On normalized convolution to measure curvature features for automatic polyp detection

Early removal of polyps has proven to decrease the incidence of colon cancer. We aim to increase the sensitivity of the screening by automatic detection of polyps. It requires accurate measurement of the colon wall curvature. This paper describes a new method which computes the curvatures using space-variant derivative operators in a strip along the edge of the colon. It optimizes the trade-off between noise reduction and mixing of adjacent image structures. The derivative operators incorporate an applicability function for regularization and interpret the strips as confidence measure; certain inside and uncertain outside. To that purpose the technique of normalized convolution is utilized and adapted to allow a local Taylor expansion of the image signal. A special scheme to compute the confidence values is also presented.