Deep learning based diagnosis of Parkinson’s Disease using diffusion magnetic resonance imaging

• Published in: Brain imaging and behavior Vol. 16; no. 4; pp. 1749 - 1760
• Main Authors: Zhao, Hengling, Tsai, Chih-Chien, Zhou, Mingyi, Liu, Yipeng, Chen, Yao-Liang, Huang, Fan, Lin, Yu-Chun, Wang, Jiun-Jie
• Format: Journal Article
• Language: English
• Published: New York Springer US 01.08.2022; Springer Nature B.V
• Subjects: Algorithms; Anisotropy; Artificial neural networks; Biomedical and Life Sciences; Biomedicine; Deep learning; Diffusion; Disease control; Greedy algorithms; Image acquisition; Machine learning; Magnetic resonance imaging; Mathematical analysis; Medical diagnosis; Medical imaging; Movement disorders; Neural networks; Neurodegenerative diseases; Neuroimaging; Neuropsychology; Neuroradiology; Neurosciences; Original Research; Parkinson's disease; Postcentral gyrus; Predictions; Psychiatry; Tensors; Differential diagnosis; Convolution neural network; Idiopathic Parkinson’s disease; Deep Learning; Diffusion tensor imaging
• ISSN: 1931-7557; 1931-7565; 1931-7565
• DOI: 10.1007/s11682-022-00631-y

The diagnostic performance of a combined architecture on Parkinson’s disease using diffusion tensor imaging was evaluated. A convolutional neural network was trained from multiple parcellated brain regions. A greedy algorithm was proposed to combine the models from individual regions into a complex one. Total 305 Parkinson’s disease patients (aged 59.9±9.7 years old) and 227 healthy control subjects (aged 61.0±7.4 years old) were enrolled from 3 retrospective studies. The participants were divided into training with ten-fold cross-validation (N = 432) and an independent blind dataset (N = 100). Diffusion-weighted images were acquired from a 3T scanner. Fractional anisotropy and mean diffusivity were calculated and was subsequently parcellated into 90 cerebral regions of interest based on the Automatic Anatomic Labeling template. A convolutional neural network was implemented which contained three convolutional blocks and a fully connected layer. Each convolutional block consisted of a convolutional layer, activation layer, and pooling layer. This model was trained for each individual region. A greedy algorithm was implemented to combine multiple regions as the final prediction. The greedy algorithm predicted the area under curve of 94.1±3.2% from the combination of fractional anisotropy from 22 regions. The model performance analysis showed that the combination of 9 regions is equivalent. The best area under curve was 74.7±5.4% from the right postcentral gyrus. The current study proposed an architecture of convolutional neural network and a greedy algorithm to combine from multiple regions. With diffusion tensor imaging, the algorithm showed the potential to distinguish patients with Parkinson’s disease from normal control with satisfactory performance.