PaperlayerView exact revision on arXiv

Source-linked AI summary

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

Hairong Ni, Jiahe Huang, Minghao Guo, Wojciech Matusik, Rose Yu

arXiv:2610.20509v1cs.CVcs.AIcs.LG

TL;DR

Vascular CT super-resolution is underdetermined, so plausible reconstructions may not preserve vessel anatomy. VascuSR iteratively combines diffusion reconstruction with explicit 3D vascular guidance. Across MSD Task08 and HVM, it improves vascular measures over learning-based baselines while maintaining image quality comparable to its SR backbone.

  • Problem

    At high upsampling factors, low-resolution observations do not uniquely determine fine detail, and plausible vascular CT reconstructions can contain missing or incorrectly connected vessels.

    Source: §1 · pp. 1–2

  • Method

    VascuSR refines an explicit 3D vascular representation and injects it into the diffusion decoder, updating the anatomical guidance from each new SR volume.

    Source: §1 · p. 2

  • Results

    Over 20% fewer class-averaged connected-component counts on both datasets and over 10% lower class-averaged MCD on MSD than learning-based baselines; vascular measures improve over the matched ResShift backbone.

    Source: §1 · p. 2, §5 · p. 10

  • Takeaways & Limitations

    The findings support decoupling 3D vascular refinement from CT-intensity reconstruction when adding anatomical conditioning to vascular CT super-resolution.

    Source: §5 · p. 10

    Slice-wise SR with a fixed vessel segmenter and separately trained 3D refiner may limit volumetric consistency and allow errors to accumulate; broader evaluation across vascular systems remains necessary.

    Source: §5 · p. 10

Abstract

from arXiv · show

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 interpolation and learning-based 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, VascuSR reduces class-averaged connected-component counts by over 20% on both datasets and mean centerline distance by over 10% on MSD.

1 INTRODUCTION

Vascular CT super-resolution is ambiguous: sharp, plausible images can still contain missing or incorrectly connected vessels. VascuSR addresses this by feeding explicit 3D vascular guidance into iterative diffusion-based reconstruction.

1 INTRODUCTION

  • At high upsampling factors, low-resolution CT poorly determines fine detail, while small, weak-contrast vessels make vascular reconstruction especially challenging.Image-space plausibility can coexist with broken, missing, or spurious vessel connections, which conventional image-quality metrics may not characterize.Source: §1 · pp. 1–2
  • VascuSR separates explicit 3D vascular refinement from CT reconstruction, then injects the refined anatomy into the diffusion decoder and updates it from each new SR volume.The 3D representation captures vascular topology beyond individual slices and makes structural refinement a simpler learning task than modeling CT appearance.Source: §1 · p. 2
  • VascuSR combines 3D vascular guidance, iterative refinement–conditioning, and evaluation across two public vascular CT datasets.The reported results include over 20% fewer class-averaged components on both datasets and over 10% lower class-averaged MCD on MSD.Source: §1 · p. 2

2 RELATED WORK

The related work reviews deterministic and generative approaches to super-resolution, then medical methods that use observation consistency or anatomy-specific guidance. It identifies vascular anatomical fidelity beyond conventional image-quality metrics as an unresolved focus.

2.1 SUPER-RESOLUTION AND GENERATIVE INVERSE PROBLEMS

  • Deterministic SR learns an LR-to-HR mapping and can achieve strong distortion-oriented performance, while pixel-wise objectives can yield overly smooth results.Generative approaches model distributions over plausible HR solutions, including diffusion-based iterative denoising.Source: §2.1 · pp. 2–3

2.2 ANATOMY- AND TASK-AWARE MEDICAL SUPER-RESOLUTION

  • Observation-consistency methods improve agreement with the measured image, but that agreement alone does not ensure recovery of fine anatomy poorly resolved in the input.Examples include alternating diffusion with image-level consistency updates, cycle consistency, iterative correction, and adaptive posterior guidance.Source: §2.2 · p. 3
  • Structural-prior methods use latent detail priors, vessel-sensitive features, or segmentation and inter-slice guidance, yet vascular fidelity beyond conventional image-quality metrics remains an important gap.The passage cites DiffMSR, DSDM, and the dual-guidance SR network as examples.Source: §2.2 · p. 3

3 METHOD

VascuSR couples slice-wise diffusion reconstruction with separate 3D vascular refinement and anatomy conditioning. Its outer loop re-extracts and refines vessel guidance, while a multi-scale controller injects that guidance into the decoder.

3.1 PROBLEM SETUP AND FRAMEWORK OVERVIEW

  • The framework combines a slice-wise diffusion SR backbone, a 3D vascular refiner, and an anatomy controller, with structural refinement decoupled from CT-intensity reconstruction.The backbone and controller receive HR supervision together; the refiner learns vascular correction separately in 3D representation space.Source: §3.1 · p. 3

3.2 ITERATIVE ANATOMY-AWARE DIFFUSION

  • Each feedback update segments the current SR volume, refines its vascular representation, and conditions the next SR volume on the refined masks.The LR input remains fixed; the experiments use K = 3 updates on MSD and K = 2 on HVM.Source: §3.2 · p. 4

3.3 FLOW-BASED 3D VASCULAR REFINEMENT

  • A conditional 3D rectified-flow refiner learns to correct imperfect vessel predictions using manual HV/PV annotations.The full three-class probability volume conditions refinement, and the HV/PV probability channels initialize the flow state; the flow time runs from a noisy initial state toward the target.Source: §3.3 · p. 4

3.4 MULTI-SCALE ANATOMY CONTROLLER

  • A lightweight convolutional controller takes axial slices of the refined HV/PV masks and produces anatomical control features at decoder injection levels.These control features are added to the corresponding SR decoder features during the next reconstruction.Source: §3.4 · p. 5

4 EXPERIMENTS AND RESULTS

On MSD, VascuSR leads the baselines in vascular Dice and clDice, and across MSD and HVM it reduces fragmentation and improves centerline agreement. Qualitative comparisons show local and 3D vascular changes after feedback rounds.

4.1 DATASETS

  • The ×8 evaluation pairs 512 × 512 HR slices with 64 × 64 LR inputs and uses patient-level splits; MSD labels were re-annotated, while HVM includes 281 complete CT–HV–PV triplets.No reported test-set patient is used to train the SR models, VSNet, or FlowSeg3D.Source: §4.1 · p. 6

4.2 BASELINES

  • Baselines span bicubic interpolation, deterministic RCAN/SwinIR/HAT, diffusion ResShift, and ResShift fine-tuned with Dice and soft-clDice losses; learned methods use the same within-dataset training pairs.Hyperparameters are selected on validation data.Source: §4.2 · p. 6

4.3 EVALUATION METRICS

  • Evaluation combines PSNR, SSIM, and LPIPS with fixed-VSNet HV/PV Dice and clDice, plus 3D component, largest-component-fraction, and centerline-distance measures.Connected components are computed separately for HV and PV using 26-connectivity; lower MCD indicates closer centerline agreement.Source: §4.3 · pp. 6–7

4.4 COMPARISON WITH SUPER-RESOLUTION BASELINES

  • VascuSR achieves the best LPIPS and leads all methods in Dice and clDice on re-annotated MSD, although SwinIR has the highest PSNR and SSIM.The results indicate improved vascular structure recovery while preserving the backbone’s image reconstruction capability.Source: §4.4 · p. 7

4.5 THREE-DIMENSIONAL VASCULAR STRUCTURE

  • Across MSD and HVM, VascuSR reduces component counts and increases largest-component fractions versus ResShift; on MSD, PV MCD falls 18.0%, from 8.0877 mm to 6.6351 mm.Bicubic’s lower PV component count but lower largest-component fraction illustrates that a low count alone does not imply better continuity.Source: §4.5 · p. 7, Table 2 · p. 7

4.6 QUALITATIVE ANALYSIS

  • In MSD case 416, round-3 VascuSR visibly improves a separated PV connection and recovers additional HR-visible contours; 3D comparisons show more continuous trees and fewer isolated components than initial SR.The figures show an axial slice and 3D segmentations for MSD cases 416 and 411.Source: §4.6 · p. 8, Figure 2 · p. 9, Figure 3 · p. 9

4.7 PARAMETER AND RUNTIME OVERHEAD

  • The anatomy controller and FlowSeg3D add modest parameter overhead relative to ResShift, and the three-round configuration processes one volume in approximately 6 minutes with eight parallel trajectories.Table 3 reports parameter totals and end-to-end GFLOPs for one SR pass, alongside runtime per volume.Source: §4.7 · p. 9, Table 3 · p. 10

5 CONCLUSION AND DISCUSSION

VascuSR separates 3D vascular refinement from CT intensity reconstruction and connects them through anatomical conditioning. Across MSD Task08 and HVM, it improves Dice and clDice over ResShift while maintaining comparable image-level quality, but its slice-wise design and separate components leave limitations in volumetric consistency and evaluation scope.

5 CONCLUSION AND DISCUSSION

  • VascuSR decouples 3D vascular refinement from CT intensity reconstruction and links them through anatomical conditioning.Across MSD Task08 and HVM, it improves Dice and clDice over ResShift while maintaining comparable image-level reconstruction quality.Source: §5 · p. 10
  • Connectivity and centerline analyses show reduced fragmentation and closer agreement with reference vessel geometry across the evaluated datasets.The authors support learning an anatomical prior in vascular label space and transferring it to SR through conditioning while retaining HR-supervised reconstruction as the image objective.Source: §5 · p. 10
  • Slice-wise SR, a fixed vessel segmenter, and a separately trained 3D refiner may limit volumetric consistency and allow errors to accumulate across feedback iterations.Although exploratory coronary CTA results are presented, larger-scale evaluation across diverse vascular systems remains necessary.Source: §5 · p. 10

A.1 TRAINING AND INFERENCE DETAILS

Training uses patient-level splits to separate vascular-refiner training from joint SR/controller training, with the refiner and SR model using distinct training and inference procedures.

A.1 TRAINING AND INFERENCE DETAILS

  • For MSD, 81 of 243 training patients train FlowSeg3D and 162 train the joint SR/controller; HVM uses 75 of 223 and 148, respectively.VSNet is trained on all training-set patients.Source: §A.1 · p. 14
  • Controller training samples raw HR labels, refined SR labels, or zero masks with probabilities 0.35/0.35/0.30 on MSD and 0.425/0.425/0.15 on HVM.Source: §A.1 · p. 14
  • The backbone and controller are jointly trained with the official ResShift objective, using a frozen VQ-f4 autoencoder and its HR round-trip reconstruction as the LPIPS reference.Source: §A.1 · p. 14
  • FlowSeg3D uses a 3D U-Net and velocity MSE with σ = 1.0; inference averages eight trajectories and applies dataset-specific vote thresholds of 0.3 on MSD and 0.4 on HVM.The thresholds correspond to at least three and four votes, respectively, and were selected separately on each validation set.Source: §A.1 · p. 14

A.2 ABLATION STUDY: ITERATIVE ANATOMICAL FEEDBACK

Iterative anatomical feedback improves vascular segmentation over ResShift on both datasets, with gains stabilizing around round 3 on MSD and strongest overall results around rounds 2–3 on HVM.

A.2 ABLATION STUDY: ITERATIVE ANATOMICAL FEEDBACK

  • All five MSD feedback rounds improve Dice and clDice over ResShift; scores largely stabilize by round 3 while image-quality metrics remain similar across rounds.Source: Table 5 · p. 15, §A.2 · p. 15
  • HVM reaches its strongest overall results around rounds 2–3.Source: Table 6 · p. 15, §A.2 · p. 15
  • The 0-mask setting does not consistently improve Dice or clDice over ResShift despite retaining the controller parameters, indicating gains depend on anatomical conditioning rather than added capacity alone.Source: §A.2 · p. 15

A.3 ADDITIONAL QUALITATIVE RESULTS

The section contrasts qualitative vessel recovery with direct segmentation-loss fine-tuning, then explains VascuSR’s separated supervision and reports its image-fidelity trade-offs near vessels.

A.3 ADDITIONAL QUALITATIVE RESULTS

  • On held-out cases, anatomical feedback regularizes a weakly delineated HV boundary and makes interrupted vessel structures more continuous or closer to the HR reference.Figures 4 and 5 compare LR, ResShift, VascuSR, and HR contours; Figure 6 shows initial and round-2 segmentations alongside HR references.Source: §A.3 · pp. 16–17, Figure 4 · p. 16, Figure 5 · p. 17

B WHY ANATOMICAL CONDITIONING RATHER THAN SEGMENTATION-LOSS FINE-TUNING?

  • Direct Dice/soft-clDice fine-tuning improves MSD segmentation over ResShift, but the comparison motivates examining its effects on local image fidelity, vascular structure, and training cost.Source: B WHY ANATOMICAL CONDITIONING RATHER… · p. 19

B.1 TWO WAYS TO INTRODUCE ANATOMICAL SUPERVISION

  • VascuSR separates HR-supervised image reconstruction from manual-reference vascular refinement, then uses anatomical conditioning to link them; direct fine-tuning instead updates one generator through image and anatomy losses with a frozen segmenter.Fine-tuning uses small learning rates and gradually increases α from 0 to 0.5 during its evaluated updates.Source: §B.1 · p. 19

B.2 IMAGE FIDELITY NEAR VESSELS

  • VascuSR R3 improves all three vessel-ROI image-quality metrics over fine-tuning, while largely preserving ResShift’s regional quality and fidelity with modest degradation in some measures; fine-tuning worsens all six regional measures.The ROI is a 3 mm dilation of the manual HV/PV union.Source: Table 7 · p. 20, Table 8 · p. 20, §B.2 · p. 20

B.3 CROSS-SEGMENTER EVALUATION AND LOCAL APPEARANCE

Cross-segmenter evaluation favors VascuSR over fine-tuning, while the local comparison shows that plausible segmentation contours can coexist with image-intensity problems.

B.3 CROSS-SEGMENTER EVALUATION AND LOCAL APPEARANCE

  • On MSD, VascuSR R3 improves frozen TotalSegmentator Dice by 2.71 percentage points, whereas fine-tuning scores slightly below ResShift.The authors note that low absolute scores may partly reflect mismatch between manual extrahepatic venous coverage and TotalSegmentator predictions.Source: Table 9 · p. 20, §B.3 · p. 20
  • In the central PV, VascuSR restores a visible connection that fine-tuning does not; fine-tuning’s blue contour spans the connection despite conspicuously bright vessel portions.Figure 7 compares LR, ResShift, VascuSR, fine-tuning, and HR images with vascular contours.Source: §B.3 · p. 21, Figure 7 · p. 21
  • The 3 mm ROI signed bias is +12.59 HU for ResShift+Dice/clDice, +1.95 HU for ResShift, and +4.14 HU for VascuSR R3; the authors suggest fine-tuning may exploit a segmenter-specific appearance shortcut.Source: §B.3 · p. 21

B.4 VASCULAR TOPOLOGY

VascuSR improves vascular connectivity and centerline accuracy over direct fine-tuning baselines, while its decoupled vascular refiner requires less training time and fewer GPU-hours. Its structural learning uses a single conditional rectified-flow loss rather than morphology-specific penalties.

B.4 VASCULAR TOPOLOGY

  • VascuSR improves both connectivity measures over ResShift+Dice/clDice, which produces more components than ResShift; MCD also favors VascuSR on MSD and HVM.Connectivity is measured using 3D 26-neighborhoods.Source: §B.4 · pp. 21–22
  • The vascular refiner learns correction in 3D label space with a single conditional rectified-flow loss, without Dice, clDice, or other morphology-specific losses.The refinement–conditioning loop transfers the learned anatomical prior to SR while HR reconstruction remains the image objective.Source: §B.4 · p. 22

B.5 COMPUTATIONAL CONSEQUENCES OF DECOUPLING

  • Direct fine-tuning is slow because each update assembles multiple SR slices into a volume, computes Dice and soft-clDice through a 3D segmenter, and backpropagates through both models.Connectivity and centerline results also highlight that manually selected differentiable penalties can favor chosen scores without fully capturing vascular plausibility.Source: §B.5 · p. 22, §B.4 · p. 22
  • VascuSR’s vascular refiner requires less training time and fewer GPU-hours than direct fine-tuning, while joint backbone-controller training takes approximately the same time as backbone-only training.The comparison reports HVM training costs, excluding validation; VascuSR’s cost is for the 3D refiner.Source: §B.5 · p. 22, Table 12 · p. 22

C EXTENSION TO CORONARY CTA ON IMAGECAS

On a 300-case ImageCAS coronary CTA subset, VascuSR achieves the highest Dice and clDice among evaluated SR methods. Qualitative comparisons show improved 3D overlap and centerline agreement, including recovery of a small annotated vessel missed by ResShift.

C EXTENSION TO CORONARY CTA ON IMAGECAS

Related papers

Loading 2610.20509v1…