Deep Learning on ^sup 18^F-FDG PET Imaging for Differential Diagnosis of Parkinsonian Syndromes

Objectives: Idiopathic Parkinson’s disease (IPD) and atypical parkinsonian syndromes may have similar symptoms at the early disease stage. Positron emission tomography (PET) with 18F-FDG was shown to be be able to assess early neuronal dysfunction of taupathies. In the past decades, machine learning on metabolic patterns has been established for the differential diagnosis of parkinsonism [1]. Recent advancement of deep learning has brought record-breaking performance in many applications. This study examined whether this artificial intelligence tool is of value in an early differential diagnosis of parkinsonism. Methods: 257 patients with clinical features suspected for parkinsonism were subjected to an 18F-FDG PET imaging. After the imaging, these patients were followed up at least 1 year by blinded movement disorders specialists before a final clinical diagnosis of IPD (n=136), multiple system atrophy (MSA, n=91), and progressive supranuclear palsy (PSP, n=30) were made. To enable the fast establishment of deep neural network with limited computational power, a tensor factorization method [2] was employed to compress 3D brain imaging data to 2D images (supplemental figure). Then, A 4-layer convolutional neural network (CNN) was established to differentiate IPD, MSA and PSP based on the tensor-factorized 2D images. To enhance the performance, the CNN network was pretrained with a large database of brain FDG PET images. To this end, 1077 subjects of 41 various neurological diseases were included. Five fold cross validation was applied for the training and test of the pretrained CNN on the 257 parkinsonian patients. The classification CNN was implemented using python on Google's TensorFlow platform and accelerated using NVIDIA Titan X graphics card. Results: After 85 iterations, all the folds for the test of the 2D CNN on tensor factorized images converges. The training took 3204 seconds. The test network has achieved 98.9% sensitivity, 90.0% specificity, 98.4% PPV and 95.0% NPV for the differentiation of PD, 98.8% sensitivity, 82.5% specificity, 96.1% PPV and 96.0% NPV for the classification of MSA and 87.1% sensitivity, 97.8% specificity, 96.1% PPV and 94.1% NPV for the classification of PSP. Conclusions: The preliminary test of deep learning on tensor factorized FDG PET images has demonstrated its ability to reach a cpmparable accuracy of state-of-the-art. Enlarging the database and further optimizing the deep neural network may enhance its potential in