LUCID: Learned Undersampling-Adaptive Consistency-Guided Inference with Deterministic Flow Matching for Sparse-View CT Reconstruction

LUCID introduces a sparsity-adaptive, consistency-guided framework for sparse-view CT reconstruction, leveraging a Flow Matching generative prior. This model is trained on high-quality CT images to create a continuous transport mechanism between Gaussian noise and the target distribution, allowing it to adaptively respond to various levels of sampling sparsity during inference. The approach demonstrates improved image quality and structural fidelity while minimizing artifacts associated with traditional generative methods, making it a significant advancement for practitioners dealing with undersampled CT data.