TITLE:
Multiclass Lesion Classification in Wireless Capsule Endoscopy: An Interpretable Deep Learning Framework
AUTHORS:
Moteb Alghamdi, Mohamed Elsersy, Salah Abdel-Mageid
KEYWORDS:
Convolutional Neural Networks (CNNs), Wireless Capsule Endoscopy, Image Classification, Deep Learning, Medical Imaging
JOURNAL NAME:
Journal of Computer and Communications,
Vol.14 No.7,
July
28,
2026
ABSTRACT: Wireless Capsule Endoscopy (WCE) has become an important tool in gastrointestinal diagnostics, yet the manual review of extensive video data remains labor-intensive and subjective. While deep learning has shown promise for automating this task, existing approaches are largely limited to binary bleeding classification and lack robustness, interpretability, and multiclass lesion analysis. This paper introduces a fully automated, CNN-based framework for robust multiclass lesion classification in WCE imagery, addressing bleeding, ulcers, and arteriovenous malformations (AVMs). Leveraging the KAUHC dataset, a novel repository of 3301 annotated small-bowel endoscopic images from Saudi Arabia comprising Normal (2156), AVM (673), and Ulcer (472) frames, our method employs a fine-tuned VGG16 architecture with advanced data augmentation and preprocessing to handle class imbalance, illumination variance, and anatomical complexity. The system achieves strong diagnostic performance, with a precision of 0.97, recall of 0.97, and an F1-score of 0.97, outperforming baseline models such as InceptionV3. Furthermore, we integrate Explainable AI (XAI) techniques (including SHAP and LIME) to provide interpretable decision support and enhance model transparency for clinical interpretation. By bridging the gap between experimental AI and real-world usability, this work offers a reliable, scalable, and interpretable tool with the potential to support computer-assisted gastrointestinal diagnostics and accelerate AI-driven automation in clinical workflows.