Applying artificial intelligence to optical coherence tomography for detection and subtyping of basal cell carcinoma

BACKGROUND AND OBJECTIVES: Basal cell carcinoma (BCC) is the most common cancer and a major public health burden. Noninvasive techniques such as optical coherence tomography (OCT) may reduce unnecessary biopsies and enable earlier detection. However, effective OCT interpretation requires expertise. Artificial intelligence (AI) may help overcome this barrier. PATIENTS AND METHODS: An AI model based on a U-Net convolutional neural network was trained using 4,680 annotated OCT image frames from 39 BCCs, 5,640 frames from 47 premalignant and benign lesions, and 12,360 frames from 103 scans of healthy skin. Detection performance was evaluated on an independent test set comprising 26 scans (3,120 frames) from 16 BCCs, 5 premalignant lesions, and 5 benign lesions. External validation used independent clinical data from 25 histologically confirmed BCCs (18 nodular, 7 superficial). RESULTS: AI-assisted OCT generated visual overlays corresponding to established BCC features and identified tumor areas. In the independent test set, the area under the ROC curve was 95.5 % for nodular and 98.2 % for superficial BCC. In clinical validation, all BCCs were detected, and subtype concordance with histology was achieved in 21 of 25 cases (83 %). CONCLUSIONS: AI-assisted OCT shows potential to improve non-invasive BCC diagnostics and warrants further clinical evaluation.