DermFusionX: An Explainable CNN–MLP Late Fusion Framework for Multimodal Skin Lesion Classification

Abstract Deep learning has led to extraordinary performance in domains like computer vision and natural language processing, leading to its expansion into fields such as healthcare, which demand high transparency. In clinical practice, dermatologists work with multiple data sources, such as patient metadata and lesion images, for diagnosis. Motivated by this, we propose a multimodal approach to enhance skin lesion classification on the HAM10000 dataset. We conducted extensive experiments comparing unimodal models (using only metadata or images) with multimodal models (combining both). Our evaluation included several pre-trained Convolutional Neural Networks (e.g., ResNet50, VGG19, XceptionNet, InceptionV3) and a novel custom architecture, DermFusionX. Results demonstrate that multimodal models significantly outperform their unimodal counterparts, with DermFusionX achieving precision and recall rates of 90%. To ensure transparency, we employed explainable AI techniques (LIME and SHAP) to interpret the models' decisions.