Abstract

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. To develop and validate a robust algorithm to measure thalamic volume using clinical routine T2-FLAIR MRI. 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. 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). 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.