Boundary delineation of reflux esophagitis lesions from endoscopic images using color and texture

Abstract

Automatic assessment of medical images and endoscopic images in particular is an attractive research topic recent years. To achieve this goal, many tasks must be conducted for example lesions detection, segmentation and classification. In order to design suitable models for such tasks, it would be preferable to know at first: i) which characteristics that differentiate a lesion from a normal region; ii) how large is the boundary of these two regions that still allows to distinguish them. This paper presents an in-depth study of the role of color and texture features for delineation of boundary between a lesion region and a background region. To this end, from the groundtruth contour of a manually segmented lesion, we first expand two margins in two directions. We name inner margin in the lesion region and outer margin in the background region. We then extract color dependent features in different color spaces (HSV, RGB, Lab) and texture features (LBP, HOG, GLCM) on these two margins. Finally we deploy the Support Vector Machine (SVM) technique to classify two classes (lesion and non-lesion). Extensive experiments conducted on a dataset of endoscopic images answer to our aforementioned questions and give some suggestions for designing suitable models of lesion detection in the future.