🧬 biology

Relative contributions of reconstruction pipeline and MRI field strength to hippocampal shape variation

In a study of healthy adults, reconstruction methodology was found to have a greater influence on hippocampal shape variation than MRI field strength (3T vs. 7T), with FSL demonstrating the closest agreement to manual segmentation and supporting the validity of whole-hippocampus morphometric analyses using 3T MRI.

Original authors: Luis Miguel Echeverry-Quiceno, Katherine A. Koenig, Álvaro Heredia-Lidón, Xavier Sevillano, Neus Martínez-Abadías

Published 2026-09-24

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Luis Miguel Echeverry-Quiceno, Katherine A. Koenig, Álvaro Heredia-Lidón, Xavier Sevillano, Neus Martínez-Abadías

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ⚕️ This is an AI-generated explanation of a preprint that has not been peer-reviewed. It is not medical advice. Do not make health decisions based on this content. Read full disclaimer

The human brain is a landscape of intricate folds and deep valleys, and among its most critical features is the hippocampus, a seahorse-shaped structure tucked deep inside the temporal lobe. This small region acts as a vital hub for memory and spatial navigation, making it a primary focus for scientists studying everything from normal aging to Alzheimer's disease. To see inside the living brain without cutting it open, researchers rely on magnetic resonance imaging, or MRI. For years, the standard tool for this work has been the 3 Tesla scanner, a powerful machine that creates detailed pictures of brain tissue. More recently, even more powerful 7 Tesla scanners have become available, offering sharper images and the ability to see finer details. The hope has been that these ultra-high-field machines would provide a clearer, more consistent view of the brain's shape, allowing doctors and scientists to spot subtle changes that might signal disease earlier.

However, getting a picture of the brain is only the first step. Turning those raw images into a usable 3D model of a specific structure like the hippocampus requires a complex series of computer steps known as a reconstruction pipeline. Different software programs use different mathematical rules to trace the boundaries of the brain tissue, and these choices can change the final shape of the model. A key question has remained unanswered: does the superior image quality of the 7 Tesla scanner actually lead to more consistent 3D models, or do the choices made by the computer software matter more? If the software introduces more variation than the scanner itself, then the expensive upgrade to a 7 Tesla machine might not improve the reliability of the measurements as much as hoped.

A team of researchers set out to settle this question by putting both the scanners and the software to the test. They recruited ten healthy adults and scanned each person's brain twice: once on a standard 3 Tesla machine and once on a 7 Tesla machine. Because the same people were scanned on both, the researchers could isolate the differences caused by the machine from the natural differences between people. They then fed the images from both scans into four different reconstruction methods. Two of these were widely used automated software packages, one was a newer tool designed specifically to unfold the complex layers of the hippocampus, and the fourth was a manual tracing done by an expert by hand, which served as the gold standard for comparison. The goal was to see how much the shape of the hippocampus changed depending on which scanner was used and which software processed the image.

The results revealed a clear hierarchy of influence. The most significant factor shaping the 3D model was simply the individual person. Just as no two people have the exact same face, no two people have the exact same hippocampus shape, and this natural biological variation was the dominant force in the data. The second most important factor was the reconstruction pipeline. The choice of software introduced a noticeable amount of variation, meaning that the same brain could look slightly different depending on which program was used to build the model. Surprisingly, the strength of the magnetic field—the difference between the 3 Tesla and 7 Tesla scanners—had the smallest effect of all. While the 7 Tesla images were sharper, this extra clarity did not translate into a significantly more consistent shape for the whole hippocampus compared to the standard 3 Tesla images.

The study also found that the impact of the scanner depended on which software was used. The interaction between the machine and the method was significant, suggesting that some software programs handled the higher-quality 7 Tesla data better than others. When the researchers looked closely at the models, they found that one specific software package, known as FSL, produced the most stable results. The 3D shapes generated by FSL from the 3 Tesla scans were very similar to those from the 7 Tesla scans, and they also matched the expert hand-traced models more closely than the other automated methods did. In contrast, another popular software package produced models that varied more between the two scanner types.

These findings suggest that for studying the overall shape of the hippocampus, the standard 3 Tesla scanner remains a highly effective tool. The extra cost and complexity of moving to a 7 Tesla scanner may not be necessary for this specific type of measurement, provided that researchers stick to a consistent and reliable software pipeline. The study highlights that in the world of brain imaging, the method used to process the data can be just as important as the quality of the image itself. While the 7 Tesla scanner offers a clearer view, the way scientists interpret that view through their software choices ultimately determines the consistency of the final result. For now, the most reliable path forward appears to be standardizing the reconstruction methods, ensuring that the subtle differences in brain shape are due to biology rather than the tools used to measure them.

Technical Summary: Relative Contributions of Reconstruction Pipeline and MRI Field Strength to Hippocampal Shape Variation

Problem Statement
While ultra-high-field 7T MRI offers superior signal-to-noise ratios and spatial resolution compared to standard 3T MRI, the extent to which field strength versus reconstruction methodology influences whole-hippocampus shape representations remains unclear. Previous research indicates that segmentation algorithms and post-processing pipelines can introduce systematic variability in volumetric and morphometric measures. However, systematic evaluations of the robustness of 3D shape measures across different magnetic field strengths and reconstruction pipelines are limited. It is unknown whether the improved image quality of 7T MRI translates into more consistent morphometric representations or if methodological variability introduced by segmentation choices outweighs potential gains in image quality.

Methodology
This study employed a within-subject design involving ten neurologically healthy adults (8 males, 2 females; mean age 44.9 ± 9.9 years). Participants underwent whole-brain T1-weighted MRI scans on both a 3T Siemens Prisma (MPRAGE sequence, 1.0 mm isotropic resolution) and a 7T Siemens Terra (MP2RAGE sequence, 0.83 mm isotropic resolution).

Right hippocampal 3D surfaces were reconstructed using four distinct pipelines:

FSL: Automated segmentation using FSL FIRST.
FreeS: Automated atlas-based segmentation using FreeSurfer (v7.4.1).
HU: Volumetric segmentation using HippUnfold.
Man: Manual segmentation following the Harmonized Hippocampal Segmentation Protocol (HARP).

To minimize variability arising from software-specific mesh generation, a standardized surface-extraction and smoothing procedure (AFNI-based IsoSurface and SurfSmooth with 1,000 iterations) was applied to the volumetric outputs of all pipelines.

Shape was quantified using geometric morphometrics based on 97 landmarks and semilandmarks (4 anatomical landmarks, 10 curve semilandmarks, 83 surface semilandmarks) automatically placed via a multi-view consensus convolutional neural network (MV-CNN) and manually corrected where necessary.

Statistical analyses included:

Procrustes ANOVA: Using a Residual Randomization Permutation Procedure (RRPP) with 10,000 permutations to quantify the contributions of subject, reconstruction pipeline, field strength, and their interaction to shape variation.
Correlation Analyses: Pearson correlations assessed the preservation of centroid size (CS) and pairwise Procrustes distances (PPD) between 3T and 7T datasets.
Consistency Metrics: Within-subject Procrustes distances between 3T and 7T shapes were calculated for each pipeline. Agreement with the manual reference was evaluated by comparing Procrustes distances between automatic pipelines and manual segmentation.
Power Analysis: Simulation-based power analysis was conducted via repeated subsampling to evaluate statistical power across sample sizes.

Key Contributions and Results
The study identified the relative magnitude of sources contributing to hippocampal shape variation:

Inter-individual Anatomy: This was the principal source of variation, explaining 19.4% of the total shape variance ( $R^2=0.194, P<0.0001$ ). Principal Component Analysis (PCA) confirmed clustering of shapes by subject.
Reconstruction Pipeline: This was the second most significant factor, explaining 5.0% of the variance ( $R^2=0.050, P=0.0003$ ).
MRI Field Strength: This contributed 2.5% of the variance ( $R^2=0.025, P<0.0001$ ).
Interaction: A significant interaction was found between field strength and reconstruction pipeline ( $R^2=0.045, P=0.0022$ ), indicating that the effect of field strength depends on the specific reconstruction method used.

Pipeline Performance:

Cross-Field Consistency: The FSL pipeline demonstrated the smallest within-subject 3T–7T shape differences (mean Procrustes distance = 0.287), followed by HippUnfold and Manual segmentation. FreeSurfer showed the largest discrepancies (0.355).
Agreement with Manual Segmentation: FSL showed the closest agreement with manual segmentation at both field strengths (lowest mean Procrustes distances).
Morphometric Preservation: Centroid size (global size) showed strong correlation between 3T and 7T ( $r=0.741$ ). However, the correlation for pairwise Procrustes distances (relative spatial relationships) was weaker ( $r=0.371$ ), suggesting that while overall size is preserved, finer-grained shape relationships are more sensitive to acquisition and reconstruction differences.

Significance and Claims
The authors conclude that in the evaluated imaging conditions, reconstruction methodology has a greater influence on hippocampal shape variation than MRI field strength. The study posits that field-strength-related differences are modest relative to inter-individual and methodological effects. Consequently, the findings support the feasibility of conducting whole-hippocampus morphometric analyses using conventional 3T MRI, provided that methodological standardization is maintained.

The paper emphasizes that while 7T offers improved anatomical delineation, these advantages do not automatically result in more consistent whole-hippocampus shape representations across different processing workflows. The authors caution that the observed differences reflect a combination of field strength, acquisition sequences (MPRAGE vs. MP2RAGE), and resolution, rather than field strength in isolation. They assert that methodological consistency remains essential when comparing shape metrics across sites or studies, and that the specific reconstruction pipeline chosen significantly impacts the resulting morphometric data.

Limitations Noted by Authors
The authors acknowledge several limitations: a modest sample size (n=10), restriction to neurologically healthy adults (limiting generalizability to clinical populations), analysis limited to the right whole hippocampus (excluding subfields and the contralateral side), and the inherent confounding of field strength with acquisition sequence and resolution differences. They note that the study does not isolate magnetic field strength as a single experimental factor.

More like this