VeCAS: Vessel-Focused Contrast-Free Angiogram Synthesis for Vascular Interventions

• Each additional 75 mL of iodinated contrast agent was associated with a 42% increase in acute kidney injury risk.Source: §I · p. 1
• Existing contrast-free strategies often require intravascular imaging, alternative contrast agents, or previously acquired angiographic information.Source: §I · p. 1
• VeCAS explicitly decouples vascular structure localization from angiographic appearance synthesis in a two-stage framework.Stage I predicts vessel masks using a discriminative model and cross-modality latent distillation; Stage II synthesizes appearance within localized vascular regions while preserving background.Source: §I · p. 2
• Cross-modality latent distillation transfers vessel-sensitive knowledge from X-ray angiograms to non-contrast X-ray images during training.Source: §I · p. 2
• VeCAS outperforms existing methods in vascular structural fidelity and image quality, with visual Turing tests and physician assessments indicating perceptual realism.Source: §I · p. 2
• 41.4% lower navigation time and 40.7% fewer operation steps were achieved versus non-contrast guidance in phantom robotic guidewire experiments.Source: §I · p. 2

Low-Contrast & Contrast-Free Interventions

• Low-contrast and contrast-free interventions include intravascular imaging, physiology-guided procedures, roadmap-based guidance, and CO2 angiography.Source: §II · p. 2, A. Low-Contrast & Contrast-Free… · pp. 2–3

Angiogram Synthesis

• VeCAS predicts vessel masks before synthesizing angiogram-like appearance within vascular regions while preserving the original non-vascular background.Source: Figure 2 · p. 3

Low-Contrast & Contrast-Free Interventions

• These clinical strategies reduce or avoid iodinated contrast but may require intravascular imaging, alternative contrast agents, or previously acquired vascular information.Source: A. Low-Contrast & Contrast-Free… · p. 3

Medical Image Translation

• Medical image translation methods have used GANs, deformable registration, and improved architectures to address anatomical variation and modality differences.Source: B. Medical Image Translation · p. 3

Stage I: Vascular Structure Localization

• VeCAS predicts a binary vessel mask from non-contrast X-ray images using a discriminative localization model with explicit spatial supervision.The formulation targets vascular structure localization directly rather than relying on a generative model to infer vessel positions.Source: A. Overview · p. 3, B. Stage I: Vascular Structure… · p. 4
• Cross-modality latent distillation transfers vessel-sensitive information from paired angiograms by predicting angiogram-domain latent features from non-contrast features.A frozen angiogram-domain segmentation model supplies the target representations, while a trainable non-contrast encoder and predictor estimate them at multiple scales.Source: B. Stage I: Vascular Structure… · p. 4, Figure 3 · p. 4
• The predicted multi-scale latent features are decoded by the frozen angiogram-domain decoder to obtain the vessel mask used as Stage II’s structural condition.The predictor operates on high-level feature maps at scales i = 3, 4, 5.Source: B. Stage I: Vascular Structure… · p. 4
• Combining latent prediction and segmentation supervision improves vascular localization over latent prediction alone.The prediction loss supervises latent features, while the segmentation loss combines binary cross-entropy and Dice loss; their trade-off is controlled by λpred.Source: B. Stage I: Vascular Structure… · p. 4

Stage II: Vessel-Focused Inpainting

• Vessel-focused inpainting uses diffusion to synthesize angiographic appearance within vascular regions while preserving the non-vascular background.Its training objective restricts denoising to vascular regions, and inference fuses denoised vascular output with the forward-diffused non-contrast image according to the predicted mask.Source: Figure 4 · p. 5

• Vessel-focused inpainting synthesizes angiographic appearance within vascular regions while preserving the non-vascular background.Source: C. Stage II: Vessel-Focused Inpainting · p. 5
• The model restricts denoising supervision to the vessel mask because task-relevant content is concentrated in vascular regions.Source: C. Stage II: Vessel-Focused Inpainting · p. 5
• During inference, reverse denoising begins from Gaussian noise and iteratively generates vessel appearance within the masked region.Source: C. Stage II: Vessel-Focused Inpainting · p. 5
• At each step, the denoised output is fused with the non-contrast image according to the vessel mask to preserve the original background.Source: C. Stage II: Vessel-Focused Inpainting · p. 5
• Inference uses the predicted mask from Stage I because ground-truth masks are unavailable, and the effect of mask quality is analyzed separately.Source: C. Stage II: Vessel-Focused Inpainting · pp. 5–6

• The dataset comprises 599 sequences from 102 patients undergoing lower-limb vascular interventions.Source: A. Clinical Dataset · p. 6
• Each sequence contributes one manually selected pre-contrast frame and one clear contrast-enhanced frame from the same sequence.Source: A. Clinical Dataset · p. 6
• Frame pairs without evident spatial displacement were treated as approximately aligned because lower-limb acquisitions had limited motion.Source: A. Clinical Dataset · p. 6
• Five-fold cross-validation split patients into 80% training and 20% testing per fold, keeping all sequences from each patient together.Source: A. Clinical Dataset · p. 6
• Final performance is reported as mean±std across the five folds.Source: A. Clinical Dataset · p. 6

• All experiments used PyTorch 2.8.0 and Python 3.11.13 under Ubuntu 20.04.3 with 256 × 256 input images.Source: B. Implementation Details · p. 6
• Stage I used U-Net for vascular localization and trained the angiogram-domain and non-contrast components in separate 100-epoch phases.Source: B. Implementation Details · p. 6
• Stage II trained an image-space diffusion inpainting model for 300 epochs with batch size 8 and used 100 sampling steps at inference.Source: B. Implementation Details · p. 6
• Stage I and Stage II reported Dice similarity coefficient and centerline Dice to assess vascular overlap and tubular topological consistency.Source: B. Implementation Details · p. 6

• 69.00% average accuracy was achieved in a blinded Visual Turing Test of 50 real and 50 synthesized angiograms.Source: C. Perceptual and Physician Assessment · p. 6
• Readers achieved average sensitivity of 77.50% and specificity of 60.50%; the low specificity indicated that synthesized images were often classified as real.Source: C. Perceptual and Physician Assessment · p. 6
• Three experienced physicians assessed 50 synthesized angiograms against corresponding real angiograms using VMR, COR, and PU criteria.Source: C. Perceptual and Physician Assessment · p. 6
• Average scores exceeded 3.00 for all three criteria across all physicians on the five-point Likert scale.Source: C. Perceptual and Physician Assessment · p. 7
• Pairwise linearly weighted Cohen’s κ ranged from 0.46 to 0.84, indicating moderate to high inter-physician agreement.Source: C. Perceptual and Physician Assessment · p. 7

Stage I

• VeCAS improves Stage I vascular localization over UNet++ by +1.17% DSC and +0.75% clDice through cross-modality latent distillation.Segmentation models outperform direct image-translation methods, and VeCAS achieves the best overall localization performance.Source: D. Main Results · p. 7

Stage I + II

• 42.27% DSC and 47.70% clDice give VeCAS the best vascular structural fidelity in the complete pipeline.VeCAS improves DSC by +5.84% over ControlNet and clDice by +5.92% over RegGAN.Source: D. Main Results · p. 7
• VeCAS achieves a better balance between image fidelity and vascular structural fidelity by preserving background while focusing synthesis on vascular regions.Whole-image metrics can be dominated by the large non-vascular background and may not reflect vascular structure adequately.Source: D. Main Results · p. 7

Qualitative results

• VeCAS produces clearer and more continuous vessels than baseline methods while maintaining the non-vascular background.Most baselines preserve background appearance but generate weak, incomplete, or barely visible vessels.Source: D. Main Results · p. 7

Distillation Loss

• Combining Lseg and Lpred achieves the best Stage I localization performance, showing complementary pixel-level and cross-modality supervision.Using only Lpred performs poorly, whereas Lseg substantially improves results.Source: E. Ablation Study · pp. 7–8

Vessel-Focused Design

• Combining vessel-focused loss and sampling yields the best overall Stage II performance.Sampling mainly improves PSNR and SSIM, while vessel-focused loss improves DSC and clDice; the two components play complementary roles.Source: E. Ablation Study · p. 8

Error Propagation from Stage I to Stage II

• 60.22% DSC and 66.54% clDice form the oracle upper bound with ground-truth masks, versus 42.27% DSC and 47.70% clDice using predicted masks.The comparison indicates that mask quality directly affects vascular fidelity.Source: E. Ablation Study · p. 8
• Random translation causes the largest degradation among mask perturbations, while erosion and dilation also reduce vascular structural fidelity.The inpainting model is sensitive to spatial misalignment, vessel boundaries, and vessel calibers.Source: E. Ablation Study · p. 8
• PSNR and SSIM remain relatively stable across mask inputs, confirming their limited sensitivity to vascular structural errors.Preserved background content dominates these whole-image fidelity metrics.Source: E. Ablation Study · pp. 8–9

Experimental Setup

• The phantom platform combines vascular phantoms, a robotic vascular system, guidewire, overhead camera, and workstation for VeCAS-guided navigation.Each phantom was based on a real angiogram’s vessel annotation and spatially aligned with the corresponding non-contrast X-ray image.Source: F. Phantom Experiment · p. 9

Protocol

• Three operators navigated guidewires to predefined distal targets under non-contrast, VeCAS, and contrast guidance conditions.Each operator repeated the task five times for each phantom and guidance condition.Source: F. Phantom Experiment · p. 9

Evaluation Metrics

• Navigation performance was measured by time to target and number of operation steps, with each step defined as one discrete control command.Time was measured from the first control command until the guidewire tip reached the target region.Source: F. Phantom Experiment · p. 9

Main Results

• 41.4% lower time to target and 40.7% fewer operation steps were achieved with VeCAS guidance than with non-contrast guidance.VeCAS guidance also achieved navigation performance comparable to contrast guidance.Source: F. Phantom Experiment · p. 9