Efficacy of a Deep Learning Convolutional Neural Network System for Melanoma Diagnosis in a Hospital Population

Abstract

(1) Background: The purpose of this study was to evaluate the efficacy in terms of sensitivity, specificity, and accuracy of the quantusSKIN system, a new clinical tool based on deep learning, to distinguish between benign skin lesions and melanoma in a hospital population. (2) Methods: A retrospective study was performed using 232 dermoscopic images from the clinical database of the Ramón y Cajal University Hospital (Madrid, Spain). The skin lesions images, previously diagnosed as nevus (n = 177) or melanoma (n = 55), were analyzed by the quantusSKIN system, which offers a probabilistic percentage (diagnostic threshold) for melanoma diagnosis. The optimum diagnostic threshold, sensitivity, specificity, and accuracy of the quantusSKIN system to diagnose melanoma were quantified. (3) Results: The mean diagnostic threshold was statistically lower (p < 0.001) in the nevus group (27.12 ± 35.44%) compared with the melanoma group (72.50 ± 34.03%). The area under the ROC curve was 0.813. For a diagnostic threshold of 67.33%, a sensitivity of 0.691, a specificity of 0.802, and an accuracy of 0.776 were obtained. (4) Conclusions: The quantusSKIN system is proposed as a useful screening tool for melanoma detection to be incorporated in primary health care systems.