acceptodds
Under review as a conference paper at ICLR 2027

AC-Diff: Anatomy–Contrast Disentangled Diffusion for Multi-Vendor Liver MRI Harmonization

Abstract

Magnetic resonance imaging (MRI) exhibits substantial scanner and protocol dependent appearance variations, making harmonization challenging without distorting patient specific anatomy, particularly in abdominal imaging. We propose AC-Diff, an anatomy-contrast disentangled diffusion framework that formulates MRI harmonization as a factor-specific generative intervention. Using aligned multi-contrast supervision and cross-patient factor swapping, AC-Diff learns to separate anatomical structure from acquisition-dependent contrast. Unlike conventional image or latent diffusion, AC-Diff restricts stochastic generation to the contrast subspace while the source anatomy bypasses diffusion and is directly reused during reconstruction. Experiments on an in-house multi-vendor cohort and the external Duke Liver MRI dataset demonstrate improved target-domain alignment with strong structural preservation. In downstream liver segmentation, AC-Diff improves Dice from 0.942 for the original inputs to 0.954 for the harmonized images. These results support contrast-specific latent generation as a promising approach to anatomy-preserving MRI harmonization.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
MRI Harmonization, Anatomy–Contrast Disentanglement, Contrast-Space Diffusion, Anatomy Preservation, Medical Image Translation, Multi-Vendor MRI

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_7FsyOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

AC-Diff proposes a two-stage framework for multi-vendor liver MRI harmonization. It learns anatomy–contrast representations from aligned multi-contrast acquisitions, then applies diffusion only to contrast features while reusing source anatomy features. Experiments show improved target-distribution alignment and modest downstream segmentation gains on an external dataset.

Critical strengths
  1. Provides a clear architectural mechanism for restricting stochastic generation to contrast features.
Critical weaknesses
  1. The factor-space analysis uses AC-Diff’s own encoder, while the ablations primarily establish the benefit of contrast-only diffusion.
  2. A baseline using TotalSegmentator-derived organ masks to condition harmonization would help assess whether learned disentanglement offers advantages over explicit anatomical guidance.
Questions for authors
  1. Could the authors independently evaluate anatomy–contrast disentanglement, particularly anatomical information leakage into contrast features, rather than relying on AC-Diff’s own encoder?
  2. Could the authors compare against a harmonization model conditioned on TotalSegmentator-derived organ masks to assess the benefit of learned disentanglement over explicit anatomical guidance?
3 comments

We thank the reviewer for the constructive suggestions. To address the concerns regarding independent validation of anatomy–contrast disentanglement and comparison with Mask-guided LDM, we conducted two additional experiments.

1. Independent Validation of Anatomy–Contrast Factorization

To assess anatomical information leakage independently of our reconstruction decoder, we froze AC-Diff checkpoint and trained separate linear and nonlinear segmentation probes on zaz_a and zcz_c. Probes were independently trained and evaluated on AMOS22 (30/9/21 subjects) and Duke (63/13/15 subjects) using subject-disjoint splits.

AMOS22

ProbeMean Dice ↑
Position Prior0.5183
zaz_a — Linear0.6011
zcz_c — Linear0.0758
zaz_a — Nonlinear0.8330
zcz_c — Nonlinear0.6941

Duke Liver MRI

ProbeMean Dice ↑
Position Prior0.5362
zaz_a — Linear0.3276
zcz_c — Linear0.0655
zaz_a — Nonlinear0.6784
zcz_c — Nonlinear0.5909

Nonlinear probes consistently recover more anatomical information from zaz_a than zcz_c, with paired Dice differences of 0.1389 (95% CI [0.1082, 0.1732]) on AMOS22 and 0.0875 ([0.0606, 0.1175]) on Duke.

However, anatomical information remains recoverable from zcz_c. Relative to the position prior, the nonlinear zcz_c probe improves Dice by 0.1758 ([0.1241, 0.2268]) on AMOS22 and 0.0547 ([-0.0074, 0.1162]) on Duke, with the latter difference not clearly distinguishable from zero.

Linear probes provide supplementary comparisons, although their interpretation is limited by substantial empty-slice false positives, particularly for zcz_c.

Overall, these results support preferential anatomical encoding in zaz_a, rather than complete anatomy–contrast disentanglement. Confidence intervals were estimated by subject-level bootstrapping conditional on the fixed trained seeds.

2. Comparison with Mask-Guided Harmonization

Following the reviewer's suggestion, we implemented a Mask-guided LDM trained on our in-house multi-vendor abdominal MRI dataset, using anatomical masks extracted by TotalSegmentator MRI as conditioning inputs. We evaluated the trained model on the external Duke Liver dataset, where conditioning masks were predicted from the original Duke images without using ground-truth annotations.

MethodPhilips Wass. ↓Liver Dice ↑HD95 (mm) ↓
Original0.18100.941820.57
Mask-guided LDM0.18480.94466.59
AC-Diff0.14610.95354.86

Although Mask-guided LDM preserves high source-image similarity (MS-SSIM: 0.9916; PSNR: 34.57 dB), it does not improve target-domain intensity alignment over the original images (Wass. 0.1848 vs. 0.1810). In contrast, AC-Diff achieves better target-domain alignment and downstream liver segmentation performance.

These results highlight the advantage of AC-Diff over explicit mask conditioning in achieving effective target-domain alignment.

Summary

Overall, our experiments show that zaz_a contains more anatomical information than zcz_c, although the two are not completely disentangled. The mask-guided comparison also shows that AC-Diff achieves better harmonization and downstream segmentation performance. We will include these results and discuss the remaining limitations in the revised manuscript.

Thank the authors for addressing my concerns. I have raised my score to lower acceptance.

Thank you for your positive reassessment and constructive feedback, which have helped us improve the manuscript.

Reviewer_8NEoOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes a harmonization method for MRI, which disentangles anatomy and contrast. They trained the anatomy and contrast encoder based on their internal dataset where each patient (patient A, B, C ...) has multiple contrasts (X, Y, Z ...) . During trianing, the contrast feature of a image of Patient B, contrast X, are used to combine with anatomy features from another Patient A's anatomy feature, and the diffusion model used the feature to generate Patient A's contrast X's image. Patient A's contrast X image is used as supervision signal. They validated on their internal test split and external Duke liver dataset. Showing they have better harmonization and segmentation performance than baselines.

Critical strengths

The paper is very easy to follow, the method is straightforward and intuitive. Their results shows that the harmonization result is good on their internal dataset.

Critical weaknesses
  1. the major novelty of disentangling anatomy and contrast is nothing new. I don't see much contribution from a method point of view. Meanwhile, the disentanglement comes from getting images from the same patient with multiple contrasts, which greatly limits the applicability of the method. Also the comparison with those existing disentangling methods like CALAMITY, CACD e.t.c. are not presented (author claims failure to reimplement). This is not convincing.
  2. The experiments are too small. Only tested on 10 in house data and 29 external duke liver dataset. Meanwhile, the harmonization result of duke liver in Table1 does not make sense. SSIM/PSNR comparing harmonized duke liver image to original image does make sense and should not be listed there.
  3. The internal dataset that the author trained the model on is very valueble. I suggest the author to use those data as validation set, while develop a model that's trained on much broader publically available MRI, with scale, and with a thorough validation on more datasets. The current contribution does not match the bar for ICLR.
Questions for authors

No.

Reviewer_rKaQOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This submission works on the MR image harmonization to relieve the image variations due to scanner and protocol differences. This is an important task in medical image analysis, benefiting the downstream tasks such as image segmentation, domain generalization, etc. The proposed approach is based on the anatomy–contrast disentanglement and diffusion model, by factoring the multi-contrast and anatomy structures, with the anatomy factor directly used for the generation. The proposed approach was evaluated with downstream liver segmentation and shows improved performance.

Critical strengths
  1. This work proposed a reasonable approach for MRI harmonization by factoring the anatomy structure from the multi-contrast information using disentanglement learning.

  2. The disentangled anatomy structure was taken as condition in the diffusion model for the harmonized image generation, demonstrating better structure preservation and downstream liver segmentation performance.

Critical weaknesses

I have several concerns on the novelty in the methodology and experimental justifications.

  1. The idea of disentanglement learning of anatomy embeddings and multi-contrast information is based on the cross-patient factor swapping and contrastive supervision. The generation is based on the diffusion model with revised conditions. Though the idea is reasonable, these techniques lacks significant novelty considering the previous disentanglement learning methods. The key novelty should be clarified.

  2. It should be clarified in the methodology sections what the contrast mean (in anatomy–contrast representation). One understanding is the different scanning protocols such as T1, T2, etc. Another understanding is the same contrast (e.g., T1) but from different imaging vendors. The paper writes "mDixon contrast A ∈ {W, F, IP,OP}", but should with clearer explanations of these contrasts.

  3. In the stage B, the training objectives seem to be complex, with several weights in the loss terms. How to tune these weights and the effect of these weights should be analyzed.

  4. The downstream segmentation task was evaluated based on two datasets, including multi-vendor cohort and Duke Liver MRI dataset. Why the generation is on the whole CT slices but only evaluated on the liver segmentation? The proposed segmentation experiments is based on finetuning TotalSegmenter which should be able to segment more organs.

  5. The comparisons with harmonization methods are limited, the authors are suggested to survey and include more MRI harmonization methods in the experiments. Moreover, more detailed analysis should be presented, e.g., the hyper-parameters analysis, the effect of introducing anatomy structure conditions in the diffusion model, etc.

Questions for authors

As discussed above, my major concerns lie on the unclear novelty considering the literatures of the disentanglement learning, the limited downstream segmentation organs and limited comparisons with other harmonization methods and hyper-parameters analysis. These concerns should be clarified in the rebuttal phase.

1 comment

Following our earlier response, we conducted additional experiments to further address the reviewer's concerns regarding anatomy conditioning, baseline comparisons, and downstream evaluation.

1. Ablation for anatomy conditioning

To assess the effect of anatomy conditioning, we retrained Stage B without the anatomy embedding in the diffusion condition, while retaining all anatomical inputs to the decoder.

ConfigurationMS-SSIM ↑PSNR ↑Wass. ↓
w/o Anatomy Conditioning0.85321.9180.132
Full AC-Diff0.89125.0540.113

The full model consistently outperforms the ablation, supporting the benefit of anatomy-conditioned denoising even when anatomical information remains available to the decoder.

2. Additional harmonization baselines

We have added CALAMITI (Zuo et al., 2021) and CACD (Scholz et al., 2025), addressing the limited comparison with existing disentanglement-based methods.

Internal cohort — image-level evaluation:

MethodMS-SSIM ↑PSNR ↑Grad. ↑Wass. ↓
ComBat0.76019.9980.9670.222
CycleGAN0.69520.7180.9560.165
StarGAN v20.52714.1880.9390.160
Harmonizing Flows0.76021.4600.9700.272
DDPM0.90323.0550.9750.145
CALAMITI0.89823.9050.9670.149
CACD0.86122.0030.9760.121
AC-Diff0.89125.0540.9790.113

MS-SSIM/PSNR assess source-image similarity, whereas Wasserstein distance measures target-domain intensity-distribution alignment.

External Duke Liver MRI — target-domain alignment and downstream segmentation:

MethodWass. ↓Dice ↑HD95 ↓
Original—0.94220.570
ComBat0.1680.9426.973
CycleGAN0.1650.87328.288
StarGAN v20.1520.9357.541
Harmonizing Flows0.1560.9409.379
DDPM0.1910.9436.518
CALAMITI0.1580.93711.613
CACD0.1740.9408.217
AC-Diff0.1460.9544.856

HD95 is reported in mm. AC-Diff achieves the lowest target-distribution distance and the best downstream liver Dice and HD95 among the evaluated methods.

3. Expanded multi-organ evaluation on AMOS22

We additionally evaluated liver, spleen, left kidney, right kidney, and pancreas segmentation on AMOS22, extending beyond the original liver-only experiments.

We used 60 labeled MRI cases with a shared patient-level 30/9/21 training/validation/test split. For each method, the segmentation model was initialized from the same TotalSegmentator MRI weights and independently fine-tuned for 20 epochs under the same protocol.

AMOS22 — five-organ segmentation Dice:

MethodLiverSpleenKidney-LKidney-RPancreasMean ↑
Original0.98260.97760.96730.96770.86820.9527
ComBat0.97830.96430.96290.96450.86350.9467
CycleGAN0.97150.95390.95270.95180.82580.9311
StarGAN v20.95600.91290.94100.94090.76340.9028
Harmonizing Flows0.98110.97520.96520.96380.87210.9515
DDPM0.98040.97330.96400.96290.86520.9492
CALAMITI0.97410.96200.95850.95700.85060.9405
CACD0.97970.97310.96490.96380.86420.9491
AC-Diff0.98360.97770.96500.96500.86910.9521

Across five abdominal organs, AC-Diff achieves a mean Dice of 0.9521, closely matching the original images (0.9527) and achieving the highest mean among the evaluated harmonization methods. This extends our downstream validation beyond liver segmentation, while showing that the effects vary across organs.

Clarification on the Previous CACD Adaptation. In our original submission, we reported an unsuccessful CACD adaptation while acknowledging possible implementation issues. Due to the memory overhead of data augmentation, we adopted multi-GPU training to accommodate larger batch sizes. However, the contrastive loss was initially computed independently on each GPU rather than across the global batch, omitting cross-GPU comparisons. After correcting the distributed loss computation, we successfully trained CACD and included it in the updated quantitative comparisons. The corrected implementation and training configurations will be documented in the revised appendix.

We appreciate the reviewer's suggestions, which have helped us strengthen the experimental evidence and clarify the contributions of AC-Diff.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of Manuscript Revisions and Additional Experiments

Dear Reviewers and Area Chair,

Thank you for your constructive feedback. We have revised the manuscript to address the concerns raised during the review process. The major revisions are summarized below.

1. Clarification of methodological contributions. We have revised the title to AC-Diff: Anatomy-Guided Contrast-Subspace Diffusion for Multi-Vendor Liver MRI Harmonization and updated the abstract, introduction, related work, and method descriptions accordingly. These revisions emphasize anatomical guidance during contrast generation while restricting stochastic modification to the contrast subspace, without altering the original framework.

2. Extended baseline comparisons. We have incorporated CALAMITI and CACD into the quantitative and qualitative evaluations, with additional implementation details provided in the appendix.

3. Additional ablation studies. We have reorganized the ablation study to better distinguish the contributions of factor swapping, diffusion scope, and anatomical conditioning. In particular, we include a new controlled experiment examining the role of anatomical guidance during contrast generation.

4. Independent representation analysis. We have added independent latent probing experiments on AMOS22 and Duke to assess anatomical information encoded in the learned representations. We also discuss the residual anatomical information in the contrast latent.

5. Extended downstream evaluation. We have added multi-organ segmentation experiments on AMOS22 and a mask-guided latent diffusion baseline to further assess anatomical utility and compare learned anatomical guidance with explicit mask-based conditioning.

6. Clarification of practical applicability and limitations. We have clarified that the aligned multi-contrast mDixon data are required for Stage-A supervision rather than inference, and expanded the discussion of the method's limitations.

We hope these revisions clarify the methodological contributions and provide more comprehensive experimental evidence for AC-Diff. We appreciate your time and consideration.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.