A 3D Deep Residual Convolutional Neural Network for Differential Diagnosis of Parkinsonian Syndromes on ¹⁸F-FDG PET Images

Purpose: Idiopathic Parkinson’s disease (IPD) and atypical parkinsonian syndromes (APS) have similar symptoms at early disease stages, making difficult their differential diagnosis. Positron emission tomography (PET) with 18F-FDG and machine learning can discriminate metabolic patterns for the differential diagnosis of parkinsonism [1]. We recently introduced deep learning to differentiate the syndromes based on projected 2D images of the 18F-FDG PET volumes [2], which may miss some characteristic features due to dimension reduction. Therefore, we aimed to develop an automated diagnosis framework operating directly on 3D image volumes based on a sufficient database. Furthermore, we depicted in saliency maps the decision mechanism of the deep learning method. Material & methods: 920 patients with evident parkinsonian features underwent 18F-FDG PET imaging with tentative diagnosis and follow-up by movement disorders specialists, yielding the diagnosis of IPD (n=502) and the APS multiple system atrophy (MSA, n=239) and progressive supranuclear palsy (PSP, n=179). We developed a 3D deep residual convolutional neural network comprising a total of 18 layers. The residual connections included in this network were helpful for simplifying its optimization. We generated saliency maps using the guided back-propagation method [3]. Five-fold cross validation was applied to evaluate the proposed network, which was implemented in the TensorFlow platform and accelerated by an NVIDIA Titan-XP GPU.

Results: After pretraining in the 378 patients with tentative diagnoses, we cross-validated in the 542 patients with definite diagnoses. The proposed framework achieved 99.1% sensitivity, 94.5% specificity, 98.6% PPV and 96.0% NPV for the classification of IPD, versus 98.9, 91.5, 96.6, and 95.8% for the classification of MSA, and 91.1, 96.8, 96.9, and 94.9% for the classification of PSP respectively. The figure shows fused saliency maps of groups of (N=5) subjects with IPD, MSA, or PSA, showing the main saliency in (A) the right prefrontal cortex, (B) the bilateral thalamus and putamen, and (C) the midbrain and left lingual gyrus.

Conclusions: We successfully developed a 3D deep residual convolutional neural network for automated differential diagnosis of IPD and atypical parkinsonism with excellent performance in diagnostic accuracy. The greatest salience was in expected regions of the basal ganglia, but initial findings also implicate high order visual cortex and prefrontal cortex. The method is currently under detailed assessment in a separate group of several hundred parkinsonian patients, and with emphasis on the interpretation of the saliency maps in diagnosis. [1] Tang et al. Lancet Neuro 2010 [2] Wu et al. SNMMI 2018 [3] Springenberg, J.T., arXiv:1412.6806

• Download figure
• Open in new tab
• Download powerpoint