Paper deep dive
Deep Learning for Biopsy-Free Subtyping of Basal Cell Carcinoma from Dermatoscopic Images
Alexandros Papadopoulos, Chrysa Episkopou, Ioannis Sarafis, Aimilios Lallas, Anastasios Delopoulos
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 95%
Last extracted: 9/13/2026, 4:56:18 AM
Summary
This paper investigates the use of deep learning, specifically Vision Transformers (ViTs), for the biopsy-free subtyping of Basal Cell Carcinoma (BCC) from dermatoscopic images. The study addresses the clinical need to differentiate aggressive BCC subtypes from non-aggressive ones to guide treatment planning. Using the AUTh BCC subtype dataset of 1,271 images, the authors fine-tuned a BEiT (BERT pre-training of image transformers) model. The proposed method achieved an AUC of 0.784, outperforming standard CNN baselines (ResNet-18, ResNet-50) and previously reported human expert performance. The findings suggest that deep learning can serve as a viable, non-invasive decision-support tool for BCC subtyping.
Entities (11)
Relation Signals (9)
BEiT → achievedperformance → AUC 0.784
confidence 98% · BEiT-16-512 0.784±0.026
BEiT → usedfor → BCC subtyping
confidence 98% · we present an automated approach for differentiating aggressive from non-aggressive BCCs based on dermatoscopic images
BEiT → trainedon → AUTh BCC subtype dataset
confidence 97% · The model is trained for a total of 10 epochs on a training subset of the AUTh BCC subtype dataset
BEiT → adaptedusing → ISIC Archive
confidence 96% · we conduct a fully supervised finetuning on the ISIC subset... to adapt it to the domain of medical imaging
BEiT → outperformed → ResNet-18
confidence 95% · BEiT... achieved the highest average AUC (0.784)... surpassing the ResNet-based baselines
BEiT → outperformed → ResNet-50
confidence 95% · BEiT... achieved the highest average AUC (0.784)... surpassing the ResNet-based baselines
Skin Biopsy → isalternativeto → Dermatoscopy-based Deep Learning
confidence 93% · provide a biopsy-free alternative for BCC subtyping
Aggressive BCC → requires → drastic measures
confidence 92% · aggressive variants requiring more drastic measures.
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract: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.
Tags
Links
- Source: https://arxiv.org/abs/2609.07180v1
- Canonical: https://arxiv.org/abs/2609.07180v1
Trouble viewing inline? Open PDF directly →
Full Text
22,511 characters extracted from source content.
Expand or collapse full text
Deep Learning for Biopsy-Free Subtyping of Basal Cell Carcinoma from Dermatoscopic Images Thanks: This work was funded by the Special Account for Research Funds of the Aristotle University of Thessaloniki (90169/2024) Alexandros Papadopoulos, Chrysa Episkopou, Ioannis Sarafis and Anastasios Delopoulos Affiliation: School of Electrical and Computer Engineering Aristotle University of Thessaloniki, Greece palexang, cepiskop, iosarafis, antelopo@ece.auth.gr Aimilios Lallas Affiliation: First Department of Dermatology Aristotle University of Thessaloniki, Greece alallas@auth.gr Abstract Basal Cell Carcinoma (BCC) is the most common type of skin cancer, accounting for nearly 80% of skin cancer diagnoses. 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 performance, 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. Index Terms: dermatoscopy, basal cell carcinoma, subtypes, deep learning, data-driven, vision transformers I Introduction Basal Cell Carcinoma (BCC) is a type of skin cancer caused primarily by ultraviolet radiation (UV). As such, it most often develops on body parts exposed to the sun[1]. Treatment options for BCC depend on its histopathologic subtype, with micrographic surgery suggested as the optimal treatment for aggressive subtypes, unlike non-aggressive and superficial tumors that may be treated more conservatively[2]. Prior to any treatment however, the subtype must first be established by a diagnostic biopsy, a procedure both costly and invasive that may also cause significant patient discomfort. Some imaging techniques, like optical coherence tomography or reflectance confocal microscopy, coupled with expert human evaluators or deep learning models, have been proposed as a viable alternative to diagnostic biopsies[3, 4]. However, only a few studies[5, 6, 7] have examined whether dermatoscopy in particular, by far the most common and accessible imaging modality in the field, can aid in BCC subtyping. Dermatoscopy (also known as dermoscopy) is a non-invasive imaging technique widely used by dermatologists for the examination and diagnosis of skin lesions [8]. It involves a handheld device called a dermatoscope, which provides magnification (typically around 10×10×) and illumination through polarized or non-polarized light sources. Dermatoscopy enables clinicians to visualize subsurface structures and morphological features not visible to the naked eye, such as pigment networks, vascular patterns, globules, and crystalline structures[9]. It significantly improves diagnostic accuracy compared to visual examination alone, facilitating early detection and differentiation of benign and malignant skin conditions. Recent advancements[10] have combined dermatoscopy with deep learning-based AI systems, to further enhance diagnostic capabilities and support clinical decision-making. Deep learning has been a natural fit for automating medical image analysis, with many applications during the past decade, and especially so in dermatology, for classification of skin lesions, like melanoma[11, 12] and other skin cancers[13, 14]. Within an abundant literature, however, little attention has been given to BCC subtyping, an essential task for treatment planning as we have discussed. The existing works focus on using either whole-slide histopathology images (WSI)[15, 16], or other imaging techniques like reflectance confocal microscopy (RCM)[3] and optical coherence tomography (OCT) [4]. To the best of our knowledge, however, no works have considered combining dermatoscopy with deep learning for subtyping basal cell carcinomas. Setting deep learning aside, few works have studied if dermatoscopy is informative enough to differentiate low-risk BCCs (superficial, nodular) from more aggressive subtypes (infiltrative, morpheaform, micronodular). The early work of[5] identified and evaluated various dermoscopic criteria for differentiating superficial BCC (sBCC) from other subtypes, resulting in an algorithm with 81.9% sensitivity and 81.8% specificity in a dataset of 77 sBCCs and 258 non-sBCCs. A follow-up work[6] investigated a more clinically relevant problem: whether human readers can accurately diagnose aggressive BCCs from a single dermoscopy image. To that end, 30 expert human evaluators with variable levels of experience were asked to perform an initial classification of 235 BCCs into superficial (sBCC), non-superficial (nsBCC) and aggressive (according to the same categorization described in SectionII-A). After a washout period of 2 months, they were explicitly instructed on the different dermatoscopic criteria for each type and repeated the classification of the 235 BCCs. The authors then compared the performance of the evaluators before and after the training and noticed substantial improvements in sensitivity (from 42.3%42.3\% to 56.4%56.4\%) with a minor decrease in specificity (from 71.7%71.7\% to 70.7%70.7\%). Here we make a first step towards filling the above gap in the literature. More specifically, we present an automated approach for differentiating aggressive from non-aggressive BCCs based on dermatoscopic images, a binary classification task identical to the one pursued in[6]. To that end, we follow a data-driven approach based on deep learning and particularly vision transformers[17], a choice dictated by the very small annotated BCC subtype dataset at our disposal (presented in Section I-A), which makes it difficult to train a data-efficient deep learning model. In doing so, we take advantage of the unsupervised pre-training of vision transformers on vast amounts of data, which has been shown to be very effective in downstream tasks with limited labeled data[18], like the learning problem we face here. This paper is structured as follows: In Section I, we present the AUTh BCC subtype dataset which motivated and informed this work, along with any other datasets that were used for auxiliary purposes. In Section I, we outline the proposed deep learning approach for building an accurate aggressive BCC subtype predictor. In Section IV, we describe the experimental setup, outline the training details and present and discuss the experimental results. Finally, we offer concluding remarks in Section V. I Data We begin with the presentation of the data at our disposal. First, we present the AUTh BCC subtype dataset which informed most of the design choices we made. We then briefly describe the ISIC archive, a public dermatoscopy dataset that was used for adapting our models to the dermatoscopy domain, prior to finetuning on our target task. I-A AUTh BCC subtype dataset The AUTh BCC subtype dataset contains 1271 dermatoscopic images of Caucasian patients with BCC, diagnosed at the First Department of Dermatology, Aristotle University of Thessaloniki, Greece, between January 2018 and December 2022. The images were captured using Dermlite Foto equipment (3Gen, Dana Point, CA) at 10-fold magnification. Mixed histopathological subtypes and incompletely excised BCCs (shave or punch biopsy) were excluded. For each tumor, the biopsy-confirmed histopathological subtype was recorded and classified into three categories: 1. Superficial BCC (sBCC) 2. Non-aggressive, non-superficial BCC (nsBCC): nodular BCC, nodulocystic BCC, and Pinkus fibroepithelioma. 3. Aggressive BCC: infiltrative, morpheaform, micronodular, and metatypical BCC. Additional metadata like age and sex of the patient, as well as tumor location were also collected. The demographic characteristics of the dataset are given in Table I, while a sample of the dataset can be seen in Figure 1 Fig. 1: Sample images from the BCC subtype dataset used in this study. First row: Superficial BCC, Second row: Non-aggressive, non-superficial BCC, Third row: Aggressive BCC TABLE I: Demographic data and tumor characteristics of the AUTh BCC subtype dataset Low Risk High Risk Superficial Non-superficial Aggressive (n=275) (n=726) (n=270) Age (Mean ± SD) 66.2 ± 14.3 70.0 ± 13.1 69.9 ± 12.8 Sex (M / F / NA) 98 / 89 / 88 327 / 212 / 187 70 / 42 / 158 Tumor location Head and neck 69 (25.1%) 373 (51.4%) 83 (30.7%) Trunk 102 (37.1%) 136 (18.7%) 25 (9.3%) Upper extremities 11 (4.0%) 15 (2.1%) 2 (0.7%) Lower extremities 6 (2.2%) 12 (1.7%) 2 (0.7%) NA 87 (31.6%) 190 (26.2%) 158 (58.5%) Total 1001 270 I-B ISIC archive The International Skin Imaging Collaboration (ISIC) Archive is an open-access repository of 503,955 publicly-available dermatology-related images, including dermatoscopic, RCM and whole-body photographs. For the purposes of this work, we used only the dermatoscopic images of the archive for which a lesion diagnosis was available. After discarding some severely under-represented classes, we ended up with the subset of Table I. TABLE I: Statistics of the subset of the publicly available ISIC dermatoscopy images that was used for adapting models pre-trained on natural images to the dermatoscopy domain. Lesion Count Actinic keratosis 1,367 Basal cell carcinoma 4,921 Benign keratosis 4,380 Dermatofibroma 420 Melanoma 7,349 Nevus 32,697 Other 972 Squamous cell carcinoma 1,372 Vascular lesion 348 Total 53,826 I Proposed approach I-A Problem statement Our goal is to build a deep learning model that will accurately differentiate the high-risk aggressive BCC subtype from other BCC subtypes considered to be low-risk, based on a single dermatoscopic image of the tumor. Given the complexity of the task (evidenced by the results of [7]) and the scarcity of labelled dermatoscopic images of BCC subtypes at our disposal, we opted for a vision transformer model, as the unsupervised pre-training of these models on vast amounts of data has rendered them very effective in downstream tasks with limited labeled data[18]. I-B Image pre-processing Since images in our dataset vary significantly in resolution due to the use of different dermatoscopic devices, we first apply a standardized preprocessing step to ensure consistency. Specifically, each image is padded to a square aspect ratio and subsequently resized to a fixed resolution of 512×512512× 512 pixels. While this resolution is higher than what is typically used in medical image analysis tasks, we deemed it to be a reasonable trade-off between preserving fine-grained image details that may be crucial for identifying subtle diagnostic features and maintaining computational feasibility. I-C Model architecture For our vision transformer backbone, we employed the BEiT model[18], a vision transformer architecture that achieved state-of-the-art performance in a number of benchmarks at the time of its release. BEiT is inspired by the success of BERT in natural language processing and adapts a similar masked image modeling strategy for vision tasks. During pre-training, BEiT learns to reconstruct discrete visual tokens (obtained from a pretrained tokenizer such as a VQ-VAE) from masked patches, enabling it to capture rich semantic representations of visual data in a highly efficient manner. Given the high resolution of our dermatoscopic images, we opted for the BEiT-Large variant, which operates on input images of size 512×512512× 512 with a patch size of 16×1616× 16. This configuration results in a grid of 32×3232× 32 input tokens and approximately 300 million parameters. The model consists of 24 transformer layers with 16 attention heads per layer and a hidden dimension of 1024. This large-scale architecture allows for substantial representational capacity, making it particularly suitable for modeling fine-grained visual distinctions such as those found between subtypes of Basal Cell Carcinoma. I-D Feature adaptation Since BEiT was originally pre-trained on natural image datasets, it is essential to first adapt it to the domain of medical imaging, specifically dermatoscopic images, before applying it to the downstream task of BCC subtyping. To facilitate the domain adaptation, we begin with the publicly available pre-trained weights of BEiT and conduct a fully supervised finetuning on the ISIC subset described in Table I, which contains approximately 53,000 dermatoscopic images spanning 9 distinct classes of skin lesions. We finetune the model for 100 epochs using 80% of the dataset for training, while reserving the remaining 20% for monitoring the model’s generalization performance. After training, we select the checkpoint with the highest validation accuracy and use it as the initialization point for our subsequent BCC subtype classification model. This intermediate step ensures that the transformer learns domain-specific features relevant to dermatoscopic imagery, thereby improving its effectiveness on the target classification task. I-E Finetuning for BCC subtype classification After feature adaptation, we proceed with finetuning the model to the downstream task of differentiating aggressive BCC from lower-risk subtypes. Following standard practice for BEiT, we employ a low initial learning rate of 10−510^-5, combined with a weight decay factor of 10−410^-4, and apply a cosine annealing schedule to gradually reduce the learning rate during training. The model is trained for a total of 10 epochs on a training subset of the AUTh BCC subtype dataset, using a batch size of 16, while its performance is continuously monitored on an independent validation set to prevent overfitting. To account for the imbalance (see Table I) between the high- and low-risk classes, we perform oversampling of the minority class, so that both classes are represented equally in a training batch. We apply moderate data augmentation during training, including random vertical and horizontal flips, subtle brightness and contrast adjustments, as well as random shifts, scalings, and rotations. No augmentation is applied during the validation or testing phases. After training, the checkpoint with the highest validation performance is selected and subsequently evaluated on a completely independent test set to assess its ability to generalize to unseen data. More details on the training/validation/test splits are provided in Section IV. IV Experimental results To allow for a reliable estimation of the model’s generalization performance and owing to the small dataset size, we opted for a stratified 5-fold cross-validation scheme which we repeated three times. In each iteration of this repeated k-fold scheme, 20% of the fold training set was set aside for validation. After training, the best performing model on the validation set (according to ROC-AUC) was evaluated on the fold test set. In addition to BEiT, we also experimented with two ResNet-based baseline models (trained with the same setup and computational budget) for comparison. We present the results of this experiment in Table I, alongside the previous human expert performance on the same task, as reported in [6]. In Table I, the reported metrics have been averaged across all training-test splits of the repeated k-fold scheme. For our models, sensitivity was calculated at the operating point yielding 0.700.70 specificity, to align with the reported specificity of human experts in [6] and to facilitate comparisons with them. TABLE I: Average performance metrics over 3 repetitions of a stratified 5-fold cross-validation experiment. For deep learning models, sensitivity was calculated at the operating point of 70% specificity to allow comparisons with [6] Method AUC Specificity Sensitivity Human experts (pre)[6] NA 71.1± 7.7 42.3± 15.9 Human experts (post)[6] NA 70.7± 6.93 56.4± 15.7 ResNet-18 0.735± 0.023 70.0* 61.6± 5.8 ResNet-50 0.755± 0.018 70.0* 67.4.±. 4.3 BEiT-16-512 0.784± 0.026 70.0* 70.0± 6.6 The results in Table I offer several insights. At first glance, deep learning models appear to outperform human experts, with the BEiT model achieving a 14% higher sensitivity compared to the human performance after the experts were instructed on the dermatoscopic predictors of BCC subtypes (70.0% vs. 56.4%). However, these comparisons should be considered preliminary, as the evaluation sets used in our experiments differ from those in the human expert study [6]. Within our dataset, the BEiT-based transformer notably surpassed the ResNet-based baselines, achieving the highest average AUC (0.784) and sensitivity (70.0%). Despite the differences in evaluation conditions, these initial findings underscore the potential advantage of deep learning (and vision transformers in particular) for addressing the clinically challenging task of BCC subtyping from dermatoscopy images, thus highlighting their promise as future decision-support tools in dermatology practice. These findings are, however, subject to important limitations. The dataset originates from a single clinical center and is limited to Caucasian patients, which constrains the generalizability of our results to more diverse populations. Future work will aim to validate the proposed approach on multi-center, multi-ethnic cohorts and to establish more rigorous, controlled comparisons with human expert performance. V Conclusions We presented an initial proof-of-concept exploration of using deep learning to classify aggressive versus non-aggressive Basal Cell Carcinoma subtypes directly from dermatoscopic images. Our approach, based on vision transformers, achieved a respectable AUC of 0.784 in a dataset of 1271 BCC cases of various subtypes. Preliminary comparisons imply that deep learning models outperform human experts, although different dataset compositions preclude definitive conclusions. Nonetheless, our results suggest that deep learning holds much promise for non-invasive BCC subtyping, although additional work is needed to further reinforce these findings based on larger datasets and standardized evaluation protocols. References [1] A. N. Crowson (2006) Basal cell carcinoma: biology, morphology and clinical implications. Modern pathology 19, p. S127–S147. Cited by: §I. [2] K. Peris et al. (2023) European consensus-based interdisciplinary guideline for diagnosis and treatment of basal cell carcinoma—update 2023. European Journal of Cancer 192, p. 113254. Cited by: §I. [3] M. Chen et al. (2022) Deep learning on reflectance confocal microscopy improves raman spectral diagnosis of basal cell carcinoma. Journal of Biomedical Optics 27 (6), p. 065004–065004. Cited by: §I, §I. [4] D. Mandache et al. (2018) Basal cell carcinoma detection in full field oct images using convolutional neural networks. In 2018 IEEE 15th International Symposium on Biomedical Imaging (ISBI 2018), p. 784–787. Cited by: §I, §I. [5] A. Lallas et al. (2014) Accuracy of dermoscopic criteria for discriminating superficial from other subtypes of basal cell carcinoma. Journal of the American Academy of Dermatology 70 (2), p. 303–311. Cited by: §I, §I. [6] E. Camela et al. (2024) Dermatoscopic predictors of histopathologically aggressive basal cell carcinoma and their positive impact of subtype prediction by human readers. Journal of the American Academy of Dermatology. Cited by: §I, §I, §I, TABLE I, TABLE I, TABLE I, TABLE I, §IV, §IV. [7] C. Longo et al. (2014) Classifying distinct basal cell carcinoma subtype by means of dermatoscopy and reflectance confocal microscopy. Journal of the American Academy of Dermatology 71 (4), p. 716–724. Cited by: §I, §I-A. [8] H. Kittler et al. (2011) Dermoscopy of pigmented skin lesions. Dermatologic Clinics 29 (2), p. 149–158. Cited by: §I. [9] G. Argenziano and H. P. Soyer (2022) Dermoscopy: a review of the structures that facilitate the recognition of malignant skin lesions. Journal of the American Academy of Dermatology 86 (2), p. S1–S14. Cited by: §I. [10] C. Magalhaes, J. Mendes, and R. Vardasca (2024) Systematic review of deep learning techniques in skin cancer detection. BioMedInformatics 4 (4), p. 2251–2270. Cited by: §I. [11] P. Yao et al. (2022) Single model deep learning on imbalanced small datasets for skin lesion classification. IEEE Transactions on Medical Imaging 41 (5), p. 1242–1254. External Links: Document Cited by: §I. [12] A. R. Lopez et al. (2017) Skin lesion classification from dermoscopic images using deep learning techniques. In 2017 13th IASTED international conference on biomedical engineering (BioMed), p. 49–54. Cited by: §I. [13] M. A. Kassem, K. M. Hosny, and M. M. Fouad (2020) Skin lesions classification into eight classes for isic 2019 using deep convolutional neural network and transfer learning. IEEE Access 8 (), p. 114822–114832. External Links: Document Cited by: §I. [14] M. A. Kassem et al. (2021) Machine learning and deep learning methods for skin lesion classification and diagnosis: a systematic review. Diagnostics 11 (8), p. 1390. Cited by: §I. [15] F. Yacob et al. (2023) Weakly supervised detection and classification of basal cell carcinoma using graph-transformer on whole slide images. Scientific Reports 13 (1), p. 7555. Cited by: §I. [16] Bungărdean et al. (2022) Deep learning with transfer learning in pathology. case study: classification of basal cell carcinoma. Romanian Journal of Morphology and Embryology 62 (4), p. 1017. Cited by: §I. [17] A. Dosovitskiy et al. (2020) An image is worth 16x16 words: transformers for image recognition at scale. arXiv preprint arXiv:2010.11929. Cited by: §I. [18] H. Bao et al. (2022) BEiT: BERT pre-training of image transformers. In International Conference on Learning Representations, Cited by: §I, §I-A, §I-C.