Hierarchical candidate recursive network for highlight restoration in endoscopic videos

Highlight restoration in endoscopic videos is crucial for enhancing the clinical value and effectiveness of endoscopic examinations. However, as an unexplored field, it poses substantial challenges including numerous and scattered highlight areas, significant inconsistency of highlight between frames, and the inability to accurately assess the effectiveness of highlight restoration. In this paper, we propose Hierarchical Candidate Recursive Network (HCRN), as a pioneering approach for endoscopic highlight restoration, preventing lesions from being obscured or impacted by highlight during endoscopic examinations. Our HCRN comprises: (1) The Progressive Candidate Discovery Module identifies the most valuable information from historical frames across diverse temporal and spatial contexts, enhancing the temporal stability and spatial consistency of highlight restoration. (2) The Hierarchical Directive Module employs a progressive approach to restore the highlight and reduce the uncertainty in the candidate pool, offering a more dynamic and adaptive method for video processing and image restoration. (3) The Dual-scale Contrast Variation Metric provides a more accurate and comprehensive performance evaluation of highlight restoration by considering both local and global contrast variations, even in the absence of ground truth. Comprehensive experiments on a generalized dataset including 1,059 endoscopic video sequences demonstrate that our HCRN achieves state-of-the-art performance. Our HCRN improves the Dual-scale Contrast Variation by at least 2.5% as compared to the seven recent advanced methods. These results demonstrate that our HCRN holds the potential to significantly improve the safety of endoscopic surgery and drive advancements in medical technology.