Colonoscopy pit-pattern classification based on shape features

BACKGROUND – With magnifying endoscopy, specialist physicians are able to make a presumptive tissue diagnosis by studying the pit pattern of a lesion in the colon (Kudo et al.), since this pattern corresponds with the histology. This could enable the endoscopist to directly remove adenomatous polyps and leave the hyperplastic polyps in situ, hereby reducing the risk and costs of colonoscopies. GOAL – Investigate the technical feasibility of a system, that supports the endoscopist in classifying the pit pattern type of colonic lesions. METHODS – We propose an algorithm, that uses shape information of the pits to identify the type of pit pattern of the lesion. We define four descriptive measures: Elongation, Circularity, Irregularity and Convexity. These measures are computed for every pit in an image obtained with magnifying colonoscopy. For each pit, the pit type is determined using a trained Support Vector Machine (SVM). The pit pattern is classified as the most occurring pit type in the image, using the occurring percentage as a reliability parameter. RESULTS – We have yet tested the proposed algorithm using 36 clinically validated pit pattern images acquired by a magnifying colonoscope. The system achieved an classification accuracy of 88.9% and similar sensitivity and specificity. CONCLUSION – Our first experiments show that the proposed measures are able to capture the shape of the mucosal pits relatively well and that the pit pattern type can be derived from classified pits. Research is ongoing to further develop and validate the proposed system.