Method for identifying tumor micro residual lesions based on pathological image ai analysis

Abstract

The present application relates to the technical field of medical image analysis, in particular to a tumor micro residual lesion identification method based on pathological image AI analysis, comprising the following steps: extracting a cell group path of a pathological section, analyzing curvature and repeated structures to obtain a cell group morphological evolution sequence atlas, judging direction alternation positioning turning regions, splicing edges to extract closed boundaries to obtain a residual lesion candidate structure set, comparing a line intersection angle to determine consistency, and obtaining a tumor residual lesion identification result. In the present application, by constructing a cell group morphological evolution atlas, the curvature change of the cell group edge path in the continuous section and the centroid turning position are extracted, the structure direction alternation change is identified to locate the key region image, the closed boundary is drawn by combining edge splicing and direction convergence, the centroid line intersection angle and convergence node coincidence are analyzed, the structure trend and morphological coherence in multiple continuous sections are compared, the tumor micro residual lesion is identified, the identification stability in the background is improved, the misjudgment is reduced, and the diagnosis reliability is enhanced.

Description

Technical Field

[0001] This invention relates to the field of medical image analysis technology, and in particular to a method for identifying minimal residual tumor lesions based on AI analysis of pathological images. Background Technology

[0002] The field of medical image analysis technology involves processing and analyzing medical image data using image processing, computer vision, machine learning, and other methods to assist in medical applications such as disease diagnosis, lesion detection, pathological classification, and efficacy evaluation. Core aspects of this field include preprocessing, feature extraction, classification and recognition, image segmentation, 3D reconstruction, and model training of medical image data. Specific image types cover MRI, CT, ultrasound, X-ray, and pathological slide images. With the development of artificial intelligence technology, learning algorithms such as deep neural networks are widely used for automatic identification and classification of lesion regions, improving the accuracy and efficiency of medical image interpretation and forming a data-driven medical image analysis system. Traditional methods for identifying minimal residual tumor lesions refer to the manual observation and analysis of pathological images after cancer patient treatment to determine the presence of residual lesions. This method typically involves professional pathologists observing morphological and structural changes in stained pathological slide images under a microscope to identify residual tumor cells. The identification method relies on the doctor's experience in judging pathological characteristics such as cell morphology, arrangement structure, and nuclear atypia. While digital images may be acquired using imaging scanning equipment, the analysis and judgment mainly depend on manual slide reading and qualitative description.

[0003] Existing technologies for identifying small residual tumor lesions in pathological images typically rely on the subjective judgment of pathologists or image algorithms based on local features. They lack dynamic analysis of the overall structural evolution of cell populations, making it difficult to accurately extract boundary morphological changes and directional turning relationships. This results in low accuracy in identifying residual lesions in complex backgrounds and an inability to effectively identify the spatial connectivity and structural continuity of small lesions across slices, leading to problems of low recognition rate and high misjudgment rate. Summary of the Invention

[0004] To address the technical problems existing in the prior art, embodiments of the present invention provide a method for identifying minimal residual tumor lesions based on AI analysis of pathological images; To achieve the above objectives, the present invention adopts the following technical solution: a method for identifying minimal residual tumor lesions based on AI analysis of pathological images, comprising the following steps: S1: Obtain continuous pathological slice images of the tumor resection area and expand the full image, extract cell population contour paths, analyze the curvature change positions, the nucleus centroid turning positions, and the repeating structure positions to obtain a cell population morphological evolution sequence map. S2: Based on the cell population morphological evolution sequence map, determine whether the characteristic directions of adjacent positions are continuously alternating, select the corresponding slice sequence of continuous alternation, locate the associated image region, and obtain the image set of structural transition region; S3: Based on the corresponding region of the image set of the structural transition region, expand the cell population edge segment and splice it along the direction, extract the direction convergence position, draw the closed boundary, analyze the connectivity of the closed boundary, and obtain the set of candidate structures of residual lesions. S4: Based on the set of candidate residual lesions, extract the coordinates of the cell nucleus centroid and the direction of the line pointing to the boundary, compare the angle between adjacent lines, select the location of directional mutation, analyze the overlap between the mutation location and the boundary convergence node, and obtain the lesion structure consistency determination result. S5: Based on the lesion structure consistency determination result, check the corresponding position of the candidate structure in the continuous slice, judge the consistency of the boundary direction and the structural continuity, and obtain the tumor residual lesion identification result.

[0005] As a further aspect of the present invention, the cell population morphological evolution sequence map includes curvature change positions, arrangement direction turning points, corresponding positions of repeating structures, and temporal relationships of positional features; the structural turning point image set includes image sequence numbers, alternating position features, and structural turning point image regions; the residual lesion candidate structure set includes splicing path direction convergence positions, closed boundaries, and closed-loop connectivity analysis results; the lesion structure consistency determination results include the angle between the direction of the line connecting the cell nuclei and centroids, the location of directional abrupt changes, and the overlap status of boundary convergence nodes; and the tumor residual lesion identification results include structural boundary direction consistency, structural continuity, and continuous slice position comparison relationships.

[0006] As a further aspect of the present invention, the cell population contour path refers to a continuous closed curve that characterizes the shape of the outer edge of the cell population; The location of curvature change refers to the point on the cell population outline path where the curvature changes.

[0007] As a further embodiment of the present invention, the closed boundary refers to a closed region enclosed by the cell population outline path; The boundary convergence node refers to the location where the boundaries of multiple cell populations tend to converge.

[0008] As a further aspect of the present invention, the specific steps of S1 are as follows: S101: Obtain continuous pathological slice images of the tumor resection area, expand the full image of each pathological slice, extract the cell population edge contour paths within the image, and obtain the cell population edge contour path set; S102: Based on the set of cell population edge contour paths, analyze the position of curvature change of cell population edge contour paths, locate the turning point of the direction of the line connecting the cell nuclei and centroids, identify the positional features of repeating structures inside the cell population, and obtain a set of positional features. S103: Based on the location feature set, obtain the pathological slide image and scan sequence number, and integrate the location features within the location feature set to obtain a cell population morphological evolution sequence map.

[0009] As a further aspect of the present invention, the specific steps of S2 are as follows: S201: Based on the cell population morphological evolution sequence map, extract the feature coordinates of adjacent positions, the direction angle sequence of the feature connection of adjacent positions, and the feature spacing sequence of adjacent positions. Compare the change signs of adjacent items in the adjacent direction angle sequence and determine whether the change signs continue to alternate within the range of three sets of consecutive alternation reference times to obtain the direction alternation segment index table. S202: Based on the alternating direction segment index table, obtain the locally stored pathological slide image acquisition sequence record, associate the segment index with the acquisition sequence record, map the pathological slide image sequence number corresponding to the segment index, and obtain the slide sequence number list. S203: Based on the list of slice numbers and the cell population morphological evolution sequence map, retrieve the feature coordinates of the corresponding position of the segment index to locate the boundary of the region in the pathological slice image, extract the image fragments covered by the region boundary and aggregate the image fragments to obtain a set of images of structural transition regions.

[0010] As a further aspect of the present invention, the specific steps of S3 are as follows: S301: Based on the pathological slide image region corresponding to the structural transition region image set, retrieve the cell population edge path within the region, extend the cell population edge path to the adjacent edge segment, and make a consistency judgment on the extended edge segment to obtain a continuous path set of edge segments. S302: Based on the continuous path set of the edge segment, analyze the connection status of adjacent path segments, aggregate path segments with consistent directions, identify the location features of the gradual convergence of path segments, and obtain the set of convergence locations. S303: Based on the set of convergence positions, draw closed boundary trajectories around each convergence position, analyze the closed-loop connectivity of the closed boundary trajectories, identify the closed-loop connectivity boundary trajectories, and obtain a set of candidate structures for residual lesions.

[0011] As a further aspect of the present invention, in the process of determining the consistency of the extended edge segments: the included angle is calculated between each pair of adjacent edge segments, and by determining whether the included angle is within a preset angle range, the direction of the adjacent edge segments is analyzed to determine whether they remain continuous. In the process of identifying the location features of the gradual convergence of path segments: the extension direction is traced on each path segment in the continuous path set of the edge segment, the direction relationship is compared with the direction relationship of the surrounding path segments, and the area where the direction converges is taken as the convergence position of the direction.

[0012] As a further aspect of the present invention, the specific steps of S4 are as follows: S401: Based on the region corresponding to the candidate structure in the set of candidate residual lesions, detect the position coordinates of the nucleus centroid within the region, detect the position coordinates of the candidate structure boundary path and the boundary convergence node, extract the direction sequence of the line connecting the nucleus centroid to the candidate structure boundary path, and obtain the centroid boundary connection direction sequence. S402: Based on the centroid boundary line connection direction sequence, read the preset direction change angle threshold, compare the direction angle of adjacent connection with the direction change angle threshold, select the position coordinates corresponding to the direction of the connection that continuously exceeds the direction change angle threshold, and obtain the direction change position set; S403: Based on the set of directional mutation locations, call the coordinates of the boundary convergence node positions in the set of candidate residual lesions, read the preset spatial overlap distance threshold, determine the relationship between the directional mutation location and the distance between the boundary convergence nodes and the spatial overlap distance threshold, and obtain the lesion structure consistency determination result.

[0013] As a further aspect of the present invention, the specific steps of S5 are as follows: S501: Based on the lesion structure consistency determination result, retrieve the candidate structure identifier in the residual lesion candidate structure set, read the pathological slice number and image region coordinates corresponding to the candidate structure, associate the candidate structure identifier with the pathological slice number and image region coordinates, and obtain the candidate structure slice location table. S502: Based on the candidate structure slice location table, check the image region coordinates of the same candidate structure identifier in consecutive pathological slices, determine whether the direction of the closed boundary in adjacent pathological slices remains the same, analyze whether the morphology of the closed boundary in adjacent pathological slices remains connected and continuous, and obtain a set of structures with consistent continuous boundaries. S503: Based on the continuous boundary consistent structure set, determine the correspondence between the candidate structure identifiers in the continuous boundary consistent structure set and the candidate structure identifiers in the residual lesion candidate structure set, identify the corresponding relationship to establish the candidate structure identifier, and obtain the tumor residual lesion identification result.

[0014] Compared with the prior art, the advantages and positive effects of the present invention are as follows: In this invention, by constructing a cell population morphological evolution map, the curvature changes and centroid turning points of the cell population edge paths in continuous slices are extracted, and alternating structural orientation changes are identified to locate key region images. Closed boundaries are drawn by combining edge splicing and directional convergence. The coincidence of the centroid connection angle and convergence node is analyzed. The structural orientation and morphological coherence are compared in multiple continuous slices to identify small residual tumor lesions, improve the stability of identification in the background, reduce misjudgment and enhance diagnostic reliability. Attached Figure Description

[0015] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0016] Figure 1 This is a schematic diagram of the steps of the present invention; Figure 2 This is a detailed schematic diagram of S1 of the present invention; Figure 3 This is a detailed schematic diagram of S2 of the present invention; Figure 4 This is a detailed schematic diagram of S3 of the present invention; Figure 5 This is a detailed schematic diagram of S4 of the present invention; Figure 6 This is a detailed schematic diagram of S5 of the present invention. Detailed Implementation

[0017] The technical solution of the present invention will now be described with reference to the accompanying drawings.

[0018] To make the technical problems, technical solutions and advantages of the present invention clearer, a detailed description will be given below in conjunction with the accompanying drawings and specific embodiments.

[0019] Please see Figure 1 This invention provides a method for identifying minimal residual tumor lesions based on AI analysis of pathological images, comprising the following steps: S1: Acquire multiple consecutive pathological slice images of the tumor resection area, expand the full image of each pathological slice, extract the edge contour path of the complete cell population in the image, analyze the position of curvature change in the edge of the cell population, the position of the change in the direction of the centroid of the cell nucleus, and the positional features corresponding to the repeating arrangement structure inside the cell cluster. Based on the sequential relationship of the pathological slice images in the acquisition process, the positional features are associated and integrated to obtain the cell population morphological evolution sequence map. S2: Based on the structural change direction and trend between adjacent position features in the cell population morphology evolution sequence map, determine whether the structural change direction occurs continuously alternating in the map. For positions where continuous alternation reaches three groups, determine the corresponding pathological slide image number, and locate the image region associated with the continuous alternation position feature in the corresponding pathological slide image to obtain the set of structural transition region images. S3: Based on the pathological slice image region corresponding to the image set of structural transition regions, extend the cell population edge path in the region to the adjacent edge segment, continuously splice the adjacent edge segments along the direction of the cell population, extract the direction convergence position in the splicing path, draw the closed boundary with the direction convergence position as the center, and analyze the closed loop connectivity of the closed boundary to obtain the candidate structure set of residual lesions. S4: Based on the region corresponding to the candidate structure in the candidate structure set of residual lesions, extract the coordinates of the centroid of the cell nucleus inside the region, extract the direction of the line connecting each centroid of the cell nucleus to the boundary of the candidate structure, compare the angle between the directions of adjacent lines, select the position where the direction change occurs continuously, analyze the spatial overlap between the position of the direction change and the convergence node of the candidate structure boundary, and obtain the consistency judgment result of the lesion structure. S5: Based on the lesion structure consistency determination results, check the corresponding positions of residual lesion candidate structures in continuous pathological sections, determine the boundary direction consistency and structural continuity of candidate structures in continuous pathological sections, select candidate structures with consistent boundary direction and structural continuity, and determine the candidate structures as residual tumor lesions to obtain the residual tumor lesion identification results.

[0020] The cell population morphological evolution sequence map includes the curvature change position, the position of the arrangement direction inversion, the corresponding position of the repeating structure, and the temporal relationship of the positional features. The image set of structural inversion regions includes the image number, alternating position features, and structural inversion image regions. The candidate structure set of residual lesions includes the splicing path direction and convergence position, closed boundary, and closed loop connectivity analysis results. The lesion structure consistency judgment results include the direction angle of the line connecting the cell nucleus centroid, the position of the direction change, and the overlap status of the boundary convergence node. The tumor residual lesion identification results include the consistency of the structural boundary direction, the structural continuity, and the comparison relationship of the continuous slice positions.

[0021] Please see Figure 2 The specific steps of S1 are as follows: S101: Obtain continuous pathological slice images of the tumor resection area, expand the full image of each pathological slice, extract the cell population edge contour paths within the image, and obtain the cell population edge contour path set; First, continuous pathological section images of the tumor resection area are acquired. This process involves fixing the surgically removed tumor tissue, then performing continuous physical sectioning using a micron-level microtome, with each section set to a thickness of 4 micrometers. These sections are then scanned using a fully automated digital pathology scanner at 40x objective magnification, generating digital pathological images with a pixel resolution of 0.25 micrometers per pixel. To ensure tissue integrity, the tissue is typically embedded in paraffin before sectioning, and after sectioning, HE staining or other tissue staining methods are used to enhance the contrast between cell nuclei and cytoplasm, thereby improving the recognizability of subsequent image analysis. After obtaining the raw image data, a full-image unfolding process is performed. An affine transformation matrix is ​​used to correct the mechanical deformation of the tissue during slide preparation, mapping the compressed image blocks to the global spatial coordinate system. Simultaneously, a local elastic registration algorithm can be used to further correct for local tissue stretching or compression, ensuring consistency in the spatial geometry of the image and avoiding errors caused by deformation during subsequent feature extraction. In each unfolded pathological section image, a gradient-based edge detection algorithm is used, utilizing the zero-crossing points of the second derivative of pixel grayscale values ​​to accurately locate tissue boundaries. The process of extracting cell population edge contour paths within an image employs an eight-neighbor chain code tracing technique. Starting from the initial pixel, the spatial positions of boundary points are recorded sequentially along the gradient direction, forming a closed path chain list. During chain code tracing, false edge structures caused by uneven staining or scanning noise can be removed by setting a minimum path length threshold and noise filtering rules. All identified cell population path coordinates are then aggregated and stored to obtain a set of cell population edge contour paths.

[0022] S102: Based on the cell population edge contour path set, analyze the position of the curvature change of the cell population edge contour path, locate the turning position of the direction of the line connecting the cell nucleus centroid, identify the positional features of the repeating structures inside the cell population, and obtain the set of positional features. First, based on the cell population edge contour path set, the locations of curvature changes along the cell population edge contour paths are analyzed. This process calculates the discrete curvature formed by five adjacent pixels along the path. When the local curvature value exceeds a preset curvature threshold of 0.15, the coordinate point is marked as a geometric deformation location. Curvature change locations typically correspond to areas where the cell population structure exhibits significant bending or protruding boundaries; therefore, these locations are often pathologically associated with changes in tissue growth direction or abnormal cell aggregation. Next, the turning points of the centroid lines connecting the nuclei are located. By calculating the gray-weighted centroid coordinates of each nucleus within the cell population, the nearest neighbor nucleus is found using Euclidean distance, and a centroid connection vector is established. The angle between adjacent connecting vectors is calculated; if the angle exceeds 30 degrees, it is recorded as a turning point. This method reflects the local trend of changes in cell arrangement direction, thereby identifying areas where cell arrangement is disordered or abnormally proliferating. The location features of repetitive structures within cell populations are identified. Energy, entropy, and correlation characteristics of local regions are extracted using a gray-level co-occurrence matrix. A repetitive structure center is identified when the correlation coefficient of the same pattern within a 100-micrometer radius is greater than 0.85 and the pattern repeats at least three times. Repetitive structures typically reflect the periodic arrangement of cell clusters or the regular patterns of tissue structure. Local destruction or abnormal repetition often occurs during tumor invasion, thus this feature has high diagnostic value. The global coordinates of all mutation points and feature points are integrated to obtain a set of location features.

[0023] S103: Based on the location feature set, obtain the pathological slide image and scan sequence number, associate and integrate the location features in the location feature set, and obtain the cell population morphological evolution sequence map. First, based on the location feature set, the pathological slide images and scan sequence numbers are obtained. By parsing the metadata tags output by the scanner, the acquisition start timestamp and physical scan index number corresponding to each slide are read. This metadata not only contains scan sequence information but may also include scan magnification, focal plane depth, and equipment calibration parameters, thus providing necessary reference for subsequent spatial alignment. Using an inter-slice spatial alignment algorithm, the translation vector and rotation angle between adjacent slide images are calculated, mapping the location feature points of different layers to a unified three-dimensional logical space. In this process, the optimal transformation matrix can be estimated through feature point matching and least squares method to improve the stability and accuracy of cross-slice registration. Location features within the location feature set are correlated and integrated by logically linking adjacent slides by finding points with the shortest distance and consistent feature types. For example, if the distance between the center point of a repeating structure in layer N and the center point of layer N+1 on the projection plane is less than 10 pixels, they are determined to be a continuation of the same structure. This correlation process can establish the correspondence between cellular structures in different layers, thereby reflecting the spatial extension trend of tissue. By fusing these cross-layer correlation paths with the time and sequence attributes of the images, a digital model that can describe the growth and deformation trends of cell population structures in three-dimensional space is generated, resulting in a cell population morphological evolution sequence map.

[0024] Please see Figure 3 The specific steps of S2 are as follows: S201: Based on the cell population morphological evolution sequence map, extract the feature coordinates of adjacent positions, the direction angle sequence of the feature connection of adjacent positions, and the feature spacing sequence of adjacent positions. Compare the change signs of adjacent items in the adjacent direction angle sequence and determine whether the change signs continue to alternate within the range of three sets of consecutive alternation reference numbers to obtain the direction alternation segment index table. First, based on the cell population morphological evolution sequence map, we extract the feature coordinates of adjacent positions, the direction angle sequence of the connecting lines of adjacent positions, and the distance sequence of adjacent positions. Specifically, we calculate the angle between the connecting lines of corresponding feature points in two adjacent layers and the horizontal axis, as well as the three-dimensional Euclidean distance. By simultaneously analyzing the changes in direction angle and spatial distance, we can more comprehensively describe the movement trend of cell structures in space. We compare the signs of changes in adjacent direction angle sequences and calculate the difference between the angle of the current layer and the angle of the previous layer. If the difference is positive, the sign is recorded as positive 1; if the difference is negative, the sign is recorded as negative 1. We determine whether the change sign continuously alternates within three sets of consecutive alternation reference numbers. The reference number is set to 3 times, meaning that the sign sequence must show at least one complete alternation cycle of positive 1, negative 1, positive 1, negative 1 to characterize the swinging or spiral characteristics of the structure during evolution. This continuous alternation often represents the periodic bending or growth trend of tissue structures in space, which may be related to the invasion and spread of tumor cells. We record all slice layer ranges that meet the continuous alternation condition, such as layers 12 to 18, to obtain an index table of alternating direction segments.

[0025] S202: Based on the alternating direction paragraph index table, obtain the locally stored pathological slide image acquisition sequence record, associate the paragraph index with the acquisition sequence record, map the pathological slide image sequence number corresponding to the paragraph index, and obtain the slide sequence number list; First, based on the alternating direction segment index table, the acquisition sequence record of pathological slide images is obtained from local storage. By retrieving the sample transfer log in the pathology department's information system, the original embedding cassette number and slide batch number associated with the current slide group are extracted. By combining the records from the sample management system, consistency between the slide order and tissue origin information can be ensured, thereby avoiding analytical errors caused by confusion between different samples. A linear mapping relationship between logical layer numbers and physical scan file names is established by associating the segment index with the acquisition sequence record. The target image number is calculated using the starting physical sequence number and the logical layer position in the index table. This mapping process allows the algorithm to directly access the raw image data without layer-by-layer retrieval, improving data processing efficiency. Mapping the segment index to the pathological slide image sequence number converts complex structural evolution segments into a list of directly readable image file indexes. For example, mapping a specific segment to slides 0045 to 0052 clarifies the target range for subsequent image segment extraction, resulting in a list of slide sequence numbers.

[0026] S203: Based on the list of slice numbers and the cell population morphological evolution sequence map, the feature coordinates of the corresponding position of the segment index are retrieved to locate the boundary of the region in the pathological slice image, the image fragments covered by the boundary of the region are extracted and the image fragments are aggregated to obtain a set of images of structural transition regions. First, based on the list of slice numbers and the cell population morphological evolution sequence map, the feature coordinates of the corresponding positions of the segment index are retrieved to locate the boundary of the region within the pathological slice image. Centered on the feature coordinates and combined with the morphological span in the evolution map, a region of interest (ROI) with a width and height of 100 pixels is determined. The size of this region can be adaptively adjusted according to the scan resolution and cell structure size to ensure that key structures are completely included in the analysis. Image fragments covering the region boundary are extracted and aggregated. The underlying image reading and writing library is used to extract the aforementioned ROI from the high-resolution image with the corresponding sequence number. These image fragments from different layers but spatially aligned are stacked according to their scan sequence number to form an image array containing local dynamic evolution information. This image array can logically be regarded as a slice sequence of local three-dimensional tissue structures. This array records the microscopic texture and boundary evolution details of the tissue at the turning points, resulting in a set of images of structural turning areas.

[0027] Please see Figure 4 The specific steps of S3 are as follows: S301: Based on the pathological slice image region corresponding to the structural transition region image set, retrieve the cell population edge path within the region, extend the cell population edge path to the adjacent edge segment, and judge the consistency of the extended edge segment to obtain the continuous path set of edge segment. First, based on the corresponding pathological slide image regions in the image set of structural transition regions, cell population edge paths within these regions are retrieved. Sub-pixel-level edge localization is then re-executed to obtain more refined initial edge segment coordinates. Sub-pixel-level edge localization can obtain higher-precision boundary coordinates through grayscale gradient interpolation, thereby reducing the impact of pixel-level quantization errors on subsequent geometric analysis. The cell population edge paths are then extended to adjacent edge segments using a predictive correction algorithm. A linear search of 3 to 5 pixels is performed along the tangent vector direction of the current edge end. If a pixel with a grayscale gradient magnitude greater than 1.5 times the average value exists within the search area, it is included in the path. During this extension process, local orientation consistency constraints can be incorporated to prevent the path from mistakenly connecting to adjacent unrelated structures, thus improving the continuity and accuracy of the edge paths. The extended edge segments are then subjected to orientation consistency judgment by calculating the cosine of the angle between the extended path segment vector and the original path segment vector. If the cosine value is greater than 0.98, the extension is considered consistent, and the two paths are logically fused. By continuously repeating the above extension and fusion process, broken boundaries appearing in cases of image noise or uneven staining can be gradually restored. This process eliminates edge breaks caused by image noise, resulting in a set of continuous path segments.

[0028] S302: Based on the continuous path set of edge segments, analyze the connection status of adjacent path segments in terms of direction, aggregate path segments with consistent direction, identify the location features of the gradual convergence of path segments, and obtain the set of direction convergence locations. First, based on the continuous path set of edge segments, the connection state of adjacent path segments is analyzed, and the spacing between the endpoints of adjacent paths is calculated. When the spacing is less than 5 pixels and the difference in the slope of the endpoints is within 5%, path segments with consistent orientation are aggregated. This method can integrate multiple local boundary segments into longer structural paths, thus more accurately reflecting the overall morphology of cell population boundaries. The location characteristics of the gradual convergence of path segments are identified, and the intersection distribution of the normals of all path segments is calculated using radial symmetry center transformation. Radial symmetry center transformation can effectively detect the geometric pattern of multiple boundaries converging towards the center, thus exhibiting high stability in identifying tissue contraction or cell aggregation regions. Within the local space, if the intersection density exceeds the threshold of 8 intersections per square micrometer, it is determined to be a morphological convergence point, representing the trend of multiple boundaries converging towards the center. The centroid coordinates of these high-density intersection point groups are recorded as the core reference location of potential lesions, resulting in a set of convergence locations.

[0029] S303: Based on the set of convergence locations, draw closed boundary trajectories around each convergence location, analyze the closed-loop connectivity of the closed boundary trajectories, identify the closed-loop connectivity boundary trajectories, and obtain a set of candidate structures for residual lesions. First, based on the set of convergence points, closed boundary trajectories are drawn around each convergence point. A level set evolution model is used, with the convergence point as the seed point, to perform region growth until the boundary energy function reaches a local minimum. The level set evolution model can adaptively adjust the boundary shape in complex backgrounds, thus accurately capturing the true contour of the cell population boundary. The closed-loop connectivity of the closed boundary trajectory is analyzed. Small voids within the boundary are filled using morphological closure operations, and the closure coefficient of the trajectory is calculated, which is four times pi multiplied by the area and divided by the square of the perimeter. If the closure coefficient is between 0.6 and 1.0, a closed-loop connected boundary trajectory is identified. The closer the closure coefficient is to 1, the closer the structure is to a regular morphology, while lower values ​​may represent irregular boundaries or broken structures. Structures that meet the requirements of closure and geometric compactness are marked as suspected lesions, resulting in a set of candidate structures for residual lesions.

[0030] Please see Figure 5 The specific steps of S4 are as follows: S401: Based on the candidate structure corresponding region in the candidate structure set of residual lesions, detect the position coordinates of the nucleus centroid within the region, detect the boundary path and boundary convergence node position coordinates of the candidate structure, extract the direction sequence of the line connecting the nucleus centroid to the boundary path of the candidate structure, and obtain the centroid boundary connection direction sequence. First, based on the candidate structures in the residual lesion candidate structure set, the coordinates of the centroid of the cell nucleus within the region are detected. The center of each cell nucleus within the candidate region is located using the difference of Gaussian operator, and its centroid coordinates are calculated. The difference of Gaussian operator can highlight the local brightness differences of the cell nucleus, thereby improving the accuracy of cell nucleus detection. The coordinates of the candidate structure boundary path and boundary convergence node are detected, and the pixel coordinates of 360 uniformly sampled points on the boundary are extracted. Uniform sampling can comprehensively describe the boundary morphology changes. The direction sequence of the path connecting the cell nucleus centroid to the candidate structure boundary is extracted. Using the total centroid of the cell nucleus as the center, a vector pointing to each boundary sampling point is established. The angle between each vector and the reference coordinate axis is calculated, and the vectors are arranged in the sampling order to form a direction sequence. This sequence quantifies the spatial constraint relationship between the centroid of the cell nucleus distribution and the geometric boundary, resulting in the centroid boundary connection direction sequence.

[0031] S402: Based on the centroid boundary connection direction sequence, read the preset direction change angle threshold, compare the direction angle of adjacent connection with the direction change angle threshold, select the position coordinates of the connection direction that continuously exceeds the direction change angle threshold, and obtain the direction change position set; First, based on the centroid boundary line direction sequence, a preset directional mutation angle threshold is read. This threshold is set to 22 degrees based on historical clinical data analysis. The absolute value of the angle between adjacent lines is compared with the directional mutation angle threshold. If the difference is greater than the threshold, the location index is marked. Directional mutations usually correspond to locations where the boundary morphology changes abruptly, and such changes are common in tumor-infiltrating areas. The coordinates of locations corresponding to the directions of consecutive lines exceeding the directional mutation angle threshold are selected. When four or more abnormal sampling points appear consecutively, their corresponding boundary pixel coordinates are extracted. Continuity judgment effectively eliminates misjudgments caused by single noise points. This step aims to filter out normal boundary undulations and accurately locate sharp transitions or irregular protrusions in the boundary caused by tumor infiltration, thus obtaining a set of directional mutation locations.

[0032] S403: Based on the set of directional mutation locations, call the coordinates of the boundary convergence node positions in the set of candidate structures of residual lesions, read the preset spatial overlap distance threshold, determine the relationship between the directional mutation location and the distance between the boundary convergence node and the spatial overlap distance threshold, and obtain the lesion structure consistency determination result. First, based on the set of directional mutation locations, the coordinates of the boundary convergence nodes in the candidate structure set of residual lesions are retrieved, and a preset spatial overlap distance threshold, set to 15 pixels, is read. The relationship between the directional mutation location, the distance between the boundary convergence nodes, and the spatial overlap distance threshold is determined by calculating the shortest Euclidean distance between the two points. If this distance is less than or equal to 15 pixels, it indicates that the geometric mutation of the boundary and the physical convergence of the structure are highly spatially coincident, conforming to the logical consistency of actual contraction or growth of pathological tissue, and is thus judged as a valid lesion structure. This spatial consistency verification can significantly reduce false positives caused by boundary noise or tissue folds. By performing this spatial overlap verification on each candidate structure, false positive interference is filtered out, and the lesion structure consistency determination result is obtained.

[0033] Please see Figure 6 The specific steps of S5 are as follows: S501: Based on the lesion structure consistency determination result, retrieve the candidate structure identifier in the residual lesion candidate structure set, read the pathological slice number and image region coordinates corresponding to the candidate structure, associate the candidate structure identifier with the pathological slice number and image region coordinates, and obtain the candidate structure slice location table. First, based on the consistency determination results of lesion structures, candidate structure identifiers are retrieved from the set of residual lesion candidate structures, and the unique IDs of all verified structures are extracted. The pathological slice number and image region coordinates corresponding to the candidate structure are read, and the physical slice number, the starting point coordinates of the ROI rectangle, and the height and width of the rectangle are retrieved from the stored metadata. This location information helps doctors quickly locate the corresponding structure in the original pathological image. A detailed physical location index mapping table is established by associating the candidate structure identifier with the pathological slice number and image region coordinates. This table records the precise pixel coordinates of each identified residual lesion in the entire large slice stack, resulting in a candidate structure slice location table.

[0034] S502: Based on the candidate structure slice location table, verify the image region coordinates of the same candidate structure identifier within consecutive pathological slices, determine whether the closed boundary direction within adjacent pathological slices maintains the same change direction, and analyze... Whether the closed boundary morphology of adjacent pathological sections remains connected and continuous is determined to obtain a set of structures with consistent continuous boundaries. First, based on the candidate structure slice location table, the coordinates of the same candidate structure identifier within the image region of consecutive pathological slices are checked, and boundary curves within the same spatial coordinate range in consecutive layers are extracted. By comparing the structures in consecutive layers, the spatial continuity trend of tissue structures can be analyzed. It is determined whether the closed boundaries in adjacent pathological slices maintain the same direction of change. By calculating the displacement vector of the center point of the boundary between adjacent layers, if the magnitude of the displacement vector is less than 12 pixels and the angle of change is within 15 degrees, it is determined to be in the same direction. It is then analyzed whether the morphology of the closed boundaries in adjacent pathological slices maintains continuity. The degree of intersection of the regions enclosed by the boundaries between two layers is evaluated using the overlap calculation formula. If the ratio exceeds 0.82, it is confirmed that the layers are connected and consistent. This step improves the reliability of cross-layer structure judgment through dual constraints of geometric consistency and spatial overlap. By verifying the stability of the three-dimensional morphology, misjudgments caused by single-layer noise are eliminated, resulting in a set of structures with consistent continuous boundaries.

[0035] S503: Based on a continuous boundary consistent structure set, determine the correspondence between the candidate structure identifiers in the continuous boundary consistent structure set and the candidate structure identifiers in the residual lesion candidate structure set, identify the correspondence to establish the candidate structure identifier, and obtain the tumor residual lesion identification result; First, based on a continuous boundary consistent structure set, the correspondence between candidate structure identifiers within the continuous boundary consistent structure set and candidate structure identifiers within the residual lesion candidate structure set is determined, and a global correlation verification is performed. To establish a candidate structure identifier, the corresponding relationship must be confirmed, requiring that the structure exhibit validated geometric features and inter-layer connectivity in at least five consecutive slices. By setting multi-layer continuity conditions, the identification results are ensured to have a true three-dimensional structural basis. The number, spatial coordinates, cumulative volume, and distribution depth of all retained lesion entities are statistically analyzed. For example, three abnormal cell population structures with significant three-dimensional continuity were identified, and their spatial coordinate distribution and distance from the surgical margin met the residual assessment criteria. These results can be further used for postoperative residual assessment, treatment planning, and pathology report generation. This process completes the deduction from local feature extraction to global pathological conclusions, resulting in the identification of residual tumor lesions.

[0036] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for identifying minimal residual tumor lesions based on AI analysis of pathological images, characterized in that, Includes the following steps: S1: Obtain continuous pathological slice images of the tumor resection area and expand the full image, extract the cell population outline path, analyze the curvature change position, the cell nucleus centroid turning position, and the position of repeating structure to obtain the cell population morphological evolution sequence map. S2: Based on the cell population morphological evolution sequence map, determine whether the characteristic directions of adjacent positions are continuously alternating, select the corresponding slice sequence of continuous alternation, locate the associated image region, and obtain the image set of structural transition region; S3: Based on the corresponding region of the image set of the structural transition region, expand the cell population edge segment and splice it along the direction, extract the direction convergence position, draw the closed boundary, analyze the connectivity of the closed boundary, and obtain the set of candidate structures of residual lesions. S4: Based on the set of candidate residual lesions, extract the coordinates of the cell nucleus centroid and the direction of the line pointing to the boundary, compare the angle between adjacent lines, select the location of the directional change, analyze the overlap between the change location and the boundary convergence node, and obtain the lesion structure consistency determination result. S5: Based on the lesion structure consistency determination result, check the corresponding position of the candidate structure in the continuous slice, judge the consistency of the boundary direction and the structural continuity, and obtain the tumor residual lesion identification result.

2. The method for identifying minimal residual tumor lesions based on AI analysis of pathological images according to claim 1, characterized in that, The cell population morphological evolution sequence map includes curvature change positions, arrangement direction inflection positions, corresponding positions of repeating structures, and temporal relationships of positional features. The set of structural inflection region images includes image sequence numbers, alternating position features, and structural inflection image regions. The set of candidate structures for residual lesions includes splicing path convergence positions, closed boundaries, and closed-loop connectivity analysis results. The lesion structure consistency determination results include the angle between the direction of the line connecting the cell nuclei and centroids, the location of directional abrupt changes, and the overlap status of boundary convergence nodes. The tumor residual lesion identification results include structural boundary orientation consistency, structural continuity, and continuous slice position comparison relationships.

3. The method for identifying minimal residual tumor lesions based on AI analysis of pathological images according to claim 1, characterized in that, The cell population contour path refers to a continuous closed curve that characterizes the shape of the outer edge of the cell population; The location of curvature change refers to the point on the cell population outline path where the curvature changes.

4. The method for identifying minimal residual tumor lesions based on AI analysis of pathological images according to claim 1, characterized in that, The closed boundary refers to a closed region enclosed by the cell population outline path; The boundary convergence node refers to the location where the boundaries of multiple cell populations tend to converge.

5. The method for identifying minimal residual tumor lesions based on AI analysis of pathological images according to claim 1, characterized in that, The specific steps of S1 are as follows: S101: Obtain continuous pathological slice images of the tumor resection area, expand the full image of each pathological slice, extract the cell population edge contour paths within the image, and obtain the cell population edge contour path set; S102: Based on the set of cell population edge contour paths, analyze the position of curvature change of cell population edge contour paths, locate the turning point of the direction of the line connecting the cell nuclei and centroids, identify the positional features of repeating structures inside the cell population, and obtain a set of positional features. S103: Based on the location feature set, obtain the pathological slide image and scan sequence number, and integrate the location features within the location feature set to obtain a cell population morphological evolution sequence map.

6. The method for identifying minimal residual tumor lesions based on AI analysis of pathological images according to claim 1, characterized in that, The specific steps of S2 are as follows: S201: Based on the cell population morphological evolution sequence map, extract the feature coordinates of adjacent positions, the direction angle sequence of the feature connection of adjacent positions, and the feature spacing sequence of adjacent positions. Compare the change signs of adjacent items in the adjacent direction angle sequence and determine whether the change signs continue to alternate within the range of three sets of consecutive alternation reference times to obtain the direction alternation segment index table. S202: Based on the alternating direction segment index table, obtain the locally stored pathological slide image acquisition sequence record, associate the segment index with the acquisition sequence record, map the pathological slide image sequence number corresponding to the segment index, and obtain the slide sequence number list. S203: Based on the list of slice numbers and the cell population morphological evolution sequence map, retrieve the feature coordinates of the corresponding position of the segment index to locate the boundary of the region in the pathological slice image, extract the image fragments covered by the region boundary and aggregate the image fragments to obtain a set of images of structural transition regions.

7. The method for identifying minimal residual tumor lesions based on AI analysis of pathological images according to claim 1, characterized in that, The specific steps for S3 are as follows: S301: Based on the pathological slide image region corresponding to the structural transition region image set, retrieve the cell population edge path within the region, extend the cell population edge path to the adjacent edge segment, and make a consistency judgment on the extended edge segment to obtain a continuous path set of edge segments. S302: Based on the continuous path set of the edge segment, analyze the connection status of adjacent path segments, aggregate path segments with consistent directions, identify the location features of the gradual convergence of path segments, and obtain the set of convergence locations. S303: Based on the set of convergence positions, draw closed boundary trajectories around each convergence position, analyze the closed-loop connectivity of the closed boundary trajectories, identify the closed-loop connectivity boundary trajectories, and obtain a set of candidate structures for residual lesions.

8. The method for identifying minimal residual tumor lesions based on AI analysis of pathological images according to claim 7, characterized in that, In the process of determining the consistency of the extended edge segments: the included angle between each pair of adjacent edge segments is calculated, and by determining whether the included angle is within a preset angle range, the direction of the adjacent edge segments is analyzed to see if they remain continuous. In the process of identifying the location features of the gradual convergence of path segments: the extension direction is traced on each path segment in the continuous path set of the edge segment, the direction relationship is compared with the direction relationship of the surrounding path segments, and the area where the direction converges is taken as the convergence position of the direction.

9. The method for identifying minimal residual tumor lesions based on AI analysis of pathological images according to claim 1, characterized in that, The specific steps of S4 are as follows: S401: Based on the region corresponding to the candidate structure in the set of candidate residual lesions, detect the position coordinates of the nucleus centroid within the region, detect the position coordinates of the candidate structure boundary path and the boundary convergence node, extract the direction sequence of the line connecting the nucleus centroid to the candidate structure boundary path, and obtain the centroid boundary connection direction sequence. S402: Based on the centroid boundary line connection direction sequence, read the preset direction change angle threshold, compare the direction angle of adjacent connection with the direction change angle threshold, select the position coordinates corresponding to the direction of the connection that continuously exceeds the direction change angle threshold, and obtain the direction change position set; S403: Based on the set of directional mutation locations, call the coordinates of the boundary convergence node positions in the set of candidate residual lesions, read the preset spatial overlap distance threshold, determine the relationship between the directional mutation location and the distance between the boundary convergence nodes and the spatial overlap distance threshold, and obtain the lesion structure consistency determination result.

10. The method for identifying minimal residual tumor lesions based on AI analysis of pathological images according to claim 1, characterized in that, The specific steps of S5 are as follows: S501: Based on the lesion structure consistency determination result, retrieve the candidate structure identifier in the residual lesion candidate structure set, read the pathological slice number and image region coordinates corresponding to the candidate structure, associate the candidate structure identifier with the pathological slice number and image region coordinates, and obtain the candidate structure slice location table. S502: Based on the candidate structure slice location table, check the image region coordinates of the same candidate structure identifier in consecutive pathological slices, determine whether the closed boundary direction in adjacent pathological slices remains in the same direction, analyze whether the closed boundary morphology in adjacent pathological slices remains connected and continuous, and obtain a set of structures with consistent continuous boundaries. S503: Based on the continuous boundary consistent structure set, determine the correspondence between the candidate structure identifiers in the continuous boundary consistent structure set and the candidate structure identifiers in the residual lesion candidate structure set, identify the corresponding relationship to establish the candidate structure identifier, and obtain the tumor residual lesion identification result.

Images