Deep learning to differentiate parkinsonian disorders separately using single midsagittal MR imaging: a proof of concept study

Objectives To evaluate the diagnostic performance of deep learning with the convolutional neural networks (CNN) to distinguish each representative parkinsonian disorder using MRI. Methods This clinical retrospective study was approved by the institutional review board, and the requirement for written informed consent was waived. Midsagittal T1-weighted MRI of a total of 419 subjects (125 Parkinson’s disease (PD), 98 progressive supranuclear palsy (PSP), and 54 multiple system atrophy with predominant parkinsonian features (MSA-P) patients, and 142 normal subjects) between January 2012 and April 2016 was retrospectively assessed. To deal with the overfitting problem of deep learning, all subjects were randomly divided into training (85%) and validation (15%) data sets with the same proportions of each disease and normal subjects. We trained the CNN to distinguish each parkinsonian disorder using single midsagittal T1-weighted MRI with a training group to minimize the differences between predicted output probabilities and the clinical diagnoses; then, we adopted the trained CNN to the validation data set. Subjects were classified into each parkinsonian disorder or normal condition according to the final diagnosis of the trained CNN, and we assessed the diagnostic performance of the CNN. Results The accuracies of diagnostic performances regarding PD, PSP, MSA-P, and normal subjects were 96.8, 93.7, 95.2, and 98.4%, respectively. The areas under the receiver operating characteristic curves for distinguishing each condition from others (PD, PSP, MSA-P, and normal subjects) were 0.995, 0.982, 0.990, and 1.000, respectively. Conclusion Deep learning with CNN enables highly accurate discrimination of parkinsonian disorders using MRI. Key Points • Deep learning convolution neural network achieves differential diagnosis of PD , PSP , MSA-P , and normal controls with an accuracy of 96.8 , 93.7 , 95.2 , and 98.4% , respectively . • The areas under the curves for distinguishing between PD , PSP , MSA-P , and normality were 0.995 , 0.982 , 0.990 , and 1.000 , respectively . • CNN may learn important features that humans not notice , and has a possibility to perform previously impossible diagnoses .