Research · Product Module

The Cancer Flag Module: Non-Diagnostic AI Triage for Skin Cancer from a Photo

Published by Ivy AI Solutions Limited · Reference DOI: 10.5281/zenodo.21225287

The Cancer Flag Module is the component of ScanSkinAI that reads an ordinary clinical photograph and returns a non-diagnostic triage signal — a flag on whether a lesion warrants closer attention from a clinician. This page explains what the module does, who it is for, and how the flag is produced. For the full training methodology, architecture, and validation, see the methodology whitepaper reference page.

Input
Ordinary clinical photograph
Output
Non-diagnostic triage flag
Skin tones covered
Fitzpatrick I–VI
Framework
ISO 13485 · UKCA Class I

What the Cancer Flag Module is

The Cancer Flag Module is a piece of software: a non-diagnostic triage aid that takes an ordinary clinical photograph of a lesion — the kind a person takes with a phone camera — and returns a flag on whether that lesion shows patterns consistent with basal cell carcinoma, melanoma, squamous cell carcinoma, or a benign or other lesion.

It is deliberately narrow. It does not diagnose. It does not replace a clinician. It does not require dermoscopy. Its job is to help a lesion that warrants closer attention get closer attention sooner — inside a teledermatology service, an insurer member journey, or a partner digital-health product.

Who the Cancer Flag Module is for

Clinicians & teledermatology

Use the flag as a triage aid on incoming photo consultations — surfacing lesions that warrant closer attention first, without replacing your clinical judgement.

Insurers & wellness platforms

Embed the module into member journeys to support earlier attention on suspicious lesions and route members to appropriate care pathways.

Digital-health integrators

Consume the flag through the ScanSkinAI API and surface it inside your own product with a bounded, non-diagnostic communication layer.

How the flag is produced

1. Image-quality gate

Every photo passes through an intake gate that rejects blur, poor lighting, and off-frame captures before the model ever sees them. Bad inputs never produce a flag.

2. DINOv2-Large visual backbone

A ViT-L/14 vision transformer processes the accepted image at native 518×518 resolution — chosen for its self-supervised representation quality on natural images and its ability to preserve fine lesion detail.

3. Bounded lesion classifier

A supervised head over the backbone maps the representation to four classes: basal cell carcinoma, melanoma, squamous cell carcinoma, and benign or other. Class imbalance for rare cancers is handled through weighted sampling and a progressive three-phase fine-tuning regimen.

4. Responsible communication layer

A second, bounded language layer translates the classifier output into a triage signal — never a diagnosis. Wording is constrained so the module cannot claim more than it should.

5. Explainability by default

Grad-CAM and SHAP make the model's attention inspectable, so its reasoning can be reviewed rather than trusted blindly.

6. Cross-tone validation

Performance is evaluated across Fitzpatrick I–VI on an internal held-out set and an independent dermatologist audit of production scans — cross-tone coverage is a first-class design requirement, not a post-hoc check.

Every quantitative claim — accuracy, sensitivity, specificity, per-Fitzpatrick-tone breakdown, dataset composition — lives in the whitepaper rather than on this page, so that the numbers you read are always the numbers in the archived version of record. Download the whitepaper for the full validation results.

Safety and non-diagnostic framing

The Cancer Flag Module is a non-diagnostic screening and triage aid. It supports earlier attention and clinician review. It does not replace dermoscopy, biopsy, histopathology, or in-person examination.

The module's language layer is bounded specifically to prevent overclaiming: it does not tell a user they have cancer, and it does not tell a user they do not. It communicates a triage signal and directs the user toward appropriate care. Any concerning, changing, bleeding, or non-healing skin change should be assessed by a qualified clinician.

Explainable, not a black box

The module supports Grad-CAM and SHAP so that clinicians and reviewers can see which regions of a lesion drove a given flag — rather than trust the output blindly. Explainability is part of how we make the module accountable to clinicians and regulators, and how partners can review edge cases when something looks wrong.

The full technical treatment of explainability sits in the whitepaper alongside the development lifecycle and the roadmap for continued validation.

Regulatory posture

The Cancer Flag Module is developed under an ISO 13485 medical-device quality management system and an ISO 27001 information-security framework. It is aligned with UKCA Class I and EU MDR requirements as software as a medical device (SaMD) positioned for non-diagnostic screening.

ISO 13485
ISO 27001
UKCA Class I
EU MDR aligned
Non-diagnostic

Frequently asked questions

What is the Cancer Flag Module?

The Cancer Flag Module is the component of ScanSkinAI that reviews an ordinary clinical photograph of a lesion and returns a non-diagnostic triage signal: whether the lesion shows patterns consistent with basal cell carcinoma, melanoma, squamous cell carcinoma, or a benign or other lesion, and whether it warrants closer attention from a clinician.

Is the Cancer Flag Module a diagnosis?

No. It is explicitly a non-diagnostic screening and triage aid. It does not replace dermoscopy, biopsy, histopathology, or in-person clinical examination. Any concerning, changing, bleeding, or non-healing skin change should be assessed by a qualified clinician.

Who is the Cancer Flag Module built for?

Three groups: clinicians and teledermatology services who want a triage aid to help prioritise reviews, insurers and wellness platforms who embed it into member journeys to support earlier attention, and digital-health integrators who consume the flag through the ScanSkinAI API to power screening in their own products.

Does the Cancer Flag Module work across skin tones?

Cross-tone coverage is a first-class design requirement. The module is evaluated across Fitzpatrick skin types I through VI using both an internal held-out set and an independent dermatologist audit of production scans. The full evaluation is documented in the whitepaper.

How is the flag produced?

A DINOv2-Large vision transformer processes the photograph at native 518×518 resolution to preserve fine lesion detail, and a bounded classifier maps the representation to lesion classes. An image-quality gate rejects unusable photos before they reach the model, and a second, bounded language layer communicates the result responsibly without ever issuing a diagnosis.

How can I see what the model looked at?

The module supports Grad-CAM and SHAP explainability, so clinicians and reviewers can inspect which regions of a lesion drove a given flag rather than trust the output blindly. Explainability outputs are described in the whitepaper.

What regulatory framework does the Cancer Flag Module sit under?

It is developed under an ISO 13485 medical-device quality management system and an ISO 27001 information-security framework, and it is aligned with UKCA Class I and EU MDR requirements as software as a medical device (SaMD) positioned for non-diagnostic screening.

Where can I read the full technical detail?

The complete training methodology, architecture, data curation strategy, class-imbalance handling, and cross-tone validation are in the whitepaper: Ivy AI Solutions Limited (2026), ScanSkinAI Cancer Flag Module: Training Methodology, Model Development, and Robustness Evaluation, Zenodo, https://doi.org/10.5281/zenodo.21225287.

Cite this work

Ivy AI Solutions Limited. (2026). ScanSkinAI Cancer Flag Module: Training Methodology, Model Development, and Robustness Evaluation. Zenodo. https://doi.org/10.5281/zenodo.21225287

Read the full whitepaper

Complete training methodology, architecture, class-imbalance strategy, and validation across Fitzpatrick I–VI.