AndreaGemignani/ANDRY-V3B-SkinScanner

ANDRY / SkinScanner

Research project for safety-oriented skin-lesion triage built on top of google/medsiglip-448.

Research only. Not clinically validated. Not a medical device. Not for diagnosis or treatment.

Current status

ANDRY evolved from raw MedSigLIP zero-shot experiments to a frozen, partially fine-tuned dermoscopic/source-domain model (V3B_SOURCE_FINAL_V1). The current product concept intentionally separates ordinary-smartphone triage from higher-information macro-lens analysis so that domain-specific training does not overwrite the validated dermoscopic representation.

Model evolution

Stage	Method	Main result / role
MedSigLIP original	Zero-shot prompts, no ANDRY training	On the 193-image ISIC2018 holdout: ROC-AUC 0.8344; MEL-vs-lower AUC 0.7293
V3A	Frozen MedSigLIP 1152-d embeddings + grouped RBF-SVM / GPU-MLP blend	ISIC2019 grouped OOF ROC-AUC 0.9166; frozen-representation baseline
V3B	Partial fine-tuning of last 6 MedSigLIP vision blocks + binary suspicious head + auxiliary 8-class head	Stage-12 pooled grouped OOF ROC-AUC 0.9185; MEL/worst-danger AUC 0.8801
V3B_SOURCE_FINAL_V1	Final fixed 4-epoch fit on all 25,331 ISIC2019 source images; PAD excluded	One-time ISIC2018 holdout ROC-AUC 0.9559; MEL-vs-lower AUC 0.9521

The 193-image holdout is intentionally small and is a sanity check, not prospective clinical validation. The more robust source-domain estimate remains the 5-fold grouped out-of-fold Stage-12 evaluation.

Canonical final checkpoint

V3B_SOURCE_FINAL_V1.pt

SHA256:

755f4ee6a6ac7f3947ab5ee83cd5d641dfec363dc578719a98edeac44e02e667

Training provenance:

• Foundation: google/medsiglip-448
• Trainable vision tail: last 6 transformer blocks
• Heads: binary suspicious + auxiliary 8-class
• Epochs: 4 fixed
• Training set: ISIC2019, n=25,331
• PAD-UFES-20 training rows: 0
• PAD-UFES-20 model-selection rows: 0
• PAD-UFES-20 calibration rows: 0
• ISIC2018 validation training rows: 0
• No product threshold frozen
• No clinical-safety claim

Safety-oriented model selection

The project treats a melanoma false negative as substantially more costly than a false positive or an abstention. Architecture selection therefore used a lexicographic safety-oriented key centered on the worst dangerous-class AUC and melanoma-vs-lower-risk AUC rather than overall accuracy alone.

Dangerous / suspicious classes: MEL, BCC, AKIEC, SCC.

Lower-risk classes: NV, BKL, DF, VASC.

Source-domain validation

Frozen FT_LAST6_AUX Stage-12 pooled grouped OOF metrics on ISIC2019:

Metric	Value
ROC-AUC	0.918537
Average Precision	0.875347
Worst-danger / MEL AUC	0.880121
Specificity @ all-danger sensitivity >=95%	0.514414
Specificity @ all-danger sensitivity >=98%	0.411544
Brier	0.112424
Log loss	0.363379
ECE (15 bins)	0.044191
Aux macro-F1	0.577890

Group-bootstrap 95% CI for MEL/worst-danger AUC: 0.869856-0.889974.

Smartphone-domain finding

PAD-UFES-20 is used only as read-only domain-shift research for the final V3B model. It does not alter the frozen source model.

Final V3B read-only PAD results:

Metric	Value
ROC-AUC suspicious	0.9172
Average Precision	0.9704
MEL-vs-lower AUC	0.7785
BCC-vs-lower AUC	0.9383
AKIEC-vs-lower AUC	0.9068
SCC-vs-lower AUC	0.9014
Brier	0.1661
Log loss	0.5138
Aux balanced accuracy	0.4576
Aux macro-F1	0.2987

The important conclusion is not the aggregate ROC-AUC. Melanoma transfer degrades substantially on ordinary smartphone images, while BCC/AKIEC/SCC transfer more strongly. This is one of the main reasons the current roadmap keeps dermoscopic and smartphone experts separate.

Current two-stage product hypothesis

ordinary smartphone image
          |
          v
     Software A
 quality / adequacy / ultra-conservative escalation
          |
          v
 controlled macro-lens acquisition
          |
          v
     Software B
 higher-information lesion analysis

The long-term target is one user-facing application with internally independent models, datasets, calibration and version histories.

Documentation

The full development history, technical decisions, rejected alternatives, unresolved questions and proposed domain-transfer architecture are preserved in:

• docs/ANDRY_Technical_Dossier_for_Supervisor_2026-08-27.docx

This document is intentionally retained so another researcher can understand why the project reached the current architecture, not only reproduce the final checkpoint.

Open questions

1. Quantifying acquisition-domain distance between dermoscopy, clinical close-up, ordinary smartphone and future macro-lens images.
2. Determining whether a smartphone-stage model should ever issue a LOW-RISK / stop-here result without a much larger independent melanoma dataset.
3. Anti-shortcut auditing of smartphone datasets.
4. Identifying intermediate-domain datasets with reliable labels and patient/lesion grouping.
5. Evaluating the frozen V3B representation on real macro-lens images once the physical accessory exists.
6. Prioritizing paired same-lesion smartphone + macro + dermoscopy + ground-truth acquisition when feasible.

MedSigLIP / HAI-DEF terms

ANDRY V3B is a model derivative of Google MedSigLIP. MedSigLIP is governed by the Health AI Developer Foundations Terms of Use:

https://developers.google.com/health-ai-developer-foundations/terms

Users of derivative model weights must comply with those terms and restrictions. See NOTICE.

Disclaimer

All reported metrics are research/development results. They do not demonstrate clinical effectiveness, safety, regulatory approval, or fitness for diagnosis. Any future physical-device use requires target-domain validation, appropriate prospective study design and applicable regulatory review.