Deep Learning-based Rare Skin Disease Classification with Metadata Fusion

The paper presents an AI-based, multimodal decision-support system for early detection of rare skin diseases using clinical and smartphone images combined with patient metadata. We fine-tune a pretrained convolutional neural network (ResNet50 backbone) and apply class-weighted training, targeted augmentation, and a late fusion module that integrates demographic and lesion metadata. An image-quality validator reduces noisy / non-skin inputs and a generative language model produces context-sensitive, non-diagnostic educational feedback. On a combined benchmark of public datasets and smartphone images the system achieved an overall accuracy of 92.4%, macro-F1 0.880, and mean AUROC 0.940, while achieving an ECE of 0.032, demonstrating well-calibrated probabilities for triage. This framework improves detection of tail classes and provides clinically-oriented decision support; it is intended as an assistive tool and not a replacement for professional diagnosis.