Low Gyration Index and eTIV Warnings in Morphologist 6.0.7 when importing FreeSurfer 7.4.1 masks

[16percentage]

Environment:

BrainViSA Version: 6.0.7

FreeSurfer Version: 7.4.1

OS: Ubuntu Linux

Description of Pipeline:
I am using a hybrid pipeline to process clinical data. My workflow is as follows:

Importing: nu.mgz from FreeSurfer.

Tissue Import: Importing Gray/White matter masks from FreeSurfer.

Referential: Adding scanner-based referential using nu.mgz for T1_orig and the initial T1 import for T1_output.

Process: Using use_morphologist graphically.

Morphologist 2025:

    Image Orientation (Unlocking).

    Talairach (Not unlocking).

    SPAM model used for sulcal recognition.

The Issue:
Despite the visual inspection looking correct (see attached SPAM sulcus extraction for Sub-1 and Sub-2), the output consistently flags the following:

Gyration Index (GI): Values are consistently low (~1.65–1.73), falling outside the 1-99% percentile of the UKB/HCP normative reference.

eTIV: Consistently flagged as abnormal/suspicious (e.g., 1,142,517mm3 for Sub-3).

Supporting Evidence:
I have compared the FreeSurfer aseg_volumes.csv with the BrainViSA outputs. There appears to be a systematic shift in how the eTIV and cortical folding are being interpreted once imported into the Morphologist environment.

morphologist_report.pdf

morphologist_report.pdf

---

Questions :

Given that I am importing FreeSurfer masks, is it possible that the "Hull" generation is too tightly constrained to the pial surface, causing the low Gyration Index?

Are there known scaling discrepancies between FreeSurfer 7.4.1 and Morphologist 6.0.7 that could explain the consistent eTIV underestimation?

Are there specific parameters in the "Hull" or "Closing" steps of the pipeline I should tune when using external masks?

[denisri]

Hi,
you raise an interesting question for which we don't have a real answer right now.

• The normative stats feature is new in Morphologist, and our normative stats have been established only on 2 datasets, UKB and HCP, none of which are very representative of the general population: one is rather old people, the other rather young, with no intermediate. We'd like to augment the normative set to other data in the future. The best would be to have different normative stats for different age ranges, and allow to select the matched one. In the meantime the normative data are likely biased already.
• Freesurfer and Morphologist produce visually lookalike segmentations, but each probably has its small consistent biases, which may be visible in statistics. Ideally normative stats for the matched method should rather be used (but we have not produced them for Freesufer segmentations imported in Morphologist).
• Is there a scaling issue ? nu.mgh is a resampled image at 1x1x1mm resolution, and not the original MRI acquisition. As far as I know (and remember) the brain scale is conserved during the resampling but right now I'm not 100% certain.
• Is your dataset "different" ? Have your data been preprocessed in any way, with a possible normalization ? Are they in a different age range (children or young for instance ?)
• Does the same source image, processed in both ways (1. Morphologist from scratch, 2. Freesurfer + Morphologist import) produce the same results in regards to the normative stats ? OK we could try this kind of things on our side but haven't taken the time to do so, but it would certainly be interesting to assess possible biases.
• I think I remember that different versions of Freesurfer produce notably different results, too, with a consistently different cortical thickness for instance.

Denis

[16percentage]

Thank you Denis for the detailed response. Since posting I have now processed 10 subjects through this pipeline and the pattern is consistent enough to characterise quantitatively. Some clarifications and new data below.

---

Pipeline Clarification: No Hull Regeneration

To clarify my original description — I am not regenerating the hull or running any closing/reconstruction step inside BrainVISA. The FreeSurfer-derived surfaces are imported directly via:

1. Data Management > Import T1 MRI — nu.mgz (converted to .nii.gz) to establish the referential anchor
2. Import > FreeSurfer > Import FreeSurfer grey/white segmentation — T1_orig = nu.mgz then T1_output = imported in the first step and then the referential generated in the step 1 in the Scanner-Based Referential slot (which i have to double select)
3. Use Morphologist (graphically) — running only: Image Orientation Handling, Talairach Transform, Cortical Fold Graph, and Sulci Recognition (SPAM)

The GI and eTIV values being flagged are therefore computed directly on FreeSurfer-derived surfaces and compared against normative stats built from Morphologist-native reconstructions. This surface-source mismatch is likely the primary driver of the systematic offset we are seeing.

---

Answers to Your Questions

Is the dataset unusual or preprocessed?
Adult clinical patients with Neurodevelopmental Bipolar disorder, ages 18–59 (mean ~36y), no spatial normalisation applied. nu.mgz is used directly — FreeSurfer's bias-field corrected, 1×1×1mm isotropic output with no further preprocessing. This age range sits squarely between UKB (older) and HCP (younger), which likely contributes to the normative offset independently of the pipeline.

FreeSurfer version consistency?
FreeSurfer 7.4.1 consistently across all subjects — the same version bundled within BrainVISA 6.0.7 installation. No cross-version inconsistency.

Same image through both pipelines?
I agree this is the right experiment and plan to run it. A full Morphologist-native reconstruction is time-intensive on the dataset; I will run it on one subject and follow up here with a direct comparison of GI, eTIV, and key sulcal metrics.

---

Updated Data: 10 Subjects — Consistent Systematic Pattern

Gyration Index — flagged in 10/10 subjects without exception.
Values range 1.39–1.76, mean 1.65 ± 0.11. This is not driven by outliers — it is a floor effect consistent across the entire cohort regardless of brain size or eTIV. The most extreme case is GI=1.39 at −5.53σ in a subject whose other volumetrics are largely normal (eTIV 1.305L, brain volume −0.70σ).

eTIV — flagged in 7/10 subjects, ranging 1.007L–1.305L. The three subjects not flagged tend to be those with larger absolute brain volumes, suggesting the issue is with the normative lower bound rather than true intracranial volume pathology.

Average cortical thickness — normal in all 10 subjects: 3.57–3.96mm, Z-scores consistently within ±1σ. This is the one metric that tracks well with the normative reference across all subjects.

Average fold depth — also largely normal: 12.01–13.63mm, mostly within ±1.5σ.

The key observation is this: GI is flagged universally while cortical thickness and fold depth remain normal throughout. These metrics are not independent — a genuinely low GI should co-occur with reduced folding depth and shorter fold length, which I do not see. This dissociation across 10 subjects is strong evidence that the GI flag is an artefact of the surface-source mismatch with the normative reference, not a true morphological signal.

---

Questions

1. GI computation and FreeSurfer mesh properties
Given that GI is flagged universally while cortical thickness and fold depth remain normal, is it possible that the GI computation in Morphologist is sensitive to a surface property that FreeSurfer meshes systematically differ on relative to Morphologist-native surfaces — for example mesh resolution, vertex density, smoothing level, or the definition of the convex hull used in the GI calculation? Understanding which specific input to the GI formula is being affected would help to determine whether this is correctable on my side.

2. Normative reference applicability for hybrid pipelines
The normative stats were established on UKB and HCP data processed through Morphologist's native pipeline. Is there any plan to either produce a separate normative reference for FreeSurfer-imported surfaces, or to document the expected offset when using the hybrid import workflow? For groups using FreeSurfer as their primary pipeline and BrainVISA for sulcal morphometry specifically, the current normative flags are systematically misleading and it would be useful to have guidance on how to interpret them.

3. eTIV estimation in the hybrid workflow
When FreeSurfer meshes are imported directly, how does Morphologist estimate eTIV — does it recompute it from the imported mesh, or does it use a value derived from the nu.mgz T1? If it recomputes from the mesh, a systematic downward bias relative to FreeSurfer's own eTIV estimate would be expected given the surface differences, and would explain the 7/10 flagging rate that is observed.

Thank you

[denisri]

The Freesurfer import process is supposed to input the mri/orig.mgz. You are using nu.mgz. I don't think it's an issue: both are the same resolution (1mm resampled by Freesurfer) but Morphologist non-uniformity correction (bias correction) will still be applied by the import process. Never mind.
Anyway all FS outputs are resampled at 1mm-iso, so we use 1mm-iso resampled images here in Morphologist. However as far as I know, it is still in native space size, and should not change notably the measurements of segmented objects.

eTIV in Morphologist is not really a TIV. From what I read in the code, it is the volume of a closed brain (both hemispheres with cerebellum and trunk, closed by a large amount: 20mm so that sulci, ventricles and inter-hemipspheric space are filled). But it does not extend up to the skull as a TIV should, and has no scaling correction for that (contrarily to Freesurfer, I think). Moreover it depends on where the image field of view cuts the bottom of the trunk (and possibly the cerebellum). So there may be a difference here.

Your cohort is a clinical patients cohort. I am not aware of massive brain volume or gyration effects in bipolar disorders, but there might still be statistical differences with controls. Do you have controls in your cohort ?

GI: it is calculated as the sum of the surface area of sulci meshes divided by the surface area of the hemispheres hulls (closed mask of the brain with sulci filled). The folds surface, especially, might be sensitive to image resolution and quality (although they are smoothed a bit): higher resolution images are more prone to capture little branches or high frequency curvature variations.
But the GI also depends (hopefully) on the subjects intrinsic folding patterns. Fold depth of course influences it, but is not the only factor: a brain with less folded patterns (with fewer branches or tertiary sulci, or with less curved sulci) will have a lower GI even then the average depth is conserved. If your cohort is patients only, there could be an effect here (or a combined effect).

It could also reflect a site effect: depending on MRI machines, acquisition sequences and parameters, contrasts may vary and produce measurable shifts in morphometric measurements. The best normative dataset will always be a dataset matched in site and acquisition parameters, age, sex etc., not a "generic" one like the ones we provide now.

Folds meshes in Morphologist are obtained from the grey and white tissues volumic segmentations, and the algorithm is also driven by the bias-corrected image intensities. They are not obtained from Freesurfer meshes. Thay might be sensitive to Freesurfer segmentations being slightly different from Morphologist ones.

We would be of course interested in re-processing the normative stats from Freesurfer segmentations, but we have not done the importations from FS to Morphologist on those datasets yet. It's a bit time consuming and needs disk storage capacity, which we have not provisioned at the moment.

Denis