Real-Time Portable Diagnostics for Seborrheic Dermatitis via Hierarchical Few-Shot Learning

Seborrheic dermatitis is a prevalent skin condition that is often misdiagnosed due to overlapping symptoms with other dermatoses, making early and accurate diagnosis essential. Conventional diagnostic methods rely heavily on expert interpretation and specialized equipment, which are typically unavailable in remote or resource-limited settings. To address these limitations, we propose a hierarchical few-shot learning model capable of automatically classifying six major variants of seborrheic dermatitis, each further assessed across three severity levels—mild, moderate, and severe. Our model integrates Prototypical Networks (ProtoNet) with a ResNet50 backbone to capture hierarchical relationships and enable fine-grained classification in data-scarce settings. The trained model achieves a classification accuracy of 98.35% on the held-out test set, and critically, retains this accuracy after deployment on a Raspberry Pi device using optimized inference pipelines. The system supports real-time, on-device diagnosis with minimal latency, enabling portable and scalable dermatological assessment in low-resource environments. Beyond its clinical utility, this work presents the first hierarchical few-shot learning framework specifically designed for real-time edge deployment in dermatology, demonstrating how advanced machine learning can be effectively translated into practical, affordable healthcare solutions.