Automatic Analysis of Endoscopic Images for Polyps Detection and Segmentation

Polyps of the gastrointestinal tract are avowed as a risk factor for cancer. Early polyps detection and removal can reduce the probability of cancer. Quality of endoscopic examination strongly depends on the physician experience and other subjective factors. Automatic classification allows to increase effectiveness of gastroscopy. The present paper discusses the polyp segmentation task and its solution using automatic two-stage classification. The first stage is binary classification (presence / absence of polyp) based on the global features taken from the initial endoscopic images. The second stage involves the use of convolutional neural networks for segmentation. The result of the method is a pathology map with polyps marked by color. Evaluations on standard public databases show that the proposed method has specificity - 0.82 and sensitivity - 0.93.