MSRCTNet: a novel multi-scale capsule triplet network for efficient redundant frame removal in wireless capsule endoscopy videos.

Wireless capsule endoscopy (WCE) examinations generate approximately 55,000 images per procedure, with a vast majority being redundant due to high structural similarity, imposing a significant burden on physicians during review. This paper introduces MSRCTNet, a novel Multi-Scale Capsule Triplet Network, to efficiently remove redundant frames while preserving clinically essential information. By addressing key challenges such as data imbalance, small sample sizes, and the need for balanced accuracy and efficiency, MSRCTNet enhances feature extraction through multi-scale processing and attention mechanisms, refines representations via capsule networks, and assesses frame similarity using an optimized triplet framework. Evaluated on a custom dataset of 257,362 WCE images (360×360 resolution) from the First Affiliated Hospital of Yangtze University, Jingzhou, China, MSRCTNet achieves 96.1% accuracy in redundancy removal, with a false detection rate of 2.84%, missing detection rate of 0.19%, and real-time processing at 0.02 seconds per frame. These advancements not only reduce physician workload and fatigue but also demonstrate superior robustness and adaptability for clinical applications, outperforming existing methods in handling diverse endoscopic scenarios.