Class Attention-aware Knowledge Distillation for Pathological Score Prediction in Endocytoscopy Images

Ulcerative colitis (UC) is a chronic inflammatory bowel disease causing irritation and ulcers in colon. To facilitate the diagnosis and treatment of UC, pathological scoring is introduced to evaluate inflammatory activity. However, manual scoring performed by pathologists is subjective, experience-dependent, and time-consuming. In this work, we propose a novel deep learning-based approach to predict pathological scores based on images collected from endocytoscopy (EC). Pathological scoring can be formulated as a multi-target prediction problem, in which either local strategy or global strategy has its own strengths and limitations. To strike a balance between them, genetic algorithm is applied to determine the optimal partition of targets based on inter-target correlations. When training the deep learning model, class attention-aware knowledge distillation is adopted to compress the model and avoid overfitting. The experimental results demonstrate that our proposed method can optimally partition the prediction targets and accurately predict pathological scores given endocytoscopy images.