Diagnostic performances of skin cancers using intensity- and gradient-based features with optical coherence...

The gold standard for diagnosing skin cancer is the skin biopsy. However, as an invasive method, the skin biopsy might be an uncomfortable process for individuals. Optical coherence tomography (OCT) is a non-invasive method capable of providing cross-sectional images of skin tissue. The intensity information in OCT images may encode information related to skin tissues. Thus, the purpose of this study is to explore the diagnostic ability of intensity- and gradient-based parameters with OCT in the diagnosis of skin cancers.

This study involved 5 patients with amelanotic melanomas (AM), 8 patients with basal cell carcinoma (BCC), 4 subjects with pigment nevi, and 3 normal subjects. The cross-sectional images were collected using the customized spectral domain OCT (SD-OCT). For each subject, 5 OCT images were used. From the OCT images, 10 intensity- and gradient-based features, and 5 texture features were calculated. Analyses of variance and receiver operating characteristics were utilized to analyze the diagnostic performances of features between study groups.

The mean of gradients showed the highest accuracy in differentiating BCC from normal skin tissue [area under receiver operating characteristic (AUROC) =0.932, sensitivity =1.000, specificity =0.775]. The variance of gradient showed high diagnostic performance (AUROC =0.719, sensitivity =0.700, specificity =0.875) to differentiate BCC from nevi. The highest accuracy was achieved by using the skewness of gradient for distinguishing AM from nevi (AUROC =0.828, sensitivity =0.750, specificity =1.000). Moreover, in differentiating AM from BCC, the skewness and kurtosis of intensity and gradient demonstrated high diagnostic performances. The highest accuracy was achieved by using the skewness and kurtosis of intensity (AUROC =0.900, sensitivity =0.875, specificity =1.000).

Our findings have suggested that intensity- and gradient-based features could capture subtle interface irregularities and optical homogeneity for classifying skin cancers.