CN-121982713-A - Multi-mode section image recognition method and system for cervical pathology

CN121982713ACN 121982713 ACN121982713 ACN 121982713ACN-121982713-A

Abstract

The application relates to the technical field of image processing, in particular to a method and a system for identifying cervical pathology multi-mode section images. The method comprises the steps of obtaining two different staining slice images of the same cervical tissue sample, respectively carrying out semantic structure recognition to extract biological structure types and geometric positions thereof, generating two groups of feature point sets with structure type labels according to the biological structure types, respectively analyzing spatial relations among feature points to construct topological descriptions, combining the two groups of topological descriptions to determine corresponding matching of cross-modal feature points under the constraint of structure type consistency, calculating sparse deformation information by matching point pairs and interpolating to generate a dense deformation field representing mapping from a second image to a first image, further carrying out iterative optimization to obtain a fine deformation field, and carrying out geometric transformation on the second slice image to realize alignment with the first slice image. The method solves the technical problem that the images of the multi-mode sections are difficult to align accurately due to tissue deformation in the pathological section making process, so that fusion analysis errors are caused.

Inventors

WANG CHUNRONG
WANG WEIYUAN
HE LANG
TIAN YIFU
HE QIONGQIONG
XIAO DESHENG
ZHOU JIANHUA

Assignees

中南大学湘雅医院

Dates

Publication Date: 20260505
Application Date: 20251230

Claims (10)

1. A method for identifying a multi-modal slice image of cervical pathology, comprising: Acquiring a first slice image and a second slice image of the same cervical tissue sample, wherein the first slice image and the second slice image respectively correspond to digitized images of different physical tissue slices of the cervical tissue sample and are slice images with different staining modes; Respectively carrying out semantic structure recognition on the first slice image and the second slice image to obtain biological structure information of the first slice image and biological structure information of the second slice image, wherein the biological structure information at least comprises the structure type of a recognized structure and geometric position parameters of a corresponding structure; Generating a first characteristic point set according to the biological structure information of the first slice image, and generating a second characteristic point set according to the biological structure information of the second slice image, wherein each characteristic point in the first characteristic point set and the second characteristic point set is determined by a central point and/or a key contour point of a corresponding biological structure and is associated with the structure type of the corresponding biological structure; Analyzing the spatial relationship between the characteristic points in the first characteristic point set, constructing a first topological relationship description, analyzing the spatial relationship between the characteristic points in the second characteristic point set, and constructing a second topological relationship description, wherein the first topological relationship description and the second topological relationship description at least comprise Euclidean distance and/or connection vector included angles between the characteristic points and connection relationships determined by adjacent relationships; under the structural type consistency constraint, combining a first topological relation description and a second topological relation description, determining a corresponding matching relation between the first characteristic point set and the second characteristic point set, wherein the structural type consistency constraint is that only characteristic points with the same structural type are allowed to form candidate matching points; Calculating sparse deformation information according to the corresponding matching relation, and constructing a dense deformation field based on the sparse deformation information through interpolation, wherein the dense deformation field is used for representing the mapping relation from the second slice image coordinate to the first slice image coordinate; Taking the dense deformation field as an initial value, and obtaining a fine deformation field through iterative optimization; And performing geometric transformation on the second slice image according to the fine deformation field to obtain a second slice image aligned with the first slice image.
2. The cervical pathology multi-modal slice image recognition method according to claim 1, wherein the second slice image is an immunohistochemical staining slice image, and the determining the corresponding matching relationship between the first feature point set and the second feature point set by combining a first topological relation description and a second topological relation description under the constraint of structural type consistency comprises: screening second feature points with the same structure type from the second feature point set as candidate matching points according to the structure type of the first feature points in the first feature point set; Verifying the structure type of the candidate matching point according to the immunohistochemical staining intensity and the distribution pattern of the region corresponding to the candidate matching point in the second slice image, correcting the structure type of the candidate matching point when the candidate matching point does not meet the preset staining characteristics, and removing the candidate matching point inconsistent with the structure type of the first characteristic point after correction to obtain a reserved candidate matching point; Calculating the local topological environment similarity of candidate point pairs by combining the first topological relation description and the second topological relation description on the candidate point pairs formed by the reserved candidate matching points and the first characteristic points; and determining the corresponding matching relation according to the local topological environment similarity.
3. The cervical pathology multi-modal slice image recognition method according to claim 1, wherein obtaining a fine deformation field by iterative optimization with the dense deformation field as an initial value, comprises: Identifying a first damage region in the first slice image and judging a first damage type corresponding to the first damage region, and identifying a second damage region in the second slice image and judging a second damage type corresponding to the second damage region; extracting biological context information of the first damaged area and biological context information of the second damaged area respectively; Performing morphological reasoning on the first damage area according to the first damage type and the biological context information thereof, and generating a first slice image subjected to morphological reasoning; performing morphological reasoning on the second damage region according to the second damage type and the biological context information thereof, and generating a second slice image subjected to morphological reasoning; and taking the first slice image subjected to morphological reasoning and the second slice image subjected to morphological reasoning as inputs, and carrying out iterative optimization adjustment on the dense deformation field under the condition of taking the dense deformation field as an initial value so as to obtain the fine deformation field.
4. A cervical pathology multi-modal slice image recognition method according to claim 3, wherein morphological reasoning is performed on the first lesion region according to the first lesion type and the biological context information thereof, and generating a morphologically-reasoning first slice image comprises: analyzing the first lesion field to identify a plurality of first lesion types present therein; when a plurality of first damage types are identified, determining a dominant damage type according to a preset damage type priority, and performing morphological reasoning on the first damage area according to the dominant damage type and biological context information of the first damage area; Classifying untrained damage types as unknown damage types when the untrained damage types are contained in the identified first damage types, and starting morphological completion operation based on local structural similarity, wherein the morphological completion operation comprises searching local areas with similarity of morphology and arrangement modes greater than a preset similarity threshold in healthy tissues around the first damage areas, and migrating and adaptively adjusting morphological characteristics of the local areas to obtain morphological completion results aiming at the unknown damage types; And generating a first slice image after morphological reasoning based on the morphological reasoning result and/or the morphological completion result.
5. The cervical pathology multi-modal slice image recognition method according to claim 4, wherein analyzing the first lesion region to identify a plurality of first lesion types present therein includes: Processing the image data of the first damaged area to obtain first image features of multiple scales; identifying a first damage type existing in the first damage area according to the first image characteristics; determining boundaries of the areas corresponding to the identified first damage types, quantifying the area, the shape regularity and the edge continuity of surrounding healthy tissues of each area, and obtaining a first quantification result; and judging whether a plurality of first damage types are overlapped in the first damage area according to the first quantification result.
6. The cervical pathology multi-modal slice image recognition method according to claim 5, wherein processing the image data of the first lesion area to obtain a plurality of scale first image features includes: preprocessing the image data of the first damage area to improve the contrast of the micro damage; Performing multi-resolution decomposition on the preprocessed image data of the first damaged area to obtain first image levels with different resolutions; extracting first features of nuclear morphology, cell gap width and local tissue texture for each first image level; and fusing the first features of each first image level to obtain first image features with multiple scales.
7. The cervical pathology multi-modality slice image recognition method of claim 6, wherein preprocessing the image data of the first lesion area, improving contrast of microscopic lesions, includes: acquiring gray value distribution of each pixel point in the image data of the first damaged area; According to the gray value distribution, determining a gray value range of a normal tissue background in the first damaged area; according to the gray value range, carrying out local contrast stretching on the image data of the first damage area so as to improve gray difference between the micro damage and the normal tissue background; And carrying out edge enhancement processing on the image data of the first damaged area after the local contrast stretching, highlighting the boundary information of the micro damage, and simultaneously inhibiting the smooth transition of the normal tissue area.
8. The cervical pathology multi-modal slice image recognition method according to claim 6, wherein the multi-resolution decomposition of the preprocessed image data of the first lesion area results in a first image hierarchy of different resolutions, comprising: determining the number of multi-resolution decomposition levels and decomposition parameters of each decomposition level according to the cell density and the heterogeneity of tissue arrangement of a local area in the preprocessed image data of the first damaged area to obtain decomposition levels; On each decomposition level, performing self-adaptive filtering processing on the image data of the first damage region, adjusting the size and the intensity of a filtering kernel according to the image characteristics and the noise level of the current decomposition level, and reserving fine structure information of micro damage to obtain a corresponding filtering image; Analyzing local gradient change of damage features in a decomposition level with relatively high resolution and corresponding macrostructure distribution in a decomposition level with relatively low resolution based on a filtering image corresponding to each decomposition level between decomposition levels with adjacent resolutions, and constructing cross-level feature association mapping; Directing the transmission of damaged features in a relatively high resolution decomposition level to a relatively low resolution decomposition level using the feature association map and correcting feature offset caused by nonlinear interactions; And separating corrected damage features in the filtered image at each resolution decomposition level to obtain first image levels with different resolutions.
9. The cervical pathology multi-modal slice image recognition method according to claim 6, wherein extracting first features of nuclear morphology, intercellular width, and local tissue texture for each of the first image levels includes: Evaluating, on each of the first image levels, a local brightness distribution and color saturation of image data of the first image level, identifying a characteristic distortion region caused by uneven staining or local differences in production; performing local color correction and brightness equalization processing on the characteristic distortion region according to the local brightness distribution and the color saturation to obtain image data of a first image level after equalization correction; when the image data of the first image level after the equalization correction is subjected to nuclear morphological feature extraction, a segmentation threshold value is adjusted according to local brightness distribution, so that first nuclear morphological features are obtained; When the image data of the first image level after the equalization correction is subjected to cell gap width characteristic extraction, adjusting identification parameters according to local pixel distribution to obtain first cell gap width characteristics; when carrying out local tissue texture feature extraction on the image data of the first image level after the balance correction, adjusting texture feature extraction parameters according to gray level distribution to obtain first local tissue texture features; And performing consistency check on the first nuclear morphological characteristic, the first cell gap width characteristic and the first local tissue texture characteristic to obtain the first characteristic.
10. A cervical pathology multi-modal slice image recognition system, comprising: the device comprises a slice image acquisition module, a first image acquisition module and a second image acquisition module, wherein the slice image acquisition module is used for acquiring a first slice image and a second slice image of the same cervical tissue sample, and the first slice image and the second slice image respectively correspond to digitized images of different physical tissue slices of the cervical tissue sample and are slice images with different staining modes; The slice image processing module is used for respectively carrying out semantic structure identification on the first slice image and the second slice image to obtain biological structure information of the first slice image and biological structure information of the second slice image, wherein the biological structure information at least comprises the structure type of an identified structure and geometric position parameters of a corresponding structure; The characteristic point set generation module is used for generating a first characteristic point set according to the biological structure information of the first slice image and generating a second characteristic point set according to the biological structure information of the second slice image, wherein each characteristic point in the first characteristic point set and the second characteristic point set is determined by a central point and/or a key contour point of a corresponding biological structure and is associated with the structure type of the corresponding biological structure; the topological relation description construction module is used for analyzing the spatial relation among the characteristic points in the first characteristic point set, constructing a first topological relation description, analyzing the spatial relation among the characteristic points in the second characteristic point set, and constructing a second topological relation description, wherein the first topological relation description and the second topological relation description at least comprise Euclidean distance and/or connection vector included angles among the characteristic points and connection relation determined by adjacent relations; The corresponding matching relation determining module is used for determining the corresponding matching relation between the first characteristic point set and the second characteristic point set by combining the first topological relation description and the second topological relation description under the structural type consistency constraint, wherein the structural type consistency constraint is that only the characteristic points with the same structural type are allowed to form candidate matching points; The computation module is used for computing sparse deformation information according to the corresponding matching relation, and constructing a dense deformation field through interpolation based on the sparse deformation information, wherein the dense deformation field is used for representing the mapping relation from the second slice image coordinate to the first slice image coordinate; the adjustment module is used for obtaining a fine deformation field through iterative optimization by taking the dense deformation field as an initial value; And the image transformation module is used for performing geometric transformation on the second slice image according to the fine deformation field to obtain a second slice image aligned with the first slice image.

Description

Multi-mode section image recognition method and system for cervical pathology Technical Field The application relates to the technical field of image processing, in particular to a method and a system for identifying multi-mode section images of cervical pathology. Background In the cervical pathology diagnosis field, methods for performing intelligent fusion analysis on various pathological information are always being explored. Such information typically includes conventional hematoxylin-eosin (H & E) stained slice images, immunohistochemical (IHC) slice images, and Human Papillomavirus (HPV) typing data. However, in practice, due to some routine manipulation in pathological section production procedures, such as "chunking" treatment of tissue wax for optimal diagnostic section acquisition, two tissue sections for H & E staining and immunohistochemical staining are not from exactly the same tissue plane, with a small three-dimensional spatial difference between the two. This discrepancy appears as a complex nonlinear tissue deformation on the two-dimensional digital image, making it difficult for existing image alignment techniques to achieve accurate correction at the cellular level. The resulting image registration errors can cause the system to falsely correlate cell morphology information on one slice with protein expression information of cells or regions that are not related on another slice during fusion analysis, thereby producing contradictory or spurious fusion characteristics. In view of the above, there is a need in the art for improvements. Disclosure of Invention The application discloses a method and a system for identifying multi-mode section images of cervical pathology, and aims to solve the technical problem that in the cervical pathology field, multi-mode section images are difficult to align accurately due to tissue deformation in a pathology sheet making process, so that fusion analysis errors are caused. The technical scheme of the application is as follows: in a first aspect, the application discloses a cervical pathology multi-mode slice image identification method, which specifically comprises the following steps: Acquiring a first slice image and a second slice image of the same cervical tissue sample, wherein the first slice image and the second slice image respectively correspond to digitized images of different physical tissue slices of the cervical tissue sample and are slice images with different staining modes; Respectively carrying out semantic structure identification on the first slice image and the second slice image to obtain biological structure information of the first slice image and biological structure information of the second slice image, wherein the biological structure information at least comprises the structure type of the identified structure and geometric position parameters of the corresponding structure; Generating a first characteristic point set according to the biological structure information of the first slice image, generating a second characteristic point set according to the biological structure information of the second slice image, wherein each characteristic point in the first characteristic point set and the second characteristic point set is determined by a central point and/or a key outline point of a corresponding biological structure and is associated with the structure type of the corresponding biological structure; analyzing the spatial relationship between the feature points in the first feature point set, constructing a first topological relationship description, analyzing the spatial relationship between the feature points in the second feature point set, and constructing a second topological relationship description, wherein the first topological relationship description and the second topological relationship description at least comprise Euclidean distance and/or connection vector included angles between the feature points and connection relationships determined by adjacent relationships; under the structural type consistency constraint, combining the first topological relation description and the second topological relation description, determining the corresponding matching relation between the first characteristic point set and the second characteristic point set, wherein the structural type consistency constraint is that only the characteristic points with the same structural type are allowed to form candidate matching points; Calculating sparse deformation information according to the corresponding matching relation, and constructing a dense deformation field based on the sparse deformation information through interpolation, wherein the dense deformation field is used for representing the mapping relation from the second slice image coordinates to the first slice image coordinates; taking the dense deformation field as an initial value, and obtaining a fine deformation field through iterative optimization; And performing geo