Explainable AI for Chest X-Rays: Combining Grad-CAM and LLMS for Improved Diagnosis

Chest radiographs (CXRs) remain one of the most widely used imaging modalities for diagnosing cardiopulmonary diseases, yet increasing imaging demand and a shortage of radiologists continue to challenge diagnostic accuracy and efficiency. Deep learning approaches offer promising solutions, but their clinical adoption is hindered by limited interpretability. In this study, we compare a baseline convolutional neural network (CNN) augmented with Gradient-weighted Class Activation Mapping (Grad-CAM) to a multimodal framework that integrates large language models (LLMs) for explainable CXR interpretation. The baseline CNN, trained on the ChestX-ray8 dataset of 108,848 images across 14 thoracic pathologies, achieved a mean Area under Curve (AUC) of 0.891 and macro F1-score of 0.80, with Grad-CAM heatmaps providing visual explanations of model predictions. In contrast, the multimodal framework combined CNN-derived features with clinical narratives using cross-attention and fewshot prompting and was fine-tuned on 68,000 paired image-report samples. This approach improved diagnostic performance (AUC=0.915, F 1=0.89), generated human-readable reports, and localized anatomical features to enhance interpretability and clinical reasoning. Our findings demonstrate that multimodal AI substantially improves precision and transparency over CNN-only systems, offering a more reliable decision-support tool to assist radiologists, reduce diagnostic errors, and optimize workflow in high-volume clinical environments.