Diffusion Posterior Sampling for Angiographic Reconstruction using Limited Angle and Sparse View data for 4D-DSA.
Authors
Affiliations (3)
Affiliations (3)
- Electrical and Computer Engineering Department at Johns Hopkins University, baltimore, MD, 21219, USA.
- Radiology Department at University of Pennsylvania, Philadelphia, PA 19104, USA.
- Biomedical Engineering Department at Johns Hopkins University, baltimore, MD, 21219, USA.
Abstract
Four-dimensional digital subtraction angiography (4D-DSA) is an effective modality for the diagnosis and treatment planning of neurovascular interventions. Single time point 4D-DSA Projection data is usually underdetermined for a 3D reconstruction. Current image formation therefore relies on approximations using a 3D-DSA. In this work, we investigate application of advanced reconstruction on limited angle and sparse view data acquired at a single time point in 4D-DSA. We adopted the diffusion posterior sampling (DPS) algorithm which combines a physics-based data fidelity with an anatomically realistic diffusion prior. The diffusion prior was trained using axial slices extracted from 3D liver vasculature generative model. A total variation (TV) penalty along the z direction was applied to encourage continuity between slices. We compared DPS with a model-based approach using a directional TV penalty (DTV). We simulated a single-arc geometry on a single-plane C-arm and inverse geometry tomosynthesis system as well as an orthogonal-arc geometry achievable using a bi-plane C-arm systems. The DPS algorithm is able to reconstruct high-fidelity vascular anatomy under the challenging geometry of 10 views over a single arc of 40 degrees. As expected, vessels with thinner branches and/or branches in directions corresponding to the missing frequencies are more challenging to reconstruct. Orthogonal arcs improve the reconstruction quality of both algorithms, indicating the importance of more distributed angular sampling. DPS outperforms DTV under all conditions evaluated. The DPS algorithm provides an effective approach to reconstruction 4D-DSA data acquired using limited angle and sparse view geometries.