Automatic Thalamus Segmentation on Unenhanced 3D T1 Weighted Images: Comparison of Publicly Available Segmentation Methods in a Pediatric Population

The anatomical structure of the thalamus renders its segmentation on 3DT1 images harder due to its low tissue contrast, and not well-defined boundaries. We aimed to investigate the differences in the precision of publicly available segmentation techniques on 3DT1 images acquired at 1.5 T and 3 T machines compared to the thalamic manual segmentation in a pediatric population. Sixty-eight subjects were recruited between the ages of one and 18 years. Manual segmentation of the thalamus was done by three junior raters, and then corrected by an experienced rater. Automated segmentation was then performed with FSL Anat, FIRST, FreeSurfer, MRICloud, and volBrain. A mask of the intersections between the manual and automated segmentation was created for each algorithm to measure the degree of similitude (DICE) with the manual segmentation. The DICE score was shown to be highest using volBrain in all subjects (0.873 ± 0.036), as well as in the 1.5 T (0.871 ± 0.037), and the 3 T (0.875 ± 0.036) groups. FSL-Anat and FIRST came in second and third. MRICloud was shown to have the lowest DICE values. When comparing 1.5 T to 3 T groups, no significant differences were observed in all segmentation methods, except for FIRST (p = 0.038). Age was not a significant predictor of DICE in any of the measurements. When using automated segmentation, the best option in both field strengths would be the use of volBrain. This will achieve results closest to the manual segmentation while reducing the amount of time and computing power needed by researchers.