Ensemble Model of Visual Transformer and CNN Helps BA Diagnosis for Doctors in Underdeveloped Areas

Zhenghao Wei; Proceedings of the Asian Conference on Computer Vision (ACCV) Workshops, 2022, pp. 68-84

The diagnosis of Biliary Atresia (BA) is still complicated and high resource consumed. Though sonographic gallbladder images can be used as an initial detection tool, lack of experienced experts limits BA infants to be treated timely, resulting in liver transplantation or even death. We developed a diagnosis tool by ViT-CNN ensemble model to help doctors in underdeveloped area to diagnose BA. It performs better than human expert (with 88.1% accuracy versus 87.4%, 0.921 AUC versus 0.837), and still has an acceptable performance on severely noised images photographed by smartphone, providing doctors in clinical facilities with outdated Ultrasound instruments a simple and feasible solution to diagnose BA with our online tool.