Pith. sign in

REVIEW 37 references

End-to-end Cortical Surface Reconstruction from Clinical Magnetic Resonance Images

T0 review · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read A domain-randomized neural network reconstructs cortical surfaces from heterogeneous clinical MRI scans, cutting cortical thickness error by roughly half versus recon-all-clinical.

arxiv 2505.14017 v1 pith:KMGHDZ5P submitted 2025-05-20 eess.IV cs.CV

Jesper Duemose Nielsen , Karthik Gopinath , Andrew Hoopes , Adrian Dalca , Colin Magdamo , Steven Arnold , Sudeshna Das , Axel Thielscher
show 2 more authors
Juan Eugenio Iglesias Oula Puonti
This is my paper · ORCID
classification eess.IVcs.CV
keywords clinicalcorticalscanssurfacemethodcontrastreconstructionresolution
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

Cortical thickness is an important measure for studying brain aging and dementia, but it is usually computed from high-resolution T1-weighted MRI scans that are standard in research studies and rare in routine hospitals. Clinical scans vary widely in contrast and resolution, with slice gaps up to 8 mm, which breaks most existing tools.

This paper trains a neural network on completely synthetic MRI data. The authors start with real brain segmentations from 5,898 subjects, then simulate thousands of different MRI looks: different intensities for each tissue, different amounts of blurring and noise, different contrasts, and different voxel sizes. Because the network only sees synthetic images, it never overfits to one scanner or protocol. At run time, the network deforms a template brain mesh onto the white matter surface, then expands it to the gray matter surface, preserving topology.

The authors compared their method with recon-all-clinical, the only other tool that claims to process heterogeneous clinical scans. On 200 ADNI FLAIR scans and 1,332 hospital scans, their thickness error fell from 0.50 to 0.24 mm, an improvement of over 50%. The age-related thinning patterns also matched FreeSurfer's results on high-resolution T1w scans more closely. The program runs in about a second on a GPU and the code is public.

Extended reading notes

Core claim

Here, we use synthetic domain-randomized data to train the first neural network for explicit estimation of cortical surfaces from scans of any contrast and resolution, without retraining. We show a approximately 50% reduction in cortical thickness error (from 0.50 to 0.24 mm) with respect to RAC. If correct, this method enables fast and accurate surface-based cortical analysis on heterogeneous clinical MRI at a scale far beyond what research-only pipelines can achieve.

Load-bearing premise

The synthetic domain-randomization generator in Section 2.3 produces training images whose intensity distributions, partial volume effects, resolution ranges, and noise models span the real clinical acquisition space, so a network trained only on synthetic data generalizes to any contrast and resolution at test time. If any real clinical acquisition (e.g., an uncommon contrast or severe anisotropy) lies outside the simulated distribution, the central claim of 'any contrast and resolution' fails.

Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Assumptions & free parameters 5 free parameters · 5 assumptions · 0 invented entities

The central claim relies on the synthetic data distribution covering real clinical acquisitions, on FreeSurfer-derived surfaces as a reference, and on the template-deformation formalism. None of these are independently validated in this paper beyond the reported experiments; they are reasonable domain assumptions for the neuroimaging community.

free parameters (5)
  • PV sigmoid steepness rho
    Controls the partial volume transfer function PV(d)=1/(1+exp(-rho*d)) in synthetic data generation (Section 2.3); chosen by hand, value not reported.
  • Minimum contrast between WM/GM/CSF
    Forced to stabilize training in synthetic data generation (Section 2.3); threshold not specified.
  • Domain randomization probabilities (gamma, bias field) = 0.33 / 0.75
    Probabilities of applying gamma transform and bias field in synthetic data generation, set by hand (Section 2.3).
  • Loss weights (L_matched, L_spring, L_chamfer, L_curv) = e.g., 1->0, 100->0, 1->1, 40->2.5
    Training loss schedule chosen by hand (Section 2.3).
  • Resolution simulation ranges
    Range of isotropic and anisotropic voxel resolutions simulated in synthetic data; not specified in the paper.
assumptions (5)
  • domain assumption FreeSurfer 7.4.1 surfaces computed from 1 mm T1w scans are a valid ground truth for cortical surfaces and thickness.
    Used as training targets and evaluation reference (Sections 2.3 and 2.4).
  • ad hoc to paper The synthetic domain-randomized images, generated from label maps and intensity distributions, sufficiently approximate real clinical MRI appearance across contrasts and resolutions.
    The central generalization claim rests on this; Section 2.3 'Domain Randomization'.
  • domain assumption Partial volume effects can be modeled by a sigmoid transfer of signed distance to the surfaces.
    Used to generate PV fractions near cortex in synthetic data (Section 2.3).
  • standard math The template mesh is topologically a sphere, and the recursive deformation preserves topology.
    Explicit deformation methods guarantee topological correctness if the template is correct; Section 2.2.
  • domain assumption Forward Euler integration with K=2 steps per mesh resolution sufficiently solves the deformation ODE.
    Discretization of Eq. (1) into Eq. (2) with h=1/K; Section 2.2.

how reviews work

0 comments
Cite this review

Pith. "Pith review of End-to-end Cortical Surface Reconstruction from Clinical Magnetic Resonance Images." pith.science (2026). https://pith.science/paper/KMGHDZ5P

@misc{pith2026250514017,
  author       = {Pith},
  title        = {Pith review of: End-to-end Cortical Surface Reconstruction from Clinical Magnetic Resonance Images},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/KMGHDZ5P}},
  note         = {Machine review of arXiv:2505.14017}
}
read the original abstract

Surface-based cortical analysis is valuable for a variety of neuroimaging tasks, such as spatial normalization, parcellation, and gray matter (GM) thickness estimation. However, most tools for estimating cortical surfaces work exclusively on scans with at least 1 mm isotropic resolution and are tuned to a specific magnetic resonance (MR) contrast, often T1-weighted (T1w). This precludes application using most clinical MR scans, which are very heterogeneous in terms of contrast and resolution. Here, we use synthetic domain-randomized data to train the first neural network for explicit estimation of cortical surfaces from scans of any contrast and resolution, without retraining. Our method deforms a template mesh to the white matter (WM) surface, which guarantees topological correctness. This mesh is further deformed to estimate the GM surface. We compare our method to recon-all-clinical (RAC), an implicit surface reconstruction method which is currently the only other tool capable of processing heterogeneous clinical MR scans, on ADNI and a large clinical dataset (n=1,332). We show a approximately 50 % reduction in cortical thickness error (from 0.50 to 0.24 mm) with respect to RAC and better recovery of the aging-related cortical thinning patterns detected by FreeSurfer on high-resolution T1w scans. Our method enables fast and accurate surface reconstruction of clinical scans, allowing studies (1) with sample sizes far beyond what is feasible in a research setting, and (2) of clinical populations that are difficult to enroll in research studies. The code is publicly available at https://github.com/simnibs/brainnet.

Figures

Figures reproduced from arXiv: 2505.14017 by the authors.

Figure 1
Figure 1. Overview of the synthetic data generation and training approach. [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Cortical surface reconstruction on 0.85×0.85×5 mm3 FLAIR (recon-all reconstruction based on 1×1×1 mm3 T1w for reference). Green arrows highlight regions of poor reconstruction by RAC. Note that RAC does not circumvent the hippocampus (blue arrows); this region, together with the medial wall, is masked out of the evaluation. Massachusetts General Hospital, which contains scans of varying contrasts (T1w, T2w, FLAIR, D… view at source ↗
Figure 3
Figure 3. Box plots for absolute value of mean cortical thickness error on the axial [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Cortical thickness variation with age on ADNI GO/2 (top) and the clinical [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

37 extracted references · 37 canonical work pages

  1. [1]

    Med Im Anal 86, 102789 (2023)

    Billot, B., Greve, D.N., Puonti, O., Thielscher, A., et al.: SynthSeg: Segmentation of brain MRI scans of any contrast and resolution without retraining. Med Im Anal 86, 102789 (2023)

    work page 2023

  2. [2]

    Med Im Anal93, 103093 (2024)

    Bongratz, F., Rickmann, A.M., Wachinger, C.: Neural deformation fields for template-based reconstruction of cortical surfaces from MRI. Med Im Anal93, 103093 (2024)

    work page 2024

  3. [3]

    NeuroIm148, 77 (2017)

    Carass, A., Roy, S., Jog, A., Cuzzocreo, J., Magrath, E., et al.: Longitudinal multi- ple sclerosis lesion segmentation: Resource and challenge. NeuroIm148, 77 (2017)

    work page 2017

  4. [4]

    In: 2021 IEEE Winter Conference on Applications of Computer Vision (WACV)

    Cruz, R.S., Lebrat, L., Bourgeat, P., Fookes, C., Fripp, J., Salvado, O.: DeepCSR: A 3D Deep Learning Approach for Cortical Surface Reconstruction. In: 2021 IEEE Winter Conference on Applications of Computer Vision (WACV). pp. 806–815 (2021)

    work page 2021

  5. [5]

    Segmen- tation and Surface Reconstruction

    Dale, A.M., Fischl, B., Sereno, M.I.: Cortical Surface-Based Analysis: I. Segmen- tation and Surface Reconstruction. NeuroImage9(2), 179–194 (1999)

    work page 1999

  6. [6]

    Mol Psychiatry19(6), 659–667 (2014)

    Di Martino, A., Yan, C.G., Li, Q., Denio, E., et al.: The autism brain imaging data exchange: Towards a large-scale evaluation of the intrinsic brain architecture in autism. Mol Psychiatry19(6), 659–667 (2014)

    work page 2014

  7. [7]

    IEEE Trans Med Im20(1), 70–80 (2001)

    Fischl, B., Liu, A., Dale, A.: Automated manifold surgery: Constructing geometri- cally accurate and topologically correct models of the human cerebral cortex. IEEE Trans Med Im20(1), 70–80 (2001)

    work page 2001

  8. [8]

    NeuroIm62(2), 774–781 (2012)

    Fischl, B.: FreeSurfer. NeuroIm62(2), 774–781 (2012)

    work page 2012

Show all 37 references
  1. [9]

    GigaSci p

    Gaser, C., Dahnke, R., Thompson, P.M., Kurth, F., et al.: CAT: A computational anatomy toolbox for the analysis of structural MRI data. GigaSci p. giae049 (2024)

    2024

  2. [10]

    NeuroImage62(2), 748–756 (2012)

    Goebel, R.: BrainVoyager — Past, present, future. NeuroImage62(2), 748–756 (2012)

    2012

  3. [11]

    Neuroinformatics11(3), 367–388 (2013)

    Gollub, R.L., Shoemaker, J.M., King, M.D., White, T., et al.: The MCIC Collec- tion: A Shared Repository of Multi-Modal, Multi-Site Brain Image Data from a Clinical Investigation of Schizophrenia. Neuroinformatics11(3), 367–388 (2013)

    2013

  4. [12]

    Gopinath, K., Greve, D.N., Magdamo, C., Arnold, S., et al.: Recon-all-clinical: Cortical surface reconstruction and analysis of heterogeneous clinical brain MRI (2024)

    2024

  5. [13]

    NeuroIm219, 117012 (2020)

    Henschel, L., Conjeti, S., Estrada, S., Diers, K., et al.: FastSurfer - A fast and accurate deep learning based neuroimaging pipeline. NeuroIm219, 117012 (2020)

    2020

  6. [14]

    IEEE Transactions on Medical Imaging41(3), 543–558 (2022)

    Hoffmann, M., Billot, B., Greve, D.N., Iglesias, J.E., Fischl, B., Dalca, A.V.: Synth- Morph: Learning Contrast-Invariant Registration Without Acquired Images. IEEE Transactions on Medical Imaging41(3), 543–558 (2022)

    2022

  7. [15]

    PMLR172, 508–520 (2022)

    Hoopes, A., Iglesias, J., Fischl, B., Greve, D., Dalca, A.: TopoFit: Rapid Recon- struction of Topologically-Correct Cortical Surfaces. PMLR172, 508–520 (2022)

    2022

  8. [16]

    Science Advances9(5), eadd3607 (2023)

    Iglesias, J.E., Billot, B., Balbastre, Y., Magdamo, C., Arnold, S.E., Das, S., et al.: SynthSR: A public AI tool to turn heterogeneous clinical brain scans into high-resolution T1-weighted images for 3D morphometry. Science Advances9(5), eadd3607 (2023)

    2023

  9. [17]

    J Magn Reson Imaging 27(4), 685–691 (2008)

    Jack Jr., C.R., Bernstein, M.A., Fox, N.C., Thompson, P., et al.: The Alzheimer’s disease neuroimaging initiative (ADNI): MRI methods. J Magn Reson Imaging 27(4), 685–691 (2008)

    2008

  10. [18]

    Schizophr Res60, 199–200 (2003) End-to-end Cortical Surface Reconstruction from Clinical MRI 11

    Kuperberg, G., Broome, M., McGuire, P., David, A., et al.: Regionally localized thinning of the cerebral cortex in schizophrenia. Schizophr Res60, 199–200 (2003) End-to-end Cortical Surface Reconstruction from Clinical MRI 11

    2003

  11. [19]

    LaMontagne, P.J., Benzinger, T.L., Morris, J.C., Keefe, S., et al.: OASIS-3: Lon- gitudinal Neuroimaging, Clinical, and Cognitive Dataset for Normal Aging and Alzheimer Disease (2019)

    2019

  12. [20]

    Medical Image Analysis98, 103292 (2024)

    Li,J.,Tuckute,G.,Fedorenko,E.,Edlow,B.L.,Dalca,A.V.,Fischl,B.:JOSA:Joint surface-based registration and atlas construction of brain geometry and function. Medical Image Analysis98, 103292 (2024)

    2024

  13. [21]

    SIGGRAPH Comput

    Lorensen, W.E., Cline, H.E.: Marching cubes: A high resolution 3D surface con- struction algorithm. SIGGRAPH Comput. Graph.21(4), 163–169 (1987)

    1987

  14. [22]

    NeuroImage12(3), 340–356 (2000)

    MacDonald, D., Kabani, N., Avis, D., Evans, A.C.: Automated 3-D Extraction of Inner and Outer Surfaces of Cerebral Cortex from MRI. NeuroImage12(3), 340–356 (2000)

    2000

  15. [23]

    Hum Brain Mapp34(9), 2302–2312 (2013)

    Mayer, A.R., Ruhl, D., Merideth, F., Ling, J., et al.: Functional imaging of the hemodynamic sensory gating response in schizophrenia. Hum Brain Mapp34(9), 2302–2312 (2013)

    2013

  16. [24]

    In: Visualization and Mathematics III, pp

    Meyer, M., Desbrun, M., Schröder, P., Barr, A.H.: Discrete Differential-Geometry Operators for Triangulated 2-Manifolds. In: Visualization and Mathematics III, pp. 35–57. Springer, Berlin, Heidelberg (2003)

    2003

  17. [25]

    Front syst neurosci6(2012)

    Milham, M.P., Fair, D., Mennes, M., Mostofsky, S.H.: The Adhd-200 consortium: A model to advance the translational potential of neuroimaging in clinical neuro- science. Front syst neurosci6(2012)

    2012

  18. [26]

    Commun Bio4(1), 1–12 (2021)

    Natu, V.S., Rosenke, M., Wu, H., Querdasi, F.R., et al.: Infants’ cortex undergoes microstructural growth coupled with myelination during development. Commun Bio4(1), 1–12 (2021)

    2021

  19. [27]

    NeuroIm168, 296–320 (2018)

    Polimeni, J.R., Renvall, V., Zaretskaya, N., Fischl, B.: Analysis strategies for high- resolution UHF-fMRI data. NeuroIm168, 296–320 (2018)

    2018

  20. [28]

    NeuroIm47, S79 (2009)

    Querbes, O., Aubry, F., Parienté, J., Lotterie, JA., et al.: Individual Early Di- agnosis of Alzheimer’s Disease using Cortical Thickness Measurement: Impact of Cognitive Reserve. NeuroIm47, S79 (2009)

    2009

  21. [29]

    In: Proceedings of MICCAI

    Ronneberger, O., Fischer, P., Brox, T.: U-Net: Convolutional Networks for Biomed- ical Image Segmentation. In: Proceedings of MICCAI. pp. 234–241 (2015)

    2015

  22. [30]

    Neurobiol Aging31(8), 1275–1283 (2010)

    Rowe,C.C.,Ellis,K.A.,Rimajova,M.,Bourgeat,P.,etal.:Amyloidimagingresults from the Australian Imaging, Biomarkers and Lifestyle (AIBL) study of aging. Neurobiol Aging31(8), 1275–1283 (2010)

    2010

  23. [31]

    In: Proceedings of MICCAI

    Santa Cruz, R., Lebrat, L., Fu, D., Bourgeat, P., et al.: CorticalFlow++: Boosting Cortical Surface Reconstruction Accuracy, Regularity, and Interoperability. In: Proceedings of MICCAI. pp. 496–505 (2022)

    2022

  24. [32]

    Medical Image Analysis6(2), 129–142 (2002)

    Shattuck, D.W., Leahy, R.M.: BrainSuite: An automated cortical surface identifi- cation tool. Medical Image Analysis6(2), 129–142 (2002)

    2002

  25. [33]

    Neurobiology of Aging39, 202–209 (2016)

    Shaw, M.E., Sachdev, P.S., Anstey, K.J., Cherbuin, N.: Age-related cortical thin- ning in cognitively healthy individuals in their 60s: The PATH Through Life study. Neurobiology of Aging39, 202–209 (2016)

    2016

  26. [34]

    In: Proceedings of IEEE International Conference on Computer Vision

    Taubin, G.: Curve and surface smoothing without shrinkage. In: Proceedings of IEEE International Conference on Computer Vision. pp. 852–857 (1995)

    1995

  27. [35]

    Journal of the American Medical Informatics Association8(5), 443–459 (2001)

    Van Essen, D.C., Drury, H.A., Dickson, J., Harwell, J., Hanlon, D., Anderson, C.H.: An Integrated Software Suite for Surface-based Analyses of Cerebral Cortex. Journal of the American Medical Informatics Association8(5), 443–459 (2001)

    2001

  28. [36]

    NeuroImage80, 62–79 (2013)

    Van Essen, D.C., Smith, S.M., Barch, D.M., Behrens, T.E.J., et al.: The WU-Minn Human Connectome Project: An overview. NeuroImage80, 62–79 (2013)

    2013

  29. [37]

    Nat Methods20, 177 (2023)

    Vogt, N.: The Chinese Human Connectome Project. Nat Methods20, 177 (2023)

    2023

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.