METHOD OF IMMUNOHISTOCHEMICAL IMAGES AUTOMATIC SEGMENTATION

Abstract

In digital pathology, cell detection and classification are often prerequisites for quantifying cell numbers and investigating tissue heterogeneity. Separation and labeling of each cell nucleus is a key task in biomedical image segmentation. Using one approach for all biomedical images is not enough, so it is necessary to develop a universal approach taking into account the input parameters of the image. Immunohistochemistry (IHC) is the primary technology used to quantify the amount of biomarkers at the protein level in tissue samples. In oncological clinical diagnosis and research, immunohistochemical imaging analyzes play a central role in tumor characterization and biomarker assessment. A preliminary diagnosis is usually formed on the basis of a visual microscopic examination by an expert. Immunohistochemistry image processing tasks are particularly challenging for multiplex immunohistochemistry images due to the high level of variability in staining, intensity, and damage resulting from preprocessing. Manual evaluation of HER2-stained microscopic images is characterized by the presence of errors, high labor intensity due to various staining, overlapping areas, and heterogeneous slide parameters.

In this work, an approach to the implementation of the mechanism of automatic selection of immunohistochemical image segmentation algorithms is proposed using the example of Her2/neu type images. Considering the need for adaptive selection of segmentation parameters, the proposed approach allows selection of segmentation algorithms and their input parameters based on input image parameters, such as average brightness level, average level of RGB image channels. Also, in this work, in addition to basic segmentation algorithms, such as k-means, water distribution. Threshold segmentation uses a deep machine learning approach based on the U-net architecture. It was found that for processing an image with a large number of micro-objects, the best results were shown by segmentation using U-net, and for objects with a small number of micro-objects, the best result is shown by a combination of water distribution algorithms in combination with threshold segmentation with values from 65 to 90 .