Colonic Polyp Identification Using Pareto Depth Anomaly Detection Algorithm

Abstract

Colon cancer prevention, diagnosis, and prognosis are directly related to the identification of colonic polyps, in colonoscopy video sequences. In addition, diagnosing colon cancer in the early stages improves significantly the chance of surviving and effective treatment. Due to the large number of images that come from colonoscopy, the identification of polyps needs to be automated for effciency. In this paper, we propose a strategy for automatic polyp recognition, based on a recent multi-objective anomaly detection concept, which itself is based on Pareto Depth Analysis (PDA). Clinically, in medical images, polyps are diagnosed based on a few criteria, such as texture, shape and color. Few works use multi-criteria classification in a systematic way for polyp detection. In the present paper we use a PDA approach, to act as a binary classifier for the identification of colonic polyps. The results obtained in a medical dataset, of conventional colonoscopy images, consisting of short videos from 34 different patients, and 34 different polyps, with a total of 1360 different polyp frames, confirm that the proposed method clearly outperforms the single performance of each criterion.