Segmentation for classification of gastroenterology images

Automatic classification of cancer lesions in tissues observed using gastroenterology imaging is a non-trivial pattern recognition task involving filtering, segmentation, feature extraction and classification. In this paper we measure the impact of a variety of segmentation algorithms (mean shift, normalized cuts, level-sets) on the automatic classification performance of gastric tissue into three classes: cancerous, pre-cancerous and normal. Classification uses a combination of color (hue-saturation histograms) and texture (local binary patterns) features, applied to two distinct imaging modalities: chromoendoscopy and narrow-band imaging. Results show that mean-shift obtains an interesting performance for both scenarios producing low classification degradations (6%), full image classification is highly inaccurate reinforcing the importance of segmentation research for Gastroenterology, and confirm that Patch Index is an interesting measure of the classification potential of small to medium segmented regions.