Semi-automatic Algorithm for Lumen Segmentation in Histological Images

In this paper we focus on a problem of lumen segmentation in histological images. A large number of annotated images are necessary for the development of diagnostic algorithms that can help to detect changes, such as lumen serration, indicating really serious health problems like cancer. We propose a semi-automatic interactive segmentation algorithm to accelerate the process of manual image annotation. The core of our annotation approach is a classical graph-cut algorithm that uses manually selected parameters. The user annotates an image with two types of scribbles corresponding to the gland lumen structure and the non-lumen area. After that, the model annotates all unlabeled pixels, providing the user with a fully annotated image based on the scribbled input data. The user can interact with the annotation algorithm and add new scribbles to adjust the result. The algorithm allows to reduce the annotation time of a typical histological image by 10 times for the PATH-DT-MSU dataset that can potentially seriously increase the number of fully annotated histological images.