A Real-Time Mobile AI-Assisted System for Skin Disease Prescreening: A Technical Feasibility Study.

Background and Aims: Skin diseases remain a widespread health concern, particularly in areas where access to dermatological care is limited. With the growing availability of smartphones and advances in deep learning, mobile-based image analysis offers a practical option for early-stage screening. This study aims to develop and evaluate a lightweight, real-time mobile AI-assisted system for skin disease prescreening on Android devices. Methods: An Android application, SkinLearn, was developed by integrating a YOLO11-based image classification model optimized for on-device inference using TensorFlow Lite. The model was trained on a publicly available seven-class skin disease dataset and further evaluated on the PAD-UFES-20 skin cancer dataset to examine its ability to generalize across datasets. A YOLOv8 model was also implemented under the same conditions for comparison. Results: The proposed YOLO11 model achieved a top-1 accuracy of 99.2% on the seven-class dataset and 71.4% on the PAD-UFES-20 dataset, outperforming the YOLOv8 baseline in both cases. The mobile implementation enabled real-time, offline inference with low computational overhead, demonstrating efficient deployment in resource-constrained environments. However, the observed performance drop across datasets highlights challenges related to domain shift, class imbalance, and variability in imaging conditions. Conclusion: These findings support the technical feasibility of lightweight deep learning models such as YOLO11 for accessible mobile skin disease prescreening. However, the observed performance drop on PAD-UFES-20 indicates that further validation across more diverse datasets and clinical settings is needed before any real-world clinical use is considered.