The expert-level distinction of systemic sclerosis from hand photographs using the deep convolutional neural network

Published: 19-01-2021| Version 2 | DOI: 10.17632/b3h99pwybf.2
Yuta Norimatsu,
Ayumi Yoshizaki,
Yoshinori Kabeya,
Takemichi Fukasawa,
Jun Omatsu,
Maiko Fukayama,
Ai Kuzumi,
Satoshi Ebata,
Asako Yoshizaki-Ogawa,
Yoshihide Asano,
Haruka Ichimura,
Sho Yonezawa,
Hiroki Nakano,
Shinichi Sato


Supplement Table 1. The diseases types of patients included in this study. Supplement Figure 1. Photography equipment and hand photographs. Photographs of the patient's hands were taken using the imaging equipment set up in the room dedicated to photography, with a constant light source, camera settings, and distance between the hand and camera (A). We took four hand photographs: palms of right hands (B), palms of left hands (C), dorsum of right hands (D), and dorsum of left hands (E) by each patient. Photographs of representative SSc patients were shown. Supplement Figure 2. The architecture of the CNN. The architecture of the CNN developed in this study was shown. The CNN was built in an Nvidia V100 environment, using Tensorflow 1.13.2 and Keras 2.2.4 as frameworks. Each of palm and dorsal of hand photographs obtained from the patients was randomly and evenly divided into four groups. Three of these groups were used to build and train the CNN. The remaining one group was used to compare the SSc distinction performance of SSc specialists with the CNN after training. SSc patient's photos LSc patient's photos SLE patient's photos DM patient's photos Vasculitis patient's photos Early SSc patient's photos Healthy Control's photos