Related Experiment Video
Updated: Nov 8, 2025
Automated Segmentation of Cortical Grey Matter from T1-Weighted MRI Images
Published on: January 7, 2019
DeepGRAI (Deep Gray Rating via Artificial Intelligence): Fast, feasible, and clinically relevant thalamic atrophy
Michael Dwyer1, Cassondra Lyman2, Hannah Ferrari2
1Buffalo Neuroimaging Analysis Center, Department of Neurology, School of Medicine and Biomedical Sciences, University at Buffalo, State University of New York, Buffalo, NY, USA; Jacobs MS Center, Department of Neurology, Jacobs School of Medicine and Biomedical Sciences, University at Buffalo, State University of New York, Buffalo, NY, USA.
Background:
Thalamic volume loss is a key marker of neurodegeneration in multiple sclerosis (MS). T2-FLAIR MRI is a common denominator in clinical routine MS imaging, but current methods for thalamic volumetry are not applicable to it.
Objective:
To develop and validate a robust algorithm to measure thalamic volume using clinical routine T2-FLAIR MRI.
Methods:
A dual-stage deep learning approach based on 3D U-net (DeepGRAI - Deep Gray Rating via Artificial Intelligence) was created and trained/validated/tested on 4,590 MRI exams (4288 2D-FLAIR, 302 3D-FLAIR) from 59 centers (80/10/10 train/validation/test split). As training/test targets, FIRST was used to generate thalamic masks from 3D T1 images. Masks were reviewed, corrected, and aligned into T2-FLAIR space. Additional validation was performed to assess inter-scanner reliability (177 subjects at 1.5 T and 3 T within one week) and scan-rescan-reliability (5 subjects scanned, repositioned, and then re-scanned). A longitudinal dataset including assessment of disability and cognition was used to evaluate the predictive value of the approach.
Results:
DeepGRAI automatically quantified thalamic volume in approximately 7 s per case, and has been made publicly available. Accuracy on T2-FLAIR relative to 3D T1 FIRST was 99.4% (r = 0.94, p < 0.001,TPR = 93.0%, FPR = 0.3%). Inter-scanner error was 3.21%. Scan-rescan error with repositioning was 0.43%. DeepGRAI-derived thalamic volume was associated with disability (r = -0.427,p < 0.001) and cognition (r = -0.537,p < 0.001), and was a significant predictor of longitudinal cognitive decline (R2 = 0.081, p = 0.024; comparatively, FIRST-derived volume was R2 = 0.080, p = 0.025).
Conclusions:
DeepGRAI provides fast, reliable, and clinically relevant thalamic volume measurement on multicenter clinical-quality T2-FLAIR images. This indicates potential for real-world thalamic volumetry, as well as quantification on legacy datasets without 3D T1 imaging.
