List of Papers Browse by Subject Areas Author List
Abstract
Cross-modality synthesis from CT to MRI provides MRI-level soft-tissue detail without additional scans, yet existing approaches face three major limitations: the substantial modality gap between CT and MRI, slice-level quality variation without explicit estimation, and underutilization of 3D anatomical continuity. While prior work has primarily focused on MRI-to-CT synthesis, CT-to-MRI generation remains challenging. We propose LGN (Learning from Good Neighbors), a framework for synthesizing MRI from CT, consisting of three components: (1) a Latent Diffusion Model (LDM) for draft MRI generation, (2) a slice-level Quality Predictor trained via LPIPS-based teacher–student learning, and (3) a 3D context-aware Cross-Attention Refiner that selectively improves suboptimal slices using high-quality neighboring slices. Experiments on 398 patients demonstrate consistent improvements in LPIPS, PSNR, and MAE with reduced artifacts and enhanced anatomical consistency, highlighting the effectiveness of quality-guided cross-slice refinement for CT-to-MRI synthesis. The source code is available at https://github.com/YouKyoung-Na/LGN.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/0206_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/YouKyoung-Na/LGN
Link to the Dataset(s)
N/A
BibTex
@InProceedings{NaYou_Learning_MICCAI2026,
author = { Na, You-Kyoung AND Choi, Kang-Ho AND Yang, Hyung-Jeong AND Kim, Jahae AND Cho, Yeong-Jun},
title = { { Learning from Good Neighbors: Slice-wise Quality-aware CT-to-Diffusion MRI Synthesis } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16890},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
The paper proposes LGN, a quality-aware CT-to-DWI synthesis framework that combines latent diffusion, slice-level quality prediction, and cross-slice refinement using higher-quality neighboring slices. The main idea is to exploit slice-wise quality variation and 3D anatomical continuity to improve synthesized MRI quality.
- 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.
1.The framework is clear and reasonably well organized. 2.The ablation studies are useful and support the proposed design choices. 3.The method improves LPIPS, PSNR, and MAE over several baselines.
- 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.
1.The evaluation is relatively limited, as the test set includes only 23 patients from a single center, which raises concerns about the robustness and generalizability of the reported results. 2.The clinical utility of the method is not quantitatively validated, as the claimed lesion-related benefits are supported only by qualitative examples rather than dedicated lesion-level or downstream clinical analyses. 3.The quality predictor is trained using LPIPS-based pseudo-labels, which may not reliably reflect true perceptual quality or clinical relevance. 4.Given that the main motivation is to improve anatomical consistency across slices, it may be more appropriate to focus on a genuinely 3D modeling strategy rather than relying primarily on slice-wise processing with neighboring-slice refinement. A more native 3D design could better capture volumetric structure and improve 3D anatomical continuity.
- 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 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
Please see the weakness section
- 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?
The problem is important and the paper is clearly written, but I find the methodological novelty modest and the experimental validation not yet strong enough for acceptance. In particular, the evaluation is limited to a small single-center test set, and the claimed clinical relevance is not supported by downstream or reader-based validation.
- Reviewer confidence
Confident but not absolutely certain (3)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
Reject
- [Post rebuttal] Please justify your final decision from above.
The rebuttal provides useful clarifications on the DWI target, dataset composition, and slice-wise refinement design. However, my main concerns are only partially addressed. The evaluation still relies on a small single-center test set of 23 patients, and multi-center validation remains future work. The claimed clinical utility is not quantitatively validated, since lesion-level evaluation, downstream clinical analysis, and reader studies are deferred to future work. The LPIPS-based quality predictor is also only shown to be consistent with its pseudo-labels, rather than clinically meaningful quality. Overall, the rebuttal improves clarity but does not fully resolve concerns about robustness, generalizability, and clinical validation.
Review #2
- Please describe the contribution of the paper
This work proposes a novel multi-component deep learning framework for CT-to-MRI (diffusion-weighted) image synthesis in patients with brain pathologies (e.g. stroke and haemorrhage). The key contribution is the integration of three complementary modules: (i) a latent diffusion model for image synthesis, (ii) a learned perceptual quality predictor based on Learned Perceptual Image Patch Similarity (LPIPS), and (iii) a cross-slice refinement mechanism using cross-attention mechanism. Unlike prior GAN-based approaches, the proposed pipeline explicitly models both perceptual quality and inter-slice consistency. Experimental validation on 23 held-out cases demonstrates improved performance over existing methods such as CycleGAN and Pix2Pix across multiple image similarity metrics.
- 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 architecture is novel and well aligned with the underlying imaging characteristics, effectively leveraging the anisotropic nature of 2D acquisitions.
- The incorporation of LPIPS as a perceptual quality measure is well motivated and provides a mechanism for improved image fidelity and potential quality assessment.
- The use of cross-slice contextual information is a strong design choice, and the authors provide convincing experimental evidence of its benefit.
- The use of a relatively large training dataset increases confidence in the robustness of the results.
- 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 paper does not sufficiently address the risk of hallucination in cross-modality synthesis (CT to DWI), where fundamentally different physical properties are being inferred. This raises concerns about clinical applicability, particularly for diagnostic use, and warrants stronger discussion and validation beyond quantitative metrics.
- The use of affine registration for creating paired training data is potentially problematic given the known geometric distortions in DWI. The impact of misregistration on model performance is not discussed.
- The choice of DWI as the target modality is not well justified, as it represents a particularly challenging synthesis task compared to more structurally aligned modalities (e.g., T1/T2).
- Key methodological details are unclear, including whether separate models were trained for different b-values or whether weights were shared or fine-tuned. One figure (3) did not render, making evaluation difficult.
- The experimental reporting lacks clarity. It is unclear whether reported metrics in tables correspond to test averages, and no information on variability (e.g., range or standard deviation across test patients) is provided.
- The study is limited to a single centre and lacks sufficient detail on dataset composition (e.g., patient diagnoses, scanner vendors, acquisition field strengths, sequence parameters), raising concerns about generalizability
- 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?
This paper presents a well-designed methodological approach to multi-slice image synthesis that is well aligned with the underlying acquisition characteristics. However, the clinical motivation and intended use case require clearer articulation, particularly given the risks of hallucination in cross-modality synthesis between imaging modalities with fundamentally different physical properties. While the methodological description is strong, limited detail on data acquisition and cohort composition limits assessment of generalizability.
- Reviewer confidence
Very confident (4)
- [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.
I beleive that the innovations within this work are presentable due to the following reasons:
- The architecture designed by the authors specifically addresses the need for 2.5D processing in images with anisotropic voxel sapcing
- The reults appear to be quite convincing in terms of qualitative and quantitaitve accuracy.
- The analysis methods for model validation are quite sound for presentation at a conference.
I would like to add a caveat that diffusion-weighted MRI is known to be distorted, and particularly problematic around the sinonasal region on MRI due to the presence of air–tissue interfaces, and independently of b-value. The authors have used affine registration, which may not completely resolve this issue, but the results seem to show that the method deals with this relatively well.
Review #3
- Please describe the contribution of the paper
The authors propose a slice-wise synthesis framework for CT-to-diffusion MRI data. The method utilizes pretrained encoders to generate latent features that the latent diffusion model employs to synthesize draft MRI slices from CT scans. The subsequent component within the framework evaluates the quality of the synthesized data and applies a perceptual metric to classify the synthesized slice in relation to the ground truth. The following module then leverages the highest-scoring neighboring slice to refine the slice of interest through a cross-attention mechanism and by fusing features from the encoded CT, the encoded draft MRI, and selected neighboring slices.
- 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 is well-written, easy to read, and well-presented. The authors combined in an intuitive manner different concepts: latent diffusion model, neighboring slice information to improve estimation or synthesis quality, and leverage an attention mechanism to find some global features at least at the neighboring slices level to correct synthesis inaccuracies. However most of these already exist, the cross attention at neighbors slices is interesting.
- 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 novelty in the paper is limited; all concepts already exist, and the paper does not propose a unified framework trained end-to-end. While proposing to improve continuity, many methods have addressed such a problem in the MRI to CT, as the author mentioned.
- 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 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
How does the proposed method consider the dependence of the diffusion MRI image on the sensitizing direction? Given that only one 3D CT scan is available, how does the model subsequently map this CT scan to each volume of the 4D diffusion-weighted imaging (DWI), such as at shell 1000?
- 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.
(2) Reject — should be rejected, independent of rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The manuscript does not present new or innovative ideas. The contributions cited by the authors are already established, and they do not present a unified or end-to-end trained framework. Furthermore, the authors appear to treat the DWI MRI data as structural MRI data and do not clarify the methodology by which their approach learns to generate DWI data for various gradient directions from the same CT scan. The paper could be improved by elaborating on how a DWI can be synthesized from a CT scan in a specified gradient direction.
- Reviewer confidence
Very confident (4)
- [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.
The authors addressed a major concern about the relationship between the DWI data and the sensitizing gradient direction.
Author Feedback
We sincerely thank the three reviewers [R1, R2, R3] for their caluable time and constructive reviews. Regarding reproducibility, we will release code, model weights, and a test-data subset within IRB constraints. We now address the main concerns by Meta-Reviewer [Meta] priority.
1.Additional DWI details [Meta, R2, R3] Synthesis target [Meta-1, R3]: Our target is trace DWI at b=0 and b=1000 s/mm², the direction-independent images that radiologists read on PACS, rather than direction-resolved 4D DW volumes. In clinical protocols, b=1000 DWI is acquired along multiple directions and combined by the scanner into a trace image, and our dataset uses these clinical trace volumes. Therefore, R3’s per-direction recovery concern does not apply, and DTI or multi-shell modeling lies outside our scope. b-value handling [R2]: As noted in Sec. 1, we train separate LDMs for b=0 and b=1000 since they differ in signal physics. All results in Tables 1–3 follow this design. Why DWI [R2]: Per Sec. 1, our scenario is emergency stroke/hemorrhage triage, where acute ischemic lesions are most reliably visible on DWI. Targeting an easier modality such as T1/T2 would undermine the clinical motivation of our work.
2.Hallucination [Meta, R2] We agree the risk is intrinsic but mitigate it via two mechanisms. The Refiner re-injects the CT condition (Eq. 3–4), anchoring outputs to the patient’s anatomy. The Quality Predictor estimates per-slice reliability (Pearson 0.986) to identify low-confidence outputs, ensuring reliability most synthesis models lack. In Fig. 3, LGN preserves true lesions while avoiding false ones in the non-bleed case. Per Sec. 1, LGN is positioned as a triage aid rather than a diagnostic replacement, with reader studies and lesion-level evaluation noted as future work (Sec. 4)
3.Registration error [Meta, R2] Preprocessing (Sec. 3.1) restricts analysis to skull-stripped intracranial tissue, which is approximately rigid. The b=0 image used for registration is free from diffusion-gradient-induced distortion. Residual misregistration affects all baselines equally, preserving Table 3’s relative comparison.
4.Per-patient variability [Meta, R2] Tables 1–3 are averages over 513 slices from 23 patients. From existing results, per-patient std at b=0 is ≈0.01 (LPIPS), ≈0.005 (MAE); LGN’s improvement is consistent across patients.
5.Novelty [R1, R3] Our contribution is the integration of LDM, neighbors, and cross-attention into a quality-aware refinement framework via slice-level quality estimation, as R3 also noted positively. Table 1 shows that adding quality-filtered neighbors improves LPIPS from 0.066 to 0.061.While 3D modeling [R1] may help volumetric continuity, full-volume 3D diffusion is costly and treats all slices uniformly, whereas our slice-wise design enables selective refinement of difficult slices.
6.Dataset and generalization [R1, R2] Our dataset comprises 398 patients at a university hospital under IRB approval (M 56% / F 44%, mean age ~70, range 26–96), centered on acute ischemic stroke including microbleed cases (Fig. 3). CT was acquired on a GE Discovery CT750 HD and MRI on a Philips Ingenia CX 3T, with DWI as trace b=0 and b=1000 s/mm². Multi-center validation is future work, and our core design (slice-wise quality estimation, neighbor-based refinement) readily extends to such settings.
7.Others Clinical utility [R1]: Lesion-level detection is feasible with our work, and as noted in Sec. 4, we plan to address it in future work with reader studies. LPIPS pseudo-label [R1]: It serves relative ranking rather than absolute scoring, and the Pearson score confirms its stability(Sec.3.2). Fig. 3 [R2]: Fig. 3 renders correctly in PC PDF viewers. We will reduce its file size in the camera-ready for mobile compatibility.
If given the opportunity for acceptance, we will incorporate the above clarifications into the camera-ready. We respectfully ask the reviewers to consider this work a meaningful contribution. Thank you.
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.
LGN proposes a quality-aware CT-to-DWI synthesis framework combining latent diffusion, an LPIPS-based quality predictor, and cross-slice refinement via cross-attention. The paper is clearly written and the ablation supports the design.
The biggest problem is one that R3 and R2 both flagged from different angles: DWI depends on specific b-values and gradient directions, yet the paper treats it as a static structural modality without explaining what the synthesized output physically represents. R2 independently flagged a related issue, noting that the choice of DWI as the target is not well justified given its fundamentally different physical properties from CT, and that hallucination risk in this cross-modality setting is not discussed. R1 and R2 both noted the evaluation is limited to 23 patients from a single center with no downstream clinical validation. R2 also questioned the use of affine registration for creating paired data given the known geometric distortions in DWI.
For the rebuttal: 1.explain what b-value and gradient configuration the synthesized output corresponds to and how the model handles this, 2.discuss hallucination risk and clinical reliability, 3.address the impact of DWI geometric distortions on the affine registration used for training data, 4.clarify whether separate models are used for different b-values and provide variability measures.
Points 1&2 will play an important role in my final comsideration.
- 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.
Two Accept and one Reject. The author(s) addressed all important issues I pointed out in the meta review. All reasons to reject not comply with MICCAI’s rebuttal rule have been ignored. This resulted no solid objection to reject the paper, so I recommend to accept it.
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.
Two reviewers recommended acceptance, while one recommended rejection. After reviewing the feedback, I believe the concerns regarding validation are valid and could be addressed in future work. However, they do not substantially undermine the main contributions of the paper. Therefore, I lean toward acceptance.
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.
The technical core of the paper is solid. The authors successfully cleared up a major misunderstanding regarding diffusion gradient directions by clarifying their use of the Trace image. While Reviewer 1 makes valid points regarding the small evaluation cohort (23 patients) and lack of downstream clinical reader studies, the other reviewers found the 2.5D neighbor-refinement architecture innovative for anisotropic data. I recommend Accept as justified above.
