Objective: The aim of this study was to develop a deep learning-based system for automatic nigrosome 1 segmentation on susceptibility-weighted imaging (SWI) from 3T magnetic resonance imaging (MRI).

Background: Nigral hyperintensity on 7T T2* MRI or 3T SWI MRI is presumed to be nigrosome 1. Its alteration has been described in Parkinson’s disease, multiple system atrophy and progressive supranuclear palsy. An intact or lost nigral hyperintensity sign are determined by visual inspection.

Method: A total of 118 patients with Parkinson’s disease were recruited, who underwent 3T brain MRI including SWI from 2017 to 2019. On their axial SWI scans, two independent board-certified neuro-radiologists manually placed region-of-interests on the bilateral nigrosome 1. We trained convolutional neural networks by combined U-Net and Squeeze-and-Excitation networks. The training and validation were performed for the labeling masks of the nigrosome 1 by two independent neuro-radiologists. Of 118 sets of SWI, we used 94 sets (80%) as a training set, and the 24 sets (20%) was utilized for a validation set.

Results: The mean voxel size for the mask was 44x36x8mm^3. The performance of dice similarity coefficient (DSC) was measured as an average of 0.650 (maximum, 0.842; minimum, 0.570) for the neuro-radiologist 1, and as an average of 0.615 (maximum 0.793; minimum, 0.514) for the neuro-radiologist 2.

Conclusion: We found a good performance of this deep learning-based automatic nigrosome 1 segmentation system. The architecture of nigrosome 1 used in this study can be applicable in the convolutional neural network-based segmentation method for the nigrosome 1. Further studies using more samples and various parkinsonism are needed to improve its clinical usefulness.

« Back to MDS Virtual Congress 2020

MDS Abstracts - https://www.mdsabstracts.org/abstract/deep-learning-based-automatic-nigrosome-1-segmentation-on-3t-susceptibility-weighted-mri/