Hippocampal Segmentation Performance on 7T MRI: Intensity-Based Accuracy Assessment with Paired 3T-7T Volume...

Clinical adoption of 7T MRI is increasing, yet the performance of commonly used hippocampal segmentation algorithms, none of which are trained on 7T data, remains largely uncharacterized. This study evaluates segmentation accuracy at 7T using a voxel intensity-based method and examines volumetric differences between paired 3T and 7T hippocampal segmentations.

Two retrospective data sets from a single center were analyzed. For the 7T-only accuracy assessment cohort, 269 brain MRI studies performed on a Siemens Magnetom Terra.X with paired pre- and postcontrast T1 MPRAGE sequences (0.6 mm isovoxel) were utilized. For the 3T-7T cross-field comparison cohort, 39 unique subjects were identified with both 3T and 7T precontrast T1 MPRAGE sequences. Hippocampal segmentation was performed on the 7T-only cohort with AssemblyNet, e2dhipseg, FastSurfer, HippMapper, hippodeep, and QuickNat. QuickNat was removed from the 3T-7T cohort due to poor performance at 7T, and NeuroQuant 5.0 was added. Voxel intensity-based correction metrics quantified segmentation accuracy at 7T, with lower total correction volumes indicating better performance. Paired 3T-7T volume differences were assessed using the Wilcoxon signed-rank test, and corresponding NeuroQuant normative percentiles were also compared.

At 7T, hippodeep achieved the lowest total correction volume (0.58 mL), followed by e2dhipseg (0.67 mL), FastSurfer (0.78 mL), HippMapper (0.84 mL), and AssemblyNet (0.89 mL). Welch ANOVA with Tukey post hoc testing confirmed significant pair-wise differences between all algorithms (P < .001). In the paired 3T-7T analysis, all algorithms yielded significantly smaller 7T volumes (P < .001), with mean absolute differences ranging from 0.19 mL (hippodeep) to 1.54 mL (HippMapper). NeuroQuant volumes differed by 0.54 mL, corresponding to a mean 41-point shift in normative percentiles.

hippodeep required the least total correction at 7T and had the smallest 3T-7T volume differences, suggesting it offers the most consistent cross-field performance among tested methods. However, consistent 7T volumetric underestimation across algorithms and the associated large normative percentile shifts from small volume changes indicate that a dedicated 7T normative database is necessary for meaningful clinical use.