A Hierarchical Multi-task Approach to Gastrointestinal Image Analysis

A large number of different lesions and pathologies can affect the human digestive system, resulting in life-threatening situations. Early detection plays a relevant role successful treatment increase current survival rates to, e.g., colorectal cancer. The standard procedure enabling detection, endoscopic video analysis, generates quantities visual data that need to be carefully analyzed by an specialist. Due wide range color, shape, general appearance pathologies, as well highly varying image quality, such process is greatly dependent on operator experience skill. In this work, we detail our solution task multi-category classification images from gastrointestinal (GI) tract within 2020 Endotect Challenge. Our approach based Convolutional Neural Network minimizing hierarchical error function takes into account not only finding category, but also its location GI (lower/upper tract), type (pathological finding/therapeutic intervention/anatomical landmark/mucosal views' quality). We describe paper for challenge polyp segmentation colonoscopies, which was addressed with pretrained double encoder-decoder network. internal cross-validation results show average performance 91.25 Mathews Correlation Coefficient (MCC) 91.82 Micro-F1 score task, 92.30 F1 task. organization provided feedback hidden test set both tasks, resulted 85.61 MCC 86.96 classification, 91.97 segmentation. At time writing no public ranking had been released.