Research On Data Generation And Quality Assessment For Improving In-vitro Diagnosis Of Chronic Neurodegenerative Diseases

Parkinson’s disease(PD)is a chronic neurodegenerative disease that mainly occurs among the elderly and severely reduces the life quality of patients and their families.Therefore,it is crucial to conduct early diagnosis of potential PD patients,which will enable doctors to take timely treatment measures to delay the exacerbation of the disease.In recent years,facial expression-based in-vitro PD diagnosis has attracted more and more attention because of its distinguishability(i.e.,PD patients always possess the symptom of facial expression disorder)and affordability.However,the performance of existing diagnostic methods is mainly constrained by: 1)the small scale of facial expression data used for training,and 2)the weak feature extraction ability of the model used for prediction.To address these problems,this paper proposes to improve the in-vitro diagnosis of PD based on facial expressions through data generation and quality assessment.Specifically,this paper focuses on the following three aspects:(1)Facial expression synthesis through generative adversarial networks.Aiming at the low generalization ability of the model due to limited training data in PD diagnosis,a deep generative adversarial network is utilized to learn the mapping among various basic expressions.As a consequence,virtual facial expression images of 6 basic emotions(i.e.,anger,disgust,surprise,happiness,sadness and fear)can be synthesized.They approximate the premorbid expressions of PD patients while preserving their identity information,which can be used as the control group data.The generated results are then incorporated to augment the training data for diagnostic model learning.(2)Quality assessment of synthesized expression images.In order to reduce the negative impact of low-quality synthesized data on the training of diagnostic models,three facial image quality assessment(FIQA)criteria are introduced to evaluate the quality of generated images.Based on the three specific FIQA criteria of facial symmetry,image sharpness and Face Qnet,a fusion screening strategy is proposed to shortlist high-quality generated expression images for model training,thereby realizing high-quality data augmentation.(3)PD prediction model learning.In order to achieve accurate in-vitro diagnosis of PD,the mixed data containing real-world patients’ expressions and high-quality generated ones is used as the training set,and the deep neural network Res Net-18 is adopted as the backbone of diagnostic model.The model is trained in a supervised manner to learn the latent mapping between facial expressions and PD.The trained model is able to extract high-level semantic features from the input expressions and perform PD prediction accordingly.In order to demonstrate the effectiveness and rationality of the method proposed in this paper,the research group of this paper cooperated with relevant hospital departments to establish a PD facial expression dataset containing multiple expressions.Experimental results on this dataset and other public facial expression datasets show that the in-vitro PD diagnostic method proposed in this paper achieves a diagnosis accuracy of 95.43%,which is superior to other comparative methods.Additionally,the results of ablation experiments also prove the effectiveness of the facial expression synthesis and the fusion screening strategy in our method.