pith. sign in

arxiv: 2509.16255 · v2 · pith:UOKF4BUVnew · submitted 2025-09-17 · 🧬 q-bio.TO · eess.IV· physics.med-ph

Segmentation of spinal rootlets across MRI contrasts with RootletSeg

Katerina Krejci , Jiri Chmelik , Sandrine Bedard , Falk Eippert , Ulrike Horn , Virginie Callot , Julien Cohen-Adad , Jan Valosek This is my paper

Pith reviewed 2026-05-21 22:11 UTC · model grok-4.3

classification 🧬 q-bio.TO eess.IVphysics.med-ph
keywords spinal rootletsMRI segmentationdeep learningRootletSegspinal levelsDice scoremulti-contrast MRInerve root segmentation
0
0 comments X

The pith

A deep learning model segments spinal rootlets on four MRI contrasts with Dice scores of 0.62 to 0.67.

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

The paper introduces RootletSeg, a deep learning approach for automatically segmenting the dorsal and ventral spinal rootlets from C2 to T1 in MRI images. It trains and tests the model on a combination of open and private datasets totaling 93 scans from 50 healthy adults, reporting Dice scores that range from 0.62 on T1w-INV1 to 0.67 on T1w-INV2. A sympathetic reader would care because this segmentation permits mapping spinal cord levels directly onto the MRI data instead of approximating from intervertebral discs. Accurate rootlet outlines could improve group analyses in functional imaging and support planning for spinal cord stimulation therapies.

Core claim

RootletSeg was developed to segment C2-T1 dorsal and ventral spinal rootlets on 3D isotropic 3T T2-weighted and 7T MP2RAGE MRI scans. On a test set of 17 scans, it reached mean Dice scores of 0.67 ± 0.09 for T1w-INV2, 0.65 ± 0.11 for UNIT1, 0.64 ± 0.08 for T2w, and 0.62 ± 0.10 for T1w-INV1. Spinal levels derived from the segmentations showed a progressively increasing rostrocaudal shift relative to vertebral levels defined by intervertebral discs, with Bland-Altman bias ranging from 0.00 to 8.15 mm.

What carries the argument

RootletSeg, a deep learning model trained to identify and outline spinal nerve rootlets in multi-contrast MRI volumes.

Load-bearing premise

Expert manual segmentations on the training and test scans serve as reliable ground truth, and the healthy adult dataset captures the full range of anatomical and contrast variations seen in broader use.

What would settle it

If RootletSeg is applied to a new collection of MRI scans from individuals with spinal cord lesions and produces Dice scores below 0.5 or fails to reproduce the expected rostrocaudal shift in level correspondence, the claim of accurate and generalizable segmentation would be falsified.

read the original abstract

Purpose: To develop a deep learning method for the automatic segmentation of spinal nerve rootlets on various MRI scans. Material and Methods: This retrospective study included MRI scans from two open-access and one private dataset, consisting of 3D isotropic 3T TSE T2-weighted (T2w) and 7T MP2RAGE (T1-weighted [T1w] INV1 and INV2, and UNIT1) MRI scans. A deep learning model, RootletSeg, was developed to segment C2-T1 dorsal and ventral spinal rootlets. Training was performed on 76 scans and testing on 17 scans. The Dice score was used to compare the model performance with an existing open-source method. Spinal levels derived from RootletSeg segmentations were compared with vertebral levels defined by intervertebral discs using Bland-Altman analysis. Results: The RootletSeg model developed on 93 MRI scans from 50 healthy adults (mean age, 28.70 years $\pm$ 6.53 [SD]; 28 [56%] males, 22 [44%] females) achieved a mean $\pm$ SD Dice score of 0.67 $\pm$ 0.09 for T1w-INV2, 0.65 $\pm$ 0.11 for UNIT1, 0.64 $\pm$ 0.08 for T2w, and 0.62 $\pm$ 0.10 for T1w-INV1 contrasts. Spinal-vertebral level correspondence showed a progressively increasing rostrocaudal shift, with Bland-Altman bias ranging from 0.00 to 8.15 mm (median difference between level midpoints). Conclusion: RootletSeg accurately segmented C2-T1 spinal rootlets across MRI contrasts, enabling the determination of spinal levels directly from MRI scans. The method is open-source and can be used for a variety of downstream analyses, including lesion classification, neuromodulation therapy, and functional MRI group analysis.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

1 major / 1 minor

Summary. The paper introduces RootletSeg, a deep learning approach for segmenting spinal nerve rootlets (C2-T1 dorsal and ventral) on multi-contrast MRI scans from 93 scans of 50 healthy adults. Training on 76 scans and testing on 17 scans yields mean Dice scores of 0.67±0.09 (T1w-INV2), 0.65±0.11 (UNIT1), 0.64±0.08 (T2w), and 0.62±0.10 (T1w-INV1). The segmentations enable spinal level determination, validated against vertebral levels using Bland-Altman analysis, and the model is released as open-source.

Significance. Should the reported performance generalize to new subjects without data leakage, RootletSeg offers a practical tool for automatic rootlet segmentation across contrasts, supporting spinal level identification independent of vertebral landmarks. This has implications for lesion localization, neuromodulation planning, and group-level fMRI analyses. The open-source availability and multi-dataset approach strengthen its potential utility in the field.

major comments (1)
  1. [Methods (Data partitioning)] The description of the train/test split (76 training scans, 17 test scans) does not specify if the partitioning was performed at the subject level to prevent scans from the same individual appearing in both sets. Given 93 scans from 50 subjects (average ~1.86 scans/subject), a scan-level split risks data leakage via subject-specific anatomical or contrast features, which would compromise the validity of the Dice scores as a measure of performance on truly unseen data.
minor comments (1)
  1. [Abstract] The abstract does not provide details on the deep learning architecture, loss function, data augmentation, or statistical methods used for comparing against the existing open-source method.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for their careful review and constructive feedback on our manuscript. We address the major comment below and will revise the manuscript to improve methodological clarity.

read point-by-point responses

  1. Referee: [Methods (Data partitioning)] The description of the train/test split (76 training scans, 17 test scans) does not specify if the partitioning was performed at the subject level to prevent scans from the same individual appearing in both sets. Given 93 scans from 50 subjects (average ~1.86 scans/subject), a scan-level split risks data leakage via subject-specific anatomical or contrast features, which would compromise the validity of the Dice scores as a measure of performance on truly unseen data.

    Authors: We thank the referee for highlighting this important detail. The train/test partitioning was performed at the subject level to avoid data leakage, with all scans from any given subject assigned exclusively to either the training or test set. This ensures the reported Dice scores reflect generalization to unseen subjects. We will revise the Methods section to explicitly state that the split was conducted at the subject level. revision: yes

Circularity Check

0 steps flagged

No circularity: empirical performance metrics are independent of model inputs

full rationale

The paper presents an empirical deep-learning segmentation study. The central results are Dice scores (0.62–0.67) computed on a held-out test set of 17 scans against expert manual segmentations that were never used in training. No equations, predictions, or first-principles derivations are offered that reduce to fitted parameters, self-referential normalizations, or self-citation chains. The reported metrics are direct, externally verifiable comparisons between model output and independent ground truth; they do not collapse to quantities defined by the training process itself. Potential concerns about subject-wise splitting affect generalization validity but do not meet the circularity criteria of self-definition or fitted-input-as-prediction.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 0 invented entities

The central performance claim rests on the assumption that expert manual labels are accurate ground truth and that the 50-subject training distribution captures the relevant anatomical and scanner variability; the model itself contains thousands of learned weights that function as free parameters.

free parameters (1)
  • deep learning model weights and hyperparameters
    All network parameters are fitted to the 76 training scans; exact count and regularization choices are not stated in the abstract.
axioms (1)
  • domain assumption Manual segmentations by human experts constitute reliable ground truth for Dice evaluation
    Dice scores are computed against these labels; any systematic bias in the reference segmentations directly affects the reported performance.

pith-pipeline@v0.9.0 · 5940 in / 1454 out tokens · 45105 ms · 2026-05-21T22:11:15.664214+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

35 extracted references · 35 canonical work pages

  1. [1]

    Location, length, and enhancement: systematic approach to differentiating intramedullary spinal cord lesions

    Mohajeri Moghaddam S, Bhatt AA. Location, length, and enhancement: systematic approach to differentiating intramedullary spinal cord lesions. Insights Imaging. 2018;9(4):511–526

    work page 2018

  2. [2]

    Traumatic spinal cord injury

    Ahuja CS, Wilson JR, Nori S, et al. Traumatic spinal cord injury. Nat Rev Dis Primers. Springer Science and Business Media LLC; 2017;3(1). doi: 10.1038/nrdp.2017.18

    work page doi:10.1038/nrdp.2017.18 2017

  3. [3]

    Neuromodulation of the cervical spinal cord in the treatment of chronic intractable neck and upper extremity pain: A case series and review of the literature

    Vallejo R MD, PhD. Neuromodulation of the cervical spinal cord in the treatment of chronic intractable neck and upper extremity pain: A case series and review of the literature. Pain Physician. American Society of Interventional Pain Physicians; 2007;2;10(3;2):305–311

    work page 2007

  4. [4]

    Activity-dependent spinal cord neuromodulation rapidly restores trunk and leg motor functions after complete paralysis

    Rowald A, Komi S, Demesmaeker R, et al. Activity-dependent spinal cord neuromodulation rapidly restores trunk and leg motor functions after complete paralysis. Nat Med. Springer Science and Business Media LLC; 2022;28(2):260–271

    work page 2022

  5. [5]

    Spinal Cord fMRI: A New Window into the Central Nervous System

    Kinany N, Pirondini E, Micera S, Van De Ville D. Spinal Cord fMRI: A New Window into the Central Nervous System. Neuroscientist. 2023;29(6):715–731

    work page 2023

  6. [6]

    Rootlets-based registration to the spinal cord PAM50 template

    Bédard S, Valošek J, Oliva V, Weber KA II, Cohen-Adad J. Rootlets-based registration to the spinal cord PAM50 template. Imaging Neuroscience. 2025; doi: 10.1162/IMAG.a.123

    work page doi:10.1162/imag.a.123 2025

  7. [7]

    Characterizing the location of spinal and vertebral levels in the human cervical spinal cord

    Cadotte DW, Cadotte A, Cohen-Adad J, et al. Characterizing the location of spinal and vertebral levels in the human cervical spinal cord. AJNR Am J Neuroradiol. 2015;36(4):803–810

    work page 2015

  8. [8]

    Morphometric Relationship between the Cervicothoracic Cord Segments and Vertebral Bodies

    Kim JH, Lee CW, Chun KS, Shin WH, Bae H-G, Chang JC. Morphometric Relationship between the Cervicothoracic Cord Segments and Vertebral Bodies. J Korean Neurosurg Soc. 2012;52(4):384–390

    work page 2012

  9. [9]

    Automatic labeling of vertebral levels using a robust template-based approach

    Ullmann E, Pelletier Paquette JF, Thong WE, Cohen-Adad J. Automatic labeling of vertebral levels using a robust template-based approach. Int J Biomed Imaging. 2014;2014:719520

    work page 2014

  10. [10]

    A Review of the Segmental Diameter of the Healthy Human Spinal Cord

    Frostell A, Hakim R, Thelin EP, Mattsson P, Svensson M. A Review of the Segmental Diameter of the Healthy Human Spinal Cord. Front Neurol. 2016;7:238

    work page 2016

  11. [11]

    Spinal Cord Anatomy and Clinical Syndromes

    Diaz E, Morales H. Spinal Cord Anatomy and Clinical Syndromes. Semin Ultrasound CT MR. 2016;37(5):360–371

    work page 2016

  12. [12]

    Segment-Specific Orientation of the Dorsal and Ventral Roots for Precise Therapeutic Targeting of Human Spinal Cord

    Mendez A, Islam R, Latypov T, et al. Segment-Specific Orientation of the Dorsal and Ventral Roots for Precise Therapeutic Targeting of Human Spinal Cord. Mayo Clin Proc. 2021;96(6):1426–1437

    work page 2021

  13. [13]

    Automatic segmentation of the spinal cord nerve rootlets

    Valošek J, Mathieu T, Schlienger R, Kowalczyk OS, Cohen-Adad J. Automatic segmentation of the spinal cord nerve rootlets. Imaging Neurosci (Camb). MIT Press; 2024;2:1–14

    work page 2024

  14. [14]

    Full cervical cord tractography: A new method for clinical use

    Dauleac C, Frindel C, Pélissou-Guyotat I, et al. Full cervical cord tractography: A new method for clinical use. Front Neuroanat. 2022;16:993464

    work page 2022

  15. [15]

    An open-access lumbosacral spine MRI 24 dataset with enhanced spinal nerve root structure resolution

    Liu J, Zhang W, Zhou Y, Xu L, Chu Y-H, Jia F. An open-access lumbosacral spine MRI 24 dataset with enhanced spinal nerve root structure resolution. Sci Data. 2024;11(1):1131

    work page 2024

  16. [16]

    A novel CNN-based image segmentation pipeline for individualized feline spinal cord stimulation modeling

    Fasse A, Newton T, Liang L, et al. A novel CNN-based image segmentation pipeline for individualized feline spinal cord stimulation modeling. J Neural Eng. IOP Publishing; 2024;21(3):036032

    work page 2024

  17. [17]

    SpIC3D imaging: Spinal In-situ Contrast 3D imaging

    Liang L, Fasse A, Damiani A, et al. SpIC3D imaging: Spinal In-situ Contrast 3D imaging. bioRxiv. 2025. doi: 10.1101/2025.03.05.641747

    work page doi:10.1101/2025.03.05.641747 2025

  18. [18]

    SCT: Spinal Cord Toolbox, an open-source software for processing spinal cord MRI data

    De Leener B, Lévy S, Dupont SM, et al. SCT: Spinal Cord Toolbox, an open-source software for processing spinal cord MRI data. Neuroimage. 2017;145(Pt A):24–43

    work page 2017

  19. [19]

    Pontomedullary junction as a reference for spinal cord cross-sectional area: validation across neck positions

    Bédard S, Bouthillier M, Cohen-Adad J. Pontomedullary junction as a reference for spinal cord cross-sectional area: validation across neck positions. Sci Rep. 2023;13(1):13527

    work page 2023

  20. [20]

    Author Correction: Open-access quantitative MRI data of the spinal cord and reproducibility across participants, sites and manufacturers

    Cohen-Adad J, Alonso-Ortiz E, Abramovic M, et al. Author Correction: Open-access quantitative MRI data of the spinal cord and reproducibility across participants, sites and manufacturers. Sci Data. 2021;8(1):251

    work page 2021

  21. [21]

    Ultra-high-field fMRI reveals layer-specific responses in the human spinal cord

    Horn U, Vannesjo SJ, Gross-Weege N, et al. Ultra-high-field fMRI reveals layer-specific responses in the human spinal cord. bioRxiv. 2025. doi: 10.1101/2025.07.17.665316

    work page doi:10.1101/2025.07.17.665316 2025

  22. [22]

    High-resolution multi-parametric quantitative magnetic resonance imaging of the human cervical spinal cord at 7T

    Massire A, Taso M, Besson P, Guye M, Ranjeva J-P, Callot V. High-resolution multi-parametric quantitative magnetic resonance imaging of the human cervical spinal cord at 7T. Neuroimage. 2016;143:58–69

    work page 2016

  23. [23]

    Generic acquisition protocol for quantitative MRI of the spinal cord

    Cohen-Adad J, Alonso-Ortiz E, Abramovic M, et al. Generic acquisition protocol for quantitative MRI of the spinal cord. Nat Protoc. 2021;16(10):4611–4632

    work page 2021

  24. [24]

    Open-access quantitative MRI data of the spinal cord and reproducibility across participants, sites and manufacturers

    Cohen-Adad J, Alonso-Ortiz E, Abramovic M, et al. Open-access quantitative MRI data of the spinal cord and reproducibility across participants, sites and manufacturers. Sci Data. 2021;8(1):219

    work page 2021

  25. [25]

    The first step for neuroimaging data analysis: DICOM to NIfTI conversion

    Li X, Morgan PS, Ashburner J, Smith J, Rorden C. The first step for neuroimaging data analysis: DICOM to NIfTI conversion. J Neurosci Methods. Elsevier; 2016;264:47–56

    work page 2016

  26. [26]

    poldracklab/pydeface: v2.0.0

    Gulban OF, Nielson D, Poldrack R, et al. poldracklab/pydeface: v2.0.0. Zenodo; 2019. doi: 10.5281/ZENODO.3524401

    work page doi:10.5281/zenodo.3524401 2019

  27. [27]

    The brain imaging data structure, a format for organizing and describing outputs of neuroimaging experiments

    Gorgolewski KJ, Auer T, Calhoun VD, et al. The brain imaging data structure, a format for organizing and describing outputs of neuroimaging experiments. Sci Data. 2016;3:160044

    work page 2016

  28. [28]

    nnU-Net: a self-configuring method for deep learning-based biomedical image segmentation

    Isensee F, Jaeger PF, Kohl SAA, Petersen J, Maier-Hein KH. nnU-Net: a self-configuring method for deep learning-based biomedical image segmentation. Nat Methods. 2021;18(2):203–211

    work page 2021

  29. [29]

    Clinical anatomy of the C1 dorsal root, ganglion, and ramus: a review and anatomical study

    Tubbs RS, Loukas M, Slappey JB, Shoja MM, Oakes WJ, Salter EG. Clinical anatomy of the C1 dorsal root, ganglion, and ramus: a review and anatomical study. Clin Anat. 2007;20(6):624–627

    work page 2007

  30. [30]

    Towards contrast-agnostic soft segmentation of the spinal cord

    Bédard S, Karthik EN, Tsagkas C, et al. Towards contrast-agnostic soft segmentation of the spinal cord. Med Image Anal. 2025;101:103473

    work page 2025

  31. [31]

    Monitoring morphometric drift in lifelong learning segmentation of the spinal cord

    Karthik EN, Sandrine B, Jan V, et al. Monitoring morphometric drift in lifelong learning segmentation of the spinal cord. arXiv [cs.CV]. 2025. http://arxiv.org/abs/2505.01364. 25

    work page arXiv 2025

  32. [32]

    Automatic spinal cord localization, robust to MRI contrasts using global curve optimization

    Gros C, De Leener B, Dupont SM, et al. Automatic spinal cord localization, robust to MRI contrasts using global curve optimization. Med Image Anal. 2018;44:215–227

    work page 2018

  33. [33]

    A cadaveric study of the cervical nerve roots and spinal segments

    Kobayashi R, Iizuka H, Nishinome M, Iizuka Y, Yorifuji H, Takagishi K. A cadaveric study of the cervical nerve roots and spinal segments. Eur Spine J. 2015;24(12):2828–2831

    work page 2015

  34. [34]

    Comparative anatomical dimensions of the complete human and porcine spine

    Busscher I, Ploegmakers JJW, Verkerke GJ, Veldhuizen AG. Comparative anatomical dimensions of the complete human and porcine spine. Eur Spine J. 2010;19(7):1104–1114

    work page 2010

  35. [35]

    Automatic segmentation of the spinal cord and intramedullary multiple sclerosis lesions with convolutional neural networks

    Gros C, De Leener B, Badji A, et al. Automatic segmentation of the spinal cord and intramedullary multiple sclerosis lesions with convolutional neural networks. Neuroimage. 2019;184:901–915. 26

    work page 2019