henhao
Sign in
BiophysicsbioRxiv
Preprint

Physics and Morphology Constrained Quantitative Susceptibility Based Segmentation of Cerebral Veins

Yousif Al-KhouryiD, Ali RezaeiiD, Vanessa Raspa, Nicole Slanina, Hendrale Gresseau, Yiming XiaoiD, Binu P. ThomasiD, Claudine Joëlle GauthieriD

Abstract

Purpose: Quantitative susceptibility mapping (QSM) provides venous contrast through the paramagnetic susceptibility of deoxyhemoglobin and can be used to estimate oxygen extraction fraction (OEF), a marker of cerebral metabolism. However, cerebral vein segmentation remains challenging due to artifacts, variability across QSM reconstruction methods, and limited sensitivity of conventional vessel-filtering approaches to small cortical veins. This study proposes a physics- and morphology-constrained deep learning framework for cerebral vein segmentation on QSM. Methods: Thirty subjects were manually segmented and used to train an attention-gated UNet. In addition to supervised Dice and cross-entropy losses, the network incorporated two selfsupervised constraints. A physics-informed loss enforced consistency between the measured local field and a field simulated from predicted veins using dipole convolution. A Frangi vesselness loss encouraged anatomically plausible tubular structures. Segmentation performance was evaluated on an independent multi-center cohort and compared with a prior vessel-filtering method and a supervised U-Net baseline. Physiological relevance was assessed by comparing QSM-derived venous OEF with calibrated fMRI-derived OEF. Results: The proposed model achieved superior segmentation performance (centreline Dice = 0.70 {+/-} 0.08) compared with MSVF (0.44 {+/-} 0.09) and the U-Net baseline (0.57 {+/-} 0.14). Improvements were most pronounced in smaller veins, resulting in greater vessel continuity. Venous OEF derived from the proposed framework demonstrated stronger agreement with calibrated fMRI-derived OEF than the previous approach. Conclusion: Combining supervised learning with physics- and morphology-based constraints improves vein segmentation and enhances the sensitivity of QSM-derived OEF measurements. Code and model weights are publicly available at https://github.com/QuantitativePhysiologyImagingLab/VeinSeg

PDFDOICode
Full-text digest ↓

02 /Specification & Classification

Click any tag to browse matching papers
Citations
0
OpenAlex Topics

In-depth summary

✦ Multi-agent verified full-text synthesis

Claude Opus 5.5 · xhigh effort · 8 Nov 2026 · from the full text of version 1 (30 Sept 2026) · CC BY 4.0

Dipole-field physics and Frangi vesselness losses lift QSM vein segmentation to centreline Dice 0.70 (U-Net 0.57); derived venous OEF tracks calibrated fMRI better than vessel filtering.

Venous quantitative susceptibility mapping (QSM) can estimate oxygen extraction fraction (OEF) from routine gradient-echo scans, but threshold and Hessian/Frangi filters rely on hand-tuned parameters and miss small cortical veins, and angiography-trained vessel foundation models may not transfer to QSM contrast. This work pairs supervised deep learning with dipole-physics and tubular-shape constraints to recover finer veins across reconstructions and field strengths, improving downstream OEF sensitivity.

Key results

  1. ShownOn held-out TGV-QSM scans, the full physics- and Frangi-constrained model outperformed a multiscale vessel-filtering method (MSVF) and a supervised vanilla U-Net. MSVF scored worst on every metric.

    Centreline Dice (clDice) 0.70 ± 0.08 for the full model vs 0.44 ± 0.09 for MSVF and 0.57 ± 0.14 for the vanilla U-Net on n = 21 held-out cases; the full model also beat U-Net on Dice, recall and ASSD. VesselFM, which was not trained on QSM, failed to recover veins and was excluded. · Abstract; 3.1 Segmentation Performance; Tables

  2. ShownVenous OEF computed from the proposed segmentations correlated with OEF from gas-calibration-model (GCM) calibrated fMRI, most strongly in the parietal and occipital lobes. OEF from MSVF segmentation showed no significant association in any region.

    Proposed: whole-brain r = +0.34 (p < 0.01), parietal r = +0.42 and occipital r = +0.40 (both p < 0.001); frontal r = +0.12 and temporal r = +0.19 not significant. MSVF: whole-brain r = +0.07, lobes r = −0.04 to +0.10, none significant. n = 74 subjects from the 3T CAD cohort (Dataset 3). · 3.4 Venous QSM-Based OEF; 4 Discussion; Figure legends

  3. SuggestedThe gains come mostly from the smallest veins. The full model had the highest Dice for vessels under 1 mm and under 2 mm, while all configurations performed similarly for veins over 4 mm.

    Diameter was estimated from skeletonized centrelines plus a Euclidean distance transform. Dice was stratified at < 1 mm, < 2 mm and > 4 mm, with n = 4 (< 1 mm, 7T only) and n = 13 (< 2 mm, > 4 mm). The vanilla U-Net was slightly lower for > 4 mm. Per-stratum Dice values are not given in the text. · 2.8 Performance Evaluation; 3.2 Ablation Study; 4 Discussion; Tables

  4. ShownThe ablation attributes the gains to the loss terms rather than the backbone. Swapping in the attention-gated U-Net changed little; the physics loss raised recall and lowered surface error; the Frangi loss raised recall further but cost precision; using both performed best.

    The full model was best on all five reported metrics, including highest Dice and lowest ASSD, across n = 21 held-out cases (Table 3A). The Frangi-only model added false-positive predictions in qualitative examples. Per-model metric values are not stated in the text. · 3.2 Ablation Study; 4 Discussion; Figure legends; Tables

  5. ShownThe model was trained on five reconstructions with FiLM conditioning, and its segmentation of held-out subjects stayed fairly stable across alternative QSM reconstructions. The choice of reconstruction mainly shifted precision and boundary error; sensitivity was nearly constant.

    Dice 0.66 ± 0.08 (iLSQR) to 0.73 ± 0.06 (MEDI, STAR-QSM); clDice 0.62 (iLSQR) to 0.68 (MEDI); recall 0.81–0.82 for all; highest precision MEDI 0.67 ± 0.09; ASSD 0.63 ± 0.32 mm (STAR-QSM) vs 0.75 ± 0.38 mm (iLSQR). n = 16 (MEDI), n = 14 (iLSQR, L1, STAR-QSM). · 3.3 Generalization Across QSM Reconstructions; 4 Discussion; Tables

Methods, data and reproducibility

An nnU-Net-based attention-gated U-Net with a prior-gating block takes QSM, local field and Frangi vesselness inputs, is FiLM-conditioned on reconstruction and field strength, and is trained in three stages (Dice plus cross-entropy, then dipole-convolution field-consistency loss, then Frangi tubularity loss); segmented veins yield OEF from susceptibility.

  • Manually annotated vein labels from Datasets 1, 3 and 5 (controls, CAD, type-2 diabetes, AD subtyping) at 7T and 3T; labels transferred to five QSM reconstructions · human · n = 30 subjects; 150 images across reconstructions
  • Independent multi-center test set: held-out controls (Dataset 1), a CAD subject (Dataset 3), and a Parkinson's disease, intracranial calcification and control cohort (Dataset 6) · human · n = 21 held-out cases (TGV-QSM)
  • Publicly available COSMOS multi-orientation QSM dataset (Dataset 4) used as an anatomical reference · human · n = five subjects
  • CAD cohort (Dataset 3) with multi-echo GRE plus calibrated fMRI (dual-echo pCASL under hypercapnic and hyperoxic challenges) for venous OEF validation · human · n = 74 subjects (OEF correlation)
  • Test subjects reconstructed with alternative QSM algorithms (MEDI, iLSQR, L1, STAR-QSM) for generalization · human · n = 16 (MEDI); 14 (iLSQR, L1, STAR-QSM)
  • Diameter-stratified evaluation subsets of the held-out cases · human · n = 13 (< 2 mm, > 4 mm); 4 (< 1 mm, 7T only)
  • Lobe-wise regional performance subset on TGV-QSM · human · n = 10 held-out cases
Methods & tools
nnU-Net (customized, attention-gated U-Net backbone), FiLM conditioning on reconstruction method and field strength, ROMEO phase unwrapping, RESHARP background field removal, TGV-QSM, MEDI, iLSQR, STAR-QSM and L1-regularized QSM reconstructions, Fourier-domain dipole convolution forward model, Frangi (Hessian) vesselness filter, MSVF multiscale vessel filtering (baseline), VesselFM (evaluated, excluded), Gas calibration model (GCM) calibrated fMRI with dual-echo pCASL, RespirAct system (Thornhill Research), MNI structural atlas
Code
https://github.com/QuantitativePhysiologyImagingLab/VeinSeg
Reproducibility gaps
Only code and model weights are linked. The text does not say whether the in-house cohorts (Datasets 1, 3, 5, 6) or the manual labels are available, and the public COSMOS dataset is cited without a URL or accession.
Limitations (6)
  • The framework is still data-driven and depends on the quality and consistency of the manual labels; annotations of fine cortical veins carry unavoidable uncertainty. · stated by the authors
  • The dataset is modest in size and variability; other vendors (e.g., GE scanners), broader acquisition protocols and more pathologies were not represented. · stated by the authors
  • In lower-resolution QSM, especially at 3T, ringing and streaking artifacts can form elongated patterns that resemble veins and get misclassified. · stated by the authors
  • Venous OEF used fixed constants (Hct = 0.40, deoxyhemoglobin susceptibility), so subject-level variation in blood composition biases absolute OEF. The pCASL reference also has coarser resolution than GRE, which adds partial-volume effects. · stated by the authors
  • OEF validation compared the proposed pipeline only with MSVF, not with the vanilla U-Net, and reported correlations rather than agreement on absolute values. The constraints' specific contribution to OEF accuracy is therefore not isolated. · noted in this summary
  • Some test subjects come from the same datasets used for training (Datasets 1 and 3), all test reconstructions were also seen in training, and sub-millimetre results rest on n = 4 cases at 7T. This limits claims of independent generalization. · noted in this summary

Who should read:Read this if you segment cerebral veins on QSM or estimate venous oxygen extraction from gradient-echo MRI and want a deep learning model with physics and vessel-shape constraints and released weights.

Reproducibility ledger

Nobody has reported trying to reproduce this paper yet. Did you replicate an experiment or run the code? Your report saves the next lab a week.

Sign in to file a reproducibility report.

Related papers

Similarity Map
SelectedSame fieldCross-field
THISBerg '21Ward '17Orooji '22Tahir '21Huck '19Xu '24
68% semantic matchNeuroscience
Ronja Berg, Christine Preibisch, David Lee Thomas+2 more·NeuroImage·Jul 14, 2021· 21 citations