Analytic validation of convolutional neural network-based classification of pigmented skin lesions (PSLs) using unseen PSL hyperspectral data for clinical applications

In this study, we aimed not only to analyze model performance of the convolutional neural network (CNN)-based pigmented skin lesion (PSL) classification, but also analyze the analytic validation of the CNN-based PSL classification using unseen PSL hyperspectral dataset with an FNR. To this end, 38 hyperspectral imaging (HSI) data samples were obtained from 19 patients diagnosed with PSLs based on biopsy results. The analytic validation dataset comprised both seen and unseen PSL datasets. The seen PSL dataset included 272,677 pixels from 32 HSI data samples, and the unseen PSL dataset included 370,820 pixels from 38 HSI data samples. A snapshot-based hyperspectral camera captured the spectral (2048 × 2048 pixels) and spatial (150 spectral bands, 470–900 nm) data. A dermatologist labeled the acquired HSI data as pigmented basal cell carcinoma (BCC), melanoma, and squamous cell carcinoma (SCC) to obtain hyperspectral data for each PSL class in software. A confusion matrix and specific performance metrics were used to evaluate CNN-based PSL classification performance. The false negative ratio (FNR) for melanoma were 0.0284 ± 0.0051 and 0.4317 ± 0.0269 for seen and unseen PSL dataset, respectively. Furthermore, 49.14% of the unseen SCC hyperspectral data was predicted as BCC. We confirmed unseen SCC hyperspectral data was most commonly confused for BCC. Therefore, we confirmed the feasibility of analytic validation of the CNN-based PSL classification using unseen PSL hyperspectral dataset for clinical applications.