An Interpretable Hybrid Deep Learning Framework for Computer-Aided Detection of Gastrointestinal Diseases in Endoscopic Imaging

Gastrointestinal (GI) illnesses, especially gastric polyps and gastroesophageal reflux disease (GERD), are still ubiquitous diagnostic challenges with their complicated presentation and high inter-observer variation on endoscopy. The current research proposes Xnception, a new dual-backbone deep learning network that synergistically combines the Xception and InceptionV3 architectures to facilitate classification robustness and feature expressivity for analysis of endoscopic images. Using transfer learning and fold wise validation, the model is optimized for small-sized medical datasets with ensured generalizability. The fusion mechanism combines deep semantic representations of both backbones through specialized dense layers to enable precise discrimination between pathological and non-pathological classes. Explainable AI (XAI) techniques, Local Interpretable Model-agnostic Explanations, Integrated gradients, Grad-CAM++, are employed to visualize important regions impacting the model's predictions, hence ensuring transparency and clinical trustworthiness. Quantitative results on publicly available datasets demonstrate that Xxception outperforms its component individual models as well as other common baselines on several measures like accuracy, precision, and AUC. The proposed framework demonstrates promise to improve real-time diagnostic pipelines in gastroenterology and provides a scalable platform for AI-augmented endoscopic screening.