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

AI summary

Overview

• Research area: Medical image computing — generative super-resolution (SR) of contrast-enhanced CT angiography, with a focus on preserving 3D vascular anatomy rather than only image appearance.
• Technical level: Advanced. The paper assumes familiarity with diffusion models, rectified flow, latent autoencoders (VQ-f4), and tubular-structure metrics such as clDice.
• Scope: The paper introduces VascuSR, an iterative framework that couples a slice-wise diffusion SR backbone with an explicit 3D vascular representation that is refined and fed back as anatomical conditioning, and evaluates it on two public hepatic-vessel CT datasets at ×8 upsampling.

What This Paper Is About

When a CT volume is upsampled aggressively (here, 64×64 low-resolution slices to 512×512), the fine detail that was never measured is not uniquely recoverable, so a generative model can produce vessels that look plausible but are missing, spurious, or wrongly connected. The authors argue that conventional image-quality scores such as SSIM cannot catch these errors, because a short break can split a vessel tree while barely changing pixel statistics. Their goal is to make super-resolution anatomically accountable by giving the reconstruction an explicit, iteratively updated 3D map of the vasculature to follow.

Key Contributions

1. Explicit 3D vascular representation as guidance. VascuSR conditions slice-wise CT SR on a volumetric hepatic-vein/portal-vein (HV/PV) representation rather than on 2D slices alone, which carries continuity and branching information across slices and separates vascular structure from CT intensity and organ texture.
2. A decoupled refinement–conditioning loop. A conditional 3D rectified-flow refiner (FlowSeg3D, a 3D U-Net velocity model) learns vessel correction in label space, while a lightweight multi-scale anatomy controller (1.44 M parameters) injects the refined masks into the diffusion decoder. At each round, the vessel representation is re-extracted from the latest SR volume with a frozen VSNet segmenter and refined again, so anatomy and image are updated in alternation (K = 3 rounds on MSD, K = 2 on HVM).
3. Evaluation beyond image metrics. The paper reports Dice and clDice alongside 3D structural measures — 26-connectivity connected-component counts, largest-component fraction, and mean centerline distance (MCD) in millimetres — plus a direct comparison against joint Dice/soft-clDice fine-tuning of the SR generator (Appendix B).
4. Reported gains. VascuSR reduces class-averaged connected-component counts by over 20% on both datasets and class-averaged MCD by over 10% on MSD relative to learning-based baselines, while keeping image quality comparable to the ResShift backbone, at a total of 189.70 M parameters versus 172.01 M for ResShift.

Main Findings

• Segmentation accuracy (MSD Task08, ×8): VascuSR reaches Dice .7072 (HV) and .6404 (PV), and clDice .6446 (HV) and .5286 (PV), the best of all compared methods; the ResShift backbone reaches .6526/.5923 Dice and .5877/.4460 clDice, and the HR reference is .7936/.7610 Dice and .7615/.7352 clDice.
• Segmentation accuracy (HVM, ×8): VascuSR reaches Dice .6206/.7117 and clDice .6033/.5731 (HV/PV), versus ResShift at .5977/.6914 and .5687/.5302; HR reference is .7177/.7917 and .7122/.7149.
• Image quality is comparable, not dominant: On MSD, SwinIR has the highest PSNR (34.44) and SSIM (.9164), while VascuSR attains the best LPIPS (.0649 versus .0660 for ResShift and .1262 for SwinIR) with PSNR 32.77 and SSIM .8876. On HVM, VascuSR's PSNR is 23.40 and LPIPS .0744, close to ResShift (23.50, .0743) and below the regression baselines in PSNR (RCAN 24.58, SwinIR 24.76).
• Less vascular fragmentation: On MSD, HV connected components drop from 31.0 (ResShift) to 22.7 and PV from 26.0 to 22.4, with largest-component fractions of .9382 (HV) and .8931 (PV). On HVM, HV drops from 48.7 to 31.8 and PV from 66.7 to 45.0, with fractions of .524 and .923.
• Centerline accuracy: On MSD, PV MCD falls from 8.0877 mm with ResShift to 6.6351 mm, a reduction of 18.0%; HV MCD falls from 6.046 mm to 5.231 mm. On HVM, HV MCD goes from 5.272 mm to 5.029 mm and PV from 5.004 mm to 4.743 mm.
• Low component counts alone are misleading: Bicubic has the lowest PV component count on MSD (13.0), but its largest-component fraction (.6432) shows this reflects under-segmentation rather than continuity.
• Anatomical conditioning, not extra capacity, drives the gain: A "0-mask" arm that keeps the controller parameters but supplies no anatomy does not consistently beat ResShift (MSD Dice .6587/.5644 and clDice .5909/.4286 versus .6526/.5923 and .5877/.4460).
• Feedback saturates quickly: On MSD, all five feedback rounds improve Dice and clDice over ResShift, with metrics largely stabilizing by round 3 (round 3 Dice .7072/.6404; round 5 clDice .6503/.5235); image-quality metrics stay essentially flat across rounds (PSNR 32.77, SSIM .8876). HVM peaks around rounds 2–3 (round 2 Dice .6206/.7117, clDice .6033/.5731) and drifts slightly downward by round 5.
• Cost: The controller adds 1.44 M and FlowSeg3D 16.25 M parameters, with end-to-end GFLOPs per SR pass rising from 6248.69 to 6284.01. A three-round configuration takes roughly 6 minutes per volume, dominated by FlowSeg3D (about 97–102 s per round at N = 8 parallel trajectories) versus about 18.4 s for SR generation and about 1.9 s for VSNet.
• Not reported in the supplied content: The results of Table 7 (SSIM and spatial LPIPS within a 3 mm dilation of the manual HV/PV union, Appendix B.2) are cut off, and the coronary CTA results in Appendix C are only mentioned, not shown.

Methodology in Plain English

The system starts by producing an ordinary super-resolution CT volume using a pretrained diffusion SR model (ResShift) with no anatomical information — an all-zero vessel mask is fed in as the initial condition. That volume is then passed to a frozen vessel segmenter (VSNet), which outputs per-voxel probabilities for background, hepatic vein, and portal vein.

Those imperfect probabilities are handed to a separate 3D refiner, FlowSeg3D. Rather than regenerating CT intensities, the refiner works purely in vessel-label space: it is trained with a rectified-flow objective to transport noisy, imperfect vessel predictions toward the manual HV/PV annotations, where the flow is a straight-line interpolation between the two states and the network predicts the velocity along that path. At inference, eight independent noisy trajectories are integrated with ten Euler steps, decoded, and combined by voting; a voxel is kept only if its class frequency exceeds a validation-selected threshold (strictly above 0.3 on MSD and 0.4 on HVM, i.e., at least three or four of eight votes).

The resulting clean 3D masks are sliced axially and passed to a small convolutional anatomy controller, which turns them into multi-scale features that are added directly to the decoder features of the diffusion SR network at each resolution level. The SR model then reconstructs the volume again with that anatomical nudge. Because the new volume changes the vessels, the segmenter and refiner are run once more on it, and the loop repeats — three rounds for MSD, two for HVM. The SR backbone and controller are trained jointly from scratch with the standard ResShift objective and HR supervision, using mixed condition sources (raw VSNet labels from HR images, refiner-corrected labels from SR images, and an all-zero mask, at probabilities 0.35/0.35/0.30 on MSD and 0.425/0.425/0.15 on HVM), so the controller sees realistic, imperfect masks at training time. This keeps structural learning and image reconstruction in separate, more tractable problems.

Why This Matters

• Research impact: The paper shifts the evaluation target for medical super-resolution from pixel statistics to anatomical topology, and shows a concrete architectural pattern — learn the prior in label space, transfer it by conditioning — that is cheaper to train than back-propagating segmentation losses through the image generator. It also provides a matched, carefully tuned joint-optimization baseline (1,200 updates on MSD at learning rate 10⁻⁵, 2,000 on HVM at 5×10⁻⁶, with the Dice/clDice weight warmed from 0 to 0.5) rather than an untuned strawman.
• Clinical imaging: Thin, low-contrast peripheral vessels are exactly where reconstruction errors can change interpretation; more continuous vessel trees bear on stenosis grading and branch identification.
• Dose reduction and low-field scanners: If SR can be trusted at ×8 in-plane, protocols can acquire fewer or noisier projections while retaining detail.
• Surgical and interventional planning: Hepatic resection, transplant, and transjugular procedures depend on hepatic-vein and portal-vein branching maps; fragmented reconstructions complicate 3D planning and image-guided navigation.
• Archival and legacy data: Retrospectively upsampling older or low-resolution CT studies for reuse in AI pipelines without introducing fabricated vasculature.
• Industry relevance: Scanner vendors and PACS/AI vendors that market "AI enhancement" features face a regulatory and liability question about hallucinated anatomy; anatomy-aware conditioning with explicit, auditable vessel masks and connectivity metrics is a plausible path toward validation frameworks for generative reconstruction.

Future Directions

• Fully volumetric reconstruction: The current pipeline is slice-wise and can break axial continuity; the authors list fully volumetric and more tightly integrated reconstruction as future work.
• Adaptive or cheaper refinement: FlowSeg3D accounts for roughly 100 s of the ~120 s per feedback round, so reducing or adaptively allocating refinement effort is needed for practical deployment.
• Error accumulation and segmenter dependence: The framework relies on a fixed vessel segmenter and a separately trained refiner, and the authors note errors may accumulate across feedback iterations and limit volumetric consistency.
• Broader anatomy coverage: The paper calls for larger-scale evaluation across vascular systems, acquisition settings, and downstream vascular-analysis tasks, offering only exploratory coronary CTA results in Appendix C. The specific number of rounds could also be made data-adaptive, since gains largely stabilize after two to three rounds and vary between datasets.

Target Audience

• Researchers in medical image reconstruction and generative models who need to move beyond PSNR/SSIM when the output has structural semantics.
• Practitioners building clinical CT enhancement or vessel-analysis pipelines who want an architecture that leaves the base SR backbone intact and adds a small, inspectable anatomical module.
• Students and engineers working on diffusion or flow-matching models who want a well-instrumented case study in conditioning, ablation design, and topology-aware evaluation.
• Readers interested in evaluating generative inverse problems in AI for Science, where domain-specific structural constraints matter more than perceptual realism.

Authors’ 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 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.