ScanSkinAI Cancer Flag Module: Training Methodology, Model Development, and Robustness Evaluation

This whitepaper documents the training methodology, model architecture, and robustness engineering behind the ScanSkinAI (www.ScanSkinAI.com) Cancer Flag Module — an AI system for consumer-facing skin-cancer risk screening from standard clinical photographs. The module uses a two-tier design: a DINOv2-Large (ViT-L/14) visual backbone operating at native 518×518 resolution, fine-tuned to distinguish basal cell carcinoma, melanoma, squamous cell carcinoma, and benign/other lesions, followed by a bounded, non-diagnostic language layer that communicates results responsibly to users.

The paper details the dermatologist-curated training corpus, a data-centric cleaning pipeline for label-noise removal, a three-phase progressive fine-tuning regimen, and the class-imbalance strategy used to protect minority malignant classes. It describes the robustness measures that support real-world consumer use — image-quality gating at intake, extensive augmentation, and test-time augmentation for inference stability — alongside measured performance on internal held-out data and an independent dermatologist audit spanning Fitzpatrick I–VI skin types.

The module is positioned as a non-diagnostic screening and triage aid, developed under an ISO 13485 / ISO 27001 quality framework and aligned with UKCA Class I and EU MDR requirements. Explainability (Grad-CAM, SHAP), the development and validation lifecycle, and the roadmap for continued validation are also presented.