra-slice structural relationship data, intra-slice connection correspondence processing and intra-slice connection merging processing are performed. Based on the cross-slice continuous relationship data, cross-slice connection correspondence processing and cross-slice connection merging processing are performed. Finally, node edge correspondence integration processing and graph data organization processing are performed on the node organization results, intra-slice edge organization results, and cross-slice edge organization results to generate liver structure semantic graph data.

[0031] The liver structure semantic map data is processed by the liver structure semantic map boundary write-back convergence recognition algorithm. Based on the slice number and spatial location, the graph nodes and connections in the liver structure semantic map data are prepared for propagation. Based on the graph propagation preparation results, graph neural network propagation is performed. Based on the node propagation update results, the initial region category discrimination results and the initial boundary consistency results are formed. Based on the boundary write-back adjustment parameters, the connection relationships corresponding to the boundary transition region results are corrected for write-back and written into the liver structure semantic map data. The graph neural network propagation process, as well as the region category correspondence sorting process and the boundary correspondence sorting process, are performed again to generate the region category discrimination results and the boundary consistency results.

[0032] Based on the slice number and spatial location, the results of liver parenchyma region, candidate abnormal region, and boundary transition region in the region category discrimination results are processed for region attribution reading. Based on the region attribution sorting results and boundary consistency results, the boundary correspondence between the boundary transition region results and adjacent liver parenchyma region results and candidate abnormal region results is processed for boundary attribution determination and boundary merging. Finally, the boundary transition region results are processed for attribution update, intra-slice identification region integration, cross-slice identification region correspondence, and identification result generation, generating the liver identification result.

[0033] The present invention has the following beneficial effects:

[0034] 1. In this invention, standardized liver image data is generated by first performing slice organization, grayscale standardization, and spatial location information association processing on liver image data. This enables multiple liver slice images to form a unified correspondence in three dimensions: slice number, grayscale expression, and spatial location. The results of liver parenchyma region, candidate abnormal region, and boundary transition region are extracted and organized into a set of liver region units. At the same time, the adjacency relationship, grayscale change relationship, and boundary continuity relationship of adjacent regions within the same slice are uniformly organized to generate structural relationship data within the slice. This improves the data orderliness, regional discrimination ability, and completeness of the expression of internal relationships within a single slice of multi-slice liver images.

[0035] 2. In this invention, position mapping, contour continuation analysis, and region correspondence processing are performed on corresponding regions in adjacent slices to generate cross-slice continuous relationship data. The liver region unit set, intra-slice structural relationship data, and cross-slice continuous relationship data are jointly organized into liver structure semantic map data. Through the liver structure semantic map boundary write-back convergence recognition algorithm, the boundary consistency results participate in the connection relationship correction and re-propagation. Then, the boundary transition region is subjected to the attribution determination, boundary merging, and region attribution update processing, thereby improving the consistency between region category discrimination and boundary connection state, and enhancing the stability, completeness, and cross-slice continuity of the final liver recognition result. Attached Figure Description

[0036] Figure 1 This is a system architecture diagram of the present invention;

[0037] Figure 2 This is a diagram illustrating the method steps of the present invention. Detailed Implementation

[0038] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0039] Reference Figure 1-2 This invention provides an embodiment of a liver image recognition system based on a graph neural network, comprising: a liver image receiving module for receiving liver image data, performing slice processing, grayscale standardization, and spatial location information association processing on the liver image data to generate standardized liver image data; a region unit generation module for performing liver parenchyma region extraction, candidate abnormal region extraction, and boundary transition region extraction processing on the standardized liver image data to generate a set of liver region units; an intra-slice relationship construction module for performing regional adjacency relationship analysis, grayscale change relationship analysis, and boundary continuity relationship analysis on the set of liver region units, and performing boundary coupling adjustment processing on the analysis results according to boundary coupling adjustment parameters to generate intra-slice structural relationship data; and a cross-slice relationship construction module. The system comprises several modules: a module for constructing a liver structure semantic map; a module for generating a liver structure semantic map; a module for generating a liver structure semantic map; a module for generating a liver structure semantic map; a module for generating a liver structure semantic map; a module for generating a liver structure semantic map; a module for generating a liver structure semantic map result; and a module for generating a liver recognition result.

[0040] In this embodiment, after receiving liver image data, the liver image receiving module performs slice numbering processing according to the acquisition order of each slice in the liver image data, and writes the sequential position of each slice in the acquisition sequence into the corresponding slice number position. At the same time, it reads and processes the original position identifiers corresponding to each slice one by one, so that each slice forms a slice number consistent with the acquisition order and a slice position relationship corresponding to the slice number, thus forming a slice processing result.

[0041] Based on the results of the slice organization, when performing grayscale range unification processing on the image grayscale in each slice, the image grayscale data is read slice by slice according to the slice number, and the image grayscale in each slice is mapped to a consistent grayscale range. Then, grayscale distribution correction processing is performed according to the grayscale distribution state within the same slice, so that the image grayscale in each slice maintains a consistent grayscale expression under the slice number correspondence, forming a grayscale standardization result.

[0042] When reading the spatial location information corresponding to each slice based on the slice organization results, the spatial location information corresponding to the current slice is extracted one by one according to the slice number, and the spatial location information is written into the corresponding slice number record. Then, the correspondence between the written spatial location information and the slice number is organized to ensure that each slice forms a spatial location record corresponding to the current slice number, thus forming a spatial location information association result.

[0043] When integrating the grayscale normalization results with the spatial location information according to the slice number, the image data of each slice in the grayscale normalization results is matched one-to-one with the corresponding spatial location record in the spatial location information association results based on the slice number. The matched image grayscale content and spatial location content are written into the same slice data structure, so that each slice contains both the image content after grayscale normalization and the spatial location information corresponding to that slice, thus generating standardized liver image data.

[0044] In this embodiment, after receiving standardized liver image data, the region unit generation module reads the image content and corresponding spatial location information of each slice according to the slice number, and performs range recognition processing on the liver region in combination with the gray distribution state of each slice. The image part located in a continuous spatial position and whose gray distribution conforms to the main characteristics of the liver is determined as the main range of the liver. Then, the main range of the liver corresponding to each slice is written in correspondence with the slice number and spatial location to form the main range result of the liver.

[0045] Based on the results of the main body of the liver, the image regions within the main body of the liver in the standardized liver image data are processed by region connectivity sorting according to the slice number. The image parts that are continuous in position and continuous in gray level within the same slice are divided into corresponding connected regions. Contour extraction processing is performed on the outer edge of each connected region. The extracted contour content is associated with the corresponding connected region, slice number and spatial position to form the initial region division result.

[0046] Based on the initial region division results, the regions located in the middle of the main body of the liver and with continuous gray-level distribution are processed one by one. Regions with continuous gray-level changes, spatial distribution located in the middle of the main body of the liver and stable connection with surrounding regions are extracted as liver parenchyma regions. The extracted region contents are then organized according to slice number and spatial location to form liver parenchyma region results.

[0047] Based on the initial region segmentation results, corresponding extraction processing is performed on regions located in the middle of the main body of the liver and whose gray-scale distribution deviates from the results of the liver parenchyma region. Regions that deviate from the results of the liver parenchyma region in gray-scale distribution but are still in the middle of the main body of the liver are separated from the initial region segmentation results. The contents of the separated regions are recorded according to the slice number and spatial location to form candidate abnormal region results.

[0048] Based on the results of the liver parenchyma region and the candidate abnormal region, edge transition recognition processing is performed at the boundary position of the two within the same slice. The edge changes at the contact position between the liver parenchyma region result and the candidate abnormal region result are read, and the transition range corresponding to the boundary position is extracted. The region that is at the boundary position and has edge transition characteristics is organized into independent region content to form the boundary transition region result.

[0049] When performing corresponding tissue processing on the liver parenchyma region results, candidate abnormal region results, and boundary transition region results according to slice number and spatial location, the three types of region results are collected one by one according to the spatial location correspondence under the same slice number, and the collected three types of region results are written into a unified region organization content. This ensures that the liver parenchyma region results, candidate abnormal region results, and boundary transition region results under the same slice number maintain the positional correspondence and category correspondence, generating a set of liver region units.

[0050] In this embodiment, after receiving the set of liver region units, the slice relationship construction module performs the same-slice aggregation processing on the liver parenchyma region results, candidate abnormal region results, and boundary transition region results in the set of liver region units according to the slice number. It reads the spatial location content corresponding to the liver parenchyma region results, candidate abnormal region results, and boundary transition region results one by one according to the same slice number, and writes the region content under the same slice number into the same slice aggregation record, so that the region results in the same slice are organized under the same slice number, forming the slice region aggregation result.

[0051] When performing region contact identification and region spacing determination on adjacent regions in the same slice based on the region aggregation results within the slice, the spatial positional relationship between each region is read one by one according to the same slice number, and the corresponding identification process is performed on whether the edges of the regions are in contact. Then, the region spacing determination process is performed on the regions that are not in direct contact but are spatially adjacent. The contact status and spacing status between each region are written into the associated record according to the slice number and the corresponding region position to form the region adjacency relationship analysis results.

[0052] When performing boundary position grayscale reading and grayscale change trend processing on adjacent regions in the same slice based on the region aggregation results within the slice, the corresponding image grayscale content is read point by point along the boundary position between adjacent regions. The grayscale change situation on both sides of the boundary position is processed according to the continuous grayscale change direction at the boundary position. The grayscale change content corresponding to the boundary position is written into the analysis record according to the slice number and the regional position relationship to form the grayscale change relationship analysis result.

[0053] When performing boundary edge extraction and edge extension direction correspondence processing on adjacent regions in the same slice based on the region aggregation results within the slice, the corresponding edge content is extracted along the boundary position between adjacent regions, and the edge extension direction is correspondingly organized according to the continuous extension state of the edge in the current slice. The extracted edge content and edge extension direction content are written into the corresponding record according to the slice number and regional position relationship to form the boundary continuity relationship analysis result.

[0054] Based on the results of regional adjacency analysis, grayscale change analysis, and boundary continuity analysis, when performing boundary coupling adjustment processing according to the boundary coupling adjustment parameters, the corresponding results of regional adjacency analysis, grayscale change analysis, and boundary continuity analysis for the same adjacent region are read according to the slice number. The correspondence between the three types of analysis results is adjusted and organized according to the boundary coupling adjustment parameters, so that the boundary relationship content corresponding to the same adjacent region forms a unified correspondence between the regional contact state, grayscale change state, and edge extension state, forming the intra-slice relationship adjustment result.

[0055] When performing connectivity processing on adjacent regions within the same slice based on the results of intra-slice relation adjustment, the content of the relations between adjacent regions after adjustment is organized one by one according to the slice number, and the connection content between adjacent regions is written into the intra-slice connectivity record. This allows the results of liver parenchyma region, candidate abnormal region, and boundary transition region to form connection content corresponding to the slice number within the same slice, generating intra-slice structural relationship data.

[0056] In this embodiment, after receiving the set of liver region units, the cross-slice relationship construction module performs front-to-back slice correspondence processing on adjacent slices in the set of liver region units according to the slice number. It reads the liver parenchyma region results, candidate abnormal region results, and boundary transition region results in the front and back slices one by one according to the slice number order, and writes the region results in adjacent slices into the same cross-slice processing record according to the front-to-back slice correspondence, so that the region results in adjacent slices are processed according to the front-to-back slice number relationship, forming the cross-slice region processing result.

[0057] When performing regional location correspondence analysis on the liver parenchyma region results, candidate abnormal region results, and boundary transition region results in adjacent slices based on the cross-slice region organization results, the spatial location content of the corresponding region in the adjacent slices is read one by one according to the correspondence between the preceding and following slices. The region position in the preceding slice is compared with the region position in the following slice. Then, the region position correspondence content after comparison is written into the location association record according to the slice number and region category, so that the same type of region in the adjacent slices forms a position correspondence between the preceding and following slices, forming a location mapping processing result.

[0058] When performing region contour extension direction reading and region contour continuous change processing on the corresponding regions in adjacent slices based on the cross-slice region processing results, the contour content of the corresponding regions in the preceding and following slices is read one by one according to the slice number, and the region contour extension direction is read along the corresponding direction from the preceding slice to the following slice. Then, the change state of the region contour between the preceding and following slices is processed continuously. The processed contour extension content and contour continuous change content are written into the analysis record according to the slice number and region correspondence to form the contour continuity analysis result.

[0059] When performing region overlap range comparison processing and region change continuity determination processing on corresponding regions in adjacent slices based on the cross-slice region organization results, the spatial coverage content of corresponding regions in the preceding and following slices is read one by one according to the slice number, and the overlap range of corresponding regions between the preceding and following slices is compared accordingly. Then, the region change is determined to maintain continuity based on the overlap range change status. The compared overlap range content and the determined continuous change content are written into the association record according to the slice number and region correspondence to form the region correspondence association result.

[0060] Based on the location mapping processing results, contour continuation analysis results, and region correspondence results, when performing continuous compensation adjustment processing according to the continuous compensation adjustment parameters, the three types of analysis results of the corresponding regions of the same preceding and following slices are read according to the slice number. The correspondence between the region location correspondence content, contour continuation content, and region continuous change content is adjusted and sorted according to the continuous compensation adjustment parameters, so that the cross-slice relationship of the corresponding regions in adjacent slices forms a unified correspondence between position change, contour change, and region continuous change, thus forming the cross-slice relationship adjustment result.

[0061] When performing continuous connectivity processing on corresponding regions in adjacent slices based on cross-slice relationship adjustment results, the corresponding region relationship content of the preceding and following slices after adjustment is sorted out one by one according to the slice number, and the continuous connectivity content between the corresponding regions is written into the cross-slice connectivity record. This enables the liver parenchyma region results, candidate abnormal region results, and boundary transition region results to form continuous connectivity content corresponding to the slice number between adjacent slices, generating cross-slice continuous relationship data.

[0062] In this embodiment, after receiving the set of liver region units, intra-slice structural relationship data, and cross-slice continuous relationship data, the liver structure semantic graph generation module performs node aggregation processing on the liver parenchyma region results, candidate abnormal region results, and boundary transition region results in the set of liver region units according to the slice number and spatial location. It reads the liver parenchyma region results, candidate abnormal region results, and boundary transition region results in each slice one by one according to the slice number, and writes the spatial location content corresponding to each region result under the same slice number into the same node organization record, so that each region result completes the node-level corresponding organization under the slice number and spatial location correspondence, forming the graph node organization result;

[0063] When performing node identifier writing and node position correspondence processing on the liver parenchyma region results, candidate abnormal region results, and boundary transition region results based on the graph node organization results, the slice number, spatial location, and region category content corresponding to each region result are read one by one according to the graph node organization results, and the corresponding node identifier content is written to each region result. Then, the node identifier content and the corresponding spatial location content are written one by one, so that each region result forms a node record in the graph node organization results corresponding to the slice number, spatial location, and region category, forming a node organization result;

[0064] When performing intra-slice connection correspondence processing and intra-slice connection merging processing on adjacent regions in the same slice based on the intra-slice structural relationship data, the connection content of adjacent regions in the same slice is read one by one according to the slice number, and the connection start position, connection corresponding position and connection end position between adjacent regions are sorted out accordingly. Then, the connection content formed under the same slice number is merged and written, so that the connection content of adjacent regions in the same slice forms an edge connection record corresponding to the slice number, forming the intra-slice edge organization result.

[0065] When performing cross-slice connection correspondence processing and cross-slice connection merging processing on corresponding regions in adjacent slices based on cross-slice continuous relationship data, the corresponding region connection content in adjacent slices is read one by one according to the correspondence between the preceding and following slices. The connection content between the region positions in the preceding slice and the region positions in the following slice is then organized and the corresponding connection content formed between adjacent slices is merged and written, so that the corresponding region connection content in adjacent slices forms edge connection records corresponding to the relationship between the preceding and following slices, thus forming a cross-slice edge organization result.

[0066] When performing node-edge correspondence integration processing based on node organization results, slice intra-edge organization results, and cross-slice edge organization results, the node identifier content and node position content in the node organization results are read one by one, and the slice intra-edge connection content in the slice intra-edge organization results and the cross-slice connection content in the cross-slice edge organization results are read. Then, the content of each node is written to the corresponding slice intra-edge connection content and cross-slice connection content, so that each node in the node organization results is uniformly associated with the corresponding slice intra-edge organization results and cross-slice edge organization results, forming a graph structure integration result.

[0067] When performing graph data organization and processing based on the graph structure integration results, the graph data is written and organized one by one according to the node content and connection content in the graph structure integration results. The intra-slice connection relationship and cross-slice connection relationship between nodes are also written into the graph data record, so that the node content and connection content in the graph structure integration results form a complete corresponding graph data expression, generating liver structure semantic graph data.

[0068] In this embodiment, after receiving the liver structure semantic graph data, the graph propagation recognition module reads the graph nodes and connections in the liver structure semantic graph data one by one according to the slice number and spatial location. It then performs corresponding processing on the region category content, slice number content, and spatial location content corresponding to each graph node, and the intra-slice connection content and cross-slice connection content corresponding to each connection relationship. This ensures that the same graph node and its associated connections are processed before propagation based on the slice number and spatial location correspondence, forming the graph propagation preparation result. The formula for forming the graph propagation preparation result is as follows: ;in, Graph nodes The corresponding graph propagation preparation results, Graph nodes The corresponding regional category content, Graph nodes The corresponding slice number content, Graph nodes The corresponding spatial location content, Graph nodes Nodes adjacent to the same slice The content of the slice inner joins between them Graph nodes Graph nodes corresponding to adjacent slices Cross-slice connection content, Graph nodes The set of adjacent graph nodes within a slice, Graph nodes The set of adjacent graph nodes across slices. : Concatenation operation, used to organize the content of graph nodes and connection relationships into a unified input expression.

[0069] When processing the liver structure semantic graph boundary write-back convergence recognition algorithm based on the graph propagation preparation results, the adjacent connection content of each graph node is read one by one according to the graph node correspondence and connection relationship in the graph propagation preparation results. Then, propagation update processing is performed on the node information between adjacent graph nodes along the intra-slice connection relationship and cross-slice connection relationship. Finally, the propagated node content is written into the corresponding graph node record, so that each graph node forms an updated node expression under the action of adjacent connection relationships, forming the node propagation update result; the formula for forming the node propagation update result is as follows: ;in, Graph nodes The results are updated at each node after the initial propagation. : A non-linear update function used to complete the mapping after node information propagation. : The update mapping matrix for the content of the current graph nodes. : The propagation mapping matrix of the content of adjacent graph nodes within a slice. : The propagation mapping matrix of content between adjacent graph nodes across slices Graph nodes With graph nodes The propagation coefficient of intra-slice connectivity between slices, Graph nodes With graph nodes The cross-slice connection propagation coefficient between them Graph nodes The corresponding graph propagation preparation results, Graph nodes The corresponding graph propagation preparation results.

[0070] Slice intra-connection propagation coefficient: Cross-slice connection propagation coefficient: ;in, : The vector for calculating the propagation coefficients of the intra-slice connectivity. : The vector for calculating the cross-slice connection propagation coefficients. : Exponential mapping function.

[0071] When performing region category correspondence processing on the liver parenchyma region results, candidate abnormal region results, and boundary transition region results based on node propagation update results, the propagation update content of each graph node is read one by one according to the slice number and spatial location. The graph nodes corresponding to the liver parenchyma region results, the candidate abnormal region results, and the boundary transition region results are then processed for category correspondence writing, ensuring that the node content after propagation maintains a consistent correspondence with the corresponding region category, thus forming the initial region category discrimination result. The formula for forming the initial region category discrimination result is as follows: ; ;in, Graph nodes The initial region category discrimination vector, Graph nodes The initial region category determination results Category normalization function Region category discrimination mapping matrix Region category discrimination bias vector : The operation that retrieves the position of the largest category.

[0072] If the results for the liver parenchyma region, the candidate abnormal region, and the boundary transition region are respectively categorized as Category 1, Category 2, and Category 3, then: Among them, 1: the category corresponding to the result of the liver parenchyma region, 2: the category corresponding to the result of the candidate abnormal region, and 3: the category corresponding to the result of the boundary transition region.

[0073] When performing boundary correspondence processing on the connection relationships between the boundary transition region results and adjacent liver parenchyma region results and candidate abnormal region results based on the node propagation update results, the connection content between the corresponding graph nodes of the boundary transition region results and adjacent graph nodes is read one by one according to the slice number and spatial location. The connection direction, connection position, and connection correspondence status between the boundary transition region results and adjacent liver parenchyma region results and candidate abnormal region results are then processed and written to ensure that the boundary transition region results and adjacent regions form boundary correspondence content consistent with the connection relationships, thus forming the initial boundary consistency result. The formula for forming the initial boundary consistency result is as follows: ;in, Graph nodes The corresponding initial boundary consistency results, : Connection direction consistency weight, : Connectivity consistency weight, : Connection to the corresponding state consistency weight, Graph nodes The connection direction vector corresponding to the boundary transition region results. Graph nodes In-neighbor reference connection direction vector, Similarity of connection directions Graph nodes Spatial location content, : Reference spatial location information corresponding to adjacent areas : L2 norm, Graph nodes The connection corresponds to the state variables. : Exponential mapping function.

[0074] If the boundary consistency result needs to be restricted to the range of 0 to 1, it can be written as: ;in, The initial boundary consistency result after normalization.

[0075] When calculating the boundary write-back adjustment parameters based on the initial region category discrimination results and initial boundary consistency results, the initial region category discrimination content corresponding to the boundary transition region results is read one by one according to the slice number and spatial location. The initial boundary consistency content between the same boundary transition region results and adjacent liver parenchymal region results and candidate abnormal region results is also read. A joint comparison is performed on the stability of the boundary transition region results in region category discrimination and the degree of agreement in boundary consistency. Then, based on the joint comparison, the adjustment amount of the corresponding connectivity relationship of the same boundary transition region results is determined, generating the boundary write-back adjustment parameters. The formula for calculating the boundary write-back adjustment parameters is as follows: ;in, Graph nodes The corresponding boundary write-back adjustment parameters, Adjustment weights for the stability of regional category discrimination. Adjustment weights for the degree of boundary consistency fit. Graph nodes The stability of regional category discrimination The initial boundary consistency result after normalization. The stability of region category discrimination can be further expressed as: ;in, The largest component in the initial region category discriminant vector. The second largest component in the initial region category discriminant vector. Graph nodes The degree of stability in the initial regional category determination.

[0076] Based on the initial region category discrimination results and initial boundary consistency results, when performing write-back correction processing on the connection relationships corresponding to the boundary transition region results according to the boundary write-back adjustment parameters, the initial region category discrimination content and initial boundary consistency content of the graph nodes corresponding to the boundary transition region results are read one by one according to the slice number and spatial location. Then, according to the boundary write-back adjustment parameters, the connection relationships between the boundary transition region results and the adjacent liver parenchyma region results and the candidate abnormal region results are corrected accordingly. Finally, the corrected connection relationships are written into the corresponding graph node records and connection records in the liver structure semantic map data, forming the write-back corrected liver structure semantic map data. The connection relationship write-back correction formula is as follows: ;in, Graph nodes With adjacent graph nodes The corrected connection relationship between them Graph nodes With adjacent graph nodes The connection relationship before correction Graph nodes The corresponding boundary write-back adjustment parameters, Graph nodes With adjacent graph nodes The target connectivity relationships are formed based on the initial region category discrimination results and the initial boundary consistency results.

[0077] To write the graph node records and connection records separately, the corrected liver structure semantic graph data can be written back as follows: ;in, : Write back the corrected semantic map data of liver structure. : Write back the corrected set of graph node records. : Write back the corrected set of connection records.

[0078] Based on the liver structure semantic map data after write-back correction, the liver structure semantic map boundary write-back convergence recognition algorithm is executed again. When performing region category correspondence and boundary correspondence processing on the propagated graph nodes, the updated node content and connection content are read one by one according to the written-back graph node records and connection records. Then, the revised connection relationships are used to perform a second propagation update process on each graph node, and the second propagation graph node content is written into the corresponding region category record and boundary correspondence record. This ensures that the written-back corrected liver structure semantic map data forms stable corresponding region category content and boundary correspondence content after the second propagation, generating region category discrimination results and boundary consistency results. The second propagation update formula is: ;in, Graph nodes The content at the node after further propagation : The propagation mapping matrix of the content of the current graph node. : The re-propagation mapping matrix of the content of adjacent graph nodes Graph nodes The set of adjacent graph nodes in the rewritten and corrected semantic graph data of liver structure. Graph nodes With graph nodes The repropagation coefficients are obtained based on the modified connection relationship. Graph nodes The initial transmission update results.

[0079] The repropagation coefficient can be expressed as: ,in, : The vector for calculating the second propagation coefficients. : Corrected connection relationship. Formula for region category determination result: Formula for generating boundary consistency results: ;in, Graph nodes The final region category discrimination vector, Graph nodes The region category discrimination results : Final region category discrimination mapping matrix : Final region category discrimination bias vector.

[0080] After the second propagation, the boundary correspondences are processed and can be written as follows: ;in, Graph nodes The final boundary consistency result, Consistency weights in the connection direction after repropagation Consistency weight of connection positions after repropagation : After repropagation, connect the corresponding state consistency weights. : Write back the corrected graph nodes The connection direction vector, : Write back and correct the reference connection direction vector : Write back the corrected graph nodes Spatial location content, : Write back and correct the reference spatial location content. : Write back the corrected graph nodes The connection corresponds to the state variables.

[0081] In this embodiment, after receiving the liver structure semantic graph data, the graph propagation recognition module reads the graph nodes and connections in the liver structure semantic graph data one by one according to the slice number and spatial location. It then performs corresponding processing on the region category content, slice number content, and spatial location content corresponding to each graph node and the intra-slice connection content and cross-slice connection content corresponding to each connection relationship to form the graph propagation preparation result. The improvement of the algorithm in this application is first reflected in the fact that the liver parenchyma region result, candidate abnormal region result, and boundary transition region result are not entered into the recognition process as separate region content, but are organized together with the intra-slice connection content and cross-slice connection content in the liver structure semantic graph data into a unified graph propagation entry point, so that the boundary transition region result participates in the connection relationship constraint before the propagation begins.

[0082] When processing the liver structure semantic graph boundary write-back convergence recognition algorithm based on the graph propagation preparation results, the adjacent connection content of each graph node is read one by one according to the graph node correspondence and connection relationship in the graph propagation preparation results. The node information between adjacent graph nodes is propagated and updated along the intra-slice connection relationship and cross-slice connection relationship. Then, the propagated node content is written into the corresponding graph node record to form the node propagation update result. The improvement of the algorithm in this application is further reflected in that the node propagation update result is not only generated based on the static connection relationship between graph nodes, but is based on the fact that the intra-slice structural relationship data and cross-slice continuous relationship data of the previous slice have been organized. This makes the boundary transition region result constrained by the intra-slice connection relationship and the cross-slice connection relationship at the same time, thereby enhancing the continuous expression ability of the boundary region in the propagation process.

[0083] When performing region category correspondence processing on the liver parenchyma region results, candidate abnormal region results, and boundary transition region results based on the node propagation update results, the propagation update content of each graph node is read one by one according to the slice number and spatial location. The graph nodes corresponding to the liver parenchyma region results, the graph nodes corresponding to the candidate abnormal region results, and the graph nodes corresponding to the boundary transition region results are respectively processed for category correspondence writing to form the initial region category discrimination results. The improvement of the algorithm in this application is also reflected in that the boundary transition region results are kept in an independent record state in the initial region category discrimination results and are not directly written into the liver parenchyma region results or candidate abnormal region results after the initial propagation, so that the boundary transition region results continue to serve as the data basis for subsequent boundary consistency analysis and connection relationship write-back correction.

[0084] When performing boundary correspondence processing on the connection relationship between the boundary transition region results and the adjacent liver parenchyma region results and the candidate abnormal region results based on the node propagation update results, the connection content between the graph nodes corresponding to the boundary transition region results and the adjacent graph nodes is read one by one according to the slice number and spatial location. The connection direction, connection position and connection correspondence status between the boundary transition region results and the adjacent liver parenchyma region results and the candidate abnormal region results are processed and written to form the initial boundary consistency result. The improvement of the algorithm in this application is also reflected in that the initial boundary consistency result is not left as a separate result verification content in the output layer, but is used as the data source for the subsequent boundary write-back adjustment parameter calculation and directly participates in the graph structure correction.

[0085] When calculating the boundary write-back adjustment parameters based on the initial region category discrimination results and the initial boundary consistency results, the initial region category discrimination content corresponding to the boundary transition region results is read one by one according to the slice number and spatial location. The initial boundary consistency content between the same boundary transition region results and the results of adjacent liver parenchyma regions and candidate abnormal regions is also read. A joint comparison process is performed on the degree of stability of the boundary transition region results in the region category discrimination and the degree of consistency in the boundary consistency. Then, the adjustment amount of the connection relationship corresponding to the same boundary transition region results is determined based on the joint comparison process, and the boundary write-back adjustment parameters are generated. The improvement of the algorithm in this application is particularly reflected in the fact that the boundary write-back adjustment parameters are jointly driven by the initial region category discrimination results and the initial boundary consistency results, so that the region category propagation results and the boundary connection state jointly participate in the determination of the subsequent connection relationship correction magnitude.

[0086] When performing write-back correction processing on the connection relationships corresponding to the boundary transition region results based on the boundary write-back adjustment parameters, the initial region category discrimination content and initial boundary consistency content of the graph nodes corresponding to the boundary transition region results are read one by one according to the slice number and spatial location. Then, the connection relationships between the boundary transition region results and the results of adjacent liver parenchyma regions and candidate abnormal regions are corrected according to the boundary write-back adjustment parameters. Finally, the corrected connection relationships are written into the corresponding graph node records and connection records in the liver structure semantic map data to form the liver structure semantic map data after write-back correction. The core improvement of the algorithm in this application is that the boundary transition region results are transformed from propagation discrimination objects into connection relationship correction trigger objects, so that the initial boundary consistency results are written back into the liver structure semantic map data and the subsequent propagation basis is changed.

[0087] Based on the liver structure semantic map data after write-back correction, the liver structure semantic map boundary write-back convergence recognition algorithm is executed again. When performing region category correspondence and boundary correspondence processing on the propagated graph nodes, the updated node content and connection content are read one by one according to the written-back graph node record and connection record. Then, the graph node is propagated and updated again along the corrected connection relationship. The propagated graph node content is written into the corresponding region category record and boundary correspondence record respectively to generate region category discrimination result and boundary consistency result. The improvement of the algorithm in this application is also reflected in that the second propagation is not a recalculation of the initial liver structure semantic map data, but a second propagation is performed on the written-back corrected liver structure semantic map data, so that the influence path between the boundary transition region result and the liver parenchyma region result and the candidate abnormal region result is redistributed along the corrected connection relationship.

[0088] After receiving the region category discrimination result and the boundary consistency result, the recognition result output module reads the liver parenchyma region result, candidate abnormal region result, and boundary transition region result from the region category discrimination result one by one according to the slice number and spatial location. It then performs boundary attribution determination processing, boundary merging processing, and region attribution update processing in combination with the boundary consistency result to form the region correction result, the recognition region sorting result, and the liver recognition result. The improvement of the algorithm in this application is ultimately reflected in the fact that the liver recognition result is not directly formed based on the initial region category discrimination result, but is based on the convergence result formed by the region category discrimination result and the boundary consistency result after boundary write-back correction and re-propagation. This ensures that the final output maintains both region category stability and boundary connection consistency.

[0089] In this embodiment, after receiving the region category discrimination result and the boundary consistency result, the identification result output module reads the liver parenchyma region result, candidate abnormal region result and boundary transition region result in the region category discrimination result one by one according to the slice number and spatial location. Then, it performs corresponding sorting processing on the slice number content, spatial location content and region category content corresponding to each region result, and writes the attribution content of each region result under the same slice number into the same region attribution record, so that each region result forms a clear region attribution corresponding content under the correspondence between slice number and spatial location, forming a region attribution sorting result.

[0090] Based on the regional attribution and boundary consistency results, when performing boundary attribution determination processing on the boundary correspondence between the boundary transition region results and the adjacent liver parenchyma region results and the candidate abnormal region results, the regional attribution content corresponding to the boundary transition region results is read one by one according to the slice number and spatial location, and the boundary correspondence content corresponding to the current boundary transition region results is read from the boundary consistency results. Then, the boundary connection position, boundary correspondence direction and boundary correspondence status between the boundary transition region results and the adjacent liver parenchyma region results and the candidate abnormal region results are compared and processed to form a clear boundary attribution correspondence between the boundary transition region results and the adjacent regions, thus forming the boundary correction result.

[0091] Based on the boundary correction results, boundary merging processing is performed on the boundary transition region results. When updating the boundary transition region results to the liver parenchyma region results or candidate abnormal region results, the boundary assignment corresponding content in the boundary correction results is read one by one according to the slice number and spatial location. The part of the boundary transition region results that is consistent with the liver parenchyma region results is merged into the liver parenchyma region results. Then, the part of the boundary transition region results that is consistent with the candidate abnormal region results is merged into the candidate abnormal region results. At the same time, the region assignment content after merging is updated and written, so that the boundary transition region results are transferred to the liver parenchyma region results or candidate abnormal region results after merging, forming the region correction results.

[0092] When performing intra-slice identification region integration processing and cross-slice identification region correspondence processing based on the region correction results and region category discrimination results, the liver parenchyma region results and candidate abnormal region results in the region correction results are read one by one according to the slice number. Combined with the region category content that has not been updated in the region category discrimination results, the position correspondence sorting processing and category merging sorting processing are performed on the corrected identification regions within the same slice. Then, the front and back slice correspondence sorting processing and continuous position correspondence sorting processing are performed on the corrected identification regions between adjacent slices. This makes the identification regions in each slice and the identification regions between adjacent slices form continuous and consistent identification region content under the slice number and spatial position correspondence relationship, forming the identification region sorting result.

[0093] When generating identification results based on the identified region organization results, the identified region content in each slice is read sequentially according to the slice number. The liver parenchyma region results and candidate abnormal region results in the same slice are then uniformly organized with the corresponding identified region content in adjacent slices. The organized identified region content is then written into the identification result record according to the slice number and spatial location. This process, after boundary correction and region assignment updates, forms a complete and corresponding identification result expression, generating the liver identification result. The formula for generating liver identification results is as follows: ;in, Liver identification results Graph nodes The corresponding slice number content, Graph nodes The corresponding spatial location content, Graph nodes The region category discrimination results Graph nodes Boundary consistency results : A set of graph nodes.

[0094] Example 1: In this specific embodiment, the liver image receiving module receives a set of liver image data arranged in the acquisition order. It sequentially performs slice numbering processing on each slice in the liver image data and writes the original location identifier corresponding to each slice into the slice number record corresponding to that slice. Then, it performs grayscale range unification processing and grayscale distribution correction processing on the image grayscale of each slice. At the same time, it performs association processing with the slice number on the spatial location information corresponding to each slice. Subsequently, it performs corresponding integration processing on the grayscale standardization result and the spatial location information association result according to the slice number to generate standardized liver image data.

[0095] After receiving standardized liver image data, the region unit generation module reads the image content and corresponding spatial location information of each slice according to the slice number. Based on the grayscale distribution and spatial location of each slice, it performs liver main body range identification processing to form the liver main body range result. Then, based on the liver main body range result, it performs region connectivity processing and region contour extraction processing on the standardized liver image data to form the initial region division result. Subsequently, it performs liver parenchyma region extraction processing on the region located in the middle of the liver main body with continuous grayscale distribution to form the liver parenchyma region result. It performs candidate abnormal region extraction processing on the region located in the middle of the liver main body with grayscale distribution deviating from the liver parenchyma region result to form the candidate abnormal region result. Then, based on the liver parenchyma region result and the candidate abnormal region result, it performs edge transition identification processing and boundary range extraction processing on the boundary position between the two to form the boundary transition region result. Finally, it performs corresponding tissue processing on the liver parenchyma region result, candidate abnormal region result, and boundary transition region result according to the slice number and spatial location to generate a set of liver region units.

[0096] After receiving the set of liver region units, the intra-slice relationship construction module performs intra-slice aggregation processing on the liver parenchyma region results, candidate abnormal region results, and boundary transition region results in the liver region unit set according to the slice number, forming intra-slice region aggregation results. Then, based on the intra-slice region aggregation results, it performs region contact recognition processing and region spacing determination processing on adjacent regions in the same slice, forming region adjacency relationship analysis results. It performs boundary position grayscale reading processing and grayscale change trend sorting processing on adjacent regions in the same slice, forming grayscale change relationship analysis results. It performs boundary edge extraction processing and edge extension direction correspondence processing on adjacent regions in the same slice, forming boundary continuity relationship analysis results. Subsequently, based on the boundary coupling adjustment parameters, it performs boundary coupling adjustment processing on the region adjacency relationship analysis results, grayscale change relationship analysis results, and boundary continuity relationship analysis results, forming intra-slice relationship adjustment results. Finally, based on the intra-slice relationship adjustment results, it performs connection relationship organization processing on adjacent regions in the same slice, generating intra-slice structural relationship data.

[0097] After receiving the set of liver region units, the cross-slice relationship construction module performs front-to-back slice correspondence processing on adjacent slices in the liver region unit set according to the slice number, forming a cross-slice region processing result. Then, based on the cross-slice region processing result, it performs region position correspondence analysis processing on the liver parenchyma region results, candidate abnormal region results, and boundary transition region results in adjacent slices, forming a position mapping processing result. It performs region contour extension direction reading processing and region contour continuous change processing processing on corresponding regions in adjacent slices, forming a contour continuity analysis result. It performs region overlap range comparison processing and region change continuity judgment processing on corresponding regions in adjacent slices, forming a region correspondence association result. Subsequently, it performs continuous compensation adjustment processing on the position mapping processing result, contour continuity analysis result, and region correspondence association result according to continuous compensation adjustment parameters, forming a cross-slice relationship adjustment result. Finally, based on the cross-slice relationship adjustment result, it performs continuous connection relationship tissue processing on corresponding regions in adjacent slices, generating cross-slice continuous relationship data.

[0098] The liver structure semantic graph generation module receives a set of liver region units, intra-slice structural relationship data, and cross-slice continuous relationship data. Based on the slice number and spatial location, it performs node aggregation processing on the liver parenchyma region results, candidate abnormal region results, and boundary transition region results within the main body of the liver, forming a graph node organization result. Then, based on the graph node organization result, it performs node identifier writing processing and node position mapping processing, forming a node organization result. Based on the intra-slice structural relationship data, it performs intra-slice connection mapping processing and intra-slice connection merging processing, forming an intra-slice edge organization result. Based on the cross-slice continuous relationship data, it performs cross-slice connection mapping processing and cross-slice connection merging processing, forming a cross-slice edge organization result. Subsequently, it performs node edge mapping integration processing on the node organization result, intra-slice edge organization result, and cross-slice edge organization result, forming a graph structure integration result. Finally, based on the graph structure integration result, it performs graph data organization processing to generate the liver structure semantic graph data.

[0099] After receiving the liver structure semantic map data, the graph propagation recognition module processes the liver structure semantic map data using a liver structure semantic map boundary write-back convergence recognition algorithm. First, it performs propagation preparation processing on the graph nodes and connections in the liver structure semantic map data based on slice numbers and spatial locations, forming a graph propagation preparation result. Then, based on the graph propagation preparation result, it performs graph neural network propagation processing, forming a node propagation update result. Next, based on the node propagation update result, it performs region category matching processing on the liver parenchyma region results, candidate abnormal region results, and boundary transition region results, forming an initial region category discrimination result. Simultaneously, based on the node propagation update result, it compares the boundary transition region results with adjacent liver parenchyma... The connection relationships between regional results and candidate abnormal regional results are processed by boundary correspondence rectification to form an initial boundary consistency result. Then, based on the initial regional category discrimination result and the initial boundary consistency result, the connection relationships corresponding to the boundary transition regional results are corrected by writing back according to the boundary write-back adjustment parameters, and the write-back correction result is written into the liver structure semantic map data to form the write-back corrected liver structure semantic map data. Finally, the graph neural network propagation process is performed again based on the write-back corrected liver structure semantic map data, and the regional category correspondence rectification and boundary correspondence rectification are performed on the propagated graph nodes to generate regional category discrimination result and boundary consistency result.

[0100] After receiving the region category discrimination results and boundary consistency results, the recognition result output module performs region attribution reading processing on the liver parenchyma region results, candidate abnormal region results, and boundary transition region results in the region category discrimination results according to the slice number and spatial location, forming a region attribution consolidation result. Then, based on the region attribution consolidation result and boundary consistency results, it performs boundary attribution determination processing on the boundary correspondence between the boundary transition region results and adjacent liver parenchyma region results and candidate abnormal region results, forming a boundary correction result. Subsequently, based on the boundary correction result, it performs boundary merging processing on the boundary transition region results and performs attribution update processing on the boundary transition region results to either the liver parenchyma region results or the candidate abnormal region results, forming a region correction result. Next, based on the region correction result and region category discrimination results, it performs intra-slice recognition region integration processing and cross-slice recognition region correspondence processing, forming a recognition region consolidation result. Finally, based on the recognition region consolidation result, it performs recognition result generation processing to generate the liver recognition result. This specific embodiment is suitable for recognition scenarios where the distribution of liver parenchyma region results and candidate abnormal region results is relatively stable in most slices and the continuity of boundary transition region results is good.

[0101] Example 2: In this specific embodiment, the liver image receiving module receives another set of liver image data, which exhibits more pronounced regional contour changes between adjacent slices. The module first performs slice numbering and positional matching processing on the liver image data according to the acquisition order, forming a slice matching result. Then, based on the slice matching result, it performs grayscale range unification and grayscale distribution correction processing on the image grayscale of each slice, forming a grayscale standardization result. Simultaneously, it reads the spatial location information corresponding to each slice and performs association writing processing between the spatial location information and the slice number, forming a spatial location information association result. Finally, it performs corresponding integration processing on the grayscale standardization result and the spatial location information association result according to the slice number, generating standardized liver image data.

[0102] After receiving standardized liver image data, the region unit generation module first performs liver main body range identification processing based on the grayscale distribution and spatial location of each slice in the standardized liver image data to form the liver main body range result. Then, based on the liver main body range result, it performs region connectivity processing and region contour extraction processing on the standardized liver image data to form the initial region division result. Subsequently, it extracts the liver parenchyma region result from the initial region division result, and extracts the region whose grayscale distribution deviates from the liver parenchyma region result but is still within the main body of the liver as the candidate abnormal region result. It further performs edge transition identification processing and boundary range extraction processing on the boundary position between the liver parenchyma region result and the candidate abnormal region result to form the boundary transition region result. Finally, it performs corresponding tissue processing on the three types of region results according to the slice number and spatial location to generate a set of liver region units. Compared with the first specific embodiment, the boundary transition region result in this specific embodiment has a higher proportion between adjacent slices, so the subsequent cross-slice continuous relationship data has a more prominent impact on the identification result.

[0103] After receiving the set of liver region units, the intra-slice relationship construction module first performs intra-slice aggregation processing on the three types of region results according to the slice number, forming intra-slice region aggregation results; then, it performs region contact recognition processing and region spacing determination processing on adjacent regions in the same slice, forming region adjacency relationship analysis results; it performs grayscale reading processing and grayscale change trend sorting processing on the boundary positions, forming grayscale change relationship analysis results; it performs extraction processing and edge extension direction correspondence processing on the boundary edges, forming boundary continuity relationship analysis results; subsequently, it performs boundary coupling adjustment processing on the above three types of analysis results according to the boundary coupling adjustment parameters, forming intra-slice relationship adjustment results; finally, it generates intra-slice structural relationship data. Because the distribution of boundary transition region results in this specific embodiment is more complex, the connection content between boundary transition region results and candidate abnormal region results, as well as the connection content between boundary transition region results and liver parenchyma region results in the intra-slice structural relationship data, will simultaneously maintain a strong correlation.

[0104] After receiving the set of liver region units, the cross-slice relationship construction module performs corresponding processing on adjacent slices based on slice numbers to form cross-slice region processing results. Then, based on the cross-slice region processing results, it performs region position correspondence analysis on corresponding regions in adjacent slices to form position mapping results. Next, it performs region contour extension direction reading and continuous contour change processing on corresponding regions in adjacent slices to form contour continuity analysis results. Finally, it performs region overlap range comparison and continuous change determination processing on corresponding regions in adjacent slices to form region correspondence association results. Subsequently, it performs continuous compensation adjustment processing on the position mapping results, contour continuity analysis results, and region correspondence association results based on continuous compensation adjustment parameters to form cross-slice relationship adjustment results. Finally, based on the cross-slice relationship adjustment results, it performs continuous connection relationship organization processing on corresponding regions in adjacent slices to generate cross-slice continuous relationship data. This specific embodiment highlights the role of cross-slice continuous relationship data in maintaining the continuous expression of boundary transition region results.

[0105] The liver structure semantic graph generation module receives a set of liver region units, intra-slice structural relationship data, and cross-slice continuous relationship data. First, it performs node aggregation processing on the three types of region results based on slice number and spatial location, forming a graph node organization result. Then, it performs node identifier writing and node position mapping processing, forming a node organization result. Based on the intra-slice structural relationship data, it forms an intra-slice edge organization result; based on the cross-slice continuous relationship data, it forms a cross-slice edge organization result. Next, it performs node edge mapping integration processing on the node organization result, intra-slice edge organization result, and cross-slice edge organization result, forming a graph structure integration result. Finally, it generates liver structure semantic graph data. In this specific embodiment, the number of connection contents corresponding to the boundary transition region results in the cross-slice edge organization result is relatively large, allowing the graph nodes to maintain a more complete continuous relationship expression between preceding and subsequent slices.

[0106] After receiving the liver structure semantic map data, the graph propagation recognition module first performs propagation preparation processing on the graph nodes and connections based on the slice number and spatial location, forming a graph propagation preparation result. Then, based on the graph propagation preparation result, it performs graph neural network propagation processing, forming a node propagation update result. Subsequently, it forms an initial region category discrimination result and an initial boundary consistency result. Next, based on the initial region category discrimination result and the initial boundary consistency result, it calculates boundary write-back adjustment parameters and performs write-back correction processing on the connections corresponding to the boundary transition region results, forming write-back corrected liver structure semantic map data. Finally, it performs graph neural network propagation processing again, as well as region category correspondence processing and boundary correspondence processing, generating region category discrimination results and boundary consistency results. In this specific embodiment, the boundary transition region results are kept independently recorded in the initial region category discrimination result and participate in connection relationship correction in reverse through the initial boundary consistency result. Therefore, it is suitable for recognition scenarios where the boundary position changes more significantly between adjacent slices.

[0107] After receiving the region category discrimination results and boundary consistency results, the recognition result output module first performs region attribution reading processing on the three types of region results in the region category discrimination results according to the slice number and spatial location, forming a region attribution consolidation result; then, based on the region attribution consolidation result and boundary consistency result, it performs boundary attribution determination processing on the boundary correspondence between the boundary transition region results and the adjacent liver parenchyma region results and the candidate abnormal region results, forming a boundary correction result; subsequently, it performs boundary merging processing on the boundary transition region results, and performs attribution update processing on the boundary transition region results to the liver parenchyma region results or the candidate abnormal region results respectively, forming a region correction result; then, based on the region correction result and the region category discrimination result, it performs intra-slice recognition region integration processing and cross-slice recognition region correspondence processing, forming a recognition region consolidation result; finally, it generates the liver recognition result. This specific embodiment is suitable for recognition scenarios where the boundary range between the candidate abnormal region results and the liver parenchyma region results is relatively wide, and the continuity of the boundary transition region results across slices is more obvious.

[0108] The liver image receiving module receives liver image data, generates standardized liver image data, and outputs the standardized liver image data to the region unit generation module. The region unit generation module receives the standardized liver image data, generates a set of liver region units, and outputs the set of liver region units to the intra-slice relationship construction module, the cross-slice relationship construction module, and the liver structure semantic map generation module, respectively. The intra-slice relationship construction module receives the set of liver region units, generates intra-slice structural relationship data, and outputs the intra-slice structural relationship data to the liver structure semantic map generation module. The cross-slice relationship construction module receives the set of liver region units, generates cross-slice continuous relationship data, and outputs the cross-slice continuous relationship data to the liver structure semantic map generation module. The liver structure semantic map generation module receives the set of liver region units, intra-slice structural relationship data, and cross-slice continuous relationship data, generates liver structure semantic map data, and outputs the liver structure semantic map data to the graph propagation recognition module. The graph propagation recognition module receives the liver structure semantic map data, generates region category discrimination results and boundary consistency results, and outputs the region category discrimination results and boundary consistency results to the recognition result output module. The recognition result output module receives the region category discrimination results and boundary consistency results and generates the liver recognition result.

[0109] A liver image recognition method based on graph neural networks, the steps of which are as follows:

[0110] The system receives liver imaging data and performs slice numbering and slice location matching processing on the liver imaging data according to the acquisition order. Based on the slice matching results, it performs grayscale range unification processing and grayscale distribution correction processing on the image grayscale of each slice. It also performs association writing processing on the spatial location information corresponding to each slice and the slice number. Finally, it performs corresponding integration processing on the grayscale standardization results and spatial location information association results according to the slice number to generate standardized liver imaging data.

[0111] Based on the grayscale distribution and spatial location of each slice in the standardized liver image data, the main body of the liver is identified. Based on the results of the main body of the liver, the standardized liver image data is processed by region connectivity sorting and region contour extraction. Based on the initial region division results, liver parenchyma region extraction, candidate abnormal region extraction, edge transition identification and boundary range extraction are performed. The liver parenchyma region results, candidate abnormal region results and boundary transition region results are processed according to the slice number and spatial location to generate a set of liver region units.

[0112] Based on the slice number, the results of the liver parenchyma region, candidate abnormal region, and boundary transition region in the liver region unit set are processed by the same slice aggregation. Based on the region aggregation results within the slice, adjacent regions in the same slice are processed by region contact recognition, region spacing determination, gray-scale reading of boundary position, gray-scale change trend sorting, boundary edge extraction, and edge extension direction correspondence processing. Based on the boundary coupling adjustment parameters, the results of region adjacency relationship analysis, gray-scale change relationship analysis, and boundary continuity relationship analysis are processed by boundary coupling adjustment to generate structural relationship data within the slice.

[0113] Based on the slice number, adjacent slices in the liver region unit set are processed to perform front-to-back slice correspondence processing. Based on the cross-slice region processing results, corresponding regions in adjacent slices are processed to perform region position correspondence analysis, region contour extension direction reading, region contour continuous change processing, region overlap range comparison, and region change continuous determination processing. Based on the continuous compensation adjustment parameters, the position mapping processing results, contour continuation analysis results, and region correspondence association results are processed to perform continuous compensation adjustment processing to generate cross-slice continuous relationship data.

[0114] Based on slice number and spatial location, node aggregation, node identifier writing, and node position correspondence processing are performed on the liver parenchyma region results, candidate abnormal region results, and boundary transition region results in the liver region unit set. Based on the intra-slice structural relationship data, intra-slice connection correspondence processing and intra-slice connection merging processing are performed. Based on the cross-slice continuous relationship data, cross-slice connection correspondence processing and cross-slice connection merging processing are performed. Finally, node edge correspondence integration processing and graph data organization processing are performed on the node organization results, intra-slice edge organization results, and cross-slice edge organization results to generate liver structure semantic graph data.

[0115] The liver structure semantic map data is processed by the liver structure semantic map boundary write-back convergence recognition algorithm. Based on the slice number and spatial location, the graph nodes and connections in the liver structure semantic map data are prepared for propagation. Based on the graph propagation preparation results, graph neural network propagation is performed. Based on the node propagation update results, the initial region category discrimination results and the initial boundary consistency results are formed. Based on the boundary write-back adjustment parameters, the connection relationships corresponding to the boundary transition region results are corrected for write-back and written into the liver structure semantic map data. The graph neural network propagation process, as well as the region category correspondence sorting process and the boundary correspondence sorting process, are performed again to generate the region category discrimination results and the boundary consistency results.

[0116] Based on the slice number and spatial location, the results of liver parenchyma region, candidate abnormal region, and boundary transition region in the region category discrimination results are processed for region attribution reading. Based on the region attribution sorting results and boundary consistency results, the boundary correspondence between the boundary transition region results and adjacent liver parenchyma region results and candidate abnormal region results is processed for boundary attribution determination and boundary merging. Finally, the boundary transition region results are processed for attribution update, intra-slice identification region integration, cross-slice identification region correspondence, and identification result generation, generating the liver identification result.

[0117] Finally, it should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A liver image recognition system based on a graph neural network, characterized in that: include: The liver image receiving module is used to receive liver image data, perform slice processing, grayscale standardization processing, and spatial location information association processing on the liver image data, and generate standardized liver image data. The region unit generation module performs liver parenchyma region extraction, candidate abnormal region extraction, and boundary transition region extraction on standardized liver imaging data to generate a set of liver region units. The intra-slice relationship construction module performs regional adjacency relationship analysis, gray-scale change relationship analysis, and boundary continuity relationship analysis on the liver region unit set, and performs boundary coupling adjustment processing on the analysis results according to the boundary coupling adjustment parameters to generate intra-slice structural relationship data; The cross-slice relationship construction module performs position mapping, contour continuation analysis, and region correspondence processing on liver region units in adjacent slices, and performs continuous compensation adjustment processing on the analysis results based on continuous compensation adjustment parameters to generate cross-slice continuous relationship data. The liver structure semantic graph generation module performs node organization processing and edge connection organization processing on the liver region unit set, intra-slice structural relationship data and cross-slice continuous relationship data to generate liver structure semantic graph data. The graph propagation recognition module performs a liver structure semantic graph boundary writeback convergence recognition algorithm on the liver structure semantic graph data to generate region category discrimination results and boundary consistency results. The recognition result output module performs boundary correction processing and recognition result generation processing on the region category discrimination result and boundary consistency result to generate liver recognition result.

2. The liver image recognition system based on a graph neural network according to claim 1, characterized in that: The liver image receiving module performs slice numbering and slice position correspondence processing on the liver image data according to the acquisition order, forming the slice processing results. Based on the slice processing results, grayscale range unification and grayscale distribution correction are performed on the image grayscale in each slice to form grayscale standardization results; Based on the slice organization results, read the spatial location information corresponding to each slice, and perform association writing processing on the spatial location information and slice number to form spatial location information association results; The grayscale standardization results and spatial location information association results are integrated according to the slice number to generate standardized liver image data.

3. The liver image recognition system based on a graph neural network according to claim 1, characterized in that: The region unit generation module performs liver main body range recognition processing based on the grayscale distribution and spatial location of each slice in the standardized liver image data to form the liver main body range result; Based on the results of the main liver extent, the standardized liver image data is processed by region connectivity sorting and region contour extraction to form the initial region division results; Based on the initial region segmentation results, liver parenchyma region extraction processing is performed on the region located in the middle of the main body of the liver and with continuous gray-level distribution to form the liver parenchyma region results; Based on the initial region segmentation results, candidate abnormal region extraction processing is performed on regions located in the middle of the main body of the liver and whose gray-level distribution deviates from the results of the liver parenchyma region to form candidate abnormal region results. Based on the results of the liver parenchyma region and the results of the candidate abnormal region, edge transition identification processing and boundary range extraction processing are performed at the junction of the two to form the boundary transition region results; The liver parenchyma region results, candidate abnormal region results, and boundary transition region results are processed according to slice number and spatial location to generate a set of liver region units.

4. The liver image recognition system based on a graph neural network according to claim 1, characterized in that: The intra-slice relation construction module performs intra-slice aggregation processing on the liver parenchyma region results, candidate abnormal region results, and boundary transition region results in the liver region unit set according to the slice number, forming intra-slice region aggregation results; Based on the region aggregation results within the slice, region contact recognition and region spacing determination are performed on adjacent regions in the same slice to form region adjacency relationship analysis results; Based on the region aggregation results within the slice, grayscale reading processing of the boundary position and grayscale change trend processing are performed on adjacent regions in the same slice to form grayscale change relationship analysis results. Based on the region aggregation results within the slice, the boundary edge extraction process and edge extension direction correspondence processing are performed on adjacent regions in the same slice to form the boundary continuity relationship analysis results; Based on the results of regional adjacency relationship analysis, grayscale change relationship analysis, and boundary continuity relationship analysis, boundary coupling adjustment processing is performed according to the boundary coupling adjustment parameters to form the intra-slice relationship adjustment results; Based on the results of intra-slice relation adjustment, connection relationship organization processing is performed on adjacent regions in the same slice to generate intra-slice structural relationship data.

5. The liver image recognition system based on a graph neural network according to claim 1, characterized in that: The cross-slice relationship construction module performs corresponding sorting processing on adjacent slices in the liver region unit set based on the slice number, forming cross-slice region sorting results; Based on the cross-slice region processing results, regional location correspondence analysis is performed on the liver parenchyma region results, candidate abnormal region results, and boundary transition region results in adjacent slices to form location mapping processing results; Based on the cross-slice region processing results, the corresponding regions in adjacent slices are processed by reading the region contour extension direction and processing the region contour continuous change, forming the contour continuity analysis results. Based on the cross-slice region processing results, the corresponding regions in adjacent slices are subjected to region overlap range comparison processing and region change continuity determination processing to form region correspondence results; Based on the location mapping processing results, contour continuation analysis results, and region correspondence results, continuous compensation adjustment processing is performed according to the continuous compensation adjustment parameters to form cross-slice relationship adjustment results. Based on the cross-slice relationship adjustment results, continuous connectivity organization processing is performed on corresponding regions in adjacent slices to generate cross-slice continuous relationship data.

6. The liver image recognition system based on a graph neural network according to claim 1, characterized in that: The liver structure semantic graph generation module performs node aggregation processing on the liver parenchyma region results, candidate abnormal region results, and boundary transition region results in the liver region unit set according to the slice number and spatial location, forming a graph node organization result; Based on the graph node organization results, node identifier writing and node position correspondence processing are performed on the liver parenchyma region results, candidate abnormal region results, and boundary transition region results to form node organization results; Based on the structural relationship data within the slice, intra-slice connection correspondence processing and intra-slice connection merging processing are performed on adjacent regions in the same slice to form intra-slice edge organization results; Based on cross-slice continuous relationship data, cross-slice connection correspondence processing and cross-slice connection merging processing are performed on corresponding regions in adjacent slices to form cross-slice edge organization results; Based on the node organization results, slice intra-slice edge organization results, and cross-slice edge organization results, perform node-edge corresponding integration processing to form a graph structure integration result; Based on the graph structure integration results, graph data organization processing is performed to generate liver structure semantic graph data.

7. The liver image recognition system based on a graph neural network according to claim 1, characterized in that: The graph propagation recognition module performs the liver structure semantic graph boundary writeback convergence recognition algorithm on the liver structure semantic graph data, and performs propagation preparation processing on the graph nodes and connections in the liver structure semantic graph data according to the slice number and spatial location, forming the graph propagation preparation result. Based on the graph propagation preparation results, perform graph neural network propagation processing to generate node propagation update results; Based on the node propagation update results, the results of liver parenchyma region, candidate abnormal region and boundary transition region are processed to form the initial region category discrimination results. Based on the node propagation update results, the connection relationship between the boundary transition region results and the adjacent liver parenchyma region results and the candidate abnormal region results is processed to organize the boundary correspondence, forming the initial boundary consistency results. Based on the initial region category discrimination results and the initial boundary consistency results, the write-back correction process is performed on the connection relationship corresponding to the boundary transition region results according to the boundary write-back adjustment parameters, and the write-back correction results are written into the liver structure semantic map data to form the write-back corrected liver structure semantic map data. Based on the revised liver structure semantic graph data, graph neural network propagation processing is performed again, and region category correspondence and boundary correspondence processing are performed on the propagated graph nodes to generate region category discrimination results and boundary consistency results.

8. The liver image recognition system based on a graph neural network according to claim 1, characterized in that: The recognition result output module performs region attribution reading processing on the liver parenchyma region results, candidate abnormal region results, and boundary transition region results in the region category discrimination results based on the slice number and spatial location, and forms the region attribution sorting result; Based on the regional attribution and boundary consistency results, the boundary attribution determination process is performed on the boundary correspondence between the boundary transition region results and the adjacent liver parenchyma region results and the candidate abnormal region results to form the boundary correction results. Based on the boundary correction results, boundary merging processing is performed on the boundary transition region results, and the boundary transition region results are respectively assigned to the liver parenchyma region results or candidate abnormal region results to form the region correction results. Based on the region correction results and region category discrimination results, perform intra-slice identification region integration processing and cross-slice identification region correspondence processing to form identification region sorting results; Based on the sorting results of the identified regions, the identification result generation process is performed to generate the liver identification result.

9. A liver image recognition system based on a graph neural network according to claim 1, characterized in that: The liver image receiving module receives liver image data, generates standardized liver image data, and outputs the standardized liver image data to the region unit generation module; The region unit generation module receives standardized liver image data, generates a set of liver region units, and outputs the set of liver region units to the intra-slice relationship construction module, the cross-slice relationship construction module, and the liver structure semantic map generation module, respectively. The intra-slice relationship construction module receives a set of liver region units, generates intra-slice structural relationship data, and outputs the intra-slice structural relationship data to the liver structural semantic map generation module; The cross-slice relationship construction module receives a set of liver region units, generates cross-slice continuous relationship data, and outputs the cross-slice continuous relationship data to the liver structure semantic map generation module; The liver structure semantic map generation module receives the set of liver region units, intra-slice structural relationship data and cross-slice continuous relationship data, generates liver structure semantic map data, and outputs the liver structure semantic map data to the graph propagation recognition module; The graph propagation recognition module receives liver structure semantic graph data, generates region category discrimination results and boundary consistency results, and outputs the region category discrimination results and boundary consistency results to the recognition result output module; The recognition result output module receives the region category discrimination result and the boundary consistency result, and generates the liver recognition result.

10. A method for applying to a liver image recognition system based on a graph neural network as described in any one of claims 1-9, characterized in that: The steps are as follows: The system receives liver imaging data and performs slice numbering and slice location matching processing on the liver imaging data according to the acquisition order. Based on the slice matching results, it performs grayscale range unification processing and grayscale distribution correction processing on the image grayscale of each slice. It also performs association writing processing on the spatial location information corresponding to each slice and the slice number. Finally, it performs corresponding integration processing on the grayscale standardization results and spatial location information association results according to the slice number to generate standardized liver imaging data. Based on the grayscale distribution and spatial location of each slice in the standardized liver image data, the main body of the liver is identified. Based on the results of the main body of the liver, the standardized liver image data is processed by region connectivity sorting and region contour extraction. Based on the initial region division results, liver parenchyma region extraction, candidate abnormal region extraction, edge transition identification and boundary range extraction are performed. The liver parenchyma region results, candidate abnormal region results and boundary transition region results are processed according to the slice number and spatial location to generate a set of liver region units. Based on the slice number, the results of the liver parenchyma region, candidate abnormal region, and boundary transition region in the liver region unit set are processed by the same slice aggregation. Based on the region aggregation results within the slice, adjacent regions in the same slice are processed by region contact recognition, region spacing determination, gray-scale reading of boundary position, gray-scale change trend sorting, boundary edge extraction, and edge extension direction correspondence processing. Based on the boundary coupling adjustment parameters, the results of region adjacency relationship analysis, gray-scale change relationship analysis, and boundary continuity relationship analysis are processed by boundary coupling adjustment to generate structural relationship data within the slice. Based on the slice number, adjacent slices in the liver region unit set are processed to perform front-to-back slice correspondence processing. Based on the cross-slice region processing results, corresponding regions in adjacent slices are processed to perform region position correspondence analysis, region contour extension direction reading, region contour continuous change processing, region overlap range comparison, and region change continuous determination processing. Based on the continuous compensation adjustment parameters, the position mapping processing results, contour continuation analysis results, and region correspondence association results are processed to perform continuous compensation adjustment processing to generate cross-slice continuous relationship data. Based on slice number and spatial location, node aggregation, node identifier writing, and node position correspondence processing are performed on the liver parenchyma region results, candidate abnormal region results, and boundary transition region results in the liver region unit set. Based on the intra-slice structural relationship data, intra-slice connection correspondence processing and intra-slice connection merging processing are performed. Based on the cross-slice continuous relationship data, cross-slice connection correspondence processing and cross-slice connection merging processing are performed. Finally, node edge correspondence integration processing and graph data organization processing are performed on the node organization results, intra-slice edge organization results, and cross-slice edge organization results to generate liver structure semantic graph data. The liver structure semantic map data is processed by the liver structure semantic map boundary write-back convergence recognition algorithm. Based on the slice number and spatial location, the graph nodes and connections in the liver structure semantic map data are prepared for propagation. Based on the graph propagation preparation results, graph neural network propagation is performed. Based on the node propagation update results, the initial region category discrimination results and the initial boundary consistency results are formed. Based on the boundary write-back adjustment parameters, the connection relationships corresponding to the boundary transition region results are corrected for write-back and written into the liver structure semantic map data. The graph neural network propagation process, as well as the region category correspondence sorting process and the boundary correspondence sorting process, are performed again to generate the region category discrimination results and the boundary consistency results. Based on the slice number and spatial location, the results of liver parenchyma region, candidate abnormal region, and boundary transition region in the region category discrimination results are processed for region attribution reading. Based on the region attribution sorting results and boundary consistency results, the boundary correspondence between the boundary transition region results and adjacent liver parenchyma region results and candidate abnormal region results is processed for boundary attribution determination and boundary merging. Finally, the boundary transition region results are processed for attribution update, intra-slice identification region integration, cross-slice identification region correspondence, and identification result generation, generating the liver identification result.

Images