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

Area of Science:

• Neuroimaging
• Magnetic Resonance Imaging (MRI)

Background:

• Clinical adoption of 7 Tesla (7T) MRI is growing.
• Existing hippocampal segmentation algorithms are not trained on 7T data, limiting their clinical utility.
• Performance of these algorithms at 7T is largely uncharacterized.

Purpose of the Study:

• Evaluate the accuracy of commonly used hippocampal segmentation algorithms at 7T MRI.
• Compare volumetric differences between paired 3T and 7T hippocampal segmentations.
• Assess the impact of 7T underestimation on normative percentile shifts.

Main Methods:

• Analyzed two retrospective datasets (269 studies for 7T accuracy, 39 subjects for 3T-7T comparison).
• Tested six algorithms (AssemblyNet, e2dhipseg, FastSurfer, HippMapper, hippodeep, QuickNat) at 7T, with QuickNat excluded from 3T-7T comparison (NeuroQuant 5.0 added).
• Quantified accuracy using voxel intensity-based correction metrics and assessed volume differences with Wilcoxon signed-rank test.

Main Results:

• Hippodeep demonstrated the lowest total correction volume (0.58 mL) at 7T, indicating superior accuracy.
• All tested algorithms produced significantly smaller hippocampal volumes at 7T compared to 3T (p < 0.001).
• Mean absolute volume differences ranged from 0.19 mL (hippodeep) to 1.54 mL (HippMapper), with NeuroQuant showing a 0.54 mL difference and a 41-point percentile shift.

Conclusions:

• Hippodeep exhibited the most consistent cross-field performance between 3T and 7T.
• Consistent volumetric underestimation at 7T across algorithms poses challenges for clinical interpretation.
• A dedicated 7T normative database is essential for accurate clinical application of hippocampal volumetry.