TempDiffReg: Temporal Diffusion Model for Non-Rigid 2D-3D Vascular Registration

TL;DR

TempDiffReg addresses the challenge of non-rigid 2D–3D vascular registration for transarterial chemoembolization by combining a structure-aware global alignment (SA-PnP) with a temporal diffusion-based local refinement (TempDiffReg). The method encodes multi-frame anatomical context per vessel branch and uses a conditional diffusion process to iteratively refine 2D vessel centerlines, providing both accuracy and uncertainty quantification. On a dataset of 626 multi-frame samples from 23 patients, TempDiffReg achieves clear improvements over state-of-the-art methods (e.g., MSE $0.63$ mm, MAE $0.51$ mm) and demonstrates better preservation of vessel topology and curvature. The work enables more reliable guidance for clinicians during complex TACE procedures and offers a framework for temporally coherent, diffusion-based registration in vascular imaging.

Abstract

Transarterial chemoembolization (TACE) is a preferred treatment option for hepatocellular carcinoma and other liver malignancies, yet it remains a highly challenging procedure due to complex intra-operative vascular navigation and anatomical variability. Accurate and robust 2D-3D vessel registration is essential to guide microcatheter and instruments during TACE, enabling precise localization of vascular structures and optimal therapeutic targeting. To tackle this issue, we develop a coarse-to-fine registration strategy. First, we introduce a global alignment module, structure-aware perspective n-point (SA-PnP), to establish correspondence between 2D and 3D vessel structures. Second, we propose TempDiffReg, a temporal diffusion model that performs vessel deformation iteratively by leveraging temporal context to capture complex anatomical variations and local structural changes. We collected data from 23 patients and constructed 626 paired multi-frame samples for comprehensive evaluation. Experimental results demonstrate that the proposed method consistently outperforms state-of-the-art (SOTA) methods in both accuracy and anatomical plausibility. Specifically, our method achieves a mean squared error (MSE) of 0.63 mm and a mean absolute error (MAE) of 0.51 mm in registration accuracy, representing 66.7\% lower MSE and 17.7\% lower MAE compared to the most competitive existing approaches. It has the potential to assist less-experienced clinicians in safely and efficiently performing complex TACE procedures, ultimately enhancing both surgical outcomes and patient care. Code and data are available at: \textcolor{blue}{https://github.com/LZH970328/TempDiffReg.git}

Figures (6)

• Figure 1: Overview of the proposed coarse-to-fine 2D–3D vessel registration framework. 3D centerlines are reconstructed from pre-operative CTA, while 2D centerlines are manually annotated on intra-operative DSA frames. In the global alignment phase, the SA-PnP module computes an initial pose for each frame. Subsequently, in the local refinement stage, TempDiffReg applies branch-wise non-rigid deformation using temporal diffusion modeling. The refined branches are combined to generate the final registered vessel structure aligned to the target DSA frame.
• Figure 2: Illustration of anatomical features used for correspondence establishment used in SA-PnP, including bifurcation points, local curvature, and branch lengths. These features are extracted from 2D and 3D centerlines and matched to establish robust initial rigid alignment between modalities.
• Figure 3: Architecture of the encoding block for each vessel branch. Inputs include 2D centerlines, projected 3D points, camera pose parameters, and frame indices. These are embedded and fused with positional embedding, then processed by a Transformer encoder to model temporal dependencies. A mean-pooled latent vector $\mathbf{y}$ is extracted and used as a conditioning input to the diffusion-based decoder during shape restoration.
• Figure 4: Architecture of the conditional diffusion-based shape decoder. At each diffusion step $t \in [1, T]$, the decoder predicts a denoised vessel shape $\hat{\mathbf{x}}^{1:N}_{0,i}$ from the noisy input $\mathbf{x}^{1:N}_{t,i}$, conditioned on the global latent vector $\mathbf{y}$. During training, intermediate predictions are optionally re-noised for supervision across multiple diffusion steps. At inference, the decoder begins from pure Gaussian noise $\mathbf{x}^{1:N}_{T,1}$ and iteratively restores anatomically plausible 2D centerlines. The conditioning vector $\mathbf{y}$ is injected at every time step to ensure temporally consistent and anatomically faithful shape generation.
• Figure 5: Comparison of registration results on intra-operative 2D DSA images from three representative cases (Case 30, 16, and 17). Ground-truth vessel annotations are shown in green, while registered projections from each method are shown in red dashed lines. Compared methods include TransMorph, uniGradICON, ViT-VNet, SIRU-Net, and the proposed TempDiffReg (bottom row).