A clinical study of two optical coherence tomography scanners – how resolution and depth affect skin cancer diagnostic accuracy classified by deep neural networks and foundation models

Early and accurate detection of skin cancer is essential to ensure effective treatment of patients. For this purpose, non-invasive bedside imaging technologies such as optical coherence tomography (OCT) are emerging in dermatology. Like ultrasound imaging, OCT generates cross-sectional in-vivo images of the skin, but with much higher resolution at the expense of imaging depth being in the range of 0.3 to 2 mm. This study investigates the impact of bedside OCT image resolution when identifying skin cancers, especially melanoma, in suspicious pigmented skin lesions. Thirty-one patients with pigmented skin lesions were imaged with a commercial high resolution OCT scanner (C-OCT) and a prototype ultrahigh resolution OCT scanner (UHR-OCT), each generating volumes of 120 and 1022 B-scans, respectively. We use the two OCT scanners to differentiate volumes of benign lesions from malignant lesions aided by deep learning (DL) methodologies. Concretely, we compare two feature extraction methods, convolutional neural networks (CNN) and foundation models, followed by a long-short term memory (LSTM) networks that capture the relation between the B-scan of a volume. With the CNN approach, we obtained 94\% accuracy, with C-OCT showing 92\% sensitivity and 95\% specificity, and UHR-OCT showing 58\% sensitivity, 89\% specificity, and 77\% accuracy, suggesting that higher resolution OCT does not necessarily provide more predictive ability for skin cancer. However, potential limitations of these outcomes are also discussed. Conclusively, this approach using DL models comparing UHR-OCT and C-OCT images, aimed to identify the optimal imaging modality for precise detection of skin cancer, and to determine the best DL approach for the automatic detection of skin cancer. This model can be used for comparing imaging methods in future diagnostic accuracy studies. Additionally, a comprehensive review of AI models used for skin cancer detection in OCT images is provided to address a gap identified in the existing literature.