Enhanced Low-Dose CT Image Reconstruction by Domain and Task Shifting Gaussian Denoisers

TL;DR

This work tackles the challenge of reconstructing high-quality CT images from low-dose projections by introducing a two-stage pipeline: first, a standard filtered backprojection (FBP) reconstruction, followed by enhancement using a pretrained Gaussian denoiser fine-tuned for LDCT. The key novelty lies in domain and task shifting: repurposing Gaussian denoisers trained on natural grayscale images to improve LDCT images, with an SSIM-based loss to guide fine-tuning. Across LoDoPaB-CT and 2016LDCTGC datasets, the method achieves state-of-the-art performance in SSIM and competitive PSNR, while offering substantially lower inference times than unrolled iterative methods like ItNet. The approach is robust to the choice of pretrained denoiser (DRUNet, KBNet, Restormer), benefits strongly from pretraining especially with limited data, and holds promise for extending to other imaging modalities or denoiser architectures as new Gaussian denoisers become available.

Abstract

Computed tomography from a low radiation dose (LDCT) is challenging due to high noise in the projection data. Popular approaches for LDCT image reconstruction are two-stage methods, typically consisting of the filtered backprojection (FBP) algorithm followed by a neural network for LDCT image enhancement. Two-stage methods are attractive for their simplicity and potential for computational efficiency, typically requiring only a single FBP and a neural network forward pass for inference. However, the best reconstruction quality is currently achieved by unrolled iterative methods (Learned Primal-Dual and ItNet), which are more complex and thus have a higher computational cost for training and inference. We propose a method combining the simplicity and efficiency of two-stage methods with state-of-the-art reconstruction quality. Our strategy utilizes a neural network pretrained for Gaussian noise removal from natural grayscale images, fine-tuned for LDCT image enhancement. We call this method FBP-DTSGD (Domain and Task Shifted Gaussian Denoisers) as the fine-tuning is a task shift from Gaussian denoising to enhancing LDCT images and a domain shift from natural grayscale to LDCT images. An ablation study with three different pretrained Gaussian denoisers indicates that the performance of FBP-DTSGD does not depend on a specific denoising architecture, suggesting future advancements in Gaussian denoising could benefit the method. The study also shows that pretraining on natural images enhances LDCT reconstruction quality, especially with limited training data. Notably, pretraining involves no additional cost, as existing pretrained models are used. The proposed method currently holds the top mean position in the LoDoPaB-CT challenge.

Figures (5)

• Figure 1: Overview of the proposed FBP-DTSGD methodology. The proposed approach consists of a two-stage process for LDCT image enhancement. A FBP is applied to the LDCT data to obtain an initial reconstruction. A pretrained Gaussian denoiser is fine-tuned for the downstream task, i.e., for the enhancement of CT-images obtained from the FBP-stage.
• Figure 2: An example highlighting the distinction between the DRUNet and FFDNet methodology of cropping versus the adapted padding strategy. (a) shows the low-dose FBP, while (b) displays the corresponding ground truth. (c) highlights the artifacts at the patch boundaries resulting from denoising on sub-images, whereas (d) exhibits artifact mitigation achieved through ommiting the partitioning into blocks.
• Figure 3: Reconstruction results for two test samples from the LoDoPaB-CT test dataset (top row) and the 2016LDCTGC dataset (bottom row). The first column displays the ground truth, followed by the FBP in the second column. The next three columns present reconstructions by the Gaussian denoisers DRUNet, KBNet, and Restormer. Each pretrained denoiser model was trained for 96 epochs on 1791 image pairs from either the LoDoPaB-CT training dataset or the 2016LDCTGC dataset. The proposed two-stage method effectively enhances the FBP reconstructions in each case, as demonstrated by the PSNR and SSIM metrics provided below each subimage.
• Figure 4: Comparison of PSNR and SSIM metrics for the pretrained DRUNet model and the DRUNet model trained from scratch on the LoDoPaB-CT test dataset. (a) shows the comparison of PSNR values, and (b) displays the comparison of the SSIM values.
• Figure 5: Comparison of network performance with and without rotational augmentation on the LoDoPaB-CT test dataset. The outputs for the network with rotational augmentation remain consistent (e, f), regardless of the input image's orientation (0° and 90°). In contrast, the network without rotational augmentation shows variation in output between normal and rotated input scenarios (c, d). (The rotated versions are presented in an unrotated fashion for easier comparison.)