Accurate visual diagnosis of pathology-proved subungual melanoma among longitudinal melanonychia by deep learning surpasses that by board certified dermatologists

Summary: Abstract Body: Subungual melanoma remains a significant issue and poses an important diagnostic dilemma to making biopsy or observation. Deep learning has revolutionized to assist in visual diagnosis for several skin diseases, including melanoma. However, less is known whether deep learning could assist in the clinical diagnosis for the longitudinal melanonychia for subungual melanoma. This study aims to develop an effective method to achieve high quality of image recognition, assisting the diagnosis for subungual melanoma. A dataset including 260 nail images of melanonychia with pathology-proved diagnoses was adopted. The data was trained by validated by ensemble deep learning (EDL) with fusing a set of convolutional neural networks (CNNs), coupling with data augmentation and transfer learning to enhance the recognition. A combination of EDL with CNNs were compared for the accuracy. The results show that among the CNNs, the accuracy for a single CNN model, ResNet50, reached an accuracy of 87.31%. Moreover, the accuracy for EDL reached 91.15%, which is superior to that of the existing model by 7.3%, and that of board-certified dermatologists by 21.06 %. Although limitation applies to the single source from a referral hospital, this study developed a good model by EDL in increase the accuracy for detecting subungual melanoma. The accuracy of this model, which is superior to the currently existing model and the dermatologists, could be used to assist to help the clinical dilemma in dealing with longitudinal melanonychia. Chih-Hung Lee^{1, 2}, Ja-Hwung Su³, Kwei-Lan Liu^{1, 2} 1. Dermatology, Kaohsiung Chang Gung Memorial Hospital, Kaohsiung, Taiwan. 2. Dermatology, Chang Gung University College of Medicine, Taoyuan, Taiwan. 3. Computer Science and Information Engineering, National University of Kaohsiung, Kaohsiung, Taiwan. Bioinformatics, Computational Biology, and Imaging