Tech Tonic

Skin cancer is often found late in rural areas because dermatologists are far away. Silent Disease Detector is an explainable AI screening aid: the user uploads a photo of a skin lesion, and the app predicts its likely type across 10 classes (melanoma, basal cell carcinoma, SCC, nevus, actinic keratosis, benign keratosis, dermatofibroma, vascular lesion, normal skin, scar). It shows a Grad-CAM++ heatmap of why the model decided, and raises a melanoma safety alert using a lowered probability threshold to reduce missed cases. An OpenCV lesion pre-filter stops unclear photos from getting a false result. Built with PyTorch, EfficientNet-B0, Streamlit, and an Android WebView app, trained on about 16,500 images (HAM10000 + ISIC 2019). It is a research prototype, not a diagnosis tool. Known limits: melanoma recall is about 0.86 (target 0.90), and accuracy drops on non-dermoscopy photos. The live demo link is https://silent-disease-detector.streamlit.app