Interpretable deep learning for gastric cancer detection: a fusion of AI architectures and explainability analysis

The rise in cases of Gastric Cancer has increased in recent times and demands accurate and timely detection to improve patients' well-being. The traditional cancer detection techniques face issues of explainability and precision posing requirement of interpretable AI based Gastric Cancer detection system.

This work proposes a novel deep-learning (DL) fusion approach to detect gastric cancer by combining three DL architectures, namely Visual Geometry Group (VGG16), Residual Networks-50 (RESNET50), and MobileNetV2. The fusion of DL models leverages robust feature extraction and global contextual understanding that is best suited for image data to improve the accuracy of cancer detection systems. The proposed approach then employs the Explainable Artificial Intelligence (XAI) technique, namely Local Interpretable Model-Agnostic Explanations (LIME), to present insights and transparency through visualizations into the model's decision-making process. The visualizations by LIME help understand the specific image section that contributes to the model's decision, which may help in clinical applications.

Experimental results show an enhancement in accuracy by 7\% of the fusion model, achieving an accuracy of 97.8\% compared to the individual stand-alone models. The usage of LIME presents the critical regions in the Image leading to cancer detection.

The enhanced accuracy of Gastric Cancer detection offers high suitability in clinical applications The usage of LIME ensures trustworthiness and reliability in predictions made by the model by presenting the explanations of the decisions, making it useful for medical practitioners. This research contributes to developing an AI-driven, trustworthy cancer detection system that supports clinical decisions and improves patient outcomes.