AI‐assisted basal cell carcinoma diagnosis with LC‐OCT: A multicentric retrospective study

Abstract

Background

Basal cell carcinoma (BCC) is the most common skin cancer, requiring an early diagnosis and accurate margin definition to prevent functional and cosmetic complications. Traditional methods using clinical and dermoscopic images (C&D) often rely on biopsies and histology for final validation. Non‐invasive techniques like LC‐OCT, enabling ‘digital biopsies’, are promising alternatives, but remain underutilized due to the expertise required. The development of Artificial Intelligence (AI) algorithms is a promising approach to assist dermatologists in their diagnosis and support the broader adoption of such technologies.

Objectives

We present a real‐time AI assistant for BCC diagnosis with LC‐OCT, which is, to date, the only real‐time AI model across all dermatological imaging modalities. The study aims to quantify the model's effectiveness when used by dermatologists with different levels of expertise and compare its performance with traditional methods and unaided LC‐OCT.

Methods

This multicenter, retrospective study involved 43 dermatologists in a double‐rounded quiz on 200 equivocal BCC lesions. Diagnoses were first made on C&D images, then with LC‐OCT or AI‐assisted LC‐OCT in a randomized manner.

Results

AI‐assisted LC‐OCT significantly improves dermatologists' diagnostic performance in detecting BCC (+25.8 points in sensitivity and +16.8 points in specificity compared to C&D), particularly benefiting those with less LC‐OCT experience, effectively bridging a 2‐year gap of expertise. These results highlight the potential for broader clinical adoption through AI assistance and underscore its promise in reducing the need for invasive procedures and improving patient outcomes.

Conclusions

These results support a broader adoption of LC‐OCT use in clinical practice thanks to AI assistance and underscore its promise in reducing the need for invasive procedures and improving patient outcomes.

Plain Language Summary

Basal cell carcinoma (BCC) is the most common skin cancer. Although rarely deadly, it can progressively invade surrounding tissues such as cartilage and bone. Early and accurate diagnosis is essential to avoid complications such as functional loss or significant cosmetic damage.

This study includes clinical and LC‐OCT data from 200 suspicious lesions for BCC, collected across four hospitals in France, Spain, Italy and Belgium. Its goal is to retrospectively quantify the effect of real‐time AI assistance with LC‐OCT on the diagnostic efficacy of dermatologists with different levels of expertise with this technology.

Forty‐three dermatologists evaluated the 200 lesions using traditional imaging methods (clinical photography and dermoscopy) as well as LC‐OCT, through a dedicated web application. They had to give a binary decision for BCC at two different stages, with traditional images alone and with LC‐OCT videos. Half of the time, LC‐OCT images were accompanied by an AI assistant.

The analysis of the diagnostic results across the different scenarios demonstrated that LC‐OCT outperforms traditional imaging for diagnosing BCC in equivocal lesions. It also showed that AI assistance provides significant help to dermatologists and more particularly those with less experience in LC‐OCT, by improving their diagnostic precision.

AI support could make LC‐OCT easier to interpret and improve BCC detection. This may reduce the need for skin biopsies, accelerate clinical decision‐making and improve patient care by guiding treatment selection and optimizing surgical procedures.

Real‐time AI assistance significantly improves dermatologists' performance in diagnosing basal cell carcinoma with LC‐OCT, outperforming traditional imaging methods. AI‐assisted novices reached expert‐level performance, bridging a 2‐year expertise gap, thus supporting a broader clinical adoption of non‐invasive technologies, reducing the need for invasive biopsies and improving patient outcomes.

Why was the study undertaken?

• This study explores whether AI‐assisted Line‐field Confocal Optical Coherence Tomography (LC‐OCT) enhances the diagnostic accuracy of basal cell carcinoma in dermatological practice compared to traditional clinical and dermoscopic methods, especially for those with limited experience with LC‐OCT.

What does this study add?

• This international, multicenter, retrospective study involved 43 dermatologists of different levels of expertise in order to statistically quantifiy the performance of AI‐assisted LC‐OCT and traditional methods on a representative set of 200 lesions. AI‐assisted LC‐OCT improved diagnostic accuracy by 20.3 percentage points, from 72.7% to 93.0%, compared to traditional methods. The AI assistant enables both novice and intermediate practitioners to attain diagnostic accuracy on par with that of experts without AI assistance, bridging a 2‐year experience gap in LC‐OCT practice.

What are the implications of this study for disease understanding and/or clinical care?

• This study shows that the integration of AI with LC‐OCT improves basal cell carcinoma detection and can facilitate wider adoption of non‐invasive ‘digital biopsies’, enhancing skin cancer management and improving patient care by reducing the need for biopsies.

INTRODUCTION

Basal cell carcinoma (BCC) is the most prevalent form of skin cancer as it accounts for 75% of all keratinocyte cancers. Although BCC rarely leads to death, it can cause severe functional and cosmetic morbidities due to its locally invasive nature, particularly on the head or neck regions, where lesions are frequent and often require surgical treatment. Following the current guidelines, ¹ clinically equivocal lesions for BCC require a biopsy for histopathological confirmation of diagnosis and subtype. The invasive nature of biopsy procedures presents several drawbacks, including patient discomfort, scarring, diagnosis delays and increased healthcare costs, which makes BCC diagnosis a major public health concern.

Dermoscopy, the most widely used technique in dermatology, achieves a sensitivity of 91.3% and a specificity of 84.3% for BCC diagnosis in non‐pigmented lesions ² but often remains followed by a biopsy. Advancements in imaging technologies, such as Reflectance Confocal Microscopy (RCM), Optical Coherence Tomography (OCT) or more recently Line‐field Confocal Optical Coherence Tomography (LC‐OCT), ³ have been pivotal in proposing non‐invasive alternatives to traditional biopsy, offering the potential for same‐day diagnosis and treatment, limiting invasive procedures, saving valuable time for dermatologists and reducing healthcare costs. ⁴ , ⁵ , ⁶ , ⁷ , ⁸ , ⁹

The field of dermatology, alongside most medical specialities, is grappling with escalating human resource shortages ¹⁰ and the continual emergence of novel technologies and imaging techniques, necessitating continuous learning efforts from specialists, making diagnostic performances highly dependent on practitioner experience and limiting clinical adoption. In this context, artificial intelligence (AI) algorithms have emerged as promising tools to assist doctors in their learning process and daily practice. As the field of medical AI is becoming more mature, the attention is shifting from human–machine competition ¹¹ , ¹² , ¹³ to human–machine collaboration ¹⁴ where dermatologists, ¹⁴ , ¹⁵ , ¹⁶ , ¹⁷ emergency medicine clinicians ¹⁸ or radiologists ¹⁹ , ²⁰ , ²¹ improve their capacity to detect cancers or fractures with AI assistance.

In this study, we analyse the efficacy of a CE‐marked AI assistant, designed to support dermatologists in identifying BCC through the use of LC‐OCT technology. Our research aims to evaluate the effect of this AI assistant on the diagnostic performance of a broad spectrum of dermatologists with different expertise levels in LC‐OCT.

Study protocol

Study objectives

The presented reader study was designed to achieve two primary outcomes in the context of AI‐assisted diagnosis of BCC in equivocal lesions, defined as challenging for traditional clinical and dermoscopic (C&D) evaluations:

• To evaluate and compare the diagnostic accuracy of traditional C&D methods with that of an AI‐assisted LC‐OCT approach.

• To assess and compare the diagnostic performance of LC‐OCT alone and AI‐assisted LC‐OCT.

In addition to these primary objectives, the study also aimed to accomplish several secondary goals:

• To compare the primary objectives among dermatologist subgroups, categorized as ‘experts’ (over 2 years or 500 LC‐OCT diagnoses), ‘intermediate’ (6 months to 2 years or 50–500 diagnoses) and ‘novices’ (less than intermediate experience).

• To compare the standalone AI model with the diagnostic capabilities of dermatologists using C&D imaging and LC‐OCT imaging.

• To compare the diagnostic confidence scores and time to diagnosis across the different methods.

Study design

Forty‐three dermatologists with different levels of expertise (Appendix A in Data S1) evaluated 200 cases through a secure web application, in a double‐round quiz format separated by a compulsory 2‐week (minimum) memory wash‐out period (Figure 1). Each case of the quiz entailed two distinct diagnostic stages, mimicking the real‐life scenario of BCC assessment in centers using LC‐OCT: first, a C&D examination, followed by LC‐OCT for an in‐depth examination (Figure S1). Initially, dermatologists were presented with C&D images, together with the body location of the lesion, requiring a binary diagnosis of ‘BCC’ or ‘not BCC’ and a confidence rating on a scale of 1–10. Subsequently, an approximately 10‐second‐long LC‐OCT video of the same lesion was shown. The reviewer was again asked to give a binary diagnosis with a confidence score. Each dermatologist was shown each LC‐OCT case with or without AI assistance, in random order, which is key to mitigate potential biases such as learning curve, boredom or fatigue.

FIGURE 1.

Study protocol: 43 dermatologists took a double‐round quiz on 200 cases, separated by a memory washout period of two weeks. Each case consisted of clinical and dermoscopy images then LC‐OCT video with or without AI assistance.

Completion of the first quiz round triggered a compulsory 2‐week (minimum) memory wash‐out period, enforced by the platform, to ensure a fresh perspective and mitigate recall bias in the second round (Figure S2). This round mirrored the first, with each previously AI‐assisted case now presented without assistance and vice versa, while maintaining the same C&D information. The sequence of cases was also randomized for each participant in the second round. Participants had the flexibility to initiate, pause and continue their quiz anytime between 8 September 2023 and 10 January 2024.

Patient cohort and data acquisition

This multicentric study, retrospective but structured to simulate the design of a prospective investigation, uses data from four European hospitals: the Hospital Clinic Barcelona, Spain, the University of Siena, Italy, the Hôpital Erasme, Belgium and the University Hospital of Saint‐Etienne, France. These centres have been acquiring LC‐OCT images as part of their clinical routine and for research purposes for several years.

To approximate the pathology distribution typically observed in hospitals, all LC‐OCT data collected from February 2020 to June 2022 were considered eligible for inclusion in the study. Nonetheless, several criteria were imposed for the selection of lesions. Each case was required to contain C&D images, LC‐OCT video or 3D imaging of sufficient image quality and histological analysis from biopsy or excision, which constitutes the ground truth for this research as it is considered today's gold standard. From this curated dataset, since LC‐OCT is used in practice only when a doubt subsists after dermoscopic examination or eventually to identify the BCC subtype, only equivocal lesions for C&D examinations were selected. To identify equivocal lesions, we engaged a group of three dermatologists who independently reviewed the C&D images for each case. They categorized the lesions into one of three groups: ‘clear‐cut BCC’, ‘clear‐cut other’ or ‘doubtful’. A lesion was classified as equivocal if at least one dermatologist deemed it ‘doubtful’ or if an incorrect diagnosis was made by one of them, representing 53% of the cases (226/426).

To ensure that the study's scope remained manageable for dermatologists to thoroughly review each case, we restricted our final study sample to the first 200 lesions and a sample of 10 s of LC‐OCT video representative of the lesion, chosen by an LC‐OCT expert at DAMAE Medical. The detailed statistics of the final study dataset can be found in Table 1.

TABLE 1.

Description of the 200 cases used for the study.

Category	Details
Number of patients	175 distinct patients
Age distribution	Average age is 65.7 years old (from 24 to 96 years old)
Gender distribution	93 males (53%), 72 females (41%), 10 unspecified genders (6%)
Lesion types	75 (37.5%) Basal Cell Carcinomas (BCCs) 15 (7.5%) Actinic Keratosis (AK) 27 (13.5%) Squamous Cell Carcinomas (SCCs) 13 (6.5%) Nevi 26 (13%) Melanomas 44 (22%) various lesions (scars, skin neoplasms, and inflammatory conditions)
BCC subtypes	11 infiltrating BCCs 4 micronodular BCCs 13 superficial BCCs 18 nodular BCCs 13 combining nodular and superficial BCCs 10 unspecified subtype 6 mixed subtype
Geographic distribution	108 cases from Hôpital Erasme (HUB), Université Libre de Bruxelles (44 BCCs) 11 cases from the Hospital Clinic Barcelona (9 BCCs) 42 cases from the University Hospital of Saint‐Etienne (13 BCCs) 39 cases from University of Siena (9 BCCs)

MATERIALS AND METHODS

LC‐OCT technology

LC‐OCT is a non‐invasive imaging modality utilized for in vivo skin analysis, capable of generating both vertical and horizontal high‐resolution images, in vivo at eight frames per second, at the cellular level ³ , ²² and 3D images. In the context of BCC diagnosis, the emphasis is placed on vertical views (from videos or 3D images) as they provide an enhanced perspective of the skin's stratified layers, very similar to a histological whole slide image ⁸ , ⁹ , ²³ and are particularly effective for the detailed observation of BCC lobules. ²³ LC‐OCT is commercially available in Europe (CE‐marked IIa) and Australia (TGA approved) with the presented AI assistance, while in the United States it is available as a class II device (FDA approved) without AI assistance, pending further approval. As of June 2025, DAMAE Medical has installed more than 100 devices worldwide, predominantly across Europe.

Live AI assistance

Effective AI assistance requires both high accuracy and explainability as physicians need to trust the model's reliability while being able to understand and verify its output, ensuring they can make informed decisions and disregard suggestions when necessary.

Focusing on these requirements, we developed a deep learning algorithm to distinguish BCCs from their most common imitators using 2D vertical images as it offers real‐time support and continuous assistance to dermatologists both during acquisition and review. It has been trained on a large diverse dataset of more than 685,000 hand‐labelled images with histological ground truth coming from more than 1000 patients across five different medical centres (Figures S4 and S5).

To avoid the pitfalls of ‘black box’ systems, a layer of interpretability is introduced through a real‐time attention heatmap overlaid on the LC‐OCT image, colour‐coded from green (0%) to red (100%) to indicate the BCC probability prediction (Figure 2). For positive detections, it highlights areas of concern of the model, helping dermatologists verify the prediction. If the focus aligns with their concerns, the diagnosis is supported: in practice, the heatmap successfully highlights BCC lobules. For negative detections, if the model identifies a concerning area according to the doctor but predicts a low BCC probability, it reassures the doctor that, despite shared concerns, the model confidently rules out the presence of BCC (Figures S1 and S6).

FIGURE 2.

Examples of different LC‐OCT frames with (bottom) and without AI assistance (top) and the associated impact on the number of correct answers among all dermatologists.

An additional layer of interpretability is added with a multiclass classification score for six potential BCC imitators: dermal nevus, melanocytic lesions, actinic keratosis (AK) and squamous cell carcinoma (SCC), sebaceous hyperplasia, healthy skin and an ‘other’ category for various benign pathologies. These differential diagnostic insights explain the model's rationale for a ‘BCC absence’ and follow the guideline of Tschandl et al. ¹⁴ for efficient human‐computer collaboration. When the dermatologist agrees with the differential diagnosis, it reinforces the likelihood of a non‐BCC diagnosis.

For an in‐depth technical analysis that adheres to international standards, ²⁴ see Appendix C in Data S1.

Statistical analysis

Confidence intervals for proportion metrics are computed using Delong's method. ²⁵ , ²⁶ To compare performances with and without AI, we use a two‐sided paired t‐test on average performances per dermatologist. All findings reported as statistically significant are meant at 95% after Bonferroni correction to account for multiple comparisons across sensitivity, specificity and accuracy metrics.

RESULTS

Enhanced diagnostic efficacy of LC‐OCT in equivocal lesions relative to C&D imaging techniques

We measured a statistically significant enhancement in diagnostic efficacy with the addition of LC‐OCT examination compared to C&D imaging. Specifically, scores improved when using LC‐OCT compared to C&D by +15.3% for sensitivity score, +14.3% for specificity and +14.8% for accuracy (Table 2).

TABLE 2.

Sensitivity, specificity and accuracy scores for clinical and dermoscopic images, LC‐OCT and LC‐OCT with AI assistance across all skill levels.

Expertise level	Method	Sensitivity	Specificity	Accuracy
Novice (n = 15)	C&D	0.631 (0.611–0.650)	0.745 (0.731–0.758)	0.702 (0.690–0.713)
	LC‐OCT	0.792 (0.767–0.815)	0.850 (0.833–0.866)	0.828 (0.814–0.841)
	LC‐OCT with AI	0.930 (0.913–0.943)	0.905 (0.890–0.917)	0.914 (0.903–0.924)
Intermediate (n = 13)	C&D	0.693 (0.672–0.713)	0.758 (0.743–0.7773)	0.734 (0.721–0.758)
	LC‐OCT	0.831 (0.806–0.853)	0.900 (0.884–0.913)	0.874 (0.861–0.886)
	LC‐OCT with AI	0.938 (0.922–0.952)	0.910 (0.895–0.923)	0.921 (0.910–0.930)
Expert (n = 15)	C&D	0.743 (0.724–0.760)	0.750 (0.736–0.763)	0.747 (0.736–0.758)
	LC‐OCT	0.901 (0.883–0.917)	0.933 (0.921–0.944)	0.921 (0.911–0.930)
	LC‐OCT with AI	0.972 (0.961–0.981)	0.942 (0.931–0.952)	0.954 (0.946–0.961)
All (n = 43)	C&D	0.689 (0.677–0.700)	0.751 (0.742–0.759)	0.727 (0.721–0.734)
	LC‐OCT	0.842 (0.829–0.854)	0.894 (0.886–0.902)	0.875 (0.867–0.881)
	LC‐OCT with AI	0.947 (0.939–0.954)	0.919 (0.912–0.926)	0.930 (0.924–0.935)

Note: Bold values indicate statistically significant improvements with AI assistance compared to LC‐OCT alone (with p‐values < 2.80 × 1e‐3).

LC‐OCT performances are consistent with the self‐declared expertise levels since experts outperform intermediates, who outperform novices (Table 2). Improvements when adding LC‐OCT imaging to C&D imaging stand regardless of levels of expertise (two‐sided paired t‐test p‐values < 1e‐5).

AI assistance improves diagnostic performance across all groups of expertise

The addition of AI assistance with LC‐OCT (Figure 2, Figure S3) significantly enhances diagnostic performances among all dermatologists with a significant increase in sensitivity, specificity and accuracy. Sensitivity increased by +10.5% from 84.2% without AI [95% CI 82.9–85.4] to 94.7% with AI [95% CI 93.9–95.4] (two‐sided paired t‐test test statistic = 9.68, p = 2.89 × 1e‐12). Specificity increased by 2.5% from 89.4% without AI [95% CI 88.6–90.2] to 91.9% with AI [95% CI 91.2–92.6] (statistic = 3.17, p = 2.80 × 1e‐3). Accuracy increased by 5.5% from 87.5% without AI [95% CI 86.7–88.1] to 93.0% with AI [95% CI 92.4–93.5] (statistic = 7.99, p = 5.74 × 1e‐10). AI assistance benefits to all subgroups of expertise with improvements in accuracy, sensitivity and specificity (Figure 3, Table 2). All improvements are significant with p‐values < 2.80 × 1e‐3 except for improvements in specificity for experts and intermediates with p‐values of 0.23 and 0.59, respectively.

FIGURE 3.

Dermatologists and groups of expertise sensitivity and specificity results with C&D images (purple), with LC‐OCT alone (red) and with AI‐assisted LC‐OCT (blue), simple aggregated AI ROC curve (green) and zoomed in version (bottom right). Small symbols (crosses, stars and circles) indicate individual scores. Large crosses represent group average scores, with the ends of the crosses showing confidence intervals. Arrows represent the average progression of each group.

LC‐OCT and AI assistance improve dermatologists' level of confidence

Dermatologists were significantly more confident when using LC‐OCT for diagnosis than C&D, with an increase of +1.50 points (on a scale from 1 to 10) overall, +1.09 points for novices, +1.66 points for intermediates and +1.86 points for experts. Interestingly, the increase in confidence with LC‐OCT is a reflection of the level of expertise with LC‐OCT.

Their level of confidence improves again when assisted by AI with LC‐OCT, with a mean confidence score increasing by 0.46 points from 8.26/10 without AI and 8.70/10 with AI (two‐sided t‐test on related scores p‐value = 3.3e‐104). This gain in confidence is statistically significant across all subgroups of expertise, with +0.58 points for novices, +0.31 points for intermediates and +0.44 points for experts (Table S1).

Intra‐operators' variability on C&D diagnosis

Dermatologists were prompted to respond to identical queries concerning C&D images on two separate occasions, with a mandatory memory wash‐out period of a minimum of 2 weeks. On average, responses varied for the exact same images and the same doctor 21.7% of the time (±5% standard deviation), with individual variability ranging from 13% to 33%. This high inconsistency level underscores both the efficacy of the memory wash‐out period, the level of ambiguity of our cases for C&D imaging methods and the variability of the human reviewers.

Comparative analysis of AI standalone performances with dermatologists

While the function of the AI tool is to assist dermatologists by providing real‐time, frame‐level predictions, its performance as a fully autonomous AI system can be simulated (see Data S1 for methodology details). When comparing Area Under the receiver operating characteristic Curve (AUC) scores (Figure S3), the AI model consistently surpassed all participating dermatologists without AI assistance. Rather than serving as a basis for comparison between an AI system and dermatologists, these findings are intended to highlight the AI model's advanced capabilities, which are mandatory to provide useful assistance.

DISCUSSION

In this investigation, we have shown that LC‐OCT is a valuable imaging modality to detect BCC on equivocal lesions for clinical and dermoscopy images, aligning with findings from prior research. ⁶ , ⁷ The utilization of LC‐OCT has led to notable enhancements in sensitivity with an increase of 15.3%, from 68.9% to 84.2%, and specificity with an increase of 14.3%, from 75.1% to 89.4%. The integration of AI assistance within the LC‐OCT diagnostic workflow has introduced further improvements, with a sensitivity increase of 10.5% to 94.7% and a specificity increase of 5.2% to 91.9% (Table 2). These concrete gains, reflective of representative clinical settings, robust to location biases (Tables S2 and S6) and skin conditions (Tables S3, S5, S7), are attributable to the performance of our AI model (Table S4) and its intuitive interface. Although the literature has extensively explored AI models for dermoscopic imaging or full‐body mapping techniques, research on newer non‐invasive imaging methods remains sparse. ²⁷ , ²⁸ , ²⁹ , ³⁰ Campanella et al. ³⁰ introduced a deep learning approach for detecting BCC using RCM, but their model was trained on a significantly smaller dataset (312 stacks from 66 lesions). They reported an AUC of 86.1% versus 98.6% in our research, without any quantification of the potential support offered to dermatologists. To our knowledge, our study represents the most comprehensive examination to date of an AI tool for BCC detection employing non‐invasive microscopic imaging.

Our research highlights the benefit of AI assistance for all user levels, with novices and intermediates improving the most on a large range of pathologies (Table S3). It also boosts dermatologists' confidence without slowing the diagnostic process. The AI assistant enables both novice (91.4% accuracy CI: 0.903–0.924) and intermediate (92.1% accuracy CI: 0.910–0.930) practitioners to attain diagnostic accuracy on par with that of unaided experts (92.1% accuracy CI: 0.911–0.930), effectively bridging a two‐year experience gap in LC‐OCT practice (Figure 3 and Figure S3). Such findings strongly suggest that AI support can facilitate the adoption of advanced modalities like LC‐OCT by dermatologists at all levels of expertise.

The AI model shows promise for automating complex and time‐consuming tasks like surgical planning for margins of BCCs, ³¹ which could enable dermatologists and surgeons to rely on AI‐generated insights for confirming and minimizing surgical margins before the procedure, potentially reducing the frequency of re‐excisions caused by overlooked margins. This concrete application ³² represents a particularly valuable advancement, as it automates a procedure that would otherwise be prohibitively time‐consuming. AI‐assisted acquisition tasks could be delegated to medical assistants; nevertheless, the current model is limited to BCC diagnosis and is not a comprehensive malignancy detector. Its differential diagnosis categories provide insights into excluding BCC but are intentionally broad to avoid misinterpretation of the model's scope, which is strictly limited to BCC.

The retrospective nature of this study introduces certain limitations to the quantitative outcomes. The performance metrics derived from the analysis of C&D imaging do not accurately reflect routine clinical performance, since the selected lesions were intentionally challenging, and an image can hardly replace physical examination. However, a recent meta‐analysis showed that it does not significantly affect dermatologists' performance when diagnosing BCC. ³³ It should also be noted that image‐based diagnosis is increasingly used with the rise of teledermatology. Additionally, AI's utility may be underappreciated in this retrospective framework. In a real‐world setting, live AI support could help capture better images for thorough assessment before diagnosis. It is also important to note that for most reviewers, this study represented their first interaction with the AI assistant. With increased familiarity with the AI system, all users should demonstrate accuracy beyond the model's capacities. Further prospective studies will be required to confirm the clinical performance and assess the impact of integrating our AI assistant into the clinical pathway.

Enhancements to the model, alongside its expansion to cover additional diseases or specific subtypes of BCCs, could be achieved through the acquisition of more labelled data and further training. Such advancements not only have the potential to enhance the precision of this initial model but also lay the groundwork for the development of other models with broader scopes of applicability in dermatology, like melanoma, AK ³⁴ or inflammatory conditions.

AUTHOR CONTRIBUTIONS

Clara Tavernier, Linda Tognetti, Josep Malvehy and Sébastien Fischman designed the study. Steven Challe implemented the web application. All members of the LC‐OCT reviewers consortium participated in the quiz. Théo Viel and Sébastien Fischman implemented and trained the AI model. Sébastien Fischman made the statistical analyses and wrote the manuscript. All authors discussed the results and commented on the manuscript.

FUNDING INFORMATION

None.

CONFLICT OF INTEREST STATEMENT

None.

ETHICAL APPROVAL

Approved by the Hospital Clinic Barcelona's Ethics Committee (No. HCB/2023/0513).

ETHICS STATEMENT

Not applicable.

Supporting information

Data S1:

Fischman S, Viel T, Perrot J‐L, Pérez‐Anker J, Suppa M, Cinotti E, et al. AI‐assisted basal cell carcinoma diagnosis with LC‐OCT: A multicentric retrospective study. J Eur Acad Dermatol Venereol. 2026;40:1059–1068. 10.1111/jdv.70099

DATA AVAILABILITY STATEMENT

All data related to the study's answers and the code needed to compute the results of this study can be accessed in our GitHub repository: https://github.com/Optimox/code‐analysis‐bcc‐classification‐results/.