An intelligent evaluation method for endoscopic images of ulcerative colitis

Abstract

The present application relates to an intelligent evaluation method for endoscopic images of ulcerative colitis, including: synchronously collecting a video stream and three-dimensional point cloud data of the colonic mucosa through an endoscopic system equipped with a structured light depth sensor, and constructing a three-dimensional coordinate system based on the colonic anatomical orientation; constructing a simplicial complex topological structure on the surface of the colonic mucosa based on the three-dimensional point cloud data, and calculating the persistent homology feature set of the lesion area; constructing a multi-scale graph neural network model specific to ulcerative colitis and Crohn's disease based on the persistent homology feature set. Implementing this solution can achieve a structured and topological modeling of the lesion area of the colonic mucosa, effectively capturing the morphological feature differences between ulcerative colitis and Crohn's disease. The multi-scale graph neural network integrates macroscopic anatomical segmentation and microscopic pathological features to achieve an efficient distinction between continuous and discontinuous lesions, and finally generates an intelligent and visual endoscopic evaluation report, improving the accuracy and real-time performance of intraoperative auxiliary diagnosis.

Description

Technical Field

[0001] This application relates to the field of medical imaging, and particularly to an intelligent evaluation method for endoscopic images of ulcerative colitis. Background Art

[0002] Ulcerative Colitis (UC) is a chronic inflammatory bowel disease characterized by starting from the rectum and continuously extending proximally. Its typical pathological changes include extensive erosion, ulcers, crypt distortion, and pseudopolyp formation in the colonic mucosa. The clinical diagnosis of ulcerative colitis mainly relies on colonoscopy. However, traditional diagnostic methods are mostly based on manual observation of video images by doctors, with significant subjectivity and limitations relying on experience.

[0003] In recent years, with the continuous deepening of the application of artificial intelligence technology in the field of medical image processing, endoscopic assisted diagnosis systems have gradually introduced deep learning algorithms, playing an important role in lesion recognition, severity assessment, etc. However, existing AI evaluation methods mainly focus on ulcer recognition or regional grading of two-dimensional images, rarely considering the three-dimensional spatial continuity structure of the colon, especially in the full-segment continuity analysis of rectal segment lesions, crypt structure distortion recognition, and topological structure difference modeling with other inflammatory bowel diseases such as Crohn's Disease (CD).

[0004] In addition, since ulcerative colitis and Crohn's disease may both present as mucosal erosion or atypical ulcers in the early stage, it is difficult to accurately distinguish them in terms of surface morphology. Traditional algorithms often ignore the anatomical extension law and homology continuity of pathological features of colonic lesions in three-dimensional structures. At the same time, existing methods usually have difficulty effectively integrating the macroscopic anatomical scale and microscopic structure scale, and also lack the ability to provide real-time and interpretable augmented reality prompts in a dynamic endoscopic environment. Summary of the Invention

[0005] The purpose of this application is to solve the problems that traditional endoscopic diagnosis relies on manual experience and has strong subjectivity in identifying colonic lesions, and that existing AI methods are difficult to express the three-dimensional spatial continuity of lesions, crypt microstructure distortion features, and topological differences between ulcerative colitis and Crohn's disease, and to improve the accuracy and intelligence level of diagnosis.

[0006] According to one aspect of this application, there is provided an intelligent evaluation method for endoscopic images of ulcerative colitis, characterized by including:

[0007] Synchronously collect the video stream and 3D point cloud data of the colonic mucosa through an endoscopic system equipped with a structured light depth sensor, and construct a 3D coordinate system based on the anatomical orientation of the colon. The 3D coordinate system includes images of the lesion area in multiple consecutive frames and their depth coordinates distributed along the taenia coli.

[0008] Based on the 3D point cloud data, construct a simplicial topological structure on the surface of the colonic mucosa, and calculate the persistent homology feature set of the lesion area.

[0009] Based on the persistent homology feature set, construct a multi-scale graph neural network model specific to ulcerative colitis and Crohn's disease, including:

[0010] Macro-scale graph: Divide the colon into several segments, and use the mean number of connected components, ulcer density, and the probability of lesion spreading along the taenia coli in each segment as node features and edge weights.

[0011] Micro-scale graph: Use individual ulcer lesions as graph nodes, and the graph nodes include the following structural pathological features:

[0012] The depth-width ratio of a single ulcer, which is used to distinguish superficial ulcers from penetrating ulcers. When the depth-width ratio is lower than a preset depth-width ratio threshold, it indicates the superficial ulcer feature of ulcerative colitis.

[0013] The sharpness of the ulcer edge, which is quantified by the boundary gradient amplitude and its standard deviation. When the sharpness is higher than a preset sharpness threshold, it indicates active inflammation.

[0014] The minimum curvature similarity between ulcers, which is used to measure the spatial continuity of the ulcer surface morphology. When the minimum curvature similarity is higher than a preset similarity threshold, it is determined to be a continuous distribution.

[0015] Preferably, the persistent homology feature set includes:

[0016] Total rectal continuity index: By analyzing the coverage ratio and spatial continuity of the lesions in the rectal segment along the taenia coli, quantify whether the lesions involve the entire rectal segment. When the coverage ratio exceeds a preset coverage threshold, it is determined to be a feature of ulcerative colitis.

[0017] Number of connected components: Used to quantify the continuous distribution pattern of lesions along the taenia coli.

[0018] Number of circular structures: Used to identify the closed-loop features of pseudopolyps.

[0019] Preferably, it further includes:

[0020] Dynamically fuse the features of the macro-scale graph and the micro-scale graph through a deformable attention mechanism to generate a joint feature vector.

[0021] Based on the combined feature vector, an endoscopic evaluation report including the differential diagnosis result of ulcerative colitis / Crohn's disease, the lesion grading, and the risk warning of pseudopolyps is output.

[0022] Preferably, the calculation of the whole-segment rectal continuity index includes:

[0023] Segment the three-dimensional point cloud data of the rectal segment, and calculate the proportion of the length covered by the lesion area to the total length of the rectum;

[0024] Analyze the topological connectivity of the lesion boundary through the Alpha complex algorithm. When the number of connected components is 1 and the coverage ratio exceeds the preset coverage threshold, it is determined as a feature of ulcerative colitis;

[0025] The distortion rate of the rectal mucosal crypt structure in the microscopic scale map is calculated through the morphological parameters of the crypt opening. If the distortion rate exceeds the preset distortion threshold, it is marked as a typical pathological change of ulcerative colitis.

[0026] Preferably, in the macroscopic scale map, when the probability of the lesion spreading along the colonic band exceeds the first preset threshold, it is determined as a continuous lesion;

[0027] When the spreading probability is lower than the second preset threshold and the number of connected components exceeds the third preset threshold, it is determined as a skip lesion of Crohn's disease;

[0028] The edge weight in the microscopic scale map is jointly determined by the minimum curvature similarity between ulcers and the correlation of the crypt distortion rate. If the correlation exceeds the fourth preset threshold, the determination of ulcerative colitis is strengthened.

[0029] Preferably, in the macroscopic scale map, the ulcer density is calculated by the product of the number of ulcers per unit area and the segmentation weight coefficient to amplify the lesion influence of a specific segment;

[0030] The edge sharpness in the microscopic scale map is the standard deviation of the gradient amplitude in the ulcer boundary region. When the gradient amplitude exceeds the fifth preset threshold, it is marked as an active inflammatory lesion. When the gradient amplitude is lower than the sixth preset threshold, it indicates the fibrotic change of Crohn's disease.

[0031] Preferably, the deformable attention mechanism includes:

[0032] Rectum-priority attention module: Based on the weight coefficient of the rectal segment, dynamically enhance the attention weight of the features in the rectal region;

[0033] Cross-scale feature fusion module: Jointly optimize the macroscopic anatomical segmentation and microscopic crypt distortion features through the multi-head attention mechanism.

[0034] Preferably, the endoscopic evaluation report generates the differential diagnosis result of ulcerative colitis / Crohn's disease according to the following rules:

[0035] If the rectal continuity index of the whole segment exceeds the preset coverage threshold and the crypt distortion rate exceeds the preset distortion threshold, it is determined as ulcerative colitis;

[0036] If the lesions are distributed in a skip pattern and the crypt distortion rate is lower than the seventh preset threshold, it is determined as Crohn's disease;

[0037] The endoscopic evaluation report combines the thermal map of rectal segment lesions and the three-dimensional reconstruction map of crypt structure to display the analysis results.

[0038] Preferably, it further includes a model optimization step based on the pathological differences between ulcerative colitis and Crohn's disease, including:

[0039] Generating adversarial samples that conform to the rectal continuity diffusion law of ulcerative colitis and the skip lesion characteristics of Crohn's disease, and the generation process is respectively constrained by rectal anatomy or random spatial distribution;

[0040] Introducing a topological consistency loss function and a discrimination loss function into the discriminator to optimize the structural rationality and category distinguishability of the generated lesions.

[0041] Preferably, the adversarial samples are generated by a generator, and the generator includes:

[0042] Ulcerative colitis branch: Input the anatomical coding of the rectal segment and output the continuous lesion point cloud;

[0043] Crohn's disease branch: Input the random position coding and output the skip lesion point cloud;

[0044] The discriminator is a multi-scale convolutional network structure, and the generation effect is jointly optimized by a topological consistency loss function, an adversarial loss function and a total loss function.

[0045] This application has the following beneficial effects: It proposes a three-dimensional endoscopic image intelligent evaluation method that integrates macroscopic and microscopic features. Aiming at the problem that it is difficult for existing two-dimensional image recognition to reflect the spatial continuity of colon lesions and the changes in crypt microstructures, a structured light depth sensor is used to obtain the video stream and three-dimensional point cloud data of the colon mucosa, and a three-dimensional coordinate system and a complex topological structure are constructed based on the anatomical trend of the colon, so as to extract the persistent homology features of the lesion area. On this basis, a graph neural network model that integrates macroscopic scales (such as lesion connectivity, diffusion probability) and microscopic scales (such as ulcer depth-width ratio, edge sharpness, crypt distortion rate) is constructed to realize the comprehensive expression of the typical features of ulcerative colitis. Through the introduction of a three-dimensional structure and multi-scale graph modeling, it is possible to more accurately distinguish ulcerative colitis from Crohn's disease and improve the intelligence and accuracy of lesion recognition and disease typing. This method effectively enhances the spatial continuity and structural interpretability of colon lesion analysis and provides a more clinically valuable auxiliary decision-making basis for endoscopic diagnosis. Brief Description of the Drawings

[0046] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings required for the description of the embodiments or the prior art. Obviously, the drawings in the following description are only some embodiments of the present application. For those of ordinary skill in the art, without creative efforts, other drawings can also be obtained based on these drawings.

[0047] Figure 1 It is a logic block diagram of the intelligent evaluation method for endoscopic images of ulcerative colitis according to an embodiment of the present application. Detailed Embodiments

[0048] To facilitate the understanding of the present application, the following will describe the present application more comprehensively with reference to the relevant drawings. The preferred embodiments of the present application are shown in the drawings. However, the present application can be implemented in many different forms and is not limited to the embodiments described herein. On the contrary, the purpose of providing these embodiments is to make the understanding of the disclosure of the present application more thorough and comprehensive.

[0049] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by those of ordinary skill in the technical field to which this application belongs. The terms used in the description of the present application in this specification are only for the purpose of describing specific embodiments and are not intended to limit the present application. The term "and / or" used herein includes any and all combinations of one or more of the related listed items.

[0050] Please refer to Figure 1 , an embodiment of the present application provides an intelligent evaluation method for endoscopic images of ulcerative colitis, including:

[0051] S10. Synchronously collect the video stream and three-dimensional point cloud data of the colon mucosa through an endoscopic system equipped with a structured light depth sensor, and construct a three-dimensional coordinate system based on the anatomical direction of the colon. The three-dimensional coordinate system includes the images of the lesion areas in multiple consecutive frames and their depth coordinates distributed along the colon bands.

[0052] The purpose of this step is to introduce three-dimensional information on the basis of traditional two-dimensional endoscopic images to solve the problem of missing spatial structure caused by relying solely on the image plane. Through the structured light depth sensor, the depth information corresponding to each frame of the image can be synchronously obtained to generate the three-dimensional point cloud of the colon mucosa. These point cloud data combined with the video frame position can be used to reconstruct the three-dimensional surface topography of the colon, and a three-dimensional coordinate system can be constructed according to the anatomical structure direction (i.e., the colon band direction) to form a continuous spatial positioning framework, thereby providing a spatial reference for subsequent topological modeling and input of the graph neural network.

[0053] S20. Based on the three-dimensional point cloud data, construct a simplicial complex topological structure on the surface of the colon mucosa, and calculate the persistent homology feature set of the lesion area.

[0054] The purpose of this step is to introduce the means of topological data analysis to quantitatively model the spatial structure of the lesion area of the colon mucosa. Through three-dimensional point cloud data, a surface mesh or simplicial complex structure can be constructed to perform topological analysis on the morphology such as depressions and protrusions on the mucosa surface. Persistent Homology is a tool in topology for characterizing high-order structures such as "holes" or "connected regions" in space. By analyzing the "persistence" of these features as the scale changes, the homology features of the lesion area can be stably extracted, such as the area, connectivity, and continuity of the spatial distribution of ulcers. These features are of great value in distinguishing the extensive, superficial, and continuous distribution of ulcerative colitis from the skip and penetrating lesions of Crohn's disease.

[0055] S30. Based on the persistent homology feature set, construct a multi-scale graph neural network model specific to ulcerative colitis and Crohn's disease. This step aims to further transform the topological features into graph-structured data and achieve the classification and reasoning of high-order spatial features through a multi-scale graph neural network (GNN). Considering that inflammatory bowel disease has different manifestations at different scales (the whole colon vs. a single lesion), two graph models, macroscopic and microscopic, are designed, including:

[0056] Macro-scale graph: Divide the colon into several segments, and use the mean number of connected components, ulcer density, and the probability of lesion spreading along the tenia coli in each segment as node features and edge weights.

[0057] At the macroscopic scale, the entire colon is divided into continuous segments (such as the rectum, sigmoid colon, etc.) in anatomical order, and each segment constitutes a node in the graph. The features of the node include:

[0058] Mean number of connected components: Reflects whether the distribution of lesions is skip or continuous;

[0059] Ulcer density: Measures the ulcer burden within this segment;

[0060] Probability of spreading along the tenia coli: Describes whether the lesion spreads proximally along the anatomical path.

[0061] These features together describe the "breadth" and "distribution pattern" of the lesion and are representative for distinguishing the continuous expansion of UC from the skip lesions of CD.

[0062] Microscopic scale graph: Each individual ulcer lesion serves as a graph node. At the microscopic scale, the focus is on the morphological characteristics of each ulcer itself, which is used for fine-grained identification of the nature of the lesion. Each ulcer is a node, and the graph nodes include the following structural pathological features:

[0063] The depth-width ratio of a single ulcer, which is used to distinguish superficial ulcers from penetrating ulcers. When the depth-width ratio is lower than the preset depth-width ratio threshold, it indicates the superficial ulcer characteristics of ulcerative colitis. It is used to distinguish the depth of the ulcer. UC is usually a superficial ulcer with a small depth-width ratio; CD is mostly a penetrating deep ulcer with a large depth-width ratio.

[0064] The sharpness of the ulcer edge, which is quantified by the boundary gradient amplitude and its standard deviation. When the sharpness is higher than the preset sharpness threshold, it indicates active inflammation. It is quantified by the boundary gradient intensity and its standard deviation, reflecting the degree of active inflammation. Sharp edges are commonly seen in active UC.

[0065] The minimum curvature similarity between ulcers, which is used to measure the spatial continuity of the ulcer surface morphology. When the minimum curvature similarity is higher than the preset similarity threshold, it is determined to be a continuous distribution. It is used to judge whether multiple ulcers are continuously distributed in three-dimensional space. A high similarity means consistent morphology, indicating the extensive mucosal lesion characteristics of UC.

[0066] Implementing the technical solution of this embodiment, a three-dimensional endoscopic image intelligent evaluation method that integrates macroscopic and microscopic features is proposed. Aiming at the problem that it is difficult for existing two-dimensional image recognition to reflect the spatial continuity and crypt microstructural changes of colon lesions, a structured light depth sensor is used to obtain the video stream and three-dimensional point cloud data of the colon mucosa, and a three-dimensional coordinate system and complex topological structure are constructed based on the colon anatomical orientation, so as to extract the persistent homology features of the lesion area. On this basis, a graph neural network model that integrates macroscopic scale (such as lesion connectivity, diffusion probability) and microscopic scale (such as ulcer depth-width ratio, edge sharpness, crypt distortion rate) is constructed to realize the comprehensive expression of the typical features of ulcerative colitis. Through the introduction of three-dimensional structure and multi-scale graph modeling, it is possible to more accurately distinguish ulcerative colitis from Crohn's disease and improve the intelligent and precise level of lesion recognition and disease typing. This method effectively enhances the spatial continuity and structural interpretability of colon lesion analysis, providing a more clinically valuable auxiliary decision-making basis for endoscopic diagnosis.

[0067] In this embodiment, the persistent homology feature set includes:

[0068] The continuity index of the entire rectum: By analyzing the coverage ratio and spatial continuity of the lesions in the rectal segment along the taenia coli, it quantifies whether the lesions involve the entire rectum. When the coverage ratio exceeds the preset coverage threshold, it is determined to be a feature of ulcerative colitis;

[0069] The number of connected components: It is used to quantify the continuous distribution pattern of the lesions along the taenia coli;

[0070] Number of circular structures: used to identify the closed-loop features of pseudopolyp formation.

[0071] It should be noted that the persistent homology feature set includes the following key indicators:

[0072] The continuity index of the entire rectal segment is obtained by quantitatively analyzing the spatial coverage ratio and continuity of rectal segment lesions along the taenia coli. The specific method is to calculate the proportion of the lesion area covered in the three-dimensional coordinate system and combine the topological structure to judge whether the lesion is continuously distributed. When the coverage ratio exceeds the preset coverage threshold, it is determined that the lesion involves the entire rectal segment, which conforms to the typical continuous and extensive distribution characteristics of ulcerative colitis (UC). This index helps to distinguish UC from Crohn's disease (CD) because CD usually does not show persistent extensive involvement of the rectum.

[0073] The number of connected components is used to quantify the continuous or discontinuous distribution of the lesion area on the taenia coli. A smaller and continuous number of connected components indicates that the lesion area is spatially connected and complete, which conforms to the characteristics of continuous lesions in UC. A larger number of connected components indicates a discontinuous spread of the lesion, which is a typical manifestation of CD.

[0074] The number of circular structures is used to identify the closed-loop structure formed by pseudopolyps in the lesion area. The calculation of the number of circular structures is based on the detection of the one-dimensional homology group (loops or "holes") in persistent homology analysis. The closed loops formed by pseudopolyps show topological circular structures, and this feature can assist in judging the pathological type and severity of the lesion.

[0075] Implementing this technical solution, by introducing the persistent homology feature set and using three-dimensional point cloud data for topological structure analysis, it is possible to quantify the spatial distribution characteristics of colonic mucosal lesions, accurately reflect the continuous, discontinuous, and circular structure characteristics of the lesions. The continuity index of the entire rectal segment effectively distinguishes the differences in the lesion scope and continuity between ulcerative colitis and Crohn's disease. The number of connected components further reveals the spatial distribution pattern of the lesions, while the number of circular structures improves the accuracy of pathological analysis for the identification of pseudopolyp formation. Overall, this method breaks through the limitations of traditional two-dimensional image analysis, enhances the three-dimensional spatial understanding and quantitative judgment ability of the lesions, thereby improving the accuracy of endoscopic image intelligent assessment and the reliability of clinical diagnosis.

[0076] This embodiment also includes:

[0077] S40. Dynamically fuse the features of the macroscopic-scale graph and the microscopic-scale graph through a deformable attention mechanism to generate a joint feature vector. The purpose of this step is to combine the macroscopic and microscopic features extracted by the multi-scale graph neural network to achieve more accurate disease characterization. The macroscopic-scale graph reflects the overall distribution pattern of the lesions, and the microscopic-scale graph focuses on the detailed features of individual ulcer lesions. By introducing a deformable attention mechanism, the system can dynamically adjust the importance weights of the two types of features according to specific cases, enhancing the flexibility and pertinence of information fusion. Through adaptive learning, this mechanism identifies the relationships between key nodes and edges, highlights the key features related to disease diagnosis, and suppresses irrelevant or noisy information, thereby generating a joint feature vector rich in semantics, providing a solid foundation for subsequent classification and risk prediction.

[0078] S50. Based on the joint feature vector, output an endoscopic evaluation report including the differential diagnosis results of ulcerative colitis / Crohn's disease, lesion grading, and false polyp risk warning. This step relies on the joint feature vector generated in S40 and uses pre-trained classification and regression models to comprehensively analyze the spatial distribution characteristics and pathological details of the patient's lesions to achieve multi-task output. Specifically, it includes:

[0079] Differential diagnosis results: Determine whether the lesion is more inclined to ulcerative colitis (UC) or Crohn's disease (CD) to assist doctors in making accurate diagnoses;

[0080] Lesion grading: Evaluate the severity of the lesion based on comprehensive indicators such as the number, size, depth, and inflammation activity of ulcers to assist in formulating treatment plans;

[0081] False polyp risk warning: Detect indicators such as the number of circular structures to identify the risk of false polyp formation in advance and remind doctors to pay attention to potential complications.

[0082] Finally, the system generates a structured endoscopic evaluation report, which is convenient for doctors to understand the characteristics of the lesions and improves the efficiency of clinical diagnosis and decision-making.

[0083] Implementing this technical solution, by fusing the multi-dimensional spatial features of the macroscopic and microscopic scales and using a deformable attention mechanism to dynamically integrate the overall distribution and detailed morphology of the lesions, high-precision differentiation between ulcerative colitis and Crohn's disease is achieved. The multi-task evaluation based on the joint feature vector not only improves the accuracy of lesion grading but also can warn of the risk of false polyps in advance, greatly enhancing the comprehensiveness and clinical guiding value of endoscopic image evaluation and improving the diagnostic efficiency and reliability.

[0084] In this embodiment, the calculation of the full-segment rectal continuity index includes:

[0085] S21. Segment the 3D point cloud data of the rectal segment and calculate the proportion of the covered length of the lesion area to the total length of the rectum. This step includes, first, spatially partitioning the collected 3D point cloud data of the colon to locate a specific anatomical segment, namely the rectum. In the 3D point cloud of this rectal segment, identify the point set of the lesion area. Calculate the covered length of the lesion area along the direction of the tenia coli, that is, the spatial length of the continuous distribution of the lesion. Divide the covered length of the lesion by the total length of the entire rectal segment to obtain the lesion coverage ratio. This ratio reflects the spatial expansion range of the lesion in the rectal segment and is an important indicator for judging the extent of the lesion.

[0086] S22. Analyze the topological connectivity of the lesion boundary through the Alpha complex algorithm. When the number of connected components is 1 and the coverage ratio exceeds a preset coverage threshold, it is determined as a characteristic of ulcerative colitis. In this step, use the Alpha complex algorithm to construct a topological structure for the 3D boundary point cloud of the lesion area, and focus on calculating the number of connected components, that is, the number of continuous lesion areas in space. If the number of connected components is 1, it means that the lesion is spatially continuous and not segmented within the rectal segment, forming a complete area. When the coverage ratio of this continuous lesion area exceeds the preset threshold at the same time, it indicates that the lesion has a wide and continuous coverage range, and this manifestation conforms to the typical characteristics of ulcerative colitis. This step combines topological connectivity and coverage ratio for determination, which helps to distinguish ulcerative colitis from other skip lesions.

[0087] S23. Calculate the distortion rate of the rectal mucosal crypt structure in the microscopic scale map through the morphological parameters of the crypt opening. If the distortion rate exceeds the preset distortion threshold, it is marked as a typical pathological change of ulcerative colitis. On the microscopic scale, analyze the morphological characteristics of the rectal mucosal crypts through a graph neural network model, especially the morphological parameters of the crypt opening (such as size, shape, symmetry, etc.). Calculate the distortion rate of the crypt structure, which reflects the degree of subtle pathological changes in the mucosa. If the distortion rate is higher than the preset threshold, it indicates that the crypt structure is significantly abnormal, which is usually a typical pathological manifestation in the active stage of ulcerative colitis. This step provides a quantitative index for the activity and severity of the lesion at the microscopic level to assist in the overall diagnosis.

[0088] Implementing this technical solution, through the introduction of 3D point cloud data and topological analysis methods, realizes the precise spatial positioning and continuous quantification of the lesion area in the rectal segment. Use the Alpha complex algorithm to determine the lesion connectivity, and effectively distinguish the continuous extensive lesion characteristics of ulcerative colitis from other skip lesions by combining the coverage ratio. In addition, based on the calculation of the distortion rate of the crypt structure at the microscopic scale, an assessment index of pathological activity at the cellular level is provided. The overall solution improves the diagnostic accuracy and disease grading ability of endoscopic images for ulcerative colitis, enhances the comprehensive understanding of the spatial distribution and microscopic pathological changes of the lesion, and has high clinical application value.

[0089] In a specific embodiment, in the macroscopic scale map, when the probability of a lesion spreading along the taenia coli exceeds a first preset threshold, it is determined as a continuous lesion; when the spreading probability is lower than a second preset threshold and the number of connected components exceeds a third preset threshold, it is determined as a skip lesion of Crohn's disease; the edge weights in the microscopic scale map are jointly determined by the minimum curvature similarity between ulcers and the correlation of crypt distortion rate. If the correlation exceeds a fourth preset threshold, the determination of ulcerative colitis is strengthened.

[0090] In the macroscopic scale map, the ulcer density is calculated by the product of the number of ulcers per unit area and the segment weight coefficient to amplify the lesion impact of a specific segment; the edge sharpness in the microscopic scale map is the standard deviation of the gradient amplitude in the ulcer boundary region. When the gradient amplitude exceeds a fifth preset threshold, it is marked as an active inflammation focus, and when the gradient amplitude is lower than a sixth preset threshold, it indicates fibrotic changes in Crohn's disease.

[0091] The deformable attention mechanism includes: a rectum - priority attention module: based on the weight coefficient of the rectal segment, dynamically enhancing the attention weight of the rectal region features; a cross - scale feature fusion module: jointly optimizing the macroscopic anatomical segments and microscopic crypt distortion features through the multi - head attention mechanism.

[0092] In this embodiment, it should be noted that this technical solution realizes the accurate identification and differentiation of the morphological, distribution characteristics, and pathological states of ulcerative colitis (UC) and Crohn's disease (CD) lesions by constructing macroscopic and microscopic scale maps and integrating multi - dimensional structural indicators and the attention mechanism. Specifically:

[0093] At the macroscopic scale map level:

[0094] The lesion spreading probability is used to quantify the degree of lesion propagation along the taenia coli. A high continuity indicates a tendency towards UC, and a high skip rate indicates a tendency towards CD;

[0095] The number of connected components combined with a low spreading probability indicates the typical skip lesion characteristics of CD;

[0096] After introducing the segment weight coefficient, the ulcer density can dynamically enhance the judgment sensitivity for specific high - risk regions (such as the rectal segment), which helps to highlight the key points of clinical concern.

[0097] At the microscopic scale map level:

[0098] The edge weights comprehensively consider the minimum curvature similarity between ulcers and the correlation of crypt distortion rate, making the model more discriminative in cell - level structure relationship modeling;

[0099] The edge sharpness characterizes the degree of inflammation activity or fibrosis trend through the standard deviation of the gradient, supporting the classification judgment of the active phase of UC and the chronic phase of CD.

[0100] Optimization of the attention mechanism:

[0101] The rectum-prior attention module dynamically enhances its feature representation through the weights of the rectal segment, fully considering the typical onset sites of UC.

[0102] The cross-scale feature fusion module utilizes multi-head attention to fuse the macroscopic topological features of the lesions and the microscopic crypt structure indicators, thereby improving the accuracy and robustness of the overall assessment.

[0103] In a specific embodiment, the endoscopic evaluation report generates the differential diagnosis results of ulcerative colitis / Crohn's disease through the following rules:

[0104] If the continuity index of the entire rectal segment exceeds the preset coverage threshold and the crypt distortion rate exceeds the preset distortion threshold, it is determined to be ulcerative colitis;

[0105] If the lesions are distributed in a skip pattern and the crypt distortion rate is lower than the seventh preset threshold, it is determined to be Crohn's disease;

[0106] The endoscopic evaluation report combines the heat map of the rectal segment lesions and the three-dimensional reconstruction map of the crypt structure to display the analysis results.

[0107] It also includes model optimization steps based on the pathological differences between ulcerative colitis and Crohn's disease, including:

[0108] Generating adversarial samples that conform to the rectal continuity diffusion law of ulcerative colitis and the skip lesion characteristics of Crohn's disease, and the generation process is respectively constrained by rectal anatomy or random spatial distribution;

[0109] Introducing a topological consistency loss function and a discrimination loss function into the discriminator to optimize the structural rationality and class discriminability of the generated lesions.

[0110] The adversarial samples are generated by a generator, and the generator includes:

[0111] The ulcerative colitis branch: Input the anatomical coding of the rectal segment and output the continuous lesion point cloud;

[0112] The Crohn's disease branch: Input the random position coding and output the skip lesion point cloud;

[0113] The discriminator is a multi-scale convolutional network structure, and the generation effect is jointly optimized through a topological consistency loss function, an adversarial loss function, and a total loss function.

[0114] It should be noted that through the introduction of a rule-based discrimination logic and an adversarial learning model optimization mechanism, the following key technological innovations and effect guarantees are achieved in endoscopic image evaluation:

[0115] The rule-driven diagnostic logic has interpretability and clinical comparability:

[0116] The combined judgment of the two indicators of the total rectal continuity index and the crypt distortion rate not only reflects the pathological characteristics of the continuous and extensive distribution of ulcerative colitis (UC) and the destruction of crypt structure, but also has clear physical significance, facilitating clinicians to verify and refer to; the skip distribution and intact crypts are consistent with the discontinuous lesions and tissue retention of Crohn's disease (CD), and the logical judgment conforms to pathological evidence.

[0117] Visualization of evaluation results enhances doctors' cognition and trust:

[0118] The thermal map of rectal lesions and the three-dimensional reconstruction map of crypt structure integrated in the endoscopic evaluation report can intuitively present the lesion range and microscopic tissue changes, improve the intuitiveness and understandability of auxiliary diagnosis, and enhance the credibility and operability of the AI-assisted system.

[0119] The model optimization mechanism combines medical prior knowledge and deep learning capabilities:

[0120] Using the adversarial sample generation mechanism with specific distribution characteristics, respectively simulating the continuous diffusion of UC and the skip characteristics of CD, effectively enriching the distribution form of training samples, and improving the generalization ability of the model to identify;

[0121] Taking the rectal anatomical coding as the input to ensure that the UC simulation process conforms to the physiological and anatomical laws; random spatial perturbation controls the CD branch generation to strengthen the model's learning ability for complex spatial structures.

[0122] The topological consistency loss function strengthens the structure recognition ability:

[0123] By introducing the topological consistency loss function (such as Betti number difference, homology structure constraint, etc.), ensuring that the generated lesion areas are consistent with real cases in terms of spatial connectivity, closed-loop structure, etc., and improving the topological rationality of the lesions generated by the model;

[0124] The multi-scale convolutional discriminator combines the adversarial loss and the class discrimination target, effectively distinguishing the morphological structures of UC and CD, and optimizing the authenticity of the lesion structure and the clarity of the class boundary.

[0125] Finally, an auxiliary evaluation model with high robustness, high interpretability and high visibility is realized:

[0126] This embodiment not only ensures the stable performance of the diagnostic model under complex lesion structures, but also integrates the dual advantages of deep learning and medical logic, forming a comprehensive solution suitable for endoscopic intelligent evaluation, providing a solid support for intraoperative auxiliary decision-making of ulcerative colitis and Crohn's disease.

[0127] The following is a complete example of using this technical solution to distinguish Crohn's disease (CD) from ulcerative colitis (UC). Specifically:

[0128] Step S10: Structured light point cloud acquisition and lesion extraction

[0129] Device: Structured light endoscope system

[0130] Data acquisition: Acquire the point cloud of the rectum and colon segments, with a total data volume of approximately 500,000 points

[0131] Point cloud processing: Use the DBSCAN clustering algorithm to extract the lesion point cloud

[0132]

[0133] Table 1

[0134] As shown in Table 1, the point cloud quantity refers to the total number of points contained in the three-dimensional point cloud data. The point cloud is a three-dimensional data form used to represent the surface or spatial structure of an object. Each "point" contains at least position coordinates (x, y, z), and sometimes additional information such as color, normal vector, reflectivity, etc.

[0135] In the structured light depth endoscope system involved in this application, the point cloud quantity directly reflects the spatial sampling density and fineness of the scanning area (such as the surface of the colon mucosa). For example: the more the point cloud quantity, the more detailed the restoration of the spatial structure, and the higher the accuracy of extracting topological features (such as crypt distortion, lesion edge morphology, curvature change, etc.); if the point cloud quantity is too small, it may lead to incomplete analysis of the lesion continuity, unstable topological connectivity or homology features, and affect the model discrimination effect.

[0136] The lesion coverage length refers to the cumulative length along the colon axis of the continuous segment detected as the lesion area (such as ulcer, inflammation) in the colon three-dimensional point cloud or image, and is an important indicator for measuring the spatial distribution range and continuity of the lesion.

[0137] Step S20: Topological homology analysis, using the Alpha complex algorithm to extract the topological structure:

[0138]

[0139] Table 2

[0140] Step S21: Continuity index calculation

[0141] Coverage ratio > 85%, connectivity number = 1, initially judged as ulcerative colitis.

[0142] Step S22: Crypt structure distortion rate evaluation

[0143] 350 crypts were sampled, 312 abnormal crypts were found, the average opening deviation was 15.4°, the distortion rate reached 89.1%, exceeding the 75% threshold, further supporting the diagnosis of UC.

[0144] Step S30: Macro-scale graph modeling, where nodes represent colon anatomical segments and edge weights are the ulcer diffusion probability;

[0145]

[0146] Table 3

[0147] Judgment: Rectal diffusion probability > 0.85 (the first preset threshold), high connectivity → UC.

[0148] Skip distribution in the ascending colon and terminal ileum → Possibility of CD.

[0149] Step S31: The edge weight W of the micro-scale graph modeling is × minimum curvature similarity + × crypt distortion correlation, = 0.4, = 0.6, and the total correlation score 0.82 > 0.75, strengthening the UC judgment.

[0150] Step S40: Attention mechanism fusion includes:

[0151] Rectum-priority attention module (boosting the weight of rectal features to 0.95);

[0152] Cross-scale fusion module (combining the macro-anatomical graph and the micro-distortion graph, outputting a 128-dimensional feature vector).

[0153] Step S50: Output of the diagnostic result

[0154] Discrimination result: UC

[0155] Grade: Severe

[0156] Visualization graph: Heat map + three-dimensional graph of crypt structure

[0157] Step S60: Adversarial sample training generator:

[0158]

[0159] Table 4

[0160] Discriminator: Topological consistency loss Ltopo = 0.043, adversarial loss Ladv = 0.217, total loss Ltotal = 0.260.

[0161] Implement this solution, utilize point cloud topological structure analysis and homology features to improve the connectivity recognition ability and enhance the recognition accuracy of UC; the multi-scale graph model combines macroscopic segmentation and microscopic crypt features to enhance the model's discriminative ability for jumping lesions; introduce a deformable attention mechanism to focus on key anatomical regions and strengthen feature expression; the adversarial sample training introduces a topological consistency loss function to improve the generation and recognition stability of the model for complex-shaped lesions; the system supports heatmap and three-dimensional visualization output to enhance the doctor's intraoperative auxiliary decision-making ability.

[0162] The above-described embodiments only represent several embodiments of the present application. The description is relatively specific and detailed, but it should not be construed as a limitation on the scope of the patent application. It should be noted that for those of ordinary skill in the art, without departing from the concept of the present application, several modifications and improvements can still be made, and these all belong to the protection scope of the present application. Therefore, the protection scope of the patent of the present application shall be subject to the appended claims.

Claims

1. An intelligent evaluation method for endoscopic images of ulcerative colitis, characterized in that, Including: Synchronously acquire the video stream and three-dimensional point cloud data of the colonic mucosa through an endoscopic system equipped with a structured light depth sensor, and construct a three-dimensional coordinate system based on the colonic anatomical orientation. The three-dimensional coordinate system includes images of the lesion area in multiple consecutive frames and their depth coordinates distributed along the taenia coli; Based on the three-dimensional point cloud data, construct a simplicial topological structure on the surface of the colonic mucosa, and calculate the persistent homology feature set of the lesion area; Based on the persistent homology feature set, construct a multi-scale graph neural network model specific to ulcerative colitis and Crohn's disease, including: Macro-scale graph: Divide the colon into several segments, and use the mean number of connected components, ulcer density, and the probability of lesion spreading along the taenia coli in each segment as node features and edge weights; Micro-scale graph: Use individual ulcer lesions as graph nodes, and the graph nodes include the following structural pathological features: The depth-width ratio of a single ulcer, which is used to distinguish superficial ulcers from penetrating ulcers. When the depth-width ratio is lower than a preset depth-width ratio threshold, it indicates the superficial ulcer feature of ulcerative colitis; The sharpness of the ulcer edge, which is quantified by the boundary gradient amplitude and its standard deviation. When the sharpness is higher than a preset sharpness threshold, it indicates active inflammation; The minimum curvature similarity between ulcers, which is used to measure the spatial continuity of the ulcer surface morphology. When the minimum curvature similarity is higher than a preset similarity threshold, it is determined to be a continuous distribution; Dynamically fuse the features of the macro-scale graph and the micro-scale graph through a deformable attention mechanism to generate a joint feature vector; Based on the joint feature vector, output an endoscopic evaluation report including the identification result of ulcerative colitis / Crohn's disease, lesion grading, and false polyp risk warning.

2. The endoscopic image intelligent evaluation method for ulcerative colitis according to claim 1, characterized in that The persistent homology feature set includes: Total rectal continuity index: By analyzing the coverage ratio and spatial continuity of the lesions in the rectal segment along the taenia coli, quantify whether the lesions involve the entire rectal segment. When the coverage ratio exceeds a preset coverage threshold, it is determined to be a feature of ulcerative colitis; Number of connected components: Used to quantify the continuous distribution pattern of the lesions along the taenia coli; Number of circular structures: Used to identify the closed-loop features of false polyp formation.

3. The intelligent endoscopic image evaluation method for ulcerative colitis according to claim 2, wherein The calculation of the total rectal continuity index includes: Segment the three-dimensional point cloud data of the rectal segment, and calculate the proportion of the covered length of the lesion area in the total length of the rectum; Analyze the topological connectivity of the lesion boundary through the Alpha complex algorithm. When the number of connected components is 1 and the coverage ratio exceeds the preset coverage threshold, it is determined to be a feature of ulcerative colitis; The distortion rate of the rectal mucosal crypt structure in the micro-scale graph is calculated through the morphological parameters of the crypt opening. If the distortion rate exceeds a preset distortion threshold, it is marked as a typical pathological change of ulcerative colitis.

4. The intelligent endoscopic image evaluation method for ulcerative colitis according to claim 3, wherein In the macro-scale graph, when the probability of the lesion spreading along the taenia coli exceeds the first preset threshold, it is determined to be a continuous lesion; When the spreading probability is lower than the second preset threshold and the number of connected components exceeds the third preset threshold, it is determined to be a skip lesion of Crohn's disease; The edge weights in the microscopic-scale image are jointly determined by the minimum curvature similarity between ulcers and the correlation of crypt distortion rate. If the correlation exceeds the fourth preset threshold, the determination of ulcerative colitis is strengthened.

5. The endoscopic image intelligent evaluation method for ulcerative colitis according to claim 4, characterized in that In the macroscopic-scale image, the ulcer density is calculated by the product of the number of ulcers per unit area and the segmentation weight coefficient to amplify the lesion influence of a specific segment. The edge sharpness in the microscopic-scale image is the standard deviation of the gradient amplitude in the ulcer boundary region. When the gradient amplitude exceeds the fifth preset threshold, it is marked as an active inflammation lesion. When the gradient amplitude is lower than the sixth preset threshold, it indicates fibrotic changes in Crohn's disease.

6. The endoscopic image intelligent evaluation method for ulcerative colitis according to claim 5, characterized in that The deformable attention mechanism includes: Rectum-priority attention module: Based on the weight coefficient of the rectum segment, dynamically enhance the attention weight of the features in the rectum region. Cross-scale feature fusion module: Jointly optimize the macroscopic anatomical segmentation and microscopic crypt distortion features through the multi-head attention mechanism.

7. The endoscopic image intelligent evaluation method for ulcerative colitis according to claim 6, characterized in that The endoscopic evaluation report generates the differential diagnosis results of ulcerative colitis / Crohn's disease according to the following rules: If the whole-segment rectum continuity index exceeds the preset coverage threshold and the crypt distortion rate exceeds the preset distortion threshold, it is determined as ulcerative colitis. If the lesions are distributed in a skip pattern and the crypt distortion rate is lower than the seventh preset threshold, it is determined as Crohn's disease. The endoscopic evaluation report combines the heat map of the lesions in the rectum segment and the three-dimensional reconstruction map of the crypt structure to display the analysis results.

8. The endoscopic image intelligent evaluation method for ulcerative colitis according to claim 7, wherein It also includes a model optimization step based on the pathological differences between ulcerative colitis and Crohn's disease, including: Generate adversarial samples that conform to the rectal continuity diffusion law of ulcerative colitis and the skip lesion characteristics of Crohn's disease. The process of generating adversarial samples is respectively restricted by rectal anatomy or random spatial distribution. Introduce a topological consistency loss function and a discrimination loss function in the discriminator to optimize the structural rationality and category discriminability of the generated lesions.

9. The endoscopic image intelligent evaluation method for ulcerative colitis according to claim 8, characterized in that The adversarial samples are generated by a generator, and the generator includes: Ulcerative colitis branch: Input the anatomical coding of the rectum segment and output the continuous lesion point cloud. Crohn's disease branch: Input the random position coding and output the skip lesion point cloud. The discriminator is a multi-scale convolutional network structure, and the generation effect is jointly optimized through the topological consistency loss function, the adversarial loss function, and the total loss function.

Images