This AI Reconstructs 3D Blood Vessels From Just a Few X-Ray Views — and It Slashes Radiation

Digital Subtraction Angiography (DSA) is a gold standard for diagnosing vascular diseases, but current commercial systems require hundreds of scanning views for 3D reconstruction, leading to substantial radiation exposure. A team of researchers from multiple institutions has developed a new neural rendering optimization framework specifically for sparse-view DSA reconstruction. Their approach, called vessel probability guided attenuation learning, models DSA imaging as a weighted combination of static and dynamic attenuation fields. The key innovation is a time-independent vessel probability field that acts as an adaptive foreground mask, providing proper gradients for both fields and enabling self-supervised decomposition between static backgrounds and dynamic contrast agent flow. This mechanism allows high-quality 3D vessel reconstruction using far fewer scanning views than traditional methods.

The model is trained by minimizing the difference between synthesized projections and real DSA images. Two additional training strategies further boost quality: coarse-to-fine progressive training refines geometry step by step, and temporal perturbed rendering loss ensures consistency across time frames. Experimental results demonstrate that the method produces high-quality 3D vessel reconstructions and 2D DSA image synthesis, even with sparse input views. Published in Medical Image Analysis (2026), this work has the potential to significantly reduce patient radiation exposure during vascular imaging procedures without compromising diagnostic accuracy. It is particularly promising for clinical settings where minimizing radiation dose is critical, such as pediatric or repeated follow-up imaging.