Research On The Key Technology Of Auto Diagnosis Of Parkinson’s Disease Based On Mixed Emotional Facial Expressions

Parkinson’s disease(PD)is a typical neurodegenerative disease,which occurs in the elderly and is mainly caused by the degeneration of dopaminergic neurons in the substantia nigra of the brain,the symptoms include facial muscle rigidity,quiescent tremor,dementia and so on,which cause a lot of life pressure to patients and their families.At present,Parkinson’s disease can not be completely cured,so early diagnosis of Parkinson’s disease for potential patients to receive timely treatment to avoid exacerbation is very important.The studies have shown that PD patients always have emotional expression disorder,and their ability to recognize facial expressions deteriorates,thus forming the characteristics of “masked faces”.Therefore,we propose an auto Parkinson’s disease diagnosis method based on mixed emotional facial expressions in this paper,which mainly includes the following three works:(1)Face expression generation based on generative adversarial learning.In this study,our group collaborated with the Second Affiliated Hospital of Nanchang University to create PDFE,a facial expression dataset for in-vitro diagnosis of Parkinson’s disease.To address the problem of wide differences in age distribution between the PDFE dataset and other public datasets,we synthesized the dataset containing six facial expressions(anger,disgust,fear,happiness,sadness,surprise)by the Star GAN model.The synthesized dataset preserved the PD patients’ identities and exhibited normal facial expression features,and it was used in the Parkinson’s disease diagnosis experiments as a control group with the same identities as the patients in the PDFE dataset.By comparing the synthesis results with the classical domain transformation model,the excellent performance of the Star GAN model in the field of multi-domain transformation was validated.(2)High-quality synthesized images screening based on deep learning.In the study,we applied the Face Qnet model based on the end-to-end framework for deep learning face recognition to achieve image quality assessment,and an effective image quality screening scheme was designed to shortlist high-quality synthesized facial expression images according to the quality scores obtained.Through the qualitative and quantitative analysis of the experiments,the effectiveness of the image quality screening scheme in shortlisting high-quality facial expression images was validated.(3)In-vitro Parkinson’s disease diagnosis based on deep learning.A new expanded dataset was constructed by mixing the original facial expression images of the PD patients,the high-quality synthesized facial expression images of the PD patients,and the normal facial expression images from other public face datasets.We trained a deep feature extractor accompanied with a facial expression classifier based on the expanded dataset.Then the well-trained deep feature extractor was used to extract and fuse the expression features of six basic facial expression images of the same person,so as to conduct the PD/non-PD classification.This study validated the effectiveness of the proposed method in Parkinson’s disease diagnosis and facial expression recognition through extensive experiments.In addition,the ablation experiment was discussed from two aspects.On the one hand,the effects of different training datasets on the performance of normal facial expression synthesis in PD patients were compared.On the other hand,the effects of facial expression synthesis module and image quality screening module on PD diagnosis performance and facial expression recognition performance were studied,the validity of the proposed method and the necessity of each step were validated again.