Authors: Research Scholar Roshan Rukshana Sulaima Lebbe, Dr.V.Priyalakshmi
Abstract: While the incorporation of deep learning in medical imaging has significantly boosted diagnostic accuracy, the "black box" problem of deep learning models continues to be an important obstacle towards their use in clinical practice. Not only are accurate predictions required by clinicians, but also clear explanations that are medically plausible. In this paper, we propose DeepVision-XAI, an explainable deep learning framework for real-time medical image diagnosis. Specifically, the framework incorporates an efficient EfficientNet-B4 model for feature extraction, MHSA self-attention for spatial information capture, and a hybrid explainability module that combines Grad-CAM, SHAP values, and Bayesian uncertainty estimates. When tested on three publicly available benchmark datasets (ChestX-ray2017 for pneumonia, ISIC 2019 for skin lesions, and APTOS 2019 for diabetic retinopathy), DeepVision-XAI attains diagnostic accuracies of 96.2%, 91.8%, and 94.5%, with an inference latency of less than 150 ms per image.
