Conditional Diffusion Posterior Alignment for Sparse-View CT Reconstruction

Researchers Luis Barba, Johannes Kirschner, and Benjamin Bejar have introduced Conditional Diffusion Posterior Alignment (CDPA), a novel method that scales diffusion-based sparse-view CT reconstruction to large 3D volumes. Traditional CT requires many projection views, leading to high radiation exposure and long scan times. Sparse-view CT reduces the number of views, but reconstruction quality often suffers. Deep neural networks, particularly diffusion models, have shown promise but struggle with 3D volumes due to memory constraints, lack of training data, and inter-slice inconsistencies when using 2D models. CDPA overcomes these by conditioning a 2D U-Net diffusion model on an initial 3D reconstruction, ensuring slice consistency, and then aligning the output with measured projections via data-consistency alignment. Experiments on synthetic and real Cone Beam CT (CBCT) data demonstrate state-of-the-art performance, with ablations confirming the synergistic effects of the pipeline. Additionally, the same principles strengthen fast denoising U-Nets, yielding near-diffusion quality at a fraction of the computational cost.

This breakthrough has significant implications for medical imaging and industrial applications. Sparse-view CT can reduce radiation exposure for patients, speed up scans, and lower costs. By enabling high-quality 3D reconstructions with fewer views, CDPA could make CT safer and more accessible. The method's ability to enhance denoising U-Nets also offers a practical path for real-time imaging, balancing quality and speed. For professionals in radiology, medical device manufacturing, and non-destructive testing, this means potentially faster diagnoses, reduced patient risk, and more efficient workflows. The research, published on arXiv, is a step toward scalable, practical AI-powered CT reconstruction.