Anomaly Detection in Endoscopic Videos: Keyframe Extraction to Designing Clinical and Synthetic Datasets

Abstract

Gastrointestinal (GI) cancers, specifically colorectal cancers (CRC), are prevalent and significant contributors to global cancer-related deaths. CRC originates from pre-malignant polyps, which can be detected through a colonoscopy procedure, during which videos of a patient's colon are captured. However, analyzing screening videos for related diagnosis and treatment faces challenges due to a large proportion of low-quality data, risking human review errors. Further, the low-quality data and the limited availability of large-scale annotated datasets pose significant hurdles in building automated computer-aided diagnostic systems. This thesis addresses these challenges while aligning with standard clinical procedures. To maintain this uniformity, we mimic these manual procedures in our proposed automated pipeline and present solutions to problems encountered at different stages.