Skin Disease Classification Using Neural Networks

Abstract

Accurate identification of dermatological conditions is essential for timely medical intervention, particularly for lesions with malignant or premalignant potential. Variability in lesion appearance, skin tone diversity, and imaging conditions often complicate traditional diagnostic approaches, creating the need for a robust automated screening system. In the present work, a hybrid artificial intelligence model is developed by integrating ResNet-50 for deep feature extraction with DenseNet-50 and LLM supported refinement for classification. The model is trained to differentiate a clinically relevant set of skin diseases including melanoma, basal cell carcinoma, actinic keratoses with intraepithelial carcinoma, melanocytic nevi, dermatofibroma, vascular lesions, and benign keratosis-like lesions chosen for their diagnostic similarity and significance in early detection workflows. The system is deployed through a backend server where the TensorFlow (.h5) model is hosted, enabling cloud-based inference and delivering predictions to the Android application through API calls. The proposed approach enhances reliability, minimizes feature redundancy, and provides confidence based predictions suitable for real-time screening applications. This work demonstrates that hybrid deep learning pipelines can serve as fast, low cost, and scalable tools for preliminary skin disease assessment, supporting both clinical environments and remote users.