Reflectance Confocal Microscopy Versus Line‐Field Confocal Optical Coherence Tomography in Facial and Scalp Lesions Suspicious for Basal Cell Carcinoma: A Histopathology‐Referenced Multi‐Reader Study
Luca Ambrosio, Emanuele Rovaldi, Claudio Conforti, Carmen Cantisani, Giovanni Di Lella, Flavia Persechino, Giorgio Marotta, Marco Virone, Camilla Chello, Steven Paul Nisticò, Giovanni PellacaniABSTRACT
Background
Non‐invasive imaging techniques can support assessment of lesions suspicious for basal cell carcinoma (BCC) when dermoscopy is inconclusive. We compared reflectance confocal microscopy (RCM), line‐field confocal optical coherence tomography (LC‐OCT), and the device‐integrated LC‐OCT artificial intelligence (AI) output.
Methods
This retrospective comparative observational study included 100 facial and scalp lesions evaluated by RCM and LC‐OCT before excisional biopsy. Two expert and two non‐expert dermatologists independently interpreted anonymized images while blinded to histopathology and recorded confidence on a four‐point scale. Diagnostic performance, inter‐reader agreement, reader‐specific receiver operating characteristic analyses, and an exploratory post hoc dermoscopically equivocal subgroup were assessed. The categorical AI output was analyzed at a single operating point.
Results
Histopathology identified 69 basal cell carcinomas (BCCs) and 31 non‐BCC lesions. Expert mean sensitivity, specificity, and accuracy were 95.7%, 85.5%, and 92.5% for RCM versus 86.2%, 67.7%, and 80.5% for LC‐OCT. The AI output yielded 87.0% sensitivity, 83.9% specificity, and 86.0% accuracy. In the equivocal subgroup ( n = 51; 31 BCCs), RCM sensitivity was 96.8% across readers, with specificity ranging from 65.0% to 85.0%. RCM AUCs exceeded LC‐OCT AUCs for all readers, but no paired comparison remained significant after Holm adjustment. Inter‐reader agreement was higher for RCM than LC‐OCT (Fleiss' κ = 0.809 vs. 0.563).
Conclusions
Both modalities provided diagnostic information as second‐level adjuncts. RCM showed higher observed performance and agreement, particularly in exploratory equivocal lesions, whereas the categorical LC‐OCT AI output showed comparatively high specificity. Prospective validation is required.