Ren Wei a , Chendan Jiang a , Jun Gao a , Ping Xu a , Debing Zhang b , Zhicheng Sun b , Xiaohai Liu a , Kan Deng a , Xinjie Bao a , Guoqiang Sun d , Yong Yao a , Lin Lu c , Huijuan Zhu c , Renzhi Wang a , Ming Feng a , *
19 July 2019
Background: Deep learning has the potential to assist the medical diagnostic process. We aimed to identify facial anomalies associated with endocrinal disorders using a deep-learning approach to facilitate the process of diagnosis and follow-up. Methods: We collected facial images of patients with hypercortisolism and acromegaly, and we augmented these images with additional negative samples from public databases. A model with a pretrained deep-learning network was constructed to automatically identify these hypersecretion statuses based on characteristic facial changes. We compared its performance to that of endocrine experts and further investigated key factors upon which the best performing model focused. Findings: The model achieved areas under the receiver operating characteristic curve of 0.9647 (Cushing’s syndrome) and 0.9556 (acromegaly), accuracies of 0.9593 (Cushing’s syndrome) and 0.9479 (acromegaly), and recalls of 0.7593 (Cushing’s syndrome) and 0.8089 (acromegaly). It performed better than any level of our endocrine experts. Furthermore, the regions of interest on the part of the machine were primarily the same as those upon which the humans focused. Interpretation: Our findings suggest that the deep-learning model learned the facial characters based merely on labeled data without learning prerequisite medical knowledge, and its performance was comparable with professional medical practitioners. The model has the potential to assist in the diagnosis and follow-up of these hypersecretion statuses.