Deep-learning segmentation of the substantia nigra from multiparametric MRI: Application to Parkinson's disease.

Abstract

Loss of dopaminergic neurons in the substantia nigra (SN) pars compacta (SNc) is a pathological hallmark of Parkinson's disease (PD). This is accompanied by a reduction of the dopamine synthesis byproduct neuromelanin (NM), which can be detected in vivo with NM-sensitive MRI, showing potential as a biomarker of PD. This relies on delineating the NM-rich region, which is achieved by applying manual or automated methods. Currently, there is a lack of publicly available tools for this task, so we trained a deep neural network intended for publishing, while exploring the effects of incorporating multiparametric MRI for segmenting the NM hyperintensity of the SN. We obtained multiple MRI contrasts, including NM-sensitive magnetization transfer contrast from 109 individuals (87 PD, 22 healthy controls) comprising a Norwegian and a Canadian cohort. The method was further evaluated on 209 MRIs from the Parkinson's Progressive Markers Initiative (PPMI). We observed that models trained naively on images from a single site tended to perform very poorly when exposed to similar data from different sites, emphasizing the importance of validating on out-of-distribution data. By applying aggressive data augmentation, we could largely attenuate the problem. We also observed a small additional regularizing effect from training the neural network on multiparametric MRIs. Volume and contrast-to-noise ratio (CNR) of the SN hyperintensity to the crus cerebri were used to distinguish patients from controls, with an area under the receiver operating characteristic (AUROC) of 0.863. CNR was found to be a better marker of disease status than volume, and we discuss a potential confusion in discerning the two measures. No contralateral association was observed between the severity of motor symptoms and volume or CNR.