acceptodds
Withdrawn from ICLR 2027

DiLaT: Semantic-Guided Volumetric Diffusion for Through-Plane Coherent Brain T1 Denoising

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Brain T1 magnitude images are typically stored after reconstruction, where thermal noise follows a Rician law rather than a Gaussian one. Most learned denoisers still treat each axial slice as an independent 2D image. That is convenient, but it leaves a gap: when a volume is scrolled, independently restored slices can flicker even if per-slice SSIM looks fine. We treat a short stack of contiguous slices as the restoration unit and try to keep anatomy stable along the through-plane axis, not only sharp in-plane. DiLaT does this with a latent diffusion model: a frozen VAE provides the working space, frozen DINO features extracted from the noisy magnitude supply structure when pixels are unreliable, and a spatiotemporal Transformer exchanges information across the clip. On held-out AOMIC volumes the clip model reduces through-plane inconsistency relative to strong slice-wise CNN, Transformer, and diffusion restorers, while remaining competitive on SSIM, especially at higher noise. A zero-shot OASIS-1 test follows the same pattern. Ablating cross-slice attention or dropping the semantic condition at sampling both hurt volumetric coherence, which is the claim we care about.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
MRI denoising, image restoration, diffusion models, medical imaging, spatiotemporal transformers

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_T4QTOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes a denoising framework for 3D T1-weighted MRI volumes that combines DINO-based conditioning with a Latte-style latent diffusion model. By jointly restoring adjacent slices through spatial and cross-slice attention, the framework aims to preserve anatomical continuity. Experiments on synthetically corrupted AOMIC and zero-shot OASIS data report improved through-plane consistency over selected slice-wise baselines, with competitive SSIM at higher noise levels.

Critical strengths

The paper is well written and easy to follow. Using frozen DINO features from noisy slices to condition a clip-level latent diffusion model is interesting. The spatiotemporal denoiser exchanges information across adjacent slices, explicitly considering anatomical continuity rather than processing each slice independently. Experiments on synthetically corrupted in vivo images show improved through-plane consistency relative to the F=1 variant and the selected slice-wise baselines, while maintaining competitive SSIM at higher tested noise levels.

Critical weaknesses

Limited comparison with MRI-specific denoising methods. Although RDDM is included, comparisons with MRI-specific methods such as DDM², Di-Fusion, and Corruption2Self are missing. These self-supervised methods are relevant for assessing the benefit of requiring clean training targets. The authors should compare against applicable methods or explain specific incompatibilities. Di-Flow may require discussion because it exploits neighboring slices for diffusion-based denoising of a single MRA volume. A suitable 2.5D or 3D baseline would further clarify the advantage of the proposed architecture.

Unclear practical relevance to low-SNR T1-weighted MRI. All experiments use synthetic Rician noise added to reference images, leaving performance on genuinely noisy acquisitions unverified. The authors should identify the intended acquisition setting and justify whether the tested noise levels and characteristics are representative of that setting. Evaluation on genuinely low-SNR in vivo T1-weighted images would substantially strengthen the practical relevance. Evaluation on another MRI contrast, such as diffusion MRI, would additionally demonstrate the framework’s broader applicability.

Limited methodological novelty. The framework largely adapts Latte’s existing spatial–temporal attention architecture to MRI slices, with DINO conditioning and task-specific regularization. While this adaptation may be useful, the current comparisons and ablations do not sufficiently establish the added value of these modifications beyond applying an existing video diffusion backbone to volumetric denoising.

Limited validation of anatomical continuity. Reporting both T-LPIPS gap and T-ΔMAE provides useful evidence of through-plane consistency. However, these metrics do not fully establish faithful preservation of 3D anatomy: similar inter-slice perceptual distances need not imply correct anatomical transitions, and errors shared across adjacent slices can cancel in T-ΔMAE. Given that anatomical continuity is a central motivation, complementary validation—such as sagittal/coronal visualizations, inspection of sliding-window boundaries, or a structure-sensitive downstream evaluation—would strengthen the claim.

Questions for authors

See Weakness

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

The paper notes that traditional MRI denoising methods process 2D slices individually, which can lead to a lack of consistency between adjacent slices. To address this, the authors employ techniques such as VAE and DINO; experimental results demonstrate that their approach offers distinct advantages over existing methods.

Critical strengths
  1. The paper notes that traditional slice-by-slice denoising methods can lead to structural discontinuities in 3D volumetric data. By introducing Latent Diffusion for joint denoising of short MRI sequences, the paper makes a constructive contribution to this field.

  2. The experimental evaluation is comprehensive, covering performance benchmarking, cross-dataset testing, ablation studies, and qualitative visualizations.

  3. The overall structure of the paper is clear, and both the content and the model architecture diagrams are well-presented.

Critical weaknesses
  1. The experiments are based on a single training run with a single random seed; no variance or confidence intervals are reported. The authors explicitly state: "The Wilcoxon tests speak to subject variation for that seed, not to whether a second run would land on the same checkpoint."

  2. The study relies on a single dataset (AOMIC) for both training and held-out evaluation. While OASIS-1 is included as an additional dataset, it is not used for training.

  3. Key hyperparameters lack ablation studies; Appendix F acknowledges that "several knobs were fixed early and not presented as ablations."

  4. There is no analysis of inference efficiency or computational costs (e.g., FLOPs, VRAM usage).

  5. The paper lacks a theoretical analysis or an explanation of the underlying mechanisms, relying solely on empirical results without providing a theoretical motivation.

  6. The components—VAE, DINO, Latte-style Transformer, and DDPM are all existing technologies. The core components lack novelty, and there is a notable absence of independent algorithmic innovation.

Questions for authors
  1. How do the authors ensure that the diffusion process does not hallucinate or remove small anatomical structures?

  2. Why was the sixth DINO transformer block selected?

  3. The input is upscaled from 256×256 to 1024×1024. This fourfold spatial expansion adds no genuine image information while increasing the computational overhead for the DINO encoder. What is the rationale for this design choice?

  4. The paper lacks baselines involving 3D deep networks. Given that the primary objective is to ensure 3D continuity, it is problematic that all external baselines are 2D-based. Why were no competitive 3D methods included for comparison?

Reviewer_pEexOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes DiLaT, a latent diffusion model that denoise six adjacent brain MRI slices together to improve consistency across slices. It uses frozen DINO features as guidance and a video style transformer to exchange information between slices. Experiments on AoMIC and OASIS show improved cross-slice consistency over slice wise baselines.

However, as providing adjunct slices as inputs are not new in 2.5 denoising/segmentations; also a lot of existing methods are already 3D denoising, i don't see enough contributions here.

Critical strengths
  1. This paper addresses consistency across MRI slices in denoising.
  2. The F=1 control provides useful evidence that neighboring slices improve restoration within the proposed model.
  3. Evaluation includes held-out AOMIC volumes and zero-shot testing on OASIS, with improved cross-slice metrics.
Critical weaknesses
  1. The contribution over existing volumetic denoising is unclear. All external baselines are slice wise; no competitive 3d or multi-slice denoisor is included. Currently experiments can not establish superiority over existing volumetric methods.
  2. The DINO contribution is not isolated. Removing DINO just removes all the condition the model gets from image, it is not surprising model drops a lot (it becomes generating from pure noise). The authors need to consider condition directly on noisy images or just condition on anatomy.
  3. The evaluation is narrow. All tested noise levels exceed the training range, while OASIS evaluates only center slices per volume. This provides limited evidence for generated full volume consistency.
  4. Anatomical fidelity is insufficiently assessed. The model generates from noise using DINO features as its image condition. I have high doubt that this feature space would lose small structures and pathology fidelity.
  5. The entire contribution of this work is unclear to this field, also with limited baselines and expeirments.
Questions for authors
  1. How does DiLaT compare with a 3D-Unet or multi-slice denosier given the same neighboring slices?
  2. Does DINO conditioning outperform raw-image conditioning or other feature conditioning? i feel a mask condition may works better for anatomical details.

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

Related papers

Discussion (0)

Sign in to comment.