Prompt Guiding Multi-Scale Adaptive Sparse Representation-driven Network for Low-Dose CT MAR

TL;DR

The paper addresses the challenge of reconstructing low-dose CT images with metal artifact reduction by identifying gaps in multi-scale information integration and dose-level generalization. It introduces PMSRNet, a prompt-guided, multi-scale sparse representation network, and embeds it into a dual-domain LDMAR framework (PDuMSRNet) via deep unfolding for interpretability. Key contributions include a prompt-guided scale-adaptive threshold generator (PSATG), a multi-scale coefficient fusion module (MSFuM), and a single model capable of handling multiple dose levels with reduced storage. Extensive ablations demonstrate the gains from dual-domain learning, PSATG, MSFuM, and prompt guidance, with clinical validation suggesting practical impact but also domain-gap limitations. Overall, the methods advance robust, storage-efficient LDMAR for LDCT, enabling better diagnostic usability and downstream segmentation.

Abstract

Low-dose CT (LDCT) is capable of reducing X-ray radiation exposure, but it will potentially degrade image quality, even yields metal artifacts at the case of metallic implants. For simultaneous LDCT reconstruction and metal artifact reduction (LDMAR), existing deep learning-based efforts face two main limitations: i) the network design neglects multi-scale and within-scale information; ii) training a distinct model for each dose necessitates significant storage space for multiple doses. To fill these gaps, we propose a prompt guiding multi-scale adaptive sparse representation-driven network, abbreviated as PMSRNet, for LDMAR task. Specifically, we construct PMSRNet inspired from multi-scale sparsifying frames, and it can simultaneously employ within-scale characteristics and cross-scale complementarity owing to an elaborated prompt guiding scale-adaptive threshold generator (PSATG) and a built multi-scale coefficient fusion module (MSFuM). The PSATG can adaptively capture multiple contextual information to generate more faithful thresholds, achieved by fusing features from local, regional, and global levels. Furthermore, we elaborate a model interpretable dual domain LDMAR framework called PDuMSRNet, and train single model with a prompt guiding strategy for multiple dose levels. We build a prompt guiding module, whose input contains dose level, metal mask and input instance, to provide various guiding information, allowing a single model to accommodate various CT dose settings. Extensive experiments at various dose levels demonstrate that the proposed methods outperform the state-of-the-art LDMAR methods.

Figures (7)

• Figure 1: Overview of our PDuMSRNet. The architecture, which is unfolded from the optimization iterations for solving the LDMAR task, comprises $T$ stages corresponding to $T$ iterative steps. Specifically, the sinogram domain variable $\tilde{\textbf{s}}^{(t+1)}$ and the image domain variable $\textbf{x}^{(t+1)}$ are updated alternately using the proxNet$_{{\tilde{\textbf{s}}}}$ for sinogram-domain processing and PMSRNet for image-domain reconstruction at the ($t+1$)-th stage. The architecture of proxNet${\tilde{\textbf{s}}}$ comprises three residual blocks in sinogram domain, whereas PMSRNet is the proposed prompt guiding multi-scale adaptive sparse representation-driven network in image domain. This design explicitly follows the update rules derived in Eqn. (\ref{['eq:9']}) and Eqn. (\ref{['eq:14']}), ensuring a model-driven deep learning framework.
• Figure 2: Overview of our prompt guiding multi-scale adaptive sparse representation-driven network named PMSRNet (top), our prompt guiding scale-adaptive threshold generator termed as PSATG (middle) and multi-scale coefficient fusion module called MSFuM (bottom). The PMSRNet contains three components: multi-scale sparsifying frames, PSATG and MSFuM. The PSATG is designed to generate faithful thresholds with the prior of prompt information and contains five parts: 1) The shallow feature extraction module relies on a convolution layer; 2) The deep feature extraction module models images at local, regional, and global levels by parallelizing convolutional neural networks, window attention, and anchored stripe attention; 3) The feature selection module utilizes channel attention mechanism to adaptively reweight information from different feature extraction module; 4) The threshold generating module enhances the performance of PMSRNet by generating faithful thresholds using information from the feature selection module; 5) The prompt guiding module utilizes prompt information, including LDCT images with metallic implants, metal masks, and dose maps, to guide the threshold generating process. The MSFuM which is used to fuse the multi-scale representations mainly contains a cross attention and a feed-forward network to effectively integrate features from different resolutions. Structurally, the PMSRNet adopts a four-scale architecture. At each scale, the processing flow consists of a sparse coding procedure by learnable sparsifying frame, a threshold processing procedure with the prior of prompt information, and an image generation procedure by learnable sparsifying frame, which work in sequence to ultimately produce the final reconstructed image.
• Figure 3: Visual evaluation of different approaches on LDCT images with metallic implants under the 1/2 dose level. Zooming in on the green screen enhances the viewing experience of the images. The PSNR and SSIM values are computed on the entire image reconstructed by the respective approaches. The metallic implants are highlighted with red masks. The blue arrow indicates the reconstructed structural information. The display window is [-1000, 1000] HU.
• Figure 4: Visual evaluation of different approaches on LDCT images with metallic implants under the 1/4 dose level. Zooming in on the green screen enhances the viewing experience of the images. The PSNR and SSIM values are computed on the entire image reconstructed by the respective approaches. The metallic implants are highlighted with red masks. The blue arrow indicates the reconstructed structural information. The display window is [-1000, 1000] HU.
• Figure 5: Visual evaluation and reconstruction error magnitude of BDuMSRNet and PDuMSRNet on LDCT images under the 1/4 dose level. Zooming in on the green screen enhances the viewing experience of the images. The PSNR and SSIM values are computed on the entire image reconstructed by the respective approaches. The metallic implants are highlighted with red masks.