TLIR: Two-layer iterative refinement model for limited-angle CT reconstruction

Limited angle reconstruction is a typical ill-posed problem in computed tomography (CT). In practical applications, due to the limited scanning angles available for fixed scan targets and the patient’s ability to tolerate radiation, complete projection data are usually not available, and images reconstructed by conventional analytical iterative methods can suffer from severe structural distortion and tilt artefacts. In this paper, we propose a deep iterative model called TLIR to recover the structural details of the missing parts of the limited angle CT images and reconstruct high quality CT images from them. Specifically, we adapt the denoising diffusion probability model to conditional image generation for the image domain recovery problem, where the model output starts from noise-blended limited-angle CT images and iteratively refines the output images using residuals U-Net trained at various noise level data. In addition, considering that the deep model corrupts the sampled part of the sinusoidal data during inference, we propose a learnable data fidelity module called DSEM to balance the data domain exchange loss and inference information loss. The two modules are executed alternately to form our two-layer iterative refinement model. The two-layer iterative structure also makes the network more robust during training and inference. TLIR shows strong reconstruction performance at different limited angles, and shows highly competitive results in all image evaluation metrics. The model proposed in this paper is open source at https://github.com/JinxTao/TLIR/tree/master.