Real-Time Viral Skin Lesion Diagnosis: A Hybrid Deep Learning Framework for On-Device Deployment

Abstract

This thesis addresses a critical public health issue by presenting a hybrid deep learning pipeline for the real-time, on-device diagnosis of viral skin lesions. The core objective was to develop a model that effectively balances high classification performance with low computational cost, enabling its deployment on mobile devices for use in resource-limitedenvironments. A comprehensive comparative study was conducted on five hybrid architectures, each combining a custom-trained Convolutional Neural Network with a powerful pre-trained backbone. Through a rigorous two-staged fine-tuning approach, the Custom CNN + EfficientNetB0 architecture was identified as the most effective, achieving an outstanding classification accuracy of 99%. The selected model was then efficiently quantized into a lightweight 5.6 MB TFLite format, demonstrating a remarkable average on-device inference time of 50 ms. This achievement culminates in the implementation of a high-performing, privacy-preserving, and low-cost model within a functional mobile application. This work underscores the feasibility of developing practical, end-to-end AI diagnostic tools that can support clinical practice and provides a scalable platform for future research in accessible visual-based diagnostic solutions.