Segmenting Glandular Biopsy Images Using the Separate Merged Objects Algorithm

Abstract

The analysis of the structure of histopathology images is crucial in determining whether biopsied tissue is benign or malignant. It is essential in pathology to be precise and, at the same time, to be able to provide a quick diagnosis. These imperatives inspired researchers to automate the process of segmenting and diagnosing biopsies. The main approach is to utilize semantic segmentation networks. Our research presents a post-processing algorithm that addresses one weakness of the semantic segmentation method - namely, the separation of close objects that have been mistakenly merged by the classification algorithm. If two or more objects have been merged, the object can be mis-classified as cancerous. This might require the pathologist to manually validate the biopsy. Our algorithm separates the objects by drawing a line along the points where they touch. We determine whether a line should be passed along the edges to separate the objects according to a loss function that is derived from probabilities based on semantic segmentation (of various classes of pixels) and pixel distances from the contour. This method is general and can be applied to different types of tissue biopsies with glandular objects. We tested the algorithm on colon biopsy images. The newly developed method was able to improve the detection rate on average from 76% to 86%.