Fully automated segmentation of substantia nigra toward longitudinal analysis of Parkinson's disease.

Abstract

Purpose: A fully automated segmentation of substantia nigra (SN) is an essential task for the development of an explainable computer-aided diagnosis system of Parkinson's disease (PD). Since anatomical alterations of SN are vital information in PD diagnosis, a precise segmentation model should have generalization ability against spatiotemporal changes. To satisfy these requirements, we propose a fully automated pipeline with several new techniques for a volumetric image obtained by neuromelanin magnetic resonance imaging.

Methods: We develop a pipeline by integrating SN-prior probability estimation into the decision of the SN-contained region of interest. An estimated SN-prior probability is further fed into a new priority attention mechanism as a gating signal in our segmentation model. Furthermore, we introduce test-time dropout to improve a segmentation model's accuracy and generalization ability. To evaluate the model's generalization ability, we collected principal and external datasets with longitudinal scans of the same PD patients.

Results: Our segmentation model achieved averaged Dice scores of 0.845 and 0.851 for SN hyperintense regions in the principal and external datasets, respectively. These results demonstrated the best generalization ability in our comparative evaluations. Thresholding the number of voxels in the SN hyperintense regions, we also evaluated the segmentation results in automated PD identification. The PD identification achieved the areas under the receiver operating characteristic curves of 0.755 and 0.726 by our pipeline's output and the ground truth, respectively.

Conclusions: The proposed pipeline, where we integrated SN-prior probability estimation, priority attention mechanism and test-time dropout to our segmentation model, achieved accurate SN segmentation with high generalization ability for our longitudinal data: the principal and external datasets. As demonstrated in the validation with the automated PD identification, our pipeline has the potential for improving the performance of PD diagnosis via further large-scale longitudinal analysis.