Severity Grading of Ulcerative Colitis Using Endoscopy Images: An Ensembled Deep Learning and Transfer Learning Approach

Ulcerative colitis (UC) is a persistent condition necessitating prompt treatment to avert potential complications. Detecting UC severity aids treatment decisions. The Mayo-endoscopic subscore is a standard for UC severity grading (UCSG). Deep learning (DL) and transfer learning (TL) have enhanced severity grading, but ensemble learning’s impact remains unexplored. This study designed DL-ensemble and TL-ensemble models for UCSG. Using the HyperKvasir dataset, we classified UCSG into two stages: initial and advanced. Three deep convolutional neural networks were trained from scratch for DL, and three pre-trained networks were trained for TL. UCSG was conducted using a majority voting ensemble scheme. A detailed comparative analysis evaluated individual networks. It is observed that TL models perform better than the DL models, and implementation of ensemble learning enhances the performance of both DL and TL models. Following a comprehensive assessment, it is observed that the TL-ensemble model has delivered the optimal outcome, boasting an accuracy of 90.58% and a MCC of 0.7624. This study highlights the efficacy of our methodology. TL-ensemble models, especially, excelled, providing valuable insights into automatic UCSG systems’ potential enhancement. Ensemble learning offers promise for enhancing accuracy and reliability in UCSG, with implications for future research in this field.