Superpixel-based Segmentation and Classification of Gastrointestinal Landmarks and Diseases

Endoscopy has long been used as a gold standard for examination and diagnosis in the gastrointestinal (GI) tract. A large volume of images is typically captured during endoscopy, particularly when a video is recorded. These images take a significant amount of endoscopist's time to review and assess. Moreover, the review process is tedious and requires substantial experience. In this paper, we propose a superpixel-based technique for segmenting and classifying GI landmarks and diseases in endoscopic images. We split each image into several superpixels and extract texture features of local binary patterns, gray-level co-occurrence matrices, and first-order statistical features. These features are extracted from grayscale and color channels of different color spaces. Different combinations of the proposed multichannel superpixel features are used to train and test support vector machines. The performance of the superpixel-based method was evaluated and compared against a pixel-based method using the area under the receiver operating characteristic curve (AUC). Our method outperformed the pixel-based one with AUC values of 82.06%, 99.6%, and 97.4% in the classification of the upper, middle, and lower GI parts, respectively. In particular, for the upper GI part, our method resulted in an AUC improvement of 4.52% compared to the pixel-based method.