Abstract
Computed Tomography (CT) imaging is a cornerstone of modern diagnostics, yet the trade-off between radiation dose and image quality remains a critical challenge. Low-Dose CT (LDCT) reduces cancer risks but introduces Poisson noise, which is further exacerbated by "streak" artifacts from metallic implants. This paper presents PDuMSRNet, a dual-domain deep unfolding framework that leverages a novel "prompt guiding" strategy. By integrating dose maps and metal masks as prompts, a single model can adapt to various radiation levels, outperforming state-of-the-art methods in reconstruction quality while significantly reducing storage overhead.
Introduction
The clinical utility of CT is often hindered by two primary noise sources: the inherent noise in low-dose protocols and the artifacts caused by high-density metallic objects like hip replacements or dental fillings. Existing deep learning solutions for Low-Dose Metal Artifact Reduction (LDMAR) typically suffer from "scale neglect"—failing to balance local textures with global structures—and the "one-model-per-dose" problem, which necessitates storing multiple large models for different clinical settings.
Methodology
The proposed solution, PDuMSRNet, addresses these issues through a dual-domain reconstruction approach. The framework is built upon the principle of deep unfolding, where a mathematical optimization algorithm is translated into a neural network architecture, enhancing interpretability.
PMSRNet and Prompt Guiding
At the heart of each stage is the Prompt Guiding Multi-Scale Adaptive Sparse Representation-driven Network (PMSRNet). This network utilizes a "prompt guiding" module that accepts a dose map, a metal mask, and the input image. These prompts allow the network to dynamically adjust its parameters based on the specific noise and artifact characteristics of the scan.
Two key components within PMSRNet ensure high-fidelity reconstruction:
- PSATG (Prompt Guiding Scale-Adaptive Threshold Generator): This module adaptively generates thresholds for sparse representation by fusing local, regional, and global contextual information, guided by the input prompts.
- MSFuM (Multi-Scale Coefficient Fusion Module): Utilizing a cross-attention mechanism, MSFuM ensures that fine-grained details and global structural consistency are effectively integrated.
Experimental Setup
The model was evaluated using both simulated and clinical data:
- Training Data: 1,200 Full-Dose CT images from the DeepLesion dataset served as ground truth, with 100 diverse metal masks used to simulate implants.
- Noise Simulation: LDCT scans were synthesized at 1/2, 1/4, and 1/8 of standard dose levels using Poisson noise models.
- Validation: The CTPelvic1K clinical dataset was used to test real-world generalization on pelvic fracture cases.
Results and Discussion
PDuMSRNet demonstrated significant improvements over existing benchmarks:
- Quantitative Performance: On the DeepLesion test set at 1/2 dose, the model achieved a Peak Signal-to-Noise Ratio (PSNR) of 42.26 dB, a 1.6 dB improvement over the previous state-of-the-art, InDuDoNet.
- Storage Efficiency: By using a single "all-in-one" model for multiple dose levels, the storage requirement was reduced to approximately 53 MB, compared to the 159 MB required for three separate dose-specific models.
- Clinical Utility: Reconstruction quality was further validated using the Segment Anything Model (SAM), where PDuMSRNet-processed images yielded more accurate automated bone segmentations.
Limitations
While highly effective, the model still faces challenges with extremely large metallic implants that cause near-total loss of anatomical information. Additionally, the "domain gap" between simulated Poisson noise and the complex noise profiles of real-world scanners remains an area for future refinement.
Conclusion
PDuMSRNet introduces a flexible and efficient approach to LDMAR. By treating dose levels and metal masks as prompts within a deep unfolding framework, the researchers have created a versatile tool that meets the demands of diverse clinical environments without sacrificing reconstruction accuracy.
Summary updated