Mengyu Li, Magnús Magnússon, Áskell Löve, Steinunn Þórðardóttir, Lotta M. Ellingsen
Introduction: Parkinson-plus syndromes (PPS) are a group of neurodegenerative diseases that are similar to Parkinson´s disease (PD) but with additional characteristics. Early diagnosis of different types of PPS is a major clinical challenge. Magnetic resonance imaging (MRI) plays an important role in the differential diagnosis of PPS from PD and from each other. Recently, fully automated brain segmentation from MRI using deep convolution neural networks (CNNs) has emerged as a major area of interest within the field of medical imaging analysis.
Methods and data: Here we present a patch-based, three-dimensional (3D) CNN brain segmentation algorithm specific for PPS imaging biomarkers. The method provides a robust and fast segmentation of 12 brain regions that have been associated with PPS, including the brainstem substructures, the four ventricular compartments, the putamen, and the caudate. For the training and prediction processes, each 3D image was divided into 125 overlapping 3D patches to extract important features and fine structural details. The CNN was trained on 123 multi-site MRIs and tested on manually labeled data.
Results: We computed the Dice similarity coefficient (DSC) to quantitatively assess segmentation accuracy. The mean DSC evaluated on 29 test images was DSC=0.92±0.03, indicating high segmentation accuracy.
Conclusions: We present a fully automated and robust MRI segmentation method that speeds up the segmentation process from hours to seconds compared with other state-of-the-art methods. We hope that this tool will contribute to the important clinical challenge of early diagnosis and differentiation of patients with PPS and related neurodegenerative diseases.