Machine-learning classification of non-melanoma skin cancers from image features obtained by optical coherence tomography

Background/purpose: A number of publications have suggested that optical coherence tomography (OCT) has the potential for non-invasive diagnosis of skin cancer. Currently, individual diagnostic features do not appear sufficiently discriminatory. The combined use of several features may however be useful. Methods: OCT is based on infrared light, photonics and fibre optics. The system used has an axial resolution of 10 mu m, lateral 20 mu m. We investigated the combined use of several OCT features from basal cell carcinomas (BCC) and actinic keratosis (AK). We studied BCC (41) and AK (37) lesions in 34 consecutive patients. The diagnostic accuracy of the combined features was assessed using a machine-learning tool. Results: OCT images of normal skin typically exhibit a layered structure, not present in the lesions imaged. BCCs showed dark globules corresponding to basaloid islands and AKs showed white dots and streaks corresponding to hyperkeratosis. Differences in OCT morphology were not sufficient to differentiate BCC from AK by the naked eye. Machine-learning analysis suggests that when a multiplicity of features is used, correct classification accuracies of 73% (AK) and 81% (BCC) are achieved. Conclusion: The data extracted from individual OCT scans included both quantitative and qualitative measures, and at the current level of resolution, these single factors appear insufficient for diagnosis. Our approach suggests that it may be possible to extract diagnostic data from the overall architecture of the OCT images with a reasonable diagnostic accuracy when used in combination.