Filter-Free Two-Stage Reconstruction for Low-Dose CT Across Regular and Irregular Sparse-Angle Sampling

Abstract Low-dose and sparse-angle computed tomography (CT) reduces radiation exposure but makes image reconstruction challenging due to noisy and limited projection data. Popular reconstruction methods are based on two-stage approaches, typically involving filtered backprojection (FBP) followed by a neural network to enhance the image. FBP, however, amplifies noise and struggles with irregular sampling. Therefore, we explore filter-free initial reconstructions, shifting the filtering step to the neural network. In particular, we investigate how two-stage methods can be adapted for cases where implementing explicit filters is difficult, such as with irregular sampling. Specifically, we propose backprojection (BP) or a small number of Landweber iterations as the initial reconstruction, followed by a fine-tuned DRUNet model, referred to as BP-DRUNet and Landweber-DRUNet, respectively. For evaluation, we consider both regular and irregular sampling conditions: For regular sampling, we compare BP-DRUNet with FBP-DRUNet (using FBP as the initial stage) in order to benchmark against standard two-stage approaches. BP-DRUNet performs comparably to FBP-DRUNet under regular sampling. In irregular sampling, Landweber-DRUNet improves reconstruction quality with more iterations, though at the cost of longer training and inference times. Experiments are carried out on synthetic and real CT datasets with parallel- and fan-beam acquisitions across different sparse-angle setups.