acceptodds
Under review as a conference paper at ICLR 2027

Spatial-Differential Guidance and Structure Anchoring for Selective Mild Motion Artifact Correction in Head CT

Abstract

Head CT scans are critical for diagnosing acute neurological conditions, yet involuntary patient motion during scanning frequently introduces motion artifacts that degrade image quality. A particularly challenging scenario is mild motion artifacts in stationary CT, where artifacts are axially sparse but spatially entangled with normal anatomical structures. Consequently, correction requires both precise artifact localization and preservation of anatomical consistency. To address these challenges, we propose a spatial-differential and structure-anchoring framework for head CT motion artifact correction. The framework introduces an explicit gradient-based discriminative response field (DRF) as a spatial prior, which is incorporated through a spatial-differential attention (SDA) module to guide the correction network toward artifact-corrupted regions. Furthermore, we propose a semantic structure anchoring (SSA) strategy that aligns intermediate representations with artifact-free anatomical features from a pretrained 3D visual encoder, encouraging structural consistency during restoration. Extensive experiments on simulated datasets with varying artifact severity, a public dataset, and a real-world dataset demonstrate our framework's effectiveness. The proposed method achieves improved restoration quality on mild motion artifacts, reducing MAE by 22.7% and achieving 37.62dB PSNR, while maintaining strong performance across diverse evaluation settings.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Diffusion Models, Representation Alignment, Image Restoration, Motion Artifact Correction

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_6K3fOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a conditional EDM-based 3D image restoration framework to address mild motion artifacts in head CT images, which are sparsely distributed and often mixed with normal anatomical structures. The method constructs a discriminator response field (DRF) using the input gradients of an artifact discriminator and employs spatial differential attention (SDA) to guide restoration. Semantic structure anchoring (SSA) and a spectral amplitude matching loss are further introduced to preserve anatomical structures and image details. Experiments on simulated data with varying artifact severity, public datasets, and real-world data demonstrate improvements in most reconstruction metrics.

Critical strengths
  1. The problem is practically important, with clear motivation. Correcting mild motion artifacts requires both local artifact removal and preservation of normal anatomical structures. The proposed method addresses these two challenges through complementary guidance and constraint mechanisms.
  2. The reconstruction results are competitive, with comprehensive ablation studies. On data with mild artifacts, the method reduces brain-region MAE from 23.43 HU (EDM) to 18.12 HU. The authors also evaluate different combinations of components and use the same backbone architecture and number of sampling steps for diffusion and flow-based models, making the comparisons more interpretable.
Critical weaknesses
  1. Limited methodological novelty. The framework combines input-gradient saliency, attention biasing, and pretrained feature alignment. While this combination is well motivated, the methodological contribution beyond these existing techniques is not sufficiently established.

  2. Insufficient validation of artifact localization. Classifier input gradients reflect sensitivity to the classification decision and do not necessarily identify artifact regions. The paper lacks quantitative localization analysis demonstrating that DRF reliably distinguishes motion artifacts from normal anatomical structures.

  3. Unclear connection between SDA and selective correction. Equation (6) introduces a query-independent bias over key positions, which changes feature aggregation but does not directly control which output regions are modified. Further analysis is needed to establish that this mechanism produces spatially selective correction.

  4. Insufficient evidence for anatomical preservation. Reference-based metrics are reported only on motion-corrupted slices, without separately evaluating unaffected slices or unaffected regions within corrupted slices. Consequently, improved reconstruction scores alone do not substantiate the central claim of preserving normal anatomy and reducing overcorrection.

  5. Limited validation on real clinical data. Real-world evaluation relies solely on FID and KID, which measure distributional similarity but cannot establish lesion preservation or anatomical fidelity. Reader studies or lesion-focused evaluations are needed to support the method’s clinical utility.

Questions for authors

The same as the weaknesses.

1 comment
AuthorsOpenReview ↗

Response to Reviewer 6K3f

We thank the reviewer for the accurate summary and the constructive weaknesses, which have directly shaped our revision. Please see our point-by-point responses below.

W1: Limited methodological novelty. We respectfully clarify that our contribution is the framework formulation for the under-explored problem of mild motion artifact correction, not the individual techniques: (1) we convert artifact-discriminative knowledge into an explicit slice-selection prior via input gradients and inject it as an attention bias (new in motion artifact correction to our knowledge); (2) we adapt representation alignment to anchor a 3D CT diffusion restoration process to artifact-free anatomical features, which empirically prevents over-correction (Tables 1, S4).

W2: Insufficient validation of artifact localization. We agree with the reviewer. Classifier input gradients reflect the sensitivity of the classification decision rather than directly identifying artifact regions, and we have revised the manuscript accordingly. Specifically, the claims throughout the paper (Abstract, §3.1–3.3, Conclusion) have been reworded: the DRF is now described as an explicit slice-selection prior with a structure-sensitive response on gated slices, instead of as a spatial localization map of artifact regions. In addition, the revision provides quantitative validation of this gating function: the slice-level artifact discriminator attains AUPRCs of 0.9942, 0.9996, and 0.9900 on the Mild, Severe, and SynthRAD2023 test splits (new Suppl. §1.3, Table S3, Fig. S2), and artifact-free slices receive an exactly zero prior, keeping them outside the SDA gate by construction.

W3: Unclear connection between SDA and selective correction. We have added direct analyses of this connection in the revision. The new whole-volume evaluation (Suppl. Table S4) quantifies modifications to unaffected anatomy: our Mild whole-volume SSIM (0.9677) exceeds the uncorrected input (0.9556), whereas all generic baselines fall below it. The new real-data difference maps (Fig. 5) visualize where the output changes: modifications concentrate at artifact-affected boundaries while parenchyma is preserved. Mechanistically, the selectivity is realized along the axial direction by the DRF prior. Artifact-free slices receive an exactly zero response, so SDA applies no bias there, and the conditional sampler, which targets the posterior mean of the clean volume given the motion input, returns those slices essentially unchanged. The bias therefore does not need to directly control output regions; it gates which slices may be modified and steers aggregation on the gated slices.

W4: Insufficient evidence for anatomical preservation. We have added this evaluation. New Suppl. §2.1 (Table S4) reports metrics over the entire volume, which is dominated by unaffected slices and regions, and §2.2 (Fig. S3) reports per-case distributions. On the Mild Dataset, our whole-volume SSIM (0.9677) is higher than the uncorrected inputs (0.9556), whereas most generic baselines fall below the inputs (e.g., EDM 0.9441); on the Public Dataset, our framework attains the best whole-volume results across both regions; on the Severe Dataset, it remains best or second in most columns (e.g., brain-region MAE 24.15 and SSIM 0.8907, the best among all methods).

W5: Limited validation on real clinical data. We agree with the reviewer and have added the following in the revision: (1) Radiologist reader study (new Table 3, §4.5.2). Two radiologists blindly rated 72 pairs of motion-corrupted vs. corrected volumes; both assigned significantly lower interference scores after correction (median 3.0→2.0 and 3.5→3.0; Holm-corrected p = 0.005 and 0.014); (2) Real-data visualization with difference maps (new Fig. 5, §4.5.3, ±50 HU scale). The difference maps show that corrections are confined to artifact-affected boundaries while parenchyma is preserved.

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

This paper introduces a spatial-differential and structure anchoring framework for correcting mild motion artifacts in head CT images. Results show gains over selected baselines. The experiments also considered severe motions.

Critical strengths
  1. The paper is straightforward to follow up with clear structure and figures.
  2. I found the semantic structure anchoring is somehow interesting idea for motion correction, while it is simple.
Critical weaknesses
  1. The novelty is limited as those components are not new to me. The gradient, attention injection, and structure anchoring are not entirely new.
  2. Some recent baselines are missing. The validation of the proposed method is fully supported.
  3. FID and KID are not gold standard for real-world clinical datasets. Evaluations from radiologists should be provided.
Questions for authors
  1. Baselines published in this year should be considered.
  2. More rigour evaluations on real word datasets should be conducted.
  3. The overall foreground with infinity seems not accurate.
1 comment
AuthorsOpenReview ↗

Response to Reviewer Gdzv

We thank the reviewer for the careful reading. We have substantially revised the manuscript to address each concern; please see the detailed responses below.

W1: Novelty is limited (gradient, attention injection, structure anchoring are not new). Our contribution lies in the framework formulation for the clinically important yet under-explored problem of mild motion artifact correction, where three design choices are specifically motivated by the problem structure:

  1. Problem-driven formulation. Mild artifacts are axially sparse and entangled with anatomy, so correction must be selective, which is a requirement that generic restoration frameworks do not address.
  2. DRF + SDA. We convert the artifact-discriminative knowledge of a classifier into an explicit spatial prior via input gradients, and inject it as an additive attention bias. To our knowledge, deriving a slice-selection prior from discriminative gradients and injecting it into a 3D diffusion restoration process is new in the motion artifact correction literature.
  3. SSA. Representation alignment itself has precedents (and we cite them), but we adapt it to a new role: anchoring the restoration trajectory of a 3D CT diffusion model to artifact-free anatomical representations, which empirically prevents over-correction (Table 1, Table S4).

W2 / Q1: Some recent baselines are missing. In the revision, we have added the recent RDBM (Wang et al., CVPR 2026) to the comparison, and Table 2, Figure 3, and Figure 4 have been updated accordingly.

W3 / Q2: FID and KID are not gold standards; radiologist evaluation is needed; more rigorous real-world evaluation.

  1. Radiologist reader study (new Table 3, §4.5.2). Two radiologists blindly rated 72 pairs of motion-corrupted vs. corrected volumes; both assigned significantly lower interference scores after correction (median 3.0→2.0 and 3.5→3.0; Holm-corrected p = 0.005 and 0.014).
  2. Real-data visualization (new Figure 5, §4.5.3). Representative real-world cases with difference maps (±50 HU scale) show that corrections concentrate at artifact-affected boundaries while parenchyma is preserved.

Q3: The overall foreground with infinity seems inaccurate. We have revised the definition to [−100, 2000] HU (bounded by the clipping range) in §4.3 and Table 2, and the supplementary is updated accordingly.

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

This paper discussed motion artifacts in head CT images which is a very important problem in medical imaging. The paper proposed a spatial-differential and structure-anchoring method to solve this problem on mild, severe and real-world datasets, and the proposed method shows the best performance over the baselines.

Critical strengths

S1 The paper focused on motion artifacts which is important for medical imaging, and proposed method shows better image quality than other methods. And the evaluation is also comprehensive.

S2 The ablation study in Table 1 is clearly presentated, which compare the single component and different combinations of the components, which clearly shows the contribution of each component.

Critical weaknesses

W1 The evidence on the real-world dataset is not convincing. Although the authors provided the FID and KID scores on the real-world dataset, its still unclear whether these scores can reflect the image quality or not, as the evaluation of motion-corrupted medical imaging is still an open problem. The authors should provide the image on the real-world dataset, and conduct a qualitative evaluation or ask the radiologists to evalaute the image quality.

W2 The evaluation on this paper is not sufficiently convincing:

  • For the real-world evaluation, the method could generate reasonable image but violate the anatomical structure, which current metrics cannot reflect.
  • Then some numerical results like PSNR in Table 2 is suspicious. See Q1 for detail.
  • And lots of baselines in Table 2 and Figure 3 even perform worse than the input motion-corrupted images, which also raises the concern that the baselines are not sufficiently tuned so that the comparison is unfair.
Questions for authors

Q1 In table 2, for the overall foreground, the PSNR of the severe dataset is better than the mild dataset, and the conclusion between PSNR and other metrics are largely different which is uncommon. Could the authors clarify the reason for this?

Q2 From Table 2 and Figure 3, lots of existing methods perform even worse than the input motion corrupted images. Could the authors clarify whether this is due to the insufficient tuning of the baseline?

Minor questions

  • Typo in Figure 2, the part (c): "pretrained visual endoder".
  • Both equation (3) and (11) use ϵ\epsilon but represent different things.
1 comment
AuthorsOpenReview ↗

Response to Reviewer Vdoo

We thank the reviewer for the positive assessment and constructive feedback. We have carefully revised the manuscript accordingly; major changes are summarized below.

W1:

  1. Real-data visualization (new Figure 5, §4.5.3). We now show representative real-world cases with motion-corrupted input, EDM baseline, our corrected result, and per-slice difference maps (corrected − motion) under a ±50 HU color scale. The modifications are concentrated at artifact-affected anatomical boundaries, while the brain parenchyma remains largely unmodified, indicating that our method improves image quality without distorting normal anatomy on real clinical data.
  2. Radiologist reader study (new Table 3, §4.5.2). Two radiologists independently rated 72 pairs of motion-corrupted and corrected volumes from the Mild test set in a randomized, blinded fashion. Both readers assigned significantly lower artifact-interference scores to the corrected volumes (median 3.0→2.0 and 3.5→3.0; Holm-corrected p = 0.005 and 0.014), with the corrected score no higher than the motion input in 94.4% and 75.0% of cases, respectively.

W2: Anatomical consistency on real data. As shown in the new Fig. 5 difference maps, our corrections are spatially confined to gated, artifact-affected regions. In addition, the new whole-volume evaluation (Suppl. Table S4) shows that on the Mild Dataset our whole-volume SSIM (0.9677) is higher than the uncorrected inputs (0.9556). In contrast, most baselines fall below the inputs, which provides evidence that our framework preserves unaffected anatomy rather than re-synthesizing it.

Suspicious PSNR values. Please see our reply to Q1 below.

Baselines performing worse than inputs. Please see our reply to Q2 below.

Q1: PSNR is computed per case in the logarithmic (dB) domain and then averaged. As shown in Suppl. Fig. S3, a few cases lift the log-domain mean disproportionately. The medians retain the expected severity ordering (mild 28.87 dB vs. severe 27.81 dB), and the inversion appears only in the log-domain mean.

Q2: The underperformance is specific to mild artifacts and is a consequence of their axial sparsity. With only 10–30% of slices corrupted per volume, generic restoration models tend to treat the sparse mild artifacts as noise and predominantly learn to reconstruct artifact-free anatomy, leaving the corrupted slices under-corrected. Hence, on the Mild Dataset, metrics drop below the input. In contrast, the same baselines mostly outperform the input on the Severe Dataset (40–60% corrupted slices), where artifact correction dominates the training signal. Moreover, all generative baselines share the same 3D U-Net backbone, 50-step sampler, and training settings as ours (§4.4), so the comparison itself is strictly controlled.

Minor questions:

The typo "endoder" → "encoder" in Fig. 2(c) has been fixed.

In Eq. (3) ϵ denoted the small constant for numerical stability, whereas in Eq. (11) ϵ denotes the Gaussian noise. We have renamed the former to ϵ′ in Eq. (3).

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.