Transparent and trustworthy interpretation of COVID-19 features in chest X-rays using explainable AI

The potential of AI-based disease prediction models for assessing COVID-19 patients outperforms conventional methods. However, their black-box nature has limited their applicability. This study explores the approach for COVID-19 identification by integrating Artificial Intelligence (AI) methods with Deep Neural Networks (DNNs), such as EfficientNet and DenseNet, applied for medical imaging. To address the black-box challenge, the study incorporates eXplainable AI (XAI) techniques, including LIME, Grad-CAM, and a novel variant of Grad-CAM++ , termed ”Modified Grad-CAM++ ”. The modification aims to enhance explanations for COVID-19 predictions, improving the interpretability and transparency of the model’s decision-making process. Collaborating with expert radiologists, the study validates the Modified Grad-CAM++ using a separate dataset of radiologist-validated chest X-ray (CXR) images. The Integrated Uncertainty Calculation (IUC) metric is introduced as an evaluation measure for the Modified Grad-CAM++ . Remarkably, both EfficientNet and DenseNet achieve high diagnostic precision, with accuracy rates of 98% and 97%, respectively. The integration of XAI algorithms enhances the interpretability and transparency of predictions, ensuring clinical relevance and validity. These precise AI models offer valuable support to healthcare professionals in decision-making and resource allocation.