Deep-THOMAS: a robust deep learning network for fast and accurate thalamic nuclei segmentation

Motivation: Robust thalamic nuclei segmentation is essential for neuroimaging studies. Traditional methods are time-intensive and less precise, necessitating an efficient, deep-learning approach for clinical applications. Goal(s): Develop a robust 3D nnUNet model for fast accurate segmentation of thalamic nuclei from structural T1-weighted MRI. Approach: Utilize a 3D nnUNet architecture to capture fine anatomical details in thalamic nuclei. By training using data from different vendors and field strengths, robustness and generalizability is insured. Results: The model achieved Dice coefficients above 0.85 for major nuclei, with GPU inferences in ~3 seconds, demonstrating efficiency and robustness across patient and scanner variations. Impact: This robust thalamic nuclei segmentation tool can be integral for clinical neuroimaging, offering fast, reliable segmentation, opening the door for analysis of large public databases to study the role of thalamic nuclei in a variety of neurodegenerative and neuropsychiatric conditions.