Automation and standardization of subject-specific region-of-interest segmentation for investigation of diffusion imaging in clinical populations

• Published in: PloS one 2022-12, Vol.17 (12), p.e0268233-e0268233
• Main Authors: Azor, Adriana M, Sharp, David J, Jolly, Amy E, Bourke, Niall J, Hellyer, Peter J
• Format: Article
• Language: English
• Subjects: Abnormalities; Accuracy; Automation; Biology and Life Sciences; Diffusion; Diffusion Magnetic Resonance Imaging; Engineering and Technology; Evaluation; Forecasts and trends; Head injuries; Humans; Image processing; Image segmentation; Magnetic resonance; Magnetic resonance imaging; Mechanization; Medical imaging; Medicine and Health Sciences; Neuroimaging; Registration; Reproducibility; Reproducibility of Results; Research and Analysis Methods; Scanners; Standardization; Statistical analysis; Substantia alba; Traumatic brain injury
• ISSN: 1932-6203; 1932-6203
• DOI: 10.1371/journal.pone.0268233

Diffusion weighted imaging (DWI) is key in clinical neuroimaging studies. In recent years, DWI has undergone rapid evolution and increasing applications. Diffusion magnetic resonance imaging (dMRI) is widely used to analyse group-level differences in white matter (WM), but suffers from limitations that can be particularly impactful in clinical groups where 1) structural abnormalities may increase erroneous inter-subject registration and 2) subtle differences in WM microstructure between individuals can be missed. It also lacks standardization protocols for analyses at the subject level. Region of Interest (ROI) analyses in native diffusion space can help overcome these challenges, with manual segmentation still used as the gold standard. However, robust automated approaches for the analysis of ROI-extracted native diffusion characteristics are limited. Subject-Specific Diffusion Segmentation (SSDS) is an automated pipeline that uses pre-existing imaging analysis methods to carry out WM investigations in native diffusion space, while overcoming the need to interpolate diffusion images and using an intermediate T1 image to limit registration errors and guide segmentation. SSDS is validated in a cohort of healthy subjects scanned three times to derive test-retest reliability measures and compared to other methods, namely manual segmentation and tract-based spatial statistics as an example of group-level method. The performance of the pipeline is further tested in a clinical population of patients with traumatic brain injury and structural abnormalities. Mean FA values obtained from SSDS showed high test-retest and were similar to FA values estimated from the manual segmentation of the same ROIs (p-value > 0.1). The average dice similarity coefficients (DSCs) comparing results from SSDS and manual segmentations was 0.8 ± 0.1. Case studies of TBI patients showed robustness to the presence of significant structural abnormalities, indicating its potential clinical application in the identification and diagnosis of WM abnormalities. Further recommendation is given regarding the tracts used with SSDS.