基于深度学习的帕金森病诊断

Parkinson’s disease diagnosis based on deep learning

为实现对帕金森病进行自动诊断,对基于Spiral和Meander手绘图的CNN和CapsNet的PD诊断方案进行研究。图像预处理过程保留了色彩信息,模型可以学习笔划压力、速度等特征。实验结果表明,CapsNet和CNN分别是基于Spiral和Meander的最优诊断方案。研究结果表明,Spiral具备比Meander更多的差异来诊断PD,CapsNet是基于Spiral的最佳方案;保留色彩信息的图像更有利于基于手绘图诊断PD,所实现结果(Acc=95.7%)优于先前最优报道(Acc=82.7%)。

To achieve automatic diagnosis of PD,the PD diagnosis scheme of CNN and CapsNet based on Spiral and Meander hand-drawings was studied.The image preprocessing process retained color information so that the model could learn stroke pressure,speed,etc.The results show that the CapsNet and CNN are the optimal diagnosis schemes based on Spiral and Meander,respectively.The research shows two conclusions,the Spiral has more differences than the Meander to diagnose PD,and the CapsNet is the optimal solution based on Spiral.The image that retains the color information is more conducive to the PD diagnosis based on hand-drawing,and the achieved result(Acc=95.7%)is better than that of the previous best report(Acc=82.7%).