Research · Technical Whitepaper
Published by Ivy AI Solutions Limited · 6 July 2026 · DOI: 10.5281/zenodo.21225287
Backbone
DINOv2-Large ViT-L/14 @ 518×518
Skin tones evaluated
Fitzpatrick I–VI
Lesion classes
BCC · Melanoma · SCC · Benign/Other
Quality framework
ISO 13485 · ISO 27001
This whitepaper sets out, in full technical detail, how the ScanSkinAI Cancer Flag Module is built, trained, and validated. It is the reference document behind our skin-cancer risk screening technology — the AI that assesses a lesion from an ordinary clinical photograph and flags whether it warrants closer attention from a clinician.
Rather than treating the model as a black box, the paper walks through every layer of its development: the curated data it learns from, the architecture that powers it, the engineering that keeps it stable in real-world conditions, and the independent evaluation that measures how well it performs across the full range of human skin tones.
Most skin-analysis AI is trained on dermoscopic images captured with specialist equipment. That works in a clinic, but it fails the moment a real person points a phone camera at a mole. ScanSkinAI is built for that real-world setting — standard clinical photographs, taken by non-experts, in uncontrolled lighting.
Designing for that reality demands more than a good model. It demands rigorous image-quality control, careful handling of rare-but-critical cancer types, and honest validation across every skin tone — not just the lighter phenotypes that dominate most public datasets. This paper documents how we address each of those challenges.
A DINOv2-Large vision transformer (ViT-L/14) forms the visual backbone, operating at native 518×518 resolution to preserve fine lesion detail. It distinguishes basal cell carcinoma, melanoma, squamous cell carcinoma, and benign or other lesions. A second, bounded language layer then communicates the result responsibly — never issuing a diagnosis, always framing output as a screening signal.
The model learns from a fine-tuning corpus of clinical (not dermoscopic) photographs, curated and quality-controlled by dermatologists. A data-centric cleaning pipeline identifies and removes label noise, so the model trains on trustworthy examples rather than inheriting errors from raw datasets.
Malignant classes like melanoma and SCC are far rarer than benign lesions in any real dataset — and missing them is the costliest error a screening tool can make. The paper details the class-imbalance strategy, weighted sampling, and progressive three-phase fine-tuning regimen used to keep the model sensitive to these minority classes.
Image-quality gating at intake rejects unusable photos before they reach the model. Extensive augmentation during training and test-time augmentation at inference keep predictions stable across lighting, angle, and camera variation — the everyday conditions of consumer use.
Performance is measured on internal held-out data and through an independent dermatologist audit of production scans spanning Fitzpatrick skin types I through VI. Evaluating across the full tone range is central to the paper — not an afterthought.
Grad-CAM and SHAP make the model's attention inspectable, so its reasoning can be reviewed rather than trusted blindly. The paper also lays out the full development and validation lifecycle and our roadmap for continued validation.
The Cancer Flag Module is positioned as a non-diagnostic screening and triage aid — it supports earlier attention and clinician review, and does not replace professional diagnosis. It is developed under an ISO 13485 and ISO 27001 quality framework and aligned with UKCA Class I and EU MDR requirements.
Ivy AI Solutions Limited. (2026). ScanSkinAI Cancer Flag Module: Training Methodology, Model Development, and Robustness Evaluation. Zenodo. https://doi.org/10.5281/zenodo.21225287
