Selig Tim ; März Thomas ; Storath Martin ; Weinmann Andreas

Abstract

Computed tomography from a low radiation dose (LDCT) is challenging due tohigh noise in the projection data. Popular approaches for LDCT imagereconstruction are two-stage methods, typically consisting of the filteredbackprojection (FBP) algorithm followed by a neural network for LDCT imageenhancement. Two-stage methods are attractive for their simplicity andpotential for computational efficiency, typically requiring only a single FBPand a neural network forward pass for inference. However, the bestreconstruction quality is currently achieved by unrolled iterative methods(Learned Primal-Dual and ItNet), which are more complex and thus have a highercomputational cost for training and inference. We propose a method combiningthe simplicity and efficiency of two-stage methods with state-of-the-artreconstruction quality. Our strategy utilizes a neural network pretrained forGaussian noise removal from natural grayscale images, fine-tuned for LDCT imageenhancement. We call this method FBP-DTSGD (Domain and Task Shifted GaussianDenoisers) as the fine-tuning is a task shift from Gaussian denoising toenhancing LDCT images and a domain shift from natural grayscale to LDCT images.An ablation study with three different pretrained Gaussian denoisers indicatesthat the performance of FBP-DTSGD does not depend on a specific denoisingarchitecture, suggesting future advancements in Gaussian denoising couldbenefit the method. The study also shows that pretraining on natural imagesenhances LDCT reconstruction quality, especially with limited training data.Notably, pretraining involves no additional cost, as existing pretrained modelsare used. The proposed method currently holds the top mean position in theLoDoPaB-CT challenge.