A deep learning framework for comprehensive segmentation of deep grey nuclei.

Background: Deep grey matter structures such as the thalamus and basal nuclei are implicated in numerous neurological disorders, yet accurate segmentation of these structures from standard T1-weighted MRI remains challenging due to poor intra-subcortical contrast, long preprocessing pipelines, and fragmented toolsets. Methods: We introduce THOMASINA a deep learning pipeline for comprehensive subcortical segmentation from standard T1-weighted (T1w) as well as white-matter-nulled (WMn) MRI. The method leverages labels derived from a recently published state-of-the-art multi-atlas segmentation method to train multiple 3D deep learning-based segmentation models including SwinUNETR, DiNTS, and SegResNet. All networks were trained on cropped volumes and tested on held-out and out-of-distribution datasets. For T1-weighted MRI, an additional synthesis step was used to generate WMn-like contrast prior to segmentation. Results: SegResNet achieved the best performance (mean Dice = 0.89 on with in-domain test data, 0.85 on out-of-domain test data), outperforming DiNTS and SwinUNETR in both accuracy and robustness. It also had the highest mean, median, and minimum Dice and lowest SD in most nuclei compared to the DiNTS and SwinUNETR. Synthetic WMn contrast provided comparable segmentation to actual WMn images. The proposed networks reduced per-subject segmentation time to the order of seconds versus tens of minutes using traditional multi-atlas segmentation. THOMASINA also generalized well across field strengths, scanner vendors, and disease cohorts. Conclusions: THOMASINA offers a fast, reproducible, and scalable solution for comprehensive subcortical segmentation from standard T1w MRI. By combining synthetic WMn contrast with state-of-the-art deep learning-based segmentation models, our method addresses key barriers to deployment and sets a foundation for biomarker discovery in clinical and population-scale imaging studies.