SUSEP-Net: Simulation-Supervised and Contrastive Learning-based Deep Neural Networks for Susceptibility Source Separation

Alan H.Wilman; Chen Chen; Feng Liu; G. Bruce Pike; Hongfu Sun; Min Li; Pengfei Rong; Peng Wu; Shanshan Shan; Yang Gao

REVIEW 3 major objections 6 minor 15 references

SUSEP-Net: Simulation-Supervised and Contrastive Learning-based Deep Neural Networks for Susceptibility Source Separation

T0 review · 3 major / 6 minor · 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 dual-branch U-net trained purely on simulated forward-model data separates paramagnetic and diamagnetic brain susceptibility maps with lower error than APART-QSM, χ-separation, and χ-sepnet.

desk verdict Solid engineering contribution to susceptibility source separation, but the headline simulation numbers partly reflect self-consistency with the APART-QSM teacher; the phantom and retrained-baseline comparisons are the more convincing evidence. read the letter →

arxiv 2506.13293 v1 pith:FMSGZSJG submitted 2025-06-16 eess.IV

Min Li , Chen Chen , Zhenghao Li , Yin Liu , Shanshan Shan , Peng Wu , Pengfei Rong , Feng Liu

show 4 more authors

G. Bruce Pike Alan H.Wilman Hongfu Sun Yang Gao

This is my paper · ORCID

classification eess.IV

keywords SusceptibilitysourceseparationQuantitativemapping(QSM)ParamagneticanddiamagneticSimulation-supervisedtrainingContrastivelearningDual-branchU-netR2'relaxometrySub-voxelQSM

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

The paper sets out to solve susceptibility source separation: in a single brain voxel, paramagnetic iron and diamagnetic myelin can cancel in standard quantitative susceptibility maps, hiding the contribution of each. SUSEP-Net is a dual-branch deep network that takes R2', local field, and QSM as inputs and outputs separate paramagnetic (χpos) and diamagnetic (χneg) maps, trained entirely on simulated data generated from the forward model rather than on large multi-orientation in vivo acquisitions. The authors report that it beats APART-QSM, χ-separation, and χ-sepnet on simulated brains, on an agarose phantom, and on patient brains with CO poisoning and MOGAD, with lower NRMSE and sharper lesion contrast. A contrastive loss keeps the two output branches disentangled by aligning each branch's latent features with a branch-specific guidance feature derived from QSM. If the results hold, the method would make iron and myelin mapping more practical and more reliable for clinical MRI, without needing special training data.

What carries the argument

The load-bearing object is the complex forward model of Eq. (1): R2'(r) + i ΔB_local(r) = A(r)·(χpos(r) − χneg(r)) + i D(r) ⊗ (χpos(r) + χneg(r)), where A(r) is a voxel-specific magnitude decay kernel relating R2' to absolute susceptibility and D(r) is the unit dipole kernel. This equation both defines what the network must invert and generates the simulation-supervised training data. The network itself is a dual-branch U-net: one shared encoder fed by the concatenated inputs, two decoders producing χpos and χneg, and two additional guidance encoders fed by QSM. The contrastive loss ties Guide_pos to F_pos and Guide_neg to F_neg while repelling cross-branch pairs. The claim is that this architecture, with pure-synthetic high-intensity lesions added to the simulated patches, is what lets the network recover sub-voxel paramagnetic and diamagnetic content without in vivo ground truth.

What would settle it

Acquire an agarose phantom with known concentrations of mixed paramagnetic and diamagnetic compounds (e.g., Fe3O4 and CaCO3) and compare SUSEP-Net's separated χpos and χneg values against gravimetrically known single-source and mixed-source values; systematic deviations in the mixture cylinders that APART-QSM does not show would indicate label bias. Alternatively, measure R2' and local field on a phantom using an A(r) map different from the APART-QSM-derived one and check whether the network's predictions follow the physical model.

Watch

Extended reading notes

Core claim

SUSEP-Net's central claim is that a simulation-supervised dual-branch U-net with contrastive feature constraints can separate susceptibility sources more accurately than the methods used to generate its own labels. The network learns the mapping from R2', local field, and QSM to χpos and χneg by training on 3024 patches synthesized from APART-QSM reconstructions of 32 brains plus purely synthetic hemorrhage and calcification lesions. On the simulated pathological brain it reports χneg NRMSE 9.66% versus 24.72% (APART-QSM), 31.97% (χ-separation), and 15.85% (χ-sepnet), and the closest lesion values (hemorrhage 0.97 ± 0.018 ppm, calcification 0.19 ± 0.017 ppm). On 10 simulated healthy brains it reports average χpos NRMSE 5.08% against 10.27–16.41% for the competitors. The authors interpret this as evidence that the network has learned the underlying physics of Eq. (1) rather than simply imitating a traditional algorithm.

Load-bearing premise

The training labels (χpos, χneg, and the A(r) decay kernel) come from APART-QSM, and Eq. (1) is assumed to describe exactly how R2' and local field are produced from those labels; if APART-QSM mis-assigns susceptibility or A(r) is wrong, SUSEP-Net inherits the bias and the simulated benchmarks cannot reveal it, since they are generated from the same labels.

Editorial extensions

If this is right

On simulated pathological brains, SUSEP-Net cuts χneg NRMSE to 9.66% versus 15.85% for the best prior deep network and recovers hemorrhage and calcification values close to ground truth (0.97 ppm and 0.19 ppm against inserted 1 ppm and 0.2 ppm).
Because training patches can be synthesized from a handful of subjects plus geometric lesions, the method avoids the expensive multi-orientation acquisitions that χ-sepnet requires.
The agarose phantom results (R2 = 0.96 for CaCO3 and 0.97 for Fe3O4 linearity) suggest the network generalizes to data with different acquisition hardware and contrast sources than its training set.
On CO poisoning and MOGAD patients, SUSEP-Net shows fewer reconstruction artifacts than APART-QSM and χ-separation and renders lesions more visible, which could improve clinical delineation of iron-related and myelin-related damage.
If the method is adopted, standard single-orientation 3T mGRE scans could yield sub-voxel iron/myelin separation, making the technique feasible for routine clinical protocols.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

Because the labels are generated by APART-QSM, SUSEP-Net's superior phantom and patient contrast likely reflects a combination of learned physics and regularization or denoising, not proof that it is closer to true sub-voxel concentrations wherever APART-QSM itself is biased.
The same simulation-supervision recipe should transfer to other ill-posed two-channel inverse problems with opposite-sign sources, such as susceptibility and R2* mapping or fat-water separation, wherever a forward model like Eq. (1) is available.
A direct test of whether SUSEP-Net truly learns physics would be to retrain it with labels from a different algorithm, such as χ-separation instead of APART-QSM: if the outputs change substantially, the network is partly imitating the label generator rather than the underlying forward model.
The method may also be adapted to estimate A(r) jointly instead of taking it as a fixed input, which would remove one of the main sources of label dependence.

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.

Desk Editor's Note

read the letter

Quick take: this is a credible engineering contribution to susceptibility source separation, but the headline simulation numbers partly reflect self-consistency with the APART-QSM teacher. The more convincing evidence is the agarose phantom and the retrained χ-sepnet baseline.

What is new: a dual-branch U-net with two QSM-derived guidance encoders and a branch-wise contrastive loss, trained with a simulation-supervised recipe that synthesizes pathological lesions. The code and trained network are public. The authors also retrain χ-sepnet on their own synthetic data, which is the right control for isolating the training strategy from the architecture. The ablation shows the contrastive loss is worth about 3% NRMSE on simulated brains, and the phantom linearity (R² = 0.96 for CaCO3, 0.97 for Fe3O4) is solid.

Where it is soft: the simulated evaluation reuses the same forward model (Eq. 1) and APART-QSM labels that generated the training data, and the simulated test lesions (1 ppm, -0.2 ppm) fall inside the training ranges ([0.4,1.2] and [-0.3,-0.1]). So the large NRMSE gains over APART-QSM and χ-separation on simulated brains partly measure how well the network inverts its own training generator, not independent accuracy. The Discussion honestly concedes the APART-QSM dependence, but the abstract's 'consistently improved results' overreaches what the simulation can support. The in vivo pathological comparisons are qualitative, and there are mechanical errors: a reference to a nonexistent Table 3 in Section 4.2 (the numbers are in Table 2), a copied typo in Eq. (6), and at least one typo in the definition of the χpos/χneg reconstructions.

The phantom experiment partially compensates. SUSEP-Net's single-source vs mixed-source regression is nearly as good as APART-QSM, and retrained χ-sepnet also improves, which indicates the synthetic training data, not just the architecture, is doing useful work.

Bottom line: the method is real, the evaluation has a self-consistency flaw that is acknowledged but not fully resolved by the current benchmarks, and the paper needs revision before publication. The right next step is a serious referee: ask for test lesions outside the training range, for a decomposition of the simulation metric into self-consistency versus true generalization, and for fixing the reporting errors.

For whom: QSM researchers and anyone building a baseline for paramagnetic/diamagnetic separation. I'd cite it with the caveat noted.

Recommendation: send to peer review, not desk reject.

Referee Report

3 major / 6 minor

Summary. The paper proposes SUSEP-Net, a dual-branch U-net with two auxiliary encoders and a contrastive learning loss, trained on simulated data generated by Eq. (1) from APART-QSM-derived χpos, χneg, and A(r) maps, with synthetic hemorrhage and calcification lesions added. The method is evaluated against APART-QSM, χ-separation, and χ-sepnet on simulated healthy and pathological brains, an agarose gel phantom, and in vivo healthy and pathological subjects, and the authors report improved NRMSE/HFEN/XSIM metrics, better lesion contrast, and reduced artifacts.

Significance. If the reported simulation accuracy reflected independent ground truth, SUSEP-Net would be a meaningful advance for susceptibility source separation, and the contrastive-learning design is a plausible contribution. The paper has concrete strengths: the code and trained networks are released, an ablation study supports the contrastive loss, and the phantom experiment provides an independent test of generalization. However, the central quantitative claim is weakened because the simulated test data are generated by the same APART-QSM-based forward model used for training, so the numerical superiority over untrained methods partly measures self-consistency rather than accuracy against independent truth; the in vivo evidence is qualitative.

major comments (3)

[§2.3, §4.2, Fig. 2] The simulation benchmark is self-referential: the test brains in §4.2 are synthesized with Eq. (1) from APART-QSM-derived χpos, χneg, and A(r) maps, which is exactly the training-data generation pipeline of Fig. 2. Consequently, SUSEP-Net's large NRMSE/HFEN advantages over APART-QSM, χ-separation, and χ-sepnet partly measure consistency with the teacher model and the generator, not independent accuracy. The concern is compounded by the test lesions: the hemorrhage value of 1 ppm and calcification of −0.2 ppm in §3.1(2) lie inside the training ranges [0.4, 1.2] and [−0.3, −0.1] given in §2.3, so the test lesions are in-distribution for SUSEP-Net. The paper itself concedes in Section 5 that training labels still rely on APART-QSM. Please add a validation setup that does not reuse the training generator, or substantially qualify the simulation-based superiority claims.
[§4.3] The agarose gel phantom is a genuinely independent test, but it uses only homogeneous CaCO3 and Fe3O4 solutions in simple cylinders. It does not exercise realistic sub-voxel coexistence of paramagnetic and diamagnetic sources in brain tissue, so it cannot by itself support the claim of improved accuracy in pathological brains. The paper should state this limitation explicitly and should not present the phantom results as sufficient evidence for the general in vivo claim.
[§4.4.2 and Abstract] The in vivo pathological evaluations are qualitative and have no independent ground truth. The abstract's claim that SUSEP-Net shows "improved high-intensity hemorrhage and calcification lesion contrasts, and reduced artifacts" is thus stronger than the evidence provided; the ROI comparisons in §4.4.1 show similarity to APART-QSM rather than superiority, and the patient results rely on visual inspection. Please temper the wording or add blinded reader scoring, lesion delineation metrics, or another quantitative outcome.

minor comments (6)

[§4.2] The text refers to "Table 3" for the 10 simulated healthy brain results, but the table shown is numbered Table 2.
[Eq. (6)] In Eq. (6), the notation "χ&#'∗ and χ&#'∗" uses the same subscript for both reconstructions; the second should be χneg∗ (i.e., χ()*∗).
[Eq. (8)] The last term in Eq. (8) is written as R1!, which appears to be a typo for R2′; please correct this to match the notation used throughout the paper.
[§4.3] The sentence referring to "AFTER-QSM" appears to mean "APART-QSM" based on the context and the methods compared in Fig. 5.
[§2.2.2] There is a typo in the sentence introducing the contrastive loss: "coxntrastive" should be "contrastive."
[Fig. 3] The figure caption and text describe the calcification lesion as 0.2 ppm, but §3.1 specifies −0.2 ppm; this sign inconsistency should be fixed.

Circularity Check

2 steps flagged · score 6.0 of 10

The headline simulated-brain gains are measured against ground truth produced by the same APART-QSM labels and Eq. (1) forward model used to train SUSEP-Net, so the numerical superiority is partly self-consistency with the teacher model rather than independent accuracy.

fitted input called prediction [Section 2.3 (Training Datasets) and Fig. 2; Section 3.1 (Evaluation Datasets); Section 4.2 (Simulated Dataset)]
"The APART-QSM (Li et al., 2023) reconstruction pipeline was conducted to obtain the corresponding full-sized χpos, χneg, and voxel-specific magnitude decay kernel (i.e., A(r) in Eq. (1) and the A map in Fig. 2), which serves the training labels for SUSEP-Net. After obtaining the χpos, χneg, and the A map patches, all the network training inputs, i.e., simulated local field maps, R2’, and the QSM patches, can be synthesized using Eq. (1)."

The 10 simulated healthy brains and the simulated pathological brain used for the headline NRMSE/HFEN/XSIM comparisons are generated by the same Fig. 2 pipeline: ground-truth χpos/χneg and A(r) come from APART-QSM, and the test inputs are synthesized with Eq. (1). SUSEP-Net is trained to invert exactly this generative process onto APART-QSM labels, and its model loss (Eq. 8) enforces the same Eq. (1) relations. Its simulated-test 'predictions' are therefore evaluated against the same teacher labels and forward model it was fitted to; the reported superiority over APART-QSM reflects APART-QSM's imperfect inversion of its own generative model, not independently verified susceptibility separation. The abstract's 'consistently showed improved results' is thus partly a self-consistency measure.
fitted input called prediction [Section 2.3 (synthetic lesion generation) versus Section 3.1(2) (simulated pathological test brain) and Fig. 3]
"Constant susceptibility was randomly assigned from uniform distributions with ranges of [0.4, 1.2] and [−0.3, −0.1] for the synthetic hemorrhage and calcification sources. ... A simulated pathological brain with a hemorrhage source (1 ppm) and calcification source (-0.2 ppm) was superimposed onto one of the 10 healthy subjects ..."

The test lesion values (1 ppm hemorrhage, -0.2 ppm calcification) lie inside the uniform ranges used to synthesize training lesions ([0.4, 1.2] and [-0.3, -0.1]). Thus the lesion-quantification comparison in Fig. 3 is an in-distribution interpolation for SUSEP-Net, which was trained on such lesions, whereas APART-QSM and χ-separation were not trained on them. The claimed 'most accurate hemorrhage and calcification measurements' is a statistically forced outcome of testing on the training distribution, compounding the same-pipeline evaluation.

full rationale

The paper's most load-bearing quantitative claim—best NRMSE/HFEN/XSIM and lesion measurements on simulated healthy and pathological brains—is evaluated on data generated by the exact APART-QSM-and-Eq. (1) pipeline used to create SUSEP-Net's training labels and inputs. Because the test ground truth is the APART-QSM decomposition itself, and the network is trained to invert that same forward model, the large simulated-data advantage over APART-QSM is largely a self-consistency result: the network learns a near-perfect inverse of the generative model, while APART-QSM's own reconstruction has residual error against its own labels. The synthetic test lesions also fall inside the training ranges, further making the lesion comparison an in-distribution fit rather than an out-of-distribution prediction. This would warrant a higher score if the paper had no external check; however, the agarose phantom experiment (Section 4.3) provides independent quantitative grounding with known CaCO3/Fe3O4 concentrations and linear regressions (R2 = 0.96-0.99), and the in vivo patient results, though qualitative, do not reuse the simulation pipeline. The paper itself acknowledges the teacher-label dependence: 'the training of the proposed SUSEP-Net still leveraged an existing algorithm (e.g., APART-QSM) to calculate the training labels.' Self-citations to iQSM/iQSM+ and to the authors' fusion module are building blocks, not circular support for the accuracy claim. On balance, the central claim has genuine independent content from the phantom and in vivo demonstrations, but the headline simulated benchmarks are partly circular, giving a score of 6.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The central claim rests on a supervised learning pipeline whose labels and forward model are imported from APART-QSM and Eq. (1). The loss weights and synthetic lesion ranges are hand-chosen. There are no newly invented physical entities. The A(r) map is an estimated kernel, not a measured quantity, which is why it is listed as a free parameter in the modeling chain.

free parameters (3)

Loss weighting coefficients (α, β, γ, δ) = α=1, β=1, γ=0.5, δ=0.1
Eq. (5); chosen empirically without reported sensitivity analysis, and they control the balance between contrastive, MSE, model-consistency, and gradient losses.
Synthetic lesion susceptibility ranges = hemorrhage U[0.4,1.2] ppm; calcification U[-0.3,-0.1] ppm
Section 2.3; hand-chosen to create high-intensity lesions in training data.
A(r) magnitude decay kernel estimate = voxel-specific values from APART-QSM reconstruction of 32 subjects
Section 2.3 and Eq. (8); imported from APART-QSM rather than measured directly; the central results depend on its accuracy.

assumptions (4)

domain assumption Eq. (1): R2'(r) + i·ΔB_local(r) = A(r)·(χpos(r)-χneg(r)) + i·D(r)⊗(χpos(r)+χneg(r)) is treated as the exact forward model.
Used to generate all simulated training inputs and in the model loss (Eq. 8); if the linear A(r) model is wrong for real tissue, the simulation supervision teaches the wrong mapping.
domain assumption APART-QSM's χpos, χneg, and A(r) outputs are sufficiently accurate to serve as training labels.
Section 2.3 uses APART-QSM reconstructions as ground truth for the network; the paper explicitly notes in the Discussion that an existing algorithm is still needed for labels.
domain assumption Uniform random synthetic lesions in ranges U[0.4,1.2] ppm (hemorrhage) and U[-0.3,-0.1] ppm (calcification) are representative of real high-intensity susceptibility sources.
Section 2.3; synthetic lesions double the training set. If real hemorrhage or calcification has different geometry or susceptibility distributions, the network may not transfer.
domain assumption The input QSM, local field, and R2' used at inference obey the same forward model used in training.
The model loss Eq. (8) and the simulation generator assume consistency; reconstruction errors in iQSM+ or iQFM could violate this and are not modeled.

how reviews work

0 comments

Cite this review

Pith. "Pith review of SUSEP-Net: Simulation-Supervised and Contrastive Learning-based Deep Neural Networks for Susceptibility Source Separation." pith.science (2026). https://pith.science/paper/FMSGZSJG

@misc{pith2026250613293,
  author       = {Pith},
  title        = {Pith review of: SUSEP-Net: Simulation-Supervised and Contrastive Learning-based Deep Neural Networks for Susceptibility Source Separation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/FMSGZSJG}},
  note         = {Machine review of arXiv:2506.13293}
}

read the original abstract

Quantitative susceptibility mapping (QSM) provides a valuable tool for quantifying susceptibility distributions in human brains; however, two types of opposing susceptibility sources (i.e., paramagnetic and diamagnetic), may coexist in a single voxel, and cancel each other out in net QSM images. Susceptibility source separation techniques enable the extraction of sub-voxel information from QSM maps. This study proposes a novel SUSEP-Net for susceptibility source separation by training a dual-branch U-net with a simulation-supervised training strategy. In addition, a contrastive learning framework is included to explicitly impose similarity-based constraints between the branch-specific guidance features in specially-designed encoders and the latent features in the decoders. Comprehensive experiments were carried out on both simulated and in vivo data, including healthy subjects and patients with pathological conditions, to compare SUSEP-Net with three state-of-the-art susceptibility source separation methods (i.e., APART-QSM, \c{hi}-separation, and \c{hi}-sepnet). SUSEP-Net consistently showed improved results compared with the other three methods, with better numerical metrics, improved high-intensity hemorrhage and calcification lesion contrasts, and reduced artifacts in brains with pathological conditions. In addition, experiments on an agarose gel phantom data were conducted to validate the accuracy and the generalization capability of SUSEP-Net.

Figures

Figures reproduced from arXiv: 2506.13293 by the authors.

**Figure 1.** Overall framework of the proposed (a) SUSEP-Net, which is developed based on a dualbranch U-net backbone and two additional encoders (i.e., Encoderpos and Encoderneg in the dashed boxes) for contrastive learning, taking the R2’, local field, and QSM images (for generating the guidance features) as inputs and producing the χ separation images (χpos and χneg) as outputs. The numbers under the feature cubes represent … view at source ↗

**Figure 2.** Training data simulation framework, including the reconstruction of full-size volumes (χpos, χneg, and the magnitude decay kernel) using APART-QSM, generation of patches (χpos and χneg) cropped from the full-size volumes as training patches, and an additional pure synthetic source generated with simple geometric shapes, then the forward calculation model for susceptibility source separation to simulate training inpu… view at source ↗

**Figure 3.** Comparison of the proposed SUSEP-Net with APART-QSM, χ-separation, and χ-sepnet on one simulated pathological brain. (a) shows the reconstruction results of different methods and 5× error maps relative to the simulated ground truth. The corresponding numerical metrics are reported below the images, with the best NRMSEs, HFENs, and XSIMs highlighted in yellow. Red arrows indicate visible reconstruction errors in χneg… view at source ↗

Figures from the paper (1 more)

**Figure 5.** Figure 5: Validation of the accuracy and generalization capability of SUSEP-Net using a specifically designed phantom. (a) illustrates the diagram of the phantom compositions, (b) compares the reconstruction results of iterative APART-QSM and the deep learning based SUSEP-Net, χ…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

15 extracted references · 13 canonical work pages

[1]

INTRODUCTION Quantitative Susceptibility Mapping (QSM) is a valuable MRI technique that can quantify tissue magnetic susceptibility distribution from MRI phase signals (Deistung et al., 2017; Liu et al., 2015; Wang & Liu, 2015). It has demonstrated great potential in studying various neurological diseases, e.g., Parkinson’s Disease (PD) (Acosta-Cabronero ...

work page 2017
[2]

#$%"(𝑟)=A(r)⋅.(𝜒&#'(𝑟)−𝜒()*(𝑟)1 + 𝑖⋅𝐷(𝑟)⨂(𝜒&#'(𝑟)+𝜒()*(𝑟)), (1) where 𝑟 is the spatial coordinate, 𝑅2′(𝑟)=𝑅2∗−𝑅2 is the reversible transverse relaxation rate, ∆𝐵

METHOD 2.1 Theoretical Model for Susceptibility Source Separation The simplified model for the magnetic source separation problem can be described with the following complex equation(Li et al., 2023; Shin et al., 2021): 𝑅2!(𝑟)+𝑖⋅∆𝐵"#$%"(𝑟)=A(r)⋅.(𝜒&#'(𝑟)−𝜒()*(𝑟)1 + 𝑖⋅𝐷(𝑟)⨂(𝜒&#'(𝑟)+𝜒()*(𝑟)), (1) where 𝑟 is the spatial coordinate, 𝑅2′(𝑟)=𝑅2∗−𝑅2 is the rever...

work page 2023
[3]

EXPERIMENTS 3.1 Evaluation Datasets To demonstrate the performance of the proposed SUSEP-Net with state-of-the-art susceptibility source separation methods, comprehensive experiments were conducted on the following datasets: (1) Ten healthy brains synthesized using the pipeline in Fig. 2 were tested in an ablation to investigate the effectiveness of the p...

work page 2023
[7]

The numbers under the feature cubes represent the channel number of the corresponding hidden feature

Overall framework of the proposed (a) SUSEP-Net, which is developed based on a dual-branch U-net backbone and two additional encoders (i.e., Encoderpos and Encoderneg in the dashed boxes) for contrastive learning, taking the R2’, local field, and QSM images (for generating the guidance features) as inputs and producing the χ separation images (χpos and χn...

work page 2025
[9]

The APART-QSM (Li et al.,

and χ-separation (Shin et al., 2021), the training datasets (32 subjects) were acquired at 3T (Phillips, Ingenia Elition X) with a 3D multi-echo GRE (mGRE) sequence with parameters: 8 unipolar echoes, first TE / ΔTE / TR = 4.1 / 4.0 / 35 ms; matrix size= 224 × 224 × 128; voxel size = 0.9 mm × 0.9 mm ×1 mm; flip angle = 20°; CS-SENSE acceleration factor = ...

work page 2021
[12]

and iQSM+ (Gao et al., 2024), respectively. The R2’ images were calculated from R2* and R2 images, which were obtained through a mono-exponential fitting function in STI-Suite based on the mGRE and mSE magnitude images, respectively. The proposed SUSEP-Net was compared with several state-of-the-art susceptibility source separation methods, including itera...

work page 2024
[15]

S., Schwesser, F., Marques, J

Milovic, C., Tejos, C., Acosta‐Cabronero, J., Özbay, P. S., Schwesser, F., Marques, J. P., Irarrazaval, P., Bilgic, B., & Langkammer, C. (2020). The 2016 QSM Challenge: Lessons learned and considerations for a future challenge design. Magnetic resonance in medicine, 84(3), 1624-1637. Milovic, C., Tejos, C., Silva, J., Shmueli, K., & Irarrazaval, P. (2025)...

2020
[128]

ACKNOWLEDGMENTS This work was supported by the National Natural Science Foundation of China under Grant No

is also greater than the simple U-net-based methods (14.31 seconds) due to the complicated network design, and we should develop more light-weight networks for susceptibility source separation in the future. ACKNOWLEDGMENTS This work was supported by the National Natural Science Foundation of China under Grant No. 62301616 and 62301352, and the Natural Sc...

2017

Show all 15 references

[2008]

and leukodystrophy (Nave, 2010). These two types of opposing susceptibility sources can coexist within a single QSM voxel due to the limited acquisition resolution, cancelling each other out and resulting in inaccurate quantification and reducing QSM’s specificity for individu...

2010
[2018]

and prompt learning (Ge et al., 2023; Zhou et al., 2022a; Zhou et al., 2022b), we propose a contrastive learning strategy to explicitly build semantic links between the guidance vectors (Guidepos, neg, i.e., the outputs of Encoderpos and Encoderneg in Fig

2023
[2019]

Different methods have been proposed for susceptibility source separation, i.e., separating QSM into paramagnetic (χpos) and diamagnetic (χneg) components (J

or iron-myelin colocalization in MS lesions (Ji et al., 2024). Different methods have been proposed for susceptibility source separation, i.e., separating QSM into paramagnetic (χpos) and diamagnetic (χneg) components (J. Chen et al., 2021; Emmerich et al., 2021; Kim et al., 2...

2024
[2021]

This method has shown great potential in the diagnosis of multiple sclerosis versus neuromyelitis optica spectrum disorder (Kim et al., 2023)

improved the specificity by jointly taking advantage of local field (or the frequency shift) data and R2’(=R2* - R2) relaxometry, which is assumed to be linearly dependent on the absolute susceptibility, and the coefficients could be determined by a spatially uniform magnitude...

2023
[2022]

realistic

and iQSM+ (Gao et al., 2024), based on 3024 paired training patches (size: 64×64×64) cropped from 36 full-size χpos and χneg volumes (size: 224×224×128) using a sliding window of size 643 with stride of 24×36×20. In addition to the “realistic” positive and negative susceptibil...

2024
[2023]

Li et al., 2025)

improved this 4 scheme by adopting voxel-specific decay kernels, accounting for regional variations in susceptibility mixtures and has been successfully adopted to study epilepsy in a recent work (Z. Li et al., 2025). DECOMPOSE-QSM (J. Chen et al., 2021), constructed based on ...

2025
[2025]

trained with in vivo acquired data (as inputs) and the training labels calculated using the traditional χ-separation method (Shin et al., 2021). Apart from the dual-branch U-net design, another major difference between the proposed SUSEP-Net and the recently proposed U-net-bas...

2021

Pith tools

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

[1] [1]

INTRODUCTION Quantitative Susceptibility Mapping (QSM) is a valuable MRI technique that can quantify tissue magnetic susceptibility distribution from MRI phase signals (Deistung et al., 2017; Liu et al., 2015; Wang & Liu, 2015). It has demonstrated great potential in studying various neurological diseases, e.g., Parkinson’s Disease (PD) (Acosta-Cabronero ...

work page 2017

[2] [2]

#$%"(𝑟)=A(r)⋅.(𝜒&#'(𝑟)−𝜒()*(𝑟)1 + 𝑖⋅𝐷(𝑟)⨂(𝜒&#'(𝑟)+𝜒()*(𝑟)), (1) where 𝑟 is the spatial coordinate, 𝑅2′(𝑟)=𝑅2∗−𝑅2 is the reversible transverse relaxation rate, ∆𝐵

METHOD 2.1 Theoretical Model for Susceptibility Source Separation The simplified model for the magnetic source separation problem can be described with the following complex equation(Li et al., 2023; Shin et al., 2021): 𝑅2!(𝑟)+𝑖⋅∆𝐵"#$%"(𝑟)=A(r)⋅.(𝜒&#'(𝑟)−𝜒()*(𝑟)1 + 𝑖⋅𝐷(𝑟)⨂(𝜒&#'(𝑟)+𝜒()*(𝑟)), (1) where 𝑟 is the spatial coordinate, 𝑅2′(𝑟)=𝑅2∗−𝑅2 is the rever...

work page 2023

[3] [3]

EXPERIMENTS 3.1 Evaluation Datasets To demonstrate the performance of the proposed SUSEP-Net with state-of-the-art susceptibility source separation methods, comprehensive experiments were conducted on the following datasets: (1) Ten healthy brains synthesized using the pipeline in Fig. 2 were tested in an ablation to investigate the effectiveness of the p...

work page 2023

[4] [7]

The numbers under the feature cubes represent the channel number of the corresponding hidden feature

Overall framework of the proposed (a) SUSEP-Net, which is developed based on a dual-branch U-net backbone and two additional encoders (i.e., Encoderpos and Encoderneg in the dashed boxes) for contrastive learning, taking the R2’, local field, and QSM images (for generating the guidance features) as inputs and producing the χ separation images (χpos and χn...

work page 2025

[5] [9]

The APART-QSM (Li et al.,

and χ-separation (Shin et al., 2021), the training datasets (32 subjects) were acquired at 3T (Phillips, Ingenia Elition X) with a 3D multi-echo GRE (mGRE) sequence with parameters: 8 unipolar echoes, first TE / ΔTE / TR = 4.1 / 4.0 / 35 ms; matrix size= 224 × 224 × 128; voxel size = 0.9 mm × 0.9 mm ×1 mm; flip angle = 20°; CS-SENSE acceleration factor = ...

work page 2021

[6] [12]

and iQSM+ (Gao et al., 2024), respectively. The R2’ images were calculated from R2* and R2 images, which were obtained through a mono-exponential fitting function in STI-Suite based on the mGRE and mSE magnitude images, respectively. The proposed SUSEP-Net was compared with several state-of-the-art susceptibility source separation methods, including itera...

work page 2024

[7] [15]

S., Schwesser, F., Marques, J

Milovic, C., Tejos, C., Acosta‐Cabronero, J., Özbay, P. S., Schwesser, F., Marques, J. P., Irarrazaval, P., Bilgic, B., & Langkammer, C. (2020). The 2016 QSM Challenge: Lessons learned and considerations for a future challenge design. Magnetic resonance in medicine, 84(3), 1624-1637. Milovic, C., Tejos, C., Silva, J., Shmueli, K., & Irarrazaval, P. (2025)...

2020

[8] [128]

ACKNOWLEDGMENTS This work was supported by the National Natural Science Foundation of China under Grant No

is also greater than the simple U-net-based methods (14.31 seconds) due to the complicated network design, and we should develop more light-weight networks for susceptibility source separation in the future. ACKNOWLEDGMENTS This work was supported by the National Natural Science Foundation of China under Grant No. 62301616 and 62301352, and the Natural Sc...

2017