Boosted Capsule Network for Gastrointestinal Disease Recognition

Abstract

Gastrointestinal diseases affect over 40% of the global population and rank among the leading causes of cancer-related deaths worldwide. ‎Wireless capsule endoscopy (WCE), though minimally invasive, generates 45,000-50,000 images per procedure, with expert endoscopists ‎missing 22-28% of pathological cases due to examination complexity and varied disease presentations. Deep learning approaches like Con-‎volutional Neural Networks require large annotated datasets, which are rarely available in medical imaging. Capsule Networks (CapsNets) ‎excel in limited-data scenarios through spatial-hierarchy inference but struggle with subtle features in complex medical images. We present a Boosted Capsule Network that incorporates dual enhancement mechanisms: modified feature boosting via intensity inversion to expose diverse low-level patterns, reducing the generalization gap by 45%, and class capsule amplification to sharpen decision boundaries and improve inter-class separation. Ablation studies on the Kvasir-V2 dataset show individual components achieve 89.3% (feature boosting) and 91.7% ‎‎(class boosting), while their synergistic combination reaches 97.90% accuracy, a 15.8% improvement over baseline CapsNet (82.10%) and ‎‎1.1% over previous state-of-the-art (96.80%). The lightweight architecture adds zero trainable parameters for class boosting and minimal ‎overhead for feature enhancement, enabling real-time clinical deployment. We experimentally deployed the model as a web-based prototype ‎with human-in-the-loop uncertainty handling for confidence scores below 0.5‎.