Specular Reflection Detection on Gastroscopic Images

Gastroscope is important in gastric cancer diagnosis. However, the specular reflection is common existed in the gastroscope images and it's easily confused with ulcer. In this paper we develop a method for detecting specular reflection in gastroscopic images. First the Intensity-Saturation joint distribution of region of interest (ROI) is obtained in HSI color space. Then based on the analysis of the Intensity-Saturation (IS) histogram, several templates of different shapes are tested. After the candidate templates are compared on true positive and false positive curves from the comparison tests, triangle template is proved more suitable for recognizing pixels of specular reflection from the others in IS histogram. The experiment results from 64 gastroscopic images show that the classifier gets a true positive rate (TPR) at 71.9%, false positive rate (FPR) at 0.75% and area under the ROC (AUC) at 0.869. There is no confusion with ulcer. It is robust in size of regions and capable of expanding to other endoscopes.