의료비디오의 인식, 인페인팅, 정량화를 위한 딥러닝 연구 개발 및 유양돌기 절제술과 대장내시경의 적용

Abstract

This study explores the deep learning technologies for medical video analysis. Applying these technologies in the medical field poses challenges, including difficulties in data collection and the unique nature of medical video. This paper addresses these challenges by applying advanced deep learning techniques such as Knowledge Distillation, Implicit Neural Representation, and Deep Metric Learning to the analysis of medical video data. We particularly focus on Surgical Phase Recognition and Video Inpainting techniques for tympanomastoidectomy (TM), and methods for quantifying colonoscopy video data. In this research, three experiments were conducted to analyze medical videos. The first study proposed a TM surgical phase recognition learning method using Teacher-Student learning, showing improved performance in situations of class imbalance and data scarcity. The second study introduced a method using implicit neural representation for high-resolution video inpainting without the need for large-scale video data collection, demonstrating enhanced visual performance compared to state-of-the-art model. The third study proposed a method integrating quality-aware metric learning with a noise-robust phase recognition model, considering the characteristics of the colon, leading to more accurate quantification in colonoscopy. We believe that the methodologies developed in this study have the potential to be applied not only to TM surgery and colonoscopy videos but also to a broader range of medical video analysis areas.