基于深度学习的帕金森疾病多病程诊断网络模型  

Deep Learning-Based Parkinson’s Disease Multi-Stage Diagnostic Systems

作　　者：陶国庆 陈曦 刘红怡 范丹丹 陈辉[1] Guoqing Tao;Da Chen;Hongyi Liu;Dandan Fan;Hui Chen(Shanghai Institute of Optics and Electronics,School of Optoelectronic Information and Computer Engineering,Shanghai University of Technology,Shanghai)

机构地区：[1]上海理工大学光电信息与计算机工程学院,上海

出　　处：《建模与仿真》2024年第3期3306-3321,共16页Modeling and Simulation

基　　金：国家自然科学基金项目(No.62275156)。

摘　　要：帕金森(Parkinson’s Disease,PD)是一种神经退行性疾病。当出现明显临床特征时,超过60%的黑质神经元已经发生不可逆的退化,早期疾病的诊断尤为关键,但疾病早期的诊断存在特异性不强、特征不明显等问题。因此,本文提出一种基于深度学习的帕金森疾病辅助诊断系统。首先,运用图像分割和生成对抗网络对原始图像进行分割和扩容;其次,引入多尺度卷积和注意力机制改进MobileNetV2网络,对PD、正常组以及特征不明显的前驱体的多病程诊断任务中,诊断准确率达到了92.4%,精确度达到91.7%,召回率达到92.4%,模型表现优于其他经典网络模型,且更聚焦帕金森病理学特征区域,具有更准确可靠的临床诊断效能。Parkinson’s Disease(PD)is a neurodegenerative disorder.When clinical features become evident,over 60%of the substantia nigra neurons have undergone irreversible degeneration.Early diagnosis of the disease is particularly crucial;however,challenges such as low specificity and inconspicuous features exist in the early detection of the disease.Therefore,this study proposes a deep learning-based Parkinson’s disease auxiliary diagnostic system.Firstly,image segmentation and generative adversarial networks are employed to segment and augment original images.Subsequently,multiscale convolution and attention mechanisms are introduced to enhance the MobileNetV2 network for multi-stage diagnosis tasks involving PD,normal groups,and indistinct precursor features.The diagnostic accuracy achieved 92.4%,precision reached 91.7%,and recall reached 92.4%.The model outperforms other classical network models and focuses more on the pathological features of Parkinson’s disease,demonstrating more accurate and reliable clinical diagnostic efficacy.

关 键 词：帕金森 医学图像 核磁共振图像 图像分割 图像增强 MobileNetV2 图像分类 

分 类 号：R74[医药卫生—神经病学与精神病学]