ResPF: Residual Poisson Flow Generative Model for Efficient and Physically Consistent Sparse-View CT Reconstruction

Sparse-view CT reduces radiation dose but leads to an ill-posed inverse problem that challenges accurate reconstruction. While deep learning and diffusion-based methods have shown promising results, they often lack physical consistency or suffer from high computational cost due to iterative sampling from random noise. Recent advances in generative modeling, especially Poisson Flow Generative Models (PFGM), offer high-fidelity synthesis by modeling the full data distribution via deterministic ODE trajectories. In this work, we propose Residual Poisson Flow (ResPF), the first application of PFGM++ to sparse-view CT reconstruction. ResPF introduces thour key innovations: (1) a conditional PFGM++ trained on paired sparse/full-view data; (2) a hijacking strategy that skips early sampling steps to accelerate convergence; (3) a data-consistency module embedded at each iteration to enforce fidelity to measured projections; (4) To preserve the stability of the generative path, we further propose a residual fusion mechanism to combine generative outputs with physics-consistent updates. Extensive experiments on synthetic and clinical datasets demonstrate that ResPF achieves state-of-the-art performance, reaching 0.964 SSIM and 39.8 dB PSNR on 63-view reconstructions (from 1000-view ground truth), significantly outperforming existing baselines. Moreover, ResPF reduces sampling time to just 1.5 seconds per image that is over 15× faster than conventional diffusion-based methods, highlighting its practicality for real-world applications.