CN-122023444-A - Intelligent segmentation method for large-field pathological section image

CN122023444ACN 122023444 ACN122023444 ACN 122023444ACN-122023444-A

Abstract

The invention relates to the technical field of image segmentation, in particular to an intelligent segmentation method of a large-visual-field pathological section image, which solves the technical problem that color deviation influences segmentation accuracy due to uneven dyeing in the prior art. The method comprises the steps of distinguishing a background area from a tissue area according to color distribution and dyeing depth of a plurality of pixels in a large-field pathological section image, classifying cell images in the tissue area, comparing each pixel with global color distribution in the cell image of each cell type, determining a color deviation area of each cell type, screening a reference area which is adjacent to the color deviation area and is similar in structure and normal in color distribution in the cell image of each cell type according to structural similarity of the image, carrying out color correction on the color deviation area according to the color of the reference area, reconstructing the large-field pathological section image according to the corrected color deviation area, carrying out cell segmentation, and outputting a cell segmentation result.

Inventors

XIAO XIAOGUANG
WEI BOWEN

Assignees

无锡晶桔数字科技有限公司

Dates

Publication Date: 20260512
Application Date: 20260131

Claims (10)

1. An intelligent segmentation method for a large-view pathological section image is characterized by comprising the following steps: Distinguishing a background area from a tissue area according to color distribution and dyeing depth of a plurality of pixels in the large-view pathological section image, and classifying cell images in the tissue area; comparing each pixel with the global color distribution in the cell image of each cell type, and determining the color deviation area of each cell type; Screening a reference area which is adjacent to the color deviation area, has similar structure and has normal color distribution in the cell image of each cell type according to the structural similarity of the images; Performing color correction on the color deviation area according to the color of the reference area; reconstructing the large-view pathological section image according to the corrected color deviation area, and performing cell segmentation on the reconstructed large-view pathological section image to output a cell segmentation result.
2. The intelligent segmentation method according to claim 1, wherein comparing each pixel with the global color distribution in the cell image of each cell type, determining the color deviation area of each cell type comprises: Calculating the difference degree between each pixel and the global color distribution according to the pixel values of a plurality of pixels in the cell image of each cell type; marking pixels with the difference degree larger than a preset difference threshold value of each type of cells in the cell image of each type of cells as color deviation pixels; and clustering based on the density of the color deviation pixels to determine the color deviation area of each cell.
3. The intelligent segmentation method according to claim 2, wherein the large-field pathological section image is dyed by a first dye and a second dye, the first dye and the second dye being used for dyeing different tissue parts respectively and having different dyeing effects; the calculating the difference degree between each pixel and the global color distribution according to the pixel values of a plurality of pixels in the cell image of each cell type comprises the following steps: decomposing the cell image of each cell type into a first channel stained with the first stain and a second channel stained with the second stain; Calculating a first pixel mean value of a plurality of pixels in the first channel, a second pixel mean value of a plurality of pixels in the second channel and a pixel covariance matrix of the first channel and the second channel, wherein the pixel mean value is used for representing a center level of normal dyeing; Determining a global mean vector of the cell image of each cell type according to the first pixel mean value and the second pixel mean value; Determining the feature vector of each pixel according to the pixel values of each pixel in the first channel and the second channel respectively; And calculating the mahalanobis distance between each pixel and the global color distribution according to the characteristic vector of each pixel, the global mean vector and the covariance matrix.
4. The intelligent segmentation method according to claim 2, wherein the screening the reference region adjacent to the color deviation region, similar in structure and normal in color distribution in the cell image of each cell type according to the structural similarity of the images comprises: Determining a coverage area covering the color deviation area in the cell image of each cell according to the edge characteristics of the color deviation area, and determining the areas, except the color deviation area, of the coverage area, which accord with the global normal color distribution of the cells, as a reference window; dividing the reference window into a plurality of subareas according to the cell edges in the reference window, and calculating the structural similarity index of each subarea and the color deviation area in a mode of eliminating or weakening the influence of the color difference; And determining the subarea with the structural similarity index larger than a preset similarity threshold value as the reference area.
5. The intelligent segmentation method according to claim 4, wherein the performing color correction on the color deviation region according to the color of the reference region includes: calculating the color offset of the color deviation area relative to the reference area according to the color vector of the reference area and the color vector of the color deviation area; And correcting the color value of the color deviation area according to the color offset to obtain a corrected new color value.
6. The intelligent segmentation method according to claim 5, wherein if the reference region includes a plurality of sub-regions, the correcting the color value of the color deviation region according to the color deviation amount to obtain a corrected new color value includes: According to the structural similarity index of each sub-region in the reference region and the color deviation region, assigning a weight to the color deviation amount of the color deviation region relative to each sub-region in the reference region; and correcting the color value of the color deviation area through weighted average to obtain the new color value.
7. The intelligent segmentation method according to claim 4, wherein the determining the coverage area covering the color deviation area in the cell image of each cell type for the edge feature of the color deviation area includes: And generating the coverage area according to a multiple scale factor based on the size of the boundary box of the color deviation area, wherein the center of the coverage area coincides with the center of the color deviation area.
8. The intelligent segmentation method according to claim 1, wherein the distinguishing the background region from the tissue region according to the color distribution and the dyeing depth of the plurality of pixels in the large-field pathological section image comprises: analyzing the color distribution of each pixel to obtain a saturation value of each pixel, and simultaneously analyzing the dyeing depth of each pixel to obtain a brightness value of each pixel; And determining pixels with saturation values lower than a preset saturation threshold and brightness values higher than a preset brightness threshold as pixels of the background area, and determining other areas except the background area in the large-view pathological section image as the tissue area.
9. The intelligent segmentation method according to claim 8, wherein the classifying of the cell image in the tissue region includes: analyzing the distribution characteristics of a plurality of pixels in the tissue region, and dividing and classifying cells in the tissue region to obtain a cell image of each type of cells, wherein the distribution characteristics at least comprise optical density distribution.
10. The intelligent segmentation method as set forth in claim 1, further comprising: scanning the pathological section to obtain an original image with resolution higher than preset resolution; And performing image preprocessing on the original image to obtain a large-field pathological section image for eliminating initial interference, wherein the image preprocessing at least comprises white balance correction, image flattening and Gaussian filtering.

Description

Intelligent segmentation method for large-field pathological section image Technical Field The invention relates to the technical field of image segmentation, in particular to an intelligent segmentation method of a large-view pathological section image. Background The large-field pathological section image (WSI) is a digitized image obtained by scanning a traditional glass pathological section at a high speed and high resolution through a full-automatic digital scanning system, and the image generally has a resolution of billions to billions pixels, can completely present the whole appearance of a tissue sample, covers multi-level information from a macroscopic tissue architecture to microscopic cell morphology, is now an important foundation for developing modern pathological diagnosis, teaching and scientific research work, and provides comprehensive and fine image data support for related fields. The current intelligent segmentation technology for large-field pathological section images mainly depends on two technical means, one is a deep learning model such as U-Net, deep Lab and the like, and the other is a traditional image processing algorithm such as threshold segmentation, level set, graph cut and the like. In practical applications, most of these methods directly classify the original image at pixel level, so as to distinguish different tissue regions, such as nuclei, cytoplasm, interstitials, etc., or identify lesion regions, such as tumor nests and inflammatory infiltrates, thereby providing a preliminary region division result for pathological analysis. At present, the large-view pathological section image segmentation process still faces a remarkable technical bottleneck, the generalization capability of a deep learning model can be remarkably reduced due to color deviation caused by uneven dyeing, meanwhile, the contrast between cell nuclei and cytoplasm can be covered by color distortion, so that the extraction of morphological characteristics is influenced, and finally, the conditions of fuzzy segmentation boundaries and high classification error rate are caused, so that the conventional segmentation method is poor in adaptability to pathological images with uneven dyeing, cells and tissue structures are difficult to accurately distinguish, and the accuracy of subsequent quantitative analysis and the reliability and consistency of clinical diagnosis are seriously influenced. Disclosure of Invention In order to solve the technical problems that cells and tissue structures are difficult to accurately distinguish due to color deviation caused by uneven dyeing when a large-view pathological section image is segmented in the prior art, and the segmentation precision is further influenced, the invention aims to provide an intelligent segmentation method of the large-view pathological section image, which adopts the following technical scheme: according to the structural similarity of the images, reference areas which are adjacent to the color deviation areas and are similar in structure and normal in color distribution are screened in the cell images of the cells of each type, color correction is conducted on the color deviation areas according to the colors of the reference areas, reconstruction is conducted on the large-view pathological section images according to the corrected color deviation areas, cell segmentation is conducted on the reconstructed large-view pathological section images, and cell segmentation results are output. Based on the technical scheme, in the intelligent segmentation method of the large-view pathological section image, the background area and the tissue area are distinguished according to the pixel color distribution and the dyeing depth, and the cell images in the tissue area are classified, so that the background interference is effectively eliminated, and the subsequent treatment can be carried out aiming at different cell characteristics. And determining a color deviation area by comparing pixels with the global color distribution in each cell image, so that the problem area of uneven dyeing is accurately positioned. And then screening out a reference area which is adjacent to the color deviation area, has similar structure and normal color distribution according to the structural similarity of the image, ensuring the consistency of the reference for correction and the tissue structure of the deviation area, and avoiding the original structure of the image from being damaged in the correction process. And the color deviation area is corrected based on the color of the reference area, so that the color deviation caused by uneven dyeing is directly eliminated, and the normal contrast between the cells and the tissue structure is recovered. Finally, by reconstructing the corrected image and carrying out cell segmentation output, the problem that in the prior art, the cell and tissue structure are difficult to accurately distinguish due to uneven dyeing, the segment