Abstract

Multi-Pool Chemical Exchange Saturation Transfer (CEST) MRI provides valuable metabolic information but is clinically limited by long acquisi-tion times. Although sparse sampling reduces scanning time, reconstructing high-resolution Z-spectra from limited data remains an ill-posed inverse problem. Conventional interpolation and generic Implicit Neural Rep-resentations (INRs) often lack physical constraints, leading to spectral artifacts and physically invalid signals. To address this, we propose Lorentz Encoding (LE), a physics-informed framework that formulates CEST recon-struction as a self-supervised reconstruction task via implicit continuous coordinate learning. Unlike generic positional encodings, LE regularizes the continuous spectral mapping by projecting sparse coordinates into a physically constrained space governed by a combination of parametric Lorentzian profiles with learnable basis functions. This mechanism effectively reduces noise and enforces consistency with physical models. Experiments on in vivo human brain data demonstrate that LE significantly outperforms state-of-the-art methods. Specifically, under a 39-point sampling strategy, LE achieves a PSNR of 57.58 dB and an SSIM of 0.9994. Furthermore, the learned physics-informed encodings form a continuous, geometrically or-dered trajectory in the latent space, ensuring accurate quantitative metabo-lite mapping (APT, NOE, MT).

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/1033_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

https://github.com/ldxFAIRYTAIL/Lorentz-Encoding.git

Link to the Dataset(s)

N/A

BibTex

@InProceedings{LiDex_SelfSupervised_MICCAI2026,
        author = { Li, Dexuan AND Wu, Yupeng AND Wang, Chenglong AND Liu, Hanlin AND Zheng, Hui AND Li, Jianqi AND Yang, Guang},
        title = { { Self-Supervised Implicit CEST Reconstruction via Physics-Informed Lorentz Encoding } },
        booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
        year = {2026},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 16888},
        month = {September},
        page = {pending}
}


Reviews

Review #1

  • Please describe the contribution of the paper

    The authors propose Lorentz Encoding, a physics-informed framework that formulates CEST reconstruction as a manifold-constrained inverse problem.

  • Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.

    By replacing generic positional encodings with Lorentz Encoding, the method explicitly incorporates the underlying CEST physics into the reconstruction process. This mechanism effectively reduces noise, promotes consistency with the physical model, and achieves superior performance over state-of-the-art methods across multiple metabolite maps, including APT, NOE, and MT.

  • Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.

    There are two major concerns. First, the training details are insufficiently described. For the 18 subjects, the manuscript should clarify the criteria used to split the data into training, validation, and testing sets. It should also report the image dimensions and the number of slices included in the analysis. For the 2D model, it remains unclear whether all slices were used for training or whether a selection strategy was adopted, such as restricting training to middle slices or slices with higher image quality. Second, it would strengthen the paper to include an ablation study that evaluates the two parallel encoding modules separately. Such an analysis would better illustrate the contribution of each module and help clarify whether spatial or spectral feature is more critical to the overall model performance.

  • Please rate the clarity and organization of this paper

    Satisfactory

  • Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.

    The authors claimed to release the source code and/or dataset upon acceptance of the submission.

  • Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?

    N/A

  • Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html

    N/A

  • Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.

    (4) Weak Accept — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal

  • Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

    The use of spectral features to construct a physics-informed model is innovative. However, the validation requires further improvement to more clearly demonstrate the effectiveness of incorporating spectral features into the framework.

  • Reviewer confidence

    Very confident (4)

  • [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.

    N/A

  • [Post rebuttal] Please justify your final decision from above.

    N/A



Review #2

  • Please describe the contribution of the paper

    In this work the authors propose a new method for fitting Lorentzian’s to Z-spectra in chemical exchange saturation transfer MRI. This method is based on a Implicit Neural Representation method in which however the authors decouple the spatial and spectral parts of the problem by performing an encoding process separately for the spatial coordinates (Through Hash Encoding) and for spectral parts through an Lorentz Encoding in which the spectras are projected onto a manifold of Lorentzian functions.

  • Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.

    The proposed Lorentz Encoding method is clearly novel for this application, also the concept with the decoupling of the spectral and spatial appears to be novel in approach for CEST MRI.

    Overall the presented work is well described and in a very understandable language. The description of the method has enough details and at the same time is not very convoluted that is can be understood by a person with a bit of experience in CEST MRI. The presented experiments are well designed, easy to follow and to understand what is going on in the results themselves.

    The limitation section identifies correctly issues of the method.

  • Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.

    There are a few weaknesses of this work. My biggest concern is about the reference which the authors use for the calculation of the PSNR/SSIM metrics in their Table 2.Based on the desciption in the text above I would understand that the calucations are performed against the maps/results from the Direct MPLF applied to the undersampled spectra. Such an approach would be wrong as the observed effects would be affected by the decreasing performance of the MPLF on the undersampled spectra. However in the Figure 3 the authors provide reference maps which are clearly obtained from direct MPLF on the fully sampled spectra. As such it’s not clear to me what is the final set up in these results. Please clarify this aspect of the results.

    Also a missing aspect of the work is how was the training performed and how were the testing cohorts created.

  • Please rate the clarity and organization of this paper

    Good

  • Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.

    The submission does not mention open access to source code or data, but provides a clear and detailed description of the algorithm to ensure reproducibility.

  • Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?

    N/A

  • Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html

    N/A

  • Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.

    (4) Weak Accept — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal

  • Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

    I believe the authors should be able to address all of the weaknesses of the work in the rebuttal phase and at the same time the presented quality of the work as well as novelty of the approach deem it fair to me to recommend a weak accept rate

  • Reviewer confidence

    Very confident (4)

  • [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.

    N/A

  • [Post rebuttal] Please justify your final decision from above.

    N/A



Review #3

  • Please describe the contribution of the paper

    This paper proposes a physics-informed INR framework for CEST reconstruction. The key idea is to treat spatial and spectral coordinates differently: spatial coordinates are encoded with hash encoding, while the frequency coordinate is encoded with a set of learnable Lorentzian basis functions. The combined representation is then decoded by an MLP to recover dense Z-spectra.

  • Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.

    The paper addresses an important problem in CEST MRI, namely reducing acquisition burden while preserving spectral fidelity. The proposed Lorentzian encoding of the frequency dimension is a reasonable way to introduce CEST specific prior knowledge into an INR framework, and this is the most interesting part of the method. The reported results against the baselines included in the paper are consistently strong across the three sampling settings, both for reconstructed spectra and for the derived APT, NOE, and MT maps. The manuscript is also generally readable, and the overall motivation is clear.

  • Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.

    The technically new part is the Lorentzian encoding of the frequency dimension. The broader ideas of using INR for CEST reconstruction, or of treating spatial and spectral dimensions differently, are not new. More importantly, closely related CEST papers have already addressed recovery or quantification from sparsely sampled data using deep learning, including Chen et al., MRM, 2024, and Cheema et al., MRM, 2025.These are more relevant comparisons than several baselines currently included in the paper, yet no direct comparison is provided.

    Terms such as “Lorentzian manifold,” “physical subspace,” and “manifold-constrained optimization” are stronger than what is actually defined. What the paper clearly specifies is a learnable Lorentzian feature map for the frequency coordinate, followed by a general INR decoder. I think the paper would be stronger and more precise if it described the method as a physics-informed spectral encoding rather than a fully constrained manifold model.

    The learning objective and optimization problem are not stated clearly enough for a methodological paper. The paper gives the forward mapping and an (L_2) loss over the sampled set, but it remains unclear how the training is organized in practice, what exactly is optimized, and how the slice-by-slice optimization mentioned later relates to the earlier mathematical description.

  • Please rate the clarity and organization of this paper

    Satisfactory

  • Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.

    The submission does not provide sufficient information for reproducibility.

  • Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?

    N/A

  • Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html

    This is a thoughtful paper on an important CEST problem, and I found the idea of using a Lorentzian encoding for the frequency dimension interesting. My main suggestion is to narrow and sharpen the claims. In my view, the specific contribution is not INR for CEST reconstruction in general, but the use of a learnable Lorentzian spectral encoding inside that framework. I would encourage the authors to explain this point more directly, to define the learning objective and optimization problem more clearly, and to moderate the language around “manifold-constrained optimization.” I also think the paper would be substantially stronger with a direct comparison to the closest recent CEST papers on sparse sampling already cited in the manuscript.

  • Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.

    (3) Weak Reject — marginally below the acceptance threshold, but would not mind if accepted, dependent on rebuttal

  • Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

    I found the paper technically sound at a high level, and the central idea is more meaningful than a generic INR variant. However, I think the contribution is still somewhat incremental for a MICCAI methodological paper. The strongest aspect is the Lorentzian encoding of the frequency dimension, but the paper does not compare directly with the most relevant CEST-specific prior work, and the validation remains limited to retrospective experiments on healthy subjects. The formulation also needs to be stated more clearly. For these reasons, I currently place the paper slightly below the acceptance threshold.

  • Reviewer confidence

    Confident but not absolutely certain (3)

  • [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.

    Accept

  • [Post rebuttal] Please justify your final decision from above.

    After considering the rebuttal, my final recommendation is to accept. The rebuttal clarified several points that were central to my initial concerns, particularly the reference used for the quantitative metrics, the subject-specific INR optimization protocol, and the pointwise L2 objective over the acquired sparse Z-spectra. These clarifications make the experimental setup and optimization procedure sufficiently clear.

    My view of the contribution remains measured. The strongest contribution is the Lorentzian physics-informed encoding of the spectral coordinate for CEST reconstruction, rather than INR for CEST in general or a fully established manifold-constrained optimization framework. I still think the manuscript would benefit from more precise language around “manifold-constrained optimization” and a clearer framing of the method as a physics-informed spectral encoding within a subject-specific INR framework.

    The submitted work contains a meaningful methodological idea, the CEST-specific spectral encoding is technically reasonable, and the reported results are promising. On balance, I view the paper as slightly above the acceptance threshold.



Author Feedback

GENERAL RESPONSE TO ALL REVIEWERS We thank the reviewers for recognizing our Lorentzian spectral encoding as a novel, meaningful contribution to INR frameworks. We prioritize addressing every clarification below. All conceptual clarifications will be incorporated into the final manuscript to ensure theoretical completeness. 1.COMMON MAJOR CONCERNS: 1.1 TABLE 2 REFERENCE METRICS (ADDRESSES R2-Q1) All quantitative metrics in Table 2 are computed against the reference maps derived from the FULLY SAMPLED (97-offset) spectra, matching the reference maps in Figure 3.The “Direct MPLF” (fitted directly on under-sampled data) is included purely as a minimum baseline. It serves as a lower bound to guarantee that all evaluated interpolation methods yield positive improvements over direct sparse fitting, preventing any negative degradation. We have modified the manuscript to clarify this rationale. 1.2 TRAINING DETAILS AND DATA SPLIT (ADDRESSES R1-Q1, R2-Q2) Regarding the data split, our framework follows the subject-specific optimization paradigm inherent to Implicit Neural Representations (INR). An independent continuous mapping f(x, y, ∆ω)→I is trained from scratch for each individual subject.

  • DATA: For all 18 participants, magnetic resonance images were acquired in the axial plane, with the field of view (FOV) specifically positioned to cover the midbrain region. A single 128x128 slice with matrix of 128x128 and 97 frequency offsets was acquired per patient.
  • SPLIT: Models are trained independently per subject using sparse offsets (21, 29, or 39 points; Fig. 2). A 21-offset strategy provides 344,064 training pairs (128x128x21). Remaining offsets form the test set.
  • OPTIMIZATION: Training involves 10,000 epochs and the final epoch model is used for evaluation. 2.SPECIFIC REVIEWER CONCERNS: REVIEWER 1
  • ABLATION STUDY: To clarify module contributions while adhering to MICCAI policy, we performed an internal check on a randomly selected single case under the 29-offset sampling condition. Ablating Spatial MHE (using Fourier PE+LE) degraded boundaries and caused PSNR to drop to 52.20 dB. Conversely, ablating Spectral LE (replacing it with generic MHE + Fourier) dropped PSNR to 42.03 dB. Our proposed joint MHE+LE approach achieves the highest fidelity (54.44 dB) on this case. These empirical findings validate our decoupled design, and we will explicitly discuss the orthogonal contributions of both modules in the revised text. REVIEWER 3
  • COMPARISON WITH PRIOR WORK: We compared our approach with previous works using a 5-fold cross-validation on an 18-subject dataset with 29-offset acquisition. First, Cheema’s 2D U-Net struggled on our single-slice data (PSNRs ~20 dB vs ours >31 dB) without their larger cross-slice training cohorts. Second, to evaluate Chen’s 1D spectral U-Net, we heavily augmented our data to match their massive training pool. While this fully-supervised U-Net achieved slightly higher Z-spectra interpolation fidelity (PSNR 57.94 dB vs our 56.53 dB), our strictly unsupervised approach generated highly competitive parameter maps (APT PSNR 31.60 vs 31.54 dB) without requiring any external training data or ground-truth labels. Importantly, the parameter maps produced by the 1D U-Net are quite noisy, whereas our unsupervised Lorentzian prior acts as a natural physical regularizer to effectively denoise the output. We will explicitly highlight this paradigm contrast in the revised text.
  • OPTIMIZATION OBJECTIVE: The objective minimizes the point-wise L_2‭‬ loss between the INR prediction and acquired sparse Z-spectra. For our slice-by-slice 2D implementation, the network takes 2D spatial coordinates (x, y) and frequency offset ∆ω ‭‬‭‬as inputs. The weights are optimized exclusively over the acquired sparse coordinates of that specific slice. The explicit loss equation and step-by-step training process will be added.‬‬‬‬‬‬




Meta-Review

Meta-review #1

  • Your recommendation

    Invite for Rebuttal

  • Please justify your decision. In case you deviate from the reviewers’ recommendations, explain in detail the reasons why. In case of an invitation for rebuttal, clarify which points are important to address in the rebuttal.

    This paper received multiple positive reviews with some key concerns that should be addressed in rebuttal. All three reviewers agree that the contribution of Lorentzian spectral encoding is a novel and meaningful way to contribute CEST-specific features into INR frameworks. Novelty and limitations are recognized and appropriate. This rebuttal should focus on a few key concerns that reviewers highlighted: 1) clarify the reference used in Table 2 2) provide data splitting and training details 3) moderating the language specific such as “Lorentzian manifold” to better reflect theory that is substantiated in the results (R3)

  • After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.

    Accept

  • Please justify your recommendation.

    The authors have addressed the reviewers’ concerns to my satisfaction. The paper makes a solid methodological contribution with a strong technical foundation and promising results. These reviews put it above the acceptance threshold in my mind.



Meta-review #2

  • After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.

    Accept

  • Please justify your recommendation.

    This paper proposes a physics-informed INR framework for CEST reconstruction that decouples spatial and spectral coordinates, encoding the frequency dimension through learnable Lorentzian basis functions to inject CEST-specific prior knowledge. All reviewers agree that the Lorentzian spectral encoding is a novel and meaningful way to incorporate CEST physics into an INR framework, that the method is clearly described, and that the reported gains across sampling settings and metabolite maps are consistent. The rebuttal clarified the points central to the reviewers’ concerns, including the reference maps used for the Table 2 metrics, the subject-specific INR optimization protocol and data split, and the point-wise L2 objective over the acquired sparse Z-spectra, and the reviewer who had initially placed the paper below the threshold confirmed these resolved the central ambiguities and moved to acceptance. The main remaining reservation is that the terminology around “manifold-constrained optimization” is stronger than what the formulation establishes, and the authors are encouraged to moderate this language toward a physics-informed spectral encoding within a subject-specific INR framework, as committed.



Meta-review #3

  • After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.

    Accept

  • Please justify your recommendation.

    This submission proposes a physics-based encoding for CEST MRI reconstruction. R1 and R2 liked the explicit integration of CEST physics with spectral approach. R3 disliked missing baseline comparisons and clarity issues. The rebuttal addressed these concerns. Consensus is on acceptance. For all these reasons, and with respect to others submissions, the recomemdnation is towards Acceptance.



back to top