3D Vessel Reconstruction from Sparse-View Dynamic DSA Images via Vessel Probability Guided Attenuation Learning

TL;DR

This work tackles sparse-view dynamic DSA for 3D vessel reconstruction by introducing a time-agnostic vessel probability field that guides attenuation learning, effectively separating static anatomy from dynamic contrast flow. Attenuation is modeled as a weighted combination $(1-p(\bm{x}))\mu_s(\bm{x}) + p(\bm{x})\mu_d(\bm{x},t)$, with $p(\bm{x})$ serving as a dynamic mask to focus gradients on the appropriate component. Training combines coarse-to-fine progressive hash-grid optimization with a temporal perturbation rendering loss to enforce geometry fidelity and temporal consistency, achieving state-of-the-art results with as few as 30 views. Experimental results on 15-patient DSA data show superior 3D CD/HD and 2D PSNR/SSIM compared to FDK, NAF, and TiAVox, highlighting significant radiation-dose reductions and potential clinical impact. The approach also provides interpretable vascular component volumes and enables high-quality 2D view synthesis through accurate dynamic modeling.

Abstract

Digital Subtraction Angiography (DSA) is one of the gold standards in vascular disease diagnosing. With the help of contrast agent, time-resolved 2D DSA images deliver comprehensive insights into blood flow information and can be utilized to reconstruct 3D vessel structures. Current commercial DSA systems typically demand hundreds of scanning views to perform reconstruction, resulting in substantial radiation exposure. However, sparse-view DSA reconstruction, aimed at reducing radiation dosage, is still underexplored in the research community. The dynamic blood flow and insufficient input of sparse-view DSA images present significant challenges to the 3D vessel reconstruction task. In this study, we propose to use a time-agnostic vessel probability field to solve this problem effectively. Our approach, termed as vessel probability guided attenuation learning, represents the DSA imaging as a complementary weighted combination of static and dynamic attenuation fields, with the weights derived from the vessel probability field. Functioning as a dynamic mask, vessel probability provides proper gradients for both static and dynamic fields adaptive to different scene types. This mechanism facilitates a self-supervised decomposition between static backgrounds and dynamic contrast agent flow, and significantly improves the reconstruction quality. Our model is trained by minimizing the disparity between synthesized projections and real captured DSA images. We further employ two training strategies to improve our reconstruction quality: (1) coarse-to-fine progressive training to achieve better geometry and (2) temporal perturbed rendering loss to enforce temporal consistency. Experimental results have demonstrated superior quality on both 3D vessel reconstruction and 2D view synthesis.

Figures (13)

• Figure 1: DSA imaging and reconstruction. DSA imaging process involves two rotational X-ray scans at identical positions: the first one before the injection of contrast agent (mask run) and the second one after injection (fill run). Subsequently, 2D DSA images are derived by subtracting X-ray images taken during fill run from those captured during mask run. This process selectively highlights the dynamic blood flow information marked by the contrast agent while removing other irrelevant tissues. The dynamic 2D DSA image sequence will be utilized to reconstruct the 3D vascular structures.
• Figure 2: X-ray attenuation process for cases (a) before injection and (b) after injection of contrast agent.
• Figure 3: Overview of our proposed method. We represent the dynamic DSA imaging as a complementary weighted combination of static and dynamic attenuation fields, where the weights assigned to each component are derived from the vessel probability field. Our model is trained by minimizing the disparity between synthesized projections $\hat{I}$ and real captured DSA images $I$.
• Figure 4: 2D toy example of coarse-to-fine progressive training with multi-resolution hash encoding. We gradually activate different levels of hash grids from coarse to fine to mitigate noisy artifacts caused by high-frequency overfitting.
• Figure 5: Illustration of temporal perturbed rendering loss. Projection value should be quite similar at slightly perturbed timestamps from the same viewpoint. The temporal perturbation is simply modeled as a Gaussian noise.