Abstract
Important parts of the visual pathway occur in relatively small subcortical structures that are often difficult to identify and segment using standard structural scans in MRI (e.g. T1 and T2 scans). Studies of the Lateral Geniculate Nucleus (LGN) often use proton density (PD) scan protocols, repeated up to 40 times, then manually segment the LGN structure from the average image. Efficiency is crucial when conducting MRI scans: minimising time spent on structural scanning can increase time available for complementary functional MRI scans and/or reduce scanning costs. In this study we asked how segmentation accuracy depended on the number of PD repeats. Four raters segmented the LGN of five participants, using different numbers of PD scans in the average image (1, 2, 4, 8, 16, 24, 32, 40), and an additional experienced expert rater segmented the LGN for just the 40PD average for all participants. We compared how the rater LGN masks at each scan average level overlapped with the expert masks. One rater performed the segmentation for the 40PD average on four separate days, to measure intra-rater variability across repeats. We also used a state-of-the-art automated segmentation process to compare the reliability to manual segmentation. We found that the average overlap between rater masks and the expert masks increased up to the 16PD scan average level, after which there was no additional benefit to including more PD scans. The automated segmentation masks were comparable to the overlap between the raters (40PD) and expert masks.