acceptodds
Under review as a conference paper at ICLR 2027

VascuSR: Iterative Anatomy-Aware Diffusion for CT Super-Resolution

Abstract

Generative super-resolution (SR) can recover sharp detail in computed tomography (CT) images from severely downsampled observations, but the missing high-frequency content is not uniquely determined by the low-resolution input. Consequently, visually plausible reconstructions may contain anatomically unsupported vascular detail, such as missing, spurious, or incorrectly connected vessels. To overcome this challenge, we introduce VascuSR, an iterative anatomy-aware diffusion framework that combines image reconstruction with explicit 3D vascular guidance. First, to provide richer anatomical guidance, we use an explicit 3D vascular representation that captures structural information beyond individual slices. This representation also simplifies vascular refinement by allowing a lightweight 3D generative refiner to focus on vascular structure rather than complex CT appearance. Second, to guide slice-wise SR, a lightweight anatomy controller injects the refined structure into the diffusion decoder, while iterative feedback updates this guidance from each new SR volume. On re-annotated MSD Task08 and HVM datasets, VascuSR outperforms evaluated SR baselines in segmentation accuracy (Dice and clDice) and 3D vascular structure (connectivity and mean centerline distance), while maintaining image quality comparable to its SR backbone. Compared with learning-based baselines, it reduces class-averaged connected-component counts by over 20% on both datasets and mean centerline distance by over 10% on MSD.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Diffusion models, CT super-resolution, Anatomical conditioning, Vascular reconstruction, Topology preservation

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

This paper studies 8× in-plane CT super-resolution with vascular anatomical guidance. VascuSR combines a slice-wise ResShift backbone, a frozen VSNet segmenter, a separately trained 3D rectified-flow refiner, and a lightweight mask controller. Refined vessel masks condition further SR passes. Experiments on re-annotated MSD and HVM, with an ImageCAS extension, report improved vessel overlap and structural metrics over the evaluated baselines, while image quality stays close to ResShift.

Critical strengths

(1)The problem and the separation of anatomy from appearance are well motivated. Learning vascular correction in 3D label space, while keeping HR-supervised image reconstruction, is a reasonable way to use volumetric information in slice-wise SR. The comparison with segmentation-loss fine-tuning is also relevant to this design choice. (2)Structural gains appear on both hepatic datasets, not only selected examples. Against ResShift, class-averaged component counts fall by about 21% on MSD and 33% on HVM. The zero-mask control and Appendix B are valuable here: regional intensity errors and the TotalSegmentator check provide evidence beyond the feedback segmenter's Dice.

Critical weaknesses

(1)The zero-mask ablation supports the use of anatomical conditioning, but does not tell me why the proposed 3D flow refiner is preferable to a simpler correction model. There is no raw-VSNet-mask conditioning control, no deterministic 3D refiner with the same probability input, and no N=1 versus N=8 comparison. The accuracy of the refined masks themselves is also not reported. The benefit of mask conditioning is therefore hard to separate from generative refinement and trajectory aggregation. This is the main issue in my assessment of the method. (2)The novelty appears incremental given the present evidence. Section 2.2 cites DSDM (Chu et al.) and the dual-guidance CT SR method of Chi et al., but neither appears in the experiments. SegSR (Xiao et al.) is also relevant because it exchanges updated information between segmentation and SR diffusion models. VascuSR’s explicit 3D label-space refinement and outer feedback loop are meaningful differences; I do not regard these methods as equivalent. Still, the paper does not establish what additional capability this particular design provides over existing structural guidance or mutual refinement. (3)The public datasets, patient-level partitions and optimizer settings are positive. However, the HR-to-LR degradation operator is not specified. Appendix A.1 uses refined labels from SR images during from-scratch SR/controller training, without identifying which SR model produces these images or whether the conditions are updated during training. The origin of the imperfect predictions used to train the refiner also needs clarification. The value and selection of lambda in Eq. 11 are missing for the key fine-tuning baseline. A further clarification is whether A.1’s VQ-f4-based LPIPS reference applies only to training or also to evaluation. These details affect both reproduction and the interpretation of the comparison. (4) Most structural measurements use VSNet, which also participates in feedback. The manual reference labels and the TotalSegmentator check mean that the evaluation is not simply circular. Still, the independent check covers only MSD Dice and has an acknowledged coverage mismatch. The aggregate structural results do not separately show missed branches, spurious branches and wrong connections. Table 2 also lacks a VSNet-on-HR structural reference, although Table 1 provides the overlap reference. I would like stronger evidence that the gains reflect correct vessel recovery, rather than mainly improved compatibility with the feedback segmenter. (5)The first feedback update does show a useful gain.On MSD, however, HV Dice changes from 0.7077 at round 1 to 0.7072 at round 3, while PV clDice improves from 0.5158 to 0.5286. Table 3 implies about 141 seconds for the first-feedback output and 380 seconds for K=3. Thus, some later gains exist, but their value is not clear relative to the extra cost. I do not require monotonic improvement or real-time inference. I need a clearer validation-based choice of K and evidence that updating the condition helps beyond the first round. (6) Quality results are reported without patient-level confidence intervals or variation over sampling seeds. The hepatic test sets contain 30 and 29 patients. Some comparisons are close: on MSD, HV Dice is 0.7072 versus 0.7010 for segmentation-loss fine-tuning; on HVM, round 1 to round 2 changes PV Dice from 0.7113 to 0.7117. These examples do not invalidate the larger structural gains. However, without uncertainty estimates I cannot tell which smaller differences and iteration choices are stable. Repeated inference and paired patient-level analysis would be more informative than additional decimal places. (7) I would not make tumor detection a major requirement for a vascular SR paper. Still, the refiner corrects vessel probabilities without directly seeing CT intensities, and its output guides reconstruction of the CT volume. There is no targeted analysis of disease-related vessel changes or lesions near vessels. Average vessel metrics leave this boundary of applicability open. I would value an audit of true narrowing, occlusion or tumor-related vessel deformation where such cases are available. This is a limited validation gap, not evidence that VascuSR erases pathology.

Questions for authors

(1)With matched input information, training data and controller capacity, how do raw VSNet-mask conditioning and a deterministic 3D refiner compare with FlowSeg3D? Please also compare N=1 with N=8 using appropriately validated decoding settings, and report the refined mask’s Dice/clDice before it is passed to SR. How much of that improvement transfers to the final reconstructed image? (2)Can you compare with one representative structure-guided CT SR method, such as DSDM or the method of Chi et al., using the same degradation and comparable tuning? Please also explain the methodological difference from SegSR’s mutual refinement. Which benefit is specific to the explicit 3D refiner and outer loop? If a direct implementation is not feasible, a comparable matched control could address the same question. (3)Please specify the degradation operator, the generation and update schedule of controller-training conditions, the source of refiner-training predictions, and lambda in Eq. 11. Does the VQ-f4 LPIPS reference describe the training loss only? What reference is used for each reported LPIPS result? Please provide patient-ID splits and executable configurations or pseudocode, and clarify the availability of component checkpoints and code. (4)Can you add a VSNet-on-HR row to Table 2 and an independent structural check on common annotated coverage, using another segmenter or a blinded review? Please separate centerline precision and recall and audit false connections for ResShift, the fine-tuning baseline and VascuSR. Do the structural improvements remain when evaluation is not mainly tied to VSNet? (5)How was K selected on each validation set, and which metric determined the choice? Please compare K=1 with the selected K in both quality and total runtime. A fixed-condition repeated-sampling control, with the SR sampling budget matched, would help distinguish the value of updated anatomy from sampling variation. (6)Please report paired patient-level confidence intervals for the principal method differences and variation across several inference seeds, especially for Tables 1, 2, 5 and 6. Are the smaller advantages over fine-tuning and over round 1 stable? Please distinguish uncertainty across patients and inference seeds from variability across independently trained models. (7)Among the existing held-out cases, have you checked whether documented disease-related vessel changes remain after feedback? A targeted abnormal-vessel audit would address the current anatomical claim. As an optional extension, the HVM dataset paper describes tumor masks: where these are available for the selected cases, a local lesion-fidelity or visibility assessment would help characterize non-target changes. Please otherwise state clearly that lesion preservation has not been established.

4 comments

(1) Raw-mask conditioning, deterministic refinement, and FlowSeg3D

Our main contribution is anatomical correction in vessel-label space, where vascular structures can be processed separately from complex tissue appearance using a relatively lightweight 3D network. Tables 5 and 6 in Appendix A.2 show limited improvements from raw VSNet-mask conditioning. Table 7 further reports mask Dice and clDice before and after FlowSeg3D refinement at each round, evaluated directly against manual annotations. We additionally trained a deterministic residual regression refiner using the same 3D U-Net architecture as FlowSeg3D. Both refiners improve mask quality before SR conditioning: FlowSeg3D achieves higher Dice for both vessel classes and higher HV clDice, while deterministic regression achieves higher PV clDice. Deterministic refinement also retains much of the subsequent SR benefit (Table 15, Appendix A.9), supporting the framework’s applicability across different refiner designs.

Preliminary tests showed limited refinement performance with N=1N=1, so we did not proceed with joint SR/controller training using this configuration. The deterministic results provide a practical alternative that requires no trajectory aggregation or choice of NN, while retaining effective anatomical guidance.

(2) Comparison with structure-guided SR and distinction from SegSR

Please see General Response G3 for our DSDM reproduction and discussion of a matched comparison.

SegSR exchanges information between image and segmentation diffusion branches through a Dual-Modality Bridge during individual reverse-diffusion steps. VascuSR instead completes an SR reconstruction, extracts volumetric vessel predictions, refines them with a separately trained 3D model, and supplies the updated anatomical condition to the next complete SR pass. The distinction lies in explicit volumetric refinement in vessel-label space and feedback between complete SR passes. SegSR

(3) Training, evaluation, and reproducibility details

The LR inputs were generated by clipping CT intensities to [−1000,600][-1000,600] HU for MSD and [−40,160][-40,160] HU for HVM, linearly mapping them to [0,255][0,255], and converting them to 8-bit images. Each axial slice was then downsampled from 512×512512 \times 512 to 64×6464 \times 64 using the PIL bicubic interpolation operator. No additional noise, blur, or JPEG corruption was introduced, and no through-plane downsampling was applied.

The three controller-training condition types and their sampling probabilities are described in Appendix A.1. These conditions were generated offline and cached, then sampled during training rather than regenerated as the model parameters changed. To train the refiner, we applied VSNet to reconstructions produced by the baseline SR model and used the resulting vessel probabilities as inputs, with manual vessel annotations as targets. During joint SR/controller training, the model receives the LR image and a sampled anatomical mask condition and learns to reconstruct the corresponding HR image.

We initially explored segmentation-loss fine-tuning as a primary approach and tested multiple hyperparameter configurations. The reported settings are the best-performing configurations we obtained for each dataset, leading to different loss coefficients. On HVM, λ=0.00191149\lambda=0.00191149 was calibrated on training data to make the weighted anatomical gradient norm approximately 10% of the SR gradient norm, then fixed. The clDice coefficient increased linearly to 0.5 over the first 400 updates. On MSD, the saved coefficients were anchor=1, Dice=1, clDice=0.5, and LPIPS=0.

The VQ-f4 reference is used only for the training perceptual loss. Evaluation LPIPS is computed against the corresponding HR image without VQ reconstruction. ROI LPIPS uses the same reference, with aggregation restricted to the specified region.

Inference starts with an initial SR reconstruction, followed by vessel segmentation, mask refinement and decoding, and SR reconstruction conditioned on the updated mask. Every SR round uses the original LR input.

We will release the code and detailed materials for reproducing our experiments after publication.

(4) Structural validation independent of VSNet

We have added the VSNet-on-HR reference to Table 2. On MSD, VascuSR produces fewer predicted components than VSNet applied to HR images. This may reflect the framework’s emphasis on vascular continuity, together with limitations of the segmentation evaluator, which can produce fragmented predictions even on HR images.

We also added independent nnU-Net evaluation on HVM (Table 10, Appendix A.4). Relative to ResShift, VascuSR improves HV/PV Dice and clDice scores, while reducing mean centerline distance by approximately 17.3%/9.8%. The paired 95% confidence intervals for these differences exclude zero. Although the magnitude of improvement varies between evaluators, the independent evaluation reproduces the direction of the VSNet-based gains.

We additionally evaluate missed branches and disconnected false positives against manual annotations (Table 12, Appendix A.6). Compared with ResShift, VascuSR has fewer missed branches, higher mean branch recall at both coverage thresholds, and fewer zero-overlap predicted components across all four dataset–vessel combinations. Together with the component statistics in Table 2, these results support reduced fragmentation accompanied by better branch preservation.

(5) Selection of KK, runtime

We address the selection of KK and the rationale for multiple rounds in General Responses G2 and G1, respectively.

Wall-clock runtime varies with hardware and GPU workload. Under a fixed configuration, each additional round repeats the same segmentation, refinement, and SR stages, so computational cost grows approximately linearly with KK. Our runtime measurements indicate approximately two minutes per patient per feedback round (Table 3).

(6) Patient-level uncertainty and inference variability

Table 14 in Appendix A.8 reports paired patient-level confidence intervals for VascuSR relative to ResShift. All reported Dice and clDice difference intervals exclude zero in favor of VascuSR. Table 13 reports the mean, sample standard deviation, and win counts across five inference runs against segmentation-loss fine-tuning. VascuSR wins nearly all comparisons, with two losses in MSD HV Dice and one in MSD HV clDice. Also, our conclusions do not rely on Dice alone: Appendix B details the limitations of direct segmentation-loss fine-tuning in local image fidelity and vascular structure.

These inference experiments keep the model weights and final anatomical conditions fixed and assess variability in the final SR sampling stage. Together, the results provide evidence of consistency across patient resampling and the tested inference settings.

Regarding gains beyond the first round, the validation and test results show broadly similar trends: R1 provides most of the improvement, while later gains in Dice and clDice are smaller (Tables 5, 6, and 8). Table 9 additionally shows reductions in connected-component counts and mean centerline distances from R1 to the selected final rounds on both datasets. The consistent trends across validation and test sets, together with improvements in structural metrics, suggest that the gains from additional feedback rounds are not merely isolated sampling fluctuations.

(7) Preservation of disease-related vascular changes and lesions

We agree that preservation of disease-related vascular changes requires dedicated clinical assessment. We currently lack access to clinical experts and a suitable dataset with clinically verified abnormal-vessel annotations for this evaluation. We would welcome suggestions for publicly available resources that could support such an assessment.

Preservation of disease-related vascular changes and lesions has therefore not yet been established. We regard targeted abnormal-vessel evaluation and lesion-local fidelity assessment as important directions for future work.

Thank you for the responses and additional experiments. The simpler-refiner comparison, independent nnU-Net evaluation, and statistical analysis address several of my earlier concerns and strengthen the paper.

However, I still see the novelty as incremental. The deterministic refiner retains much of the benefit, so the extra value of generative refinement and multiple-trajectory aggregation remains unclear. Later feedback rounds provide some structural gains, but their benefit relative to the added cost and their stability across the full pipeline still need stronger evidence. I recognize that the work is solid overall, but these additions have not changed my main assessment. Therefore, I maintain my original rating.

Thank you for your follow-up and for recognizing that our work is solid overall. However, we believe there may still be a misunderstanding regarding the central contribution of our work and we would therefore like to clarify this point.

As emphasized in our contribution statement, we first “introduce explicit 3D vascular representations as anatomical guidance” for slice-wise CT super-resolution. Correcting fine vascular structures directly in image space is challenging because they are intertwined with complex tissue appearance. Extracting these structures into an explicit 3D representation allows anatomical correction to focus on vascular geometry and continuity, making the task accessible to a relatively lightweight network.

Our second contribution is the refinement–conditioning mechanism that brings these 3D corrections back into image reconstruction. Appendix B further explains the motivation for separating anatomical correction from HR-supervised image reconstruction, including the limitations observed when directly fine-tuning the generator with segmentation losses.

We recognize that our presentation of FlowSeg3D may not have sufficiently distinguished these framework-level contributions from the particular refiner used to implement them. Our central claim is not that anatomical correction must use rectified flow. The deterministic refiner retaining much of the benefit supports the broader applicability of the proposed design: the same separation of vascular correction and image reconstruction can provide effective guidance with different refiner architectures. We therefore view this result as supporting the framework’s flexibility, while clarifying the scope of the contribution.

Reviewer_dpZCOpenReview ↗
Rating
2(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes a method for generative super-resolution in computer tomography based on an iterative anatomy-aware diffusion framework. The method, namely VascuSR, decouples three-dimensional vascular refinement and anatomical conditioning. VascuSR is demonstrated in two different datasets across several competing methods for image reconstruction fidelity, vessel segmentation, and vascular structure.

Critical strengths

The authors addresses an important problem of generative super-resolution in computer tomography. The paper is well-written, it is clear and easy to understand. The proposed method benefits from recent concepts in computer vision such as diffusion models, rectified-flow methods that can help provide enhanced performance in this task. The results are shown across different datasets, and ablation studies and compute metrics are provided.

Critical weaknesses
  1. As dicussed by the authors, the proposed idea performs slice wise superresolution and then relies on a 3D refiner. This inherently limits the contextual information and volumetric consistency. It is not clear how the feedback mechanism provides contextual information, and as reported the errors might accumulate across iterations.

  2. While the authors demonstrate the proposed approach against several competing methods, there are concerns regarding whether these baselines can represent the current state-of-the-art. The competing methods were mainly published in 2023 or earlier. The authors should include comparisons with more recent state-of-the-art methods to provide a more comprehensive and fresh evaluation of their approach.

  3. Based on the results displayed in Figure 2 and Figure 3, the improvements by the proposed VascuSR method appear to be limited (the recovered images look blurry and not accurate). Why are the results for all competing methods not included in the comparison (comparisons are only against ResShift)? Also, how does the non-learning-based bicubic interpolation method outperform the proposed approach across a large number of evaluation metrics in Table 2? Also, as the authors reported, SwinIR achieves highest PSNR and SSIM, where the proposed method is unable to get top performance in image quality. Overall, these results raise concerns regarding the effectiveness of the proposed approach.

  4. The proposed approach should be evaluated across a wider range of super-resolution scaling factors to demonstrate its robustness and effectiveness under different conditions.

  5. Table 1 is located at the very beginning of the paper. It is hard to analyze it while reading the results section.

  6. When looking at the image quality of VascuSR across iterative feedback rounds, we observe that PSNR and SSIM measurements are not changing across iterations. Meanwhile, there's no certain pattern on where the best dice scores are achieved. This raises concerns regarding the necessity of the iterative feedback loop.

Questions for authors
  1. How does the proposed method compare against more recent state-of-the-art approaches?

  2. What are the potential benefits of the proposed method in terms of increasing image quality in super-resolution, other than downstream performance?

  3. How can the authors mitigate the issues regarding limited contextual information due to 2D processing and error accumulation in the iterative feedback mechanism?

  4. How does the proposed method work across different superresolution ratios?

2 comments

1. Choice of ResShift and comparison with recent methods

We selected ResShift as a reliable and effective backbone for medical SR, supported by its competitive performance on the ×4 and ×8 medical CT benchmarks reported in QDM. We additionally evaluated MambaIRv2 (CVPR 2025), a strong regression-based SR model primarily benchmarked on natural images. Its HVM validation results are shown below.

MetricPSNR (dB)SSIMDice (HV)Dice (PV)clDice (HV)clDice (PV)
Score24.16710.83970.58290.63760.54250.4949

Although MambaIRv2 achieves strong PSNR and SSIM, its vessel-segmentation performance is less competitive. Based on our experimental experience throughout this project, we highlight two points: (i) a model achieving higher PSNR/SSIM on natural-image benchmarks does not necessarily achieve higher PSNR/SSIM on medical images; and (ii) higher PSNR/SSIM on medical images does not necessarily translate into better anatomy-sensitive metrics, such as Dice and clDice. These distinctions motivate our anatomical-feedback framework.

Moreover, we emphasize that our anatomical-feedback framework is not restricted to a single reconstruction model and can be integrated with different backbones. Beyond ResShift, we have adapted it to CoreDiff for simulated low-dose CT restoration, obtaining improvements in vessel reconstruction on MSD and HVM. This provides evidence of applicability beyond both the ResShift backbone and spatial SR (Table 25, Appendix E).

Nevertheless, we appreciate the reviewer’s suggestion and agree that our framework should be evaluated on a broader range of models. We will therefore include one or two additional recent, competitive SR baselines in the revised manuscript and evaluate the effectiveness of our anatomical-feedback framework when integrated with these models.

2. Reconstruction quality and qualitative comparisons

Our primary objective is to improve the reconstruction of local vascular structures, rather than to maximize global image-quality metrics. Recovering small vessels and their connections is itself an aspect of reconstruction quality; segmentation-based measurements provide a way to assess these structures against manual annotations.

Figures 2 and 3 originally focused on ResShift to isolate the effect of adding vascular guidance to the same SR backbone. We have expanded the visual comparisons to include additional baselines in Appendix A.7, Figures 4 and 5. The improvements over ResShift are localized, particularly around small vessels and their connections. Recovering these fine structures at ×8\times 8 super-resolution is particularly challenging because much of their high-frequency information is lost during downsampling. Although the visual changes are modest, the accompanying improvements in vessel recovery and structural metrics indicate meaningful progress in this setting. We nevertheless acknowledge that some fine branches remain unresolved.

We acknowledge that SwinIR achieves the highest PSNR and SSIM. However, the expanded visual comparisons show smoother reconstructions in which fine vessel boundaries are less distinct. Pixel-wise regression objectives can favor smooth predictions under the ambiguity of high-factor super-resolution, so higher PSNR and SSIM do not necessarily imply better recovery of small vascular structures.

Beyond segmentation metrics, the vessel-region evaluation on MSD shows lower LPIPS with VascuSR than with ResShift (0.1554→0.14700.1554 \rightarrow 0.1470; Table 17, Appendix B.2), while PSNR and SSIM are slightly lower. We therefore claim improved measured vascular reconstruction, rather than uniformly better image fidelity.

Bicubic interpolation. In Table 2, bicubic’s advantage is confined to component counts; VascuSR achieves higher largest-component fractions and lower mean centerline distances for both vessel classes on both datasets. Low component counts alone do not establish better vascular connectivity, as incomplete vessel predictions can also contain fewer components. Table 12 in Appendix A.6 further shows that bicubic misses the most annotated branches and has the lowest branch recall at both thresholds across all four dataset–vessel combinations. These results explain why component counts should be interpreted alongside branch recovery.

3. Limited 3D context and error accumulation

Although SR generation is slice-wise, FlowSeg3D processes volumetric vessel predictions and supplies cross-slice anatomical guidance to subsequent rounds. Each round remains conditioned on the original LR input, and multiple refinement trajectories are aggregated to reduce reliance on a single stochastic prediction. These mechanisms mitigate the limitations of slice-wise processing but do not replace fully volumetric generation.

We acknowledge that erroneous anatomical conditions can propagate through the feedback loop. The original LR input cannot fully constrain fine structures lost during downsampling, and trajectory aggregation cannot eliminate systematic errors shared across predictions.

Nevertheless, Table 11 in Appendix A.5 shows that VascuSR and zero-mask conditioning have nearly identical LR-consistency errors, indicating that vascular guidance introduces little additional global inconsistency with the observed input. Moreover, despite three feedback rounds, VascuSR achieves better vessel-region image quality and lower intensity errors than the single-pass segmentation-loss fine-tuning baseline on MSD (Tables 17 and 18, Appendix B.2). Together, these results support limited observed fidelity degradation in the evaluated setting, although they do not rule out local feedback errors.

4. Performance at different super-resolution scales

Due to time constraints, we conducted the additional ×4\times4 SR experiments only on MSD (Table 26, Appendix F). At R2, VascuSR improves Dice, clDice, and all reported vascular structural metrics over the baseline. The gains are smaller than those observed at ×8\times8. This is reasonable because ×4\times4 downsampling preserves more detail and the baseline is already closer to the HR reference, leaving less room for improvement.

5. Necessity and limitations of iterative feedback

Please see General Response G1 for our discussion of the incremental structural benefits, computational cost, and limitations of additional feedback rounds.

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

This paper introduces VascuSR, which performs super-resolution for contrast-enhanced vascular CT, where a visually plausible reconstruction may contain missing, spurious, or incorrectly connected vessels. VascuSR combines a slice-wise diffusion SR backbone, a 3D vascular refiner, and an anatomy controller. The backbone and controller are trained jointly with HR supervision, while the refiner learns vascular correction separately in the 3D vascular representation space. During inference, the vascular representation extracted from the current SR CT is refined and passed through the controller to guide the next reconstruction, forming an iterative feedback loop. On re-annotated MSD Task08 and HVM datasets, VascuSR outperforms SR baselines in Dice and clDice for segmentation accuracy while maintaining similar image quality.

Critical strengths

I like the analysis of why anatomical conditioning is used rather than segmentation-loss fine-tuning. In Tables 7&8, the "ResShift+Dice/clDice" worsens all six regional measures, indicating that direct fine-tuning improves Dice and clDice but reduces local vascular image fidelity. In Table 9, when evaluated with TotalSegmentator, fine-tuning shows no Dice improvement and scores even below ResShift. So the model may learn a segmenter-specific appearance shortcut. I find this decoupled design, together with the analysis of the alternative, fairly interesting.

The evaluation seems fair within its setting. For example, the patient-level splits are clean, and there is a 0-mask control for controller capacity.

Critical weaknesses

It is not clear where the gains come from. (a) The 3D refiner: no row conditions the controller on the raw VSNet mask without FlowSeg3D, and the refined mask's own Dice against the manual labels, vs the thresholded Q(r), is never reported, so I cannot tell whether the gain comes from the rectified-flow refiner or from conditioning on VSNet's own prediction. (b) Iteration: in Tables 5&6, round 1 already gives almost all the gain (HV Dice at round 1 vs the chosen K: MSD .7077 vs .7072 at K=3; HVM .6191 vs .6206 at K=2); later rounds add at most 0.013 (which is a very small gain) in any Dice or clDice at about two more minutes per volume each, and K seems to be chosen from these test-set tables (K=3 is the best PV score on MSD, K=2 on HVM). (c) The main metric comes from the same VSNet that produces the conditioning mask inside the loop, so it may score vessels from its earlier prediction. The checks that do not use VSNet do not support the gain: TotalSegmentator Dice is about 0.3 for all methods (Table 9), and VascuSR's image quality and fidelity within 3 mm of the vessels are slightly worse than ResShift's on five of six measures (Tables 7&8).

No clinical or other verifiable evaluation. At x8 from 64 x 64 slices, a 1 to 2 mm branch is below one LR pixel, so the vessel drawn there comes from the learned prior, and there is no check of the output against the LR input. Also, the current evaluation cannot show whether the result is correct for a patient: every LR input is a synthetic downsampling of the HR CT, the HR CT is the only reference, and there is no reader study, no downstream task like identifying segmental branches for resection planning, no test on real thick-slice or low-dose acquisitions, and no count of hallucinated vessels (predicted components with no overlap with the manual mask) or of true branches lost. #CC and largest-CC fraction improve whenever the output becomes a clean connected tree, even if small true branches are dropped, and Figure 3 shows that round-3 segmentations still miss many small HV branches present in the HR. For surgical planning, the atypical cases matter most (for example, accessory hepatic veins and variant branching); a prior-driven refiner would correct toward the typical case, and no variant case is tested. Finally, slice thickness and z resampling are never stated although connectivity and mean centerline distance are 3D measures, DSDM [1] is not compared with, and all results are reported on 30 and 29 patients with no confidence intervals. Without a clinical or otherwise verifiable evaluation, I cannot tell whether the added vessels are real or drawn by the learned prior.

[1] Topology-Preserving Computed Tomography Super-Resolution Based on Dual-Stream Diffusion Model (MICCAI 2023).

Questions for authors

I would like to see two more results: VascuSR conditioned on the raw VSNet mask (no FlowSeg3D), and the Dice/clDice of the refined mask vs the manual labels, compared with the thresholded Q(r), at each round. How was K chosen? If on the test set, please report K selected on the validation set. Since round 1 gives almost all the gain, why use more than one round? Is there a clinical use case where a 64 x 64 in-plane CT is the input? Can you evaluate on real thick-slice or low-dose acquisitions with thin-slice references? Or add a verifiable downstream task? Or test cases with known anatomical variants, and report hallucinated vessels and lost branches per case? Does downsampling the SR output reproduce the LR input (please report this consistency error), and can the eight-trajectory vote frequency be shown as an uncertainty map so the reader can see where the refiner guessed? How do you get the 64 x 64 input? What are the slice thicknesses? Were volumes resampled in z before the 3D refiner, connectivity and mean centerline distance computations? Why is the HR-reference Dice (0.794/0.761) below VSNet's reported 0.824/0.807 on the same dataset?

3 comments

Response

We thank the reviewer for these suggestions. Below, we summarize the additional analyses and clarifications.

1. Effect of raw VSNet feedback on SR.

To examine feedback without refinement, we feed raw VSNet masks from the initial SR reconstruction directly to the controller for one additional SR round. Tables 5 and 6 in Appendix A.2 show modest gains over the ResShift baseline on MSD and little change from zero-mask conditioning on HVM. The FlowSeg3D-guided R1 results achieve higher Dice and clDice across both datasets, supporting the value of refined anatomical feedback.

2. Per-round mask quality before and after FlowSeg3D.

Table 7 in Appendix A.2 compares the segmentation of fixed SR outputs with the corresponding FlowSeg3D-refined masks. From R0 to R1, SR segmentation scores closely follow the improvements in the refined R0 masks. Later SR gains are smaller, although refinement still improves mask quality within each round, particularly on MSD. One possible explanation is that the refined SR masks used during controller training come from the initial reconstruction stage. Later-round conditions may therefore differ from those encountered during training, limiting how effectively the model uses further corrections. Training with conditions collected across multiple feedback rounds could potentially improve this transfer.

3. Selection of KK.

Please see General Response G2 for the reported values of KK and the corresponding validation results.

4. Additional gains beyond the first round.

Please see General Response G1 for the incremental structural benefits, computational cost, and limitations of additional feedback rounds.

5. Independent segmentation evaluation.

We added an independent nnU-Net v2 evaluation on HVM (Appendix A.4, Table 10). Under this evaluator, VascuSR improves Dice and clDice and reduces mean centerline distance relative to ResShift for both vessel classes. The 95% confidence intervals for all six paired differences exclude zero.

We acknowledge that the magnitude of improvement varies across segmenters. On HVM, the relative HV Dice improvement over ResShift is approximately 3.83% with VSNet and 2.69% with nnU-Net. Conversely, VSNet-based HV/PV MCD decreases by 4.6%/5.2%, whereas nnU-Net-based MCD decreases by 17.3%/9.8% (Tables 1, 2, and 10). These results show that the improvements persist under an independent evaluator, although their magnitude depends on the segmentation model.

As discussed in Appendix B, our separate training of vascular refinement and HR-supervised reconstruction avoids directly optimizing the SR generator through a segmenter’s Dice/clDice loss. This reduces the incentive to learn segmenter-specific appearance shortcuts and provides a rationale for the observed improvements across evaluators. While these results do not establish complete evaluator independence, they support that VascuSR’s gains are not confined to VSNet.

6. Interpretation of TotalSegmentator results.

The low absolute TotalSegmentator Dice may partly reflect a mismatch in anatomical coverage: our manual annotations include large-caliber extrahepatic venous segments that are not captured by these predictions. Under the same evaluation protocol, VascuSR nevertheless increases Dice from 0.2912 to 0.3183 relative to ResShift (Appendix B.3, Table 19). We interpret this as supporting the direction of improvement, while acknowledging the limited absolute overlap.

7. Local image fidelity.

We agree that VascuSR does not improve every local image-fidelity measure over ResShift. On MSD, within the 3 mm vessel ROI, PSNR decreases by 0.155 dB and SSIM by 0.000445, while MAE and RMSE increase by 0.24 and 0.25 HU, respectively. Signed bias increases from +1.95 to +4.14 HU, whereas spatial LPIPS improves by approximately 5.4% (Appendix B.2, Tables 17 and 18). Together with the structural improvements, these results indicate a trade-off between vascular recovery and local image fidelity. We therefore acknowledge the remaining intensity errors rather than claiming uniform improvements in reconstruction quality.

8. Patient-level confidence intervals.

We have added paired 95% bootstrap confidence intervals for the VascuSR–ResShift differences corresponding to Table 1 (Appendix A.8, Table 14). All eight intervals for HV/PV Dice and clDice on MSD and HVM are above zero, supporting positive mean improvements on both datasets. These intervals quantify patient-sampling uncertainty; variability across repeated inference runs is reported separately in Appendix A.8, Table 13.

9. HR-reference Dice and the published VSNet results.

The MSD HR-reference Dice scores in Table 1 (0.7936 for HV and 0.7610 for PV) were obtained by applying our reproduced VSNet to HR scans and comparing its predictions with manual annotations. We used the authors’ public code and selected the checkpoint with the highest validation Dice. Differences in model configuration or evaluation splits may contribute to the gap from the published scores of 0.824 and 0.807. If you can point us to a VSNet checkpoint that reproduces the published Dice scores, we would be happy to rerun our evaluation using that checkpoint.

10. Clinical relevance of the ×8\times 8 setting.

We chose ×8\times 8 in-plane SR as a challenging stress test of vascular reconstruction under substantial loss of fine detail, rather than as a representation of routine clinical CT inputs. QDM’s Med-SR8 benchmark also evaluates reconstruction from 64×6464 \times 64 inputs to 512×512512 \times 512 outputs. To complement this stress test, we provide evaluations on thin-slice CT under low- and high-noise conditions (Appendix E, Table 25).

Also, due to time constraints, we conducted the additional ×4\times4 SR experiments only on MSD (Table 26, Appendix F). At R2, VascuSR improves Dice, clDice, and all reported vascular structural metrics over the baseline. The gains are smaller than those observed at ×8\times8. This is reasonable because ×4\times4 downsampling preserves more detail and the baseline is already closer to the HR reference, leaving less room for improvement.

11. Evaluation under simulated low-dose conditions.

We adapted anatomical conditioning to CoreDiff for simulated low-dose CT restoration, with implementation details provided in Appendix E. Table 25 in Appendix E reports the test-set results. Our method also improves vascular reconstruction under simulated low-dose conditions, supporting its generalizability across restoration backbones and degradation settings.

12. Branch recovery and disconnected false positives.

To assess whether lower component counts arise from missing true branches, we additionally evaluate branch recovery against manual annotations (Appendix A.6, Table 12). Compared with ResShift, VascuSR misses fewer branches and achieves higher mean recall at both thresholds across all four dataset–vessel combinations. It also produces fewer zero-overlap components, indicating fewer disconnected false positives. Together with the lower total component counts in Table 2, these findings support reduced fragmentation accompanied by better branch preservation.

13. Anatomical variants and clinical downstream validation.

We agree that evaluation on known anatomical variants and clinically meaningful downstream tasks would strengthen the study. We currently lack access to specialist clinicians who could verify variant anatomy, as well as suitable public datasets with the annotations required for such downstream evaluation. We therefore cannot reliably assess preservation of anatomical variants or establish clinical utility in the present study, and we apologize for this limitation. We would welcome suggestions for suitable public datasets or annotated resources and would be willing to extend our evaluation accordingly.

14. LR input generation, slice thickness, and 3D evaluation.

For the main ×8\times 8 SR experiments, CT intensities are clipped to [−1000,600][-1000, 600] HU on MSD and [−40,160][-40, 160] HU on HVM, linearly mapped to [0,255][0, 255], and converted to 8-bit images. Each 512×512512 \times 512 axial slice is downsampled to 64×6464 \times 64 using PIL bicubic interpolation, and image intensities are normalized to [−1,1][-1, 1] for network input. No additional noise, blur, or JPEG compression is introduced in this benchmark.

Slice thickness ranges from 0.8 to 8.0 mm on MSD and is 1.0 or 1.25 mm on HVM. SR slices retain their original order, count, and spacing; no z-axis resampling is applied before refinement or 3D evaluation.

Connectivity is measured using 3D 26-neighborhoods. We extract centerlines using Kimimaro’s TEASAR-based skeletonization, processing HV and PV separately and each connected component independently. Original voxel spacing is retained, and centerline distances are computed in millimeters.

15. Consistency with the LR input.

We downsampled each SR reconstruction to the input resolution and measured MAE and RMSE against the observed LR input. Table 11 in Appendix A.5 shows slightly higher errors than ResShift on MSD and lower errors on HVM. On both datasets, VascuSR and zero-mask conditioning yield nearly identical errors, indicating that vascular guidance has little effect on global LR-input consistency.

16. Eight-trajectory vote-frequency and uncertainty maps.

We have added vote-frequency and uncertainty visualizations for MSD and HVM in Appendix D, Figures 10 and 11. For each vessel class, the vote frequency p=n/8p=n/8 is the fraction of refinement trajectories predicting that class, and uncertainty is visualized as u=4p(1−p)u=4p(1-p).

17. Comparison with DSDM.

Please see General Response G3 for our DSDM reproduction results and the limitations of this comparison.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response to Common Reviewer Concerns

We thank all reviewers for their helpful feedback. Below we address three recurring questions. All table and appendix references refer to the revised manuscript.

G1. Benefits and limitations of iterative feedback

We agree that R1 provides most of the improvement and that subsequent gains in Dice and clDice are modest. Nevertheless, Table 9 in Appendix A.3 shows further reductions in connected-component counts and mean centerline distances at the selected final rounds on both datasets. Additional rounds therefore offer incremental structural improvements at increased inference cost. We consider this trade-off reasonable when structural refinement is prioritized, while R1 provides an efficient option when computational resources are limited.

Our current results motivate exploring this potential, without establishing that more rounds are always beneficial. Future work could also investigate an agent-based policy that adapts the stopping round to the available computational budget and estimated reconstruction quality.

G2. Number of feedback rounds

The reported configurations use K=3K=3 for MSD and K=2K=2 for HVM. Validation results are provided in Table 8 in Appendix A.3.

G3. DSDM reproduction

We reproduced DSDM and evaluated validation slices using exactly the same PSNR and SSIM evaluation code as for our other baselines, obtaining 25.6182 dB and 0.6452, respectively. These results fell below our expectations. We would welcome a publicly available checkpoint and configuration that reproduce the reported results and would be happy to evaluate them under matched conditions.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.