GI-ScreenNet v2: A Modular Framework for Gastrointestinal Disease Detection Based on an Integrated Transfer Learning.

BACKGROUND: Gastrointestinal diseases pose a significant global health challenge, and their early screening relies on wireless capsule endoscopy (WCE). However, analysing the massive volume of WCE images is time-consuming and prone to human error. Although deep learning offers solutions, existing systems are often inflexible and technically complex, limiting clinical adoption. METHODS: We propose GI-ScreenNet v2, a multi-backbone network framework based on ensemble and transfer learning. It supports arbitrary backbones through a standardised interface and leverages a cross-attention mechanism to dynamically integrate multi-model features for sophisticated representation learning. RESULTS: In KvasirV2, GI-ScreenNet v2 achieves 94.87% accuracy, 3.31% higher than traditional methods. This high-performance result enables efficient GI screening and paves the way for practical AI-assisted diagnostics. CONCLUSIONS: We present a unified framework for GI disease detection, with an integrated workflow for dynamic model selection and cross-attention fusion. This design enables efficient integration of novel models and techniques, advancing robust diagnostic systems.