Browser-Based Dermoscopy Triage with Calibrated MobileNet: A Privacy-First Approach

Early melanoma detection saves lives, yet rural clinics often lack dermatologists and reliable connectivity. We present a browser-based, client-side skin-lesion classifier that returns top-three diagnostic suggestions in under 1 s on commodity smartphones. Using a leakage-safe lesion-level split of HAM10000 (10 015 images, seven classes) [1], a fine-tuned MobileNet attains 97.9 % top-3 accuracy, macro-F1 0.78, and melanoma sensitivity 0.908 at specificity 0.95. Temperature scaling [2] reduces expected calibration error from 0.089 to 0.043 (Brier 0.098) when evaluated on a held-out test subset. Warm inference latency ranges 285 ms to 920 ms. The 2.8 MB Brotli-compressed TensorFlow.js model runs entirely on-device, avoiding image uploads and enabling rapid triage in connectivitylimited settings.