Beyond the Surface: AI‐Assisted Basal Cell Carcinoma Detection via Optical Coherence Tomography
Aneri B. Patel, Olamide Sonuga Finney, Nicole M. Golbari, Jon Holmes, Kristen M. Kelly, Christopher ZacharyABSTRACT
Objectives
To determine whether integrating an artificial intelligence (AI) decision‐support system (VivoAID) into optical coherence tomography (OCT) interpretation improves diagnostic accuracy and efficiency for novice users evaluating basal cell carcinomas (BCCs).
Methods
Two dermatology research fellows completed a structured OCT training module for BCC identification using the VivoSight Dx system. Each independently reviewed 45 de‐identified OCT scans (31 BCC‐positive and 14 BCC‐negative, negative ones included squamous cell carcinomas, keratoacanthomas, actinic keratoses and scar tissue), with ground truth established by expert clinical interpretation and histopathology, in two sessions: one unassisted and one AI‐assisted. Session order was counterbalanced between learners, and scans were re‐randomized for AI sessions. In unassisted sessions, learners manually classified scans; in AI‐assisted sessions, they reviewed AI‐annotated scans, recorded agreement/disagreement, and noted additional findings. Time‐to‐decision per scan was measured. A linear mixed‐effects model in SAS assessed the effect of AI on decision time while accounting for scan difficulty.
Results
Without AI, correct BCC identification was 62% for Learner 1% and 64% for Learner 2 (combined 63%). With AI assistance, accuracy increased to 78% and 76%, respectively (combined 77%), reflecting relative improvements of 25% and 17% (combined +21%). Mean decision time decreased from 45 to 28 s for Learner 1 and from 35 to 22 s for Learner 2, yielding a combined reduction from 40 to 25 s (~38% faster). AI use significantly reduced decision time by an average of 15.1 s per image ( p < 0.0001). AI standalone accuracy versus expert opinion was 64%, comparable to novice baseline performance.
Conclusions
AI‐enhanced OCT with VivoAID significantly improved novice users’ diagnostic accuracy and efficiency, despite AI performance alone being similar to baseline novice interpretation. These data support VivoAID's role as a decision‐support and potential educational tool in dermatology. Findings are limited by small sample size, single‐institution design, and repeated review of the same scans.
Trial Registration
None.