SkinSense: Dermatological Screening and Skin-Care Advisory System

Access to dermatological assessment in many regions is limited less by the availability of imaging than by the availability of trained specialists, and the resulting delay affects precisely those lesions for which early attention matters most. Most reported systems terminate at a single classification score, give no indication of how ambiguous that score is, and provide the end user with no actionable guidance. This paper describes SkinSense, a web-deployed educational screening and preventive advisory system joining seven-class lesion classification with EfficientNetV2, lesion delineation with U-Net, entropy-based uncertainty quantification, Grad-CAM visual attribution, heuristic risk banding, disease-specific preventive guidance and a locally hosted retrieval-based educational assistant inside one authenticated request pipeline built on FastAPI, SQLAlchemy, SQLite and a React front end. Classification is trained and evaluated on 10,013 de-duplicated ISIC 2018 Task 3 images under a fixed 7,009 / 1,502 / 1,502 partition whose test portion is withheld from every design and tuning decision. On that partition the baseline network records 52.86% accuracy, 46.49% balanced accuracy, 32.30% macro F1, 58.65% weighted F1 and 84.57% ROC-AUC; a compute-constrained improved configuration reaches 61.65% accuracy, 52.06% balanced accuracy and 39.81% macro F1. The limited-training U-Net attains a Dice coefficient of 67.68%, an IoU of 53.29% and pixel accuracy of 86.85%. A variational-autoencoder augmentation trial using thirty generated minority-class images reduced macro F1 to 16.68% and is reported as a negative result rather than suppressed. The contribution claimed is the integration itself — an uncertainty-gated advisory workflow with transparent reporting — not a new accuracy record. The system is explicitly educational and does not perform clinical diagnosis.