Deep Learning for Biopsy-Free Subtyping of Basal Cell Carcinoma from Dermatoscopic Images
Alexandros Papadopoulos, Chrysa Episkopou, Ioannis Sarafis, Aimilios Lallas, Anastasios Delopoulos
cs.CV · cs.AI
Quick verdict
ViT (BEiT) hits AUC 0.784 for aggressive-vs-low-risk BCC subtyping from dermatoscopy (n=1271, single center); promising but preliminary, no significance tests.
Summary
This proof-of-concept study investigates whether a single dermatoscopic image can predict whether a basal cell carcinoma (BCC) is aggressive, potentially replacing diagnostic biopsies. The authors assembled the AUTh dataset: 1,271 biopsy-confirmed BCC images from a single Greek center (2018–2022), grouped into superficial, non-aggressive non-superficial, and aggressive subtypes, with the task cast as binary classification of aggressive versus low-risk. A BEiT-Large vision transformer (~300M parameters, 512×512 input), first adapted to dermatoscopy by supervised finetuning on 53,826 public ISIC images and then finetuned for 10 epochs on the BCC data, was evaluated with stratified 5-fold cross-validation repeated three times. It achieved AUC 0.784±0.026, versus 0.755 for ResNet-50 and 0.735 for ResNet-18, and 70.0% sensitivity at 70% specificity — above the 56.4% sensitivity of trained human readers reported by Camela et al., though on a different 235-image evaluation set, which is why the authors themselves call this comparison preliminary. Main limitations: a single-center, Caucasian-only cohort; no statistical significance testing of the model ranking; unstated patient-level splitting; and no code or data release. The takeaway is a promising direction requiring multi-center validation, not a deployable biopsy-free subtyping tool.
Review complete
Assessment at a glance
4
risk flags identified
● high 0
● medium 3
● low 1
5 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- This is the first work to combine dermatoscopy with deep learning for BCC subtype classification (Section I).
- A BEiT-Large ViT reaches AUC 0.784±0.026 for aggressive-vs-low-risk BCC classification on 1,271 images under 3×5-fold cross-validation, beating ResNet-18/50 baselines trained with the same budget (Table III).
- At 70% specificity the model's 70.0% sensitivity exceeds the 56.4% sensitivity of trained human readers reported by Camela et al. on the same task definition, though on a different dataset.
- These results support dermatoscopy-based deep learning as a potential biopsy-free decision-support tool, pending multi-center validation.
Strengths
Clinically underexplored task with a clear gap claim
Section I argues no prior work combines dermatoscopy with deep learning for BCC subtyping, while existing DL efforts use WSI, RCM or OCT (refs [3], [4], [15], [16]); the task (aggressive vs low-risk) directly maps to treatment choice per the 2023 European guideline [2].
Matched-operating-point comparison design
Sensitivity in Table III is reported at a fixed 0.70 specificity specifically to align with the human-reader study of [6] (experts: 70.7±6.93% specificity), an unusually careful choice for cross-study comparison.
Appropriate small-data protocol
Well-motivated two-stage transfer pipeline
Honest in-paper hedging
Weaknesses
No statistical testing of the model ranking
Cross-dataset human comparison presented as superiority
Split granularity unstated — leakage risk
Class-correlated metadata missingness unexamined
No code/data release and no ablation of the adaptation stage
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
The paper presents a proof-of-concept deep learning system that classifies a basal cell carcinoma (BCC) as **aggressive** (infiltrative, morpheaform, micronodular, metatypical) versus **low-risk** (superficial; nodular/nodulocystic; Pinkus fibroepithelioma) from a single dermatoscopic image. The motivation is clinical: subtype determines treatment (micrographic surgery for aggressive variants, per the 2023 European guideline [2]) but currently requires an invasive biopsy. The authors make three claims: (1) this is the first work combining dermatoscopy with deep learning for BCC subtyping; (2) a BEiT-Large vision transformer outperforms ResNet baselines on their data; (3) performance exceeds previously reported trained human readers. Claims (1) and (2) are supported as stated; claim (3) is hedged inside Sections IV–V but asserted more bluntly in the abstract — the paper's main overclaiming risk.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 4 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
Can this AI replace a biopsy for basal cell carcinoma subtyping?
No, not yet. This is a proof of concept trained and tested on 1,271 dermatoscopic images from a single Greek center, reaching AUC 0.784. The authors themselves call for multi-center, multi-ethnic validation before clinical use; BCC subtyping still requires histopathology in current practice.
What performance does the model achieve?
The BEiT-Large vision transformer reached AUC 0.784±0.026 (repeated 5-fold cross-validation), versus 0.755 for ResNet-50 and 0.735 for ResNet-18. At 70% specificity, its sensitivity was 70.0%; trained human readers achieved 56.4% sensitivity on a different 235-image dataset, so the human comparison is not controlled.
Is the dataset or code publicly available?
No. The AUTh BCC dataset is a private clinical dataset from Aristotle University of Thessaloniki and the paper mentions no code release. Only the ISIC archive images used for domain adaptation are publicly available.
How is 'aggressive BCC' defined in this study?
The aggressive class groups infiltrative, morpheaform, micronodular, and metatypical BCC, each confirmed by biopsy. The low-risk class groups superficial BCC together with nodular, nodulocystic BCC and Pinkus fibroepithelioma; mixed subtypes and incompletely excised tumors were excluded.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figure 1 (sample dataset images) is referenced in the text but its visual content cannot be assessed from the text extraction; all quoted numbers come from Tables I–III, which extracted cleanly.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.