False Positive Reduction in Colonic Polyp Detection Using Glocal Information

Abstract

This paper proposes an explicit parametric model for colonic polyps. The model captures the overall shape of the polyp and is then used to derive the probability distribution of features relevant for polyp detection. The probability distribution represents the glocal properties of the polyp candidates, where the glocal properties capture both global and local information of an object. The probability distribution is implemented on the unit sphere, which is divided into 26 partitions, and each partition captures the local properties of a polyp candidate. From the partitions on the sphere, an observation sequence also defines global properties of the polyp candidate and the observation sequence is assessed by explicit models for classification. When it represents glocal parameters of a polyp candidate, we call the unit sphere a brilliant sphere. The parametric models are estimated from 20 geometric models typifying the various cap shapes of colonic polyps.