Dermoscopic diagnostic performance of Japanese dermatologists for skin tumors differs by patient origin: A deep learning convolutional neural network closes the gap.

ABSTRACT

In the dermoscopic diagnosis of skin tumors, it remains unclear whether a deep neural network (DNN) trained with images from fair-skinned-predominant archives is helpful when applied for patients with darker skin. This study compared the performance of 30 Japanese dermatologists with that of a DNN for the dermoscopic diagnosis of International Skin Imaging Collaboration (ISIC) and Shinshu (Japanese only) datasets to classify malignant melanoma, melanocytic nevus, basal cell carcinoma and benign keratosis on the non-volar skin. The DNN was trained using 12 254 images from the ISIC set and 594 images from the Shinshu set. The sensitivity for malignancy prediction by the dermatologists was significantly higher for the Shinshu set than for the ISIC set (0.853 [95% confidence interval, 0.820-0.885] vs 0.608 [0.553-0.664], P < 0.001). The specificity of the DNN at the dermatologists' mean sensitivity value was 0.962 for the Shinshu set and 1.00 for the ISIC set and significantly higher than that for the human readers (both P < 0.001). The dermoscopic diagnostic performance of dermatologists for skin tumors tended to be less accurate for patients of non-local populations, particularly in relation to the dominant skin type. A DNN may help close this gap in the clinical setting.