Abstract
Accurate diagnosis of middle ear diseases, such as acute otitis media (AOM), remains a clinical challenge due to the reliance on subjective visual assessment through otoscopy. While deep learning has shown promise in improving diagnostic accuracy using digital otoscopy videos, existing models often rely on manually selected still frames, a step that reduces their practicality in real-world clinical workflows. In this study, we present the first systematic method (LUCID) for automatically identifying the most informative frame (MIF) selection in otoscopy videos. Through analyzing of 713 videos, we identified three key factors that impact frame informativeness: eardrum visibility, eardrum coverage, and image clarity. We then develop a novel MIF pipeline that integrates (1) a ResNet-50 classifier trained on over 38,000 labeled frames to assess eardrum visibility, (2) a binary-adversarial CAM (BC-AdvCAM) method for weakly supervised eardrum segmentation and coverage estimation, and (3) a specialized blur and focus detection algorithm tailored to otoscope imagery. These components are combined into an "informative score" to rank frames automatically. Comparative evaluations using human reviewers and diagnostic AI models show that frames selected by our AI method perform comparably to expert-selected frames-achieving similar classification accuracy across multiple deep learning architectures. Notably, using the top four frames per video identified by our method significantly improves diagnostic accuracy over using a single expert-selected frame. This framework offers a scalable, expert-level tool for automating key frame selection and enhancing AI-based otoscopy diagnosis.