PaperlayerView exact revision on arXiv

Source-linked AI summary

Deep Learning for Biopsy-Free Subtyping of Basal Cell Carcinoma from Dermatoscopic Images

Alexandros Papadopoulos, Chrysa Episkopou, Ioannis Sarafis, Aimilios Lallas, Anastasios Delopoulos

arXiv:2609.07180v1cs.CVcs.AI

TL;DR

BCC subtyping informs treatment, but biopsy is costly and invasive, and dermatoscopy-based deep-learning subtyping has received little attention. This paper tests pretrained vision transformers on single dermatoscopic images to distinguish aggressive from non-aggressive BCC. BEiT achieved an average AUC of 0.784, with comparisons to human readers remaining preliminary because evaluation sets differed.

  • Problem

    BCC treatment depends on histopathologic subtype, which is established by costly, invasive biopsy; few studies have examined dermatoscopy for BCC subtyping.

    Source: §I · p. 1

  • Method

    The authors use pretrained vision transformers to classify aggressive versus non-aggressive BCC from dermatoscopic images, adapting the model to dermatoscopy before target-task training.

    Source: §I · p. 2, D. Feature adaptation · p. 3

  • Results

    BEiT achieved the highest average AUC (0.784) and sensitivity (70.0%), surpassing ResNet baselines; its apparent sensitivity advantage over instructed human readers is preliminary because evaluation sets differed.

    Source: §IV · p. 4

  • Takeaways & Limitations

    The findings suggest potential for deep learning, particularly vision transformers, as future decision-support tools for BCC subtyping from dermatoscopy images.

    Source: §IV · p. 4, §V · p. 4

    The dataset comes from a single clinical center and includes Caucasian patients, constraining generalizability to more diverse populations.

    Source: §IV · p. 4

Abstract

from arXiv · show

Basal Cell Carcinoma (BCC) is the most common type of skin cancer, accounting for nearly 80% of skin cancer di- agnoses. Its optimal clinical management is guided by the distinct histopathologic subtype, with aggressive variants requiring more drastic measures. In current clinical practice, subtyping relies on skin biopsies, a procedure both costly and invasive. In this paper, we conduct a preliminary investigation into using deep learning for BCC subtyping, solely from a single dermatoscopic image of the lesion. Given the limited data at our disposal, we employ pre-trained vision transformers (ViTs), a state-of-the-art family of models highly effective for challenging downstream tasks with limited labeled data. Through repeated stratified k-fold cross-validation, we demonstrate that ViTs can achieve superior performance (AUC 0.784 on a dataset of 1271 dermatoscopic images of various BCC subtypes) over standard CNN-based baselines as well as previously-reported human reader perfor- mance, on the task of differentiating aggressive BCCs from other subtype families. These initial findings highlight the potential of combining deep learning and dermatoscopy to provide a biopsy- free alternative for BCC subtyping, thus aiding in improving treatment planning and patient outcomes.

I. INTRODUCTION

BCC subtype guides treatment, but establishing it requires a costly, invasive biopsy. The introduction motivates dermatoscopy as an accessible imaging route and proposes vision transformers for aggressive-versus-non-aggressive BCC classification despite limited labeled data.

I. INTRODUCTION

  • Aggressive BCC subtypes may require micrographic surgery, whereas non-aggressive and superficial tumors may be treated more conservatively; biopsy is costly and invasive.Dermatoscopy is described as the most common and accessible imaging modality, but only a few studies have examined its use for BCC subtyping.Source: §I · p. 1
  • To the authors’ knowledge, prior deep-learning BCC-subtyping work used histopathology, reflectance confocal microscopy, or optical coherence tomography, but not dermatoscopy.The introduction frames dermatoscopy-based BCC subtyping as a gap despite its accessibility.Source: §I · p. 1
  • After instruction on dermatoscopic criteria, human readers’ sensitivity rose from 42.3% to 56.4%, while specificity shifted from 71.7% to 70.7%.The readers classified 235 BCCs before and after a two-month washout period.Source: §I · pp. 1–2
  • The paper proposes vision transformers for classifying aggressive versus non-aggressive BCC from dermatoscopic images, motivated by limited annotated data and transformer pre-training.The task is the same binary classification pursued in.Source: §I · p. 2

II. DATA

The data section introduces the AUTh BCC subtype dataset and the ISIC dermatoscopy subset used for domain adaptation before target-task fine-tuning.

II. DATA

  • The paper presents the AUTh BCC subtype dataset and briefly describes the ISIC dermatoscopy dataset used to adapt models to that domain before fine-tuning on the target task.Source: §II · p. 2

A. AUTh BCC subtype dataset

The AUTh dataset comprises biopsy-confirmed BCC cases grouped into three subtype categories, with demographic and tumor-characteristic information collected for the cohort.

A. AUTh BCC subtype dataset

  • The AUTh dataset contains 1,271 dermatoscopic images from Caucasian patients diagnosed at one Thessaloniki department between 2018 and 2022.Images were captured at 10-fold magnification; mixed histopathologic subtypes and incompletely excised tumors were excluded.Source: A. AUTh BCC subtype dataset · p. 2
  • Biopsy-confirmed subtypes were grouped as superficial BCC, non-aggressive non-superficial BCC, or aggressive BCC.The non-aggressive non-superficial group includes nodular, nodulocystic, and Pinkus fibroepithelioma; the aggressive group includes infiltrative, morpheaform, micronodular, and metatypical BCC.Source: A. AUTh BCC subtype dataset · p. 2
  • Table I covers demographic data and tumor characteristics, while Figure 1 shows sample images from each of the three subtype categories.Source: Table I · p. 2, A. AUTh BCC subtype dataset · p. 2, Figure 1 · p. 2

B. ISIC archive

The ISIC Archive supplies public dermatology images; the study filters it to diagnosed dermatoscopic images and uses a subset for domain adaptation.

B. ISIC archive

  • The ISIC Archive contains 503,955 publicly available dermatology-related images, including dermatoscopic, reflectance confocal microscopy, and whole-body photographs.Source: B. ISIC archive · p. 2
  • For domain adaptation, the study retained diagnosed dermatoscopic images and discarded severely under-represented classes; Table II reports statistics for the resulting subset.Source: B. ISIC archive · p. 2, Table II · p. 3

A. Problem statement

The study targets aggressive-versus-lower-risk BCC classification from a single dermatoscopic image, using a vision transformer because labeled subtype images are scarce. Images are standardized to balance detail preservation with computational feasibility.

A. Problem statement

  • The target is to distinguish high-risk aggressive BCC from lower-risk subtypes using one dermatoscopic image.The paper selects a vision transformer because labeled BCC subtype images are scarce and the task is complex.Source: A. Problem statement · p. 3
  • Unsupervised pre-training on large datasets is the stated reason for choosing vision transformers for limited-label downstream tasks.The paper applies this rationale to its scarce labeled dermatoscopic BCC subtype data.Source: A. Problem statement · p. 3

B. Image pre-processing

  • Images are padded to square aspect ratios and resized to 512×512 pixels.The authors describe this resolution as a trade-off between preserving fine-grained diagnostic details and computational feasibility.Source: B. Image pre-processing · p. 3

C. Model architecture

The paper uses BEiT, a vision transformer with masked image modeling, as its dermatoscopic backbone. It adapts pretrained weights on a larger dermatoscopy dataset before initializing BCC subtype classification.

C. Model architecture

  • BEiT learns visual representations by reconstructing discrete visual tokens from masked image patches during pre-training.The paper describes this as a masked image modeling strategy inspired by BERT.Source: C. Model architecture · p. 3
  • The BEiT-Large model uses 512×512 inputs, 16×16 patches, a 32×32 token grid, and approximately 300 million parameters.It comprises 24 transformer layers, each with 16 attention heads, and a hidden dimension of 1024.Source: C. Model architecture · p. 3

D. Feature adaptation

  • The pretrained BEiT weights are fine-tuned on about 53,000 ISIC dermatoscopic images across nine lesion classes before initializing BCC subtype classification.The model is trained for 100 epochs using 80% of the data for training and 20% for validation; the checkpoint with the highest validation accuracy is selected.Source: D. Feature adaptation · p. 3

E. Finetuning for BCC subtype classification

The model is fine-tuned for aggressive-versus-lower-risk BCC classification with validation-based checkpoint selection. Training addresses class imbalance through oversampling and uses moderate image augmentation.

E. Finetuning for BCC subtype classification

  • BEiT is fine-tuned for 10 epochs with a 10^-5 initial learning rate, 10^-4 weight decay, cosine annealing, and batch size 16.Performance is monitored on an independent validation set during training.Source: E. Finetuning for BCC subtype… · p. 3
  • Minority-class oversampling balances class representation in each training batch, alongside moderate flips, brightness and contrast changes, shifts, scaling, and rotations.No augmentation is applied during validation or testing.Source: E. Finetuning for BCC subtype… · p. 3
  • The highest-validation-performance checkpoint is evaluated on a completely independent test set to assess generalization to unseen data.The validation performance criterion is not specified in this passage.Source: E. Finetuning for BCC subtype… · p. 3

IV. EXPERIMENTAL RESULTS

Across repeated stratified cross-validation, BEiT had the strongest reported performance among the tested models for aggressive BCC classification. The comparison with human readers is preliminary, and the single-center, Caucasian-only dataset limits generalizability.

IV. EXPERIMENTAL RESULTS

  • The study used stratified 5-fold cross-validation repeated three times, setting aside 20% of each fold’s training set for validation.The best validation model by ROC-AUC was evaluated on that fold’s test set.Source: §IV · pp. 3–4
  • ResNet-based baselines used the same training setup and computational budget, and model sensitivity was measured at 70% specificity for comparison with human experts.The human expert results are reported from.Source: §IV · p. 4
  • 0.784 average AUC and 70.0% sensitivity were BEiT’s highest reported results among the tested models; its sensitivity exceeded instructed experts’ 56.4% by 14%, a preliminary comparison given differing evaluation sets.Table III reports average metrics over three repetitions of stratified 5-fold cross-validation.Source: §IV · p. 4, Table III · p. 4
  • The findings suggest potential for future dermatology decision-support tools, but the single-center, Caucasian-only dataset constrains generalizability to more diverse populations.The authors call for multi-center, multi-ethnic validation and more controlled comparisons with human experts.Source: §IV · p. 4

V. CONCLUSIONS

The paper presents an initial proof of concept for classifying aggressive versus non-aggressive BCC subtypes from dermatoscopic images. Its vision-transformer approach achieved an AUC of 0.784, while comparisons with human experts remain preliminary.

  • AUC 0.784 was achieved by the vision-transformer approach on a dataset of 1271 BCC cases of various subtypes.Source: §V · p. 4
  • Preliminary comparisons suggest deep-learning models outperform human experts, but different dataset compositions prevent definitive conclusions.Source: §V · p. 4
  • The findings suggest promise for non-invasive BCC subtyping, with larger datasets and standardized evaluation protocols needed to reinforce them.Source: §V · p. 4
Loading 2609.07180v1…