A Novel Face-based Approach for the Early Diagnosis of Parkinson's Disease

Parkinson's disease (PD) is a chronic neurological disorder commonly seen in the elder population, and it severely impacts the lives of patients, their families and caregivers. The early PD diagnosis is essential to alleviate its symptoms and delay its progressions. In recent years, the PD diagnosis based on facial expressions begins to receive increasing research attentions, but contemporary related studies often suffer from the problem of incomplete inclusions of all 6 basic facial expressions and identity's factors. As a result, inconsistencies and ambiguities often exist among contemporary related studies, and their diagnosis performances are far from satisfaction. In this study, a novel PD diagnosis method based on synthesized identity-aware facial expression images is proposed to solve the above problems. First, the identity's factor is taken into consideration and all 6 basic facial expression images are synthesized to reflect "non-PD scenario"of PD patients, for the first time in the PD diagnosis field. Then, latent features are learned and automatically extracted from real / synthesized images of both PD and non-PD patients. Finally, a new triplet loss-based metric learning network is constructed to differentiate PD and non-PD patients. For experimental evaluations, a new facial expression image dataset composed of 95 PD patients is constructed in this study. The new dataset has been associated with three other public facial expression image datasets with non-PD patients. A number of popular or state-of-the-art methods in related studies have been compared with the new approach based on these datasets. The experimental results demonstrated the superiority of our method.