https://doi.org/10.1140/epjp/s13360-025-06588-3
Regular Article
MSACNet: an advanced multi-scale artifact correction network with motion-adaptive attention for the high-fidelity low-dose x-ray imaging
1
Department of Electronics and Communication Engineering, Saveetha Engineering College, 602 105, Thandalam, Chennai, Tamil Nadu, India
2
Department of Electronics and Communication Engineering, Karpagam Institute of Technology, Coimbatore, Tamil Nadu, India
3
Department of Computer Science and Engineering, Koneru Lakshmaiah Education Foundation, Vaddeswaram, Andhra Pradesh, India
4
Department of Mathematics, Vel Tech Rangarajan Dr. Sagunthala R&D Institute of Science and Technology, Avadi, Chennai, Tamil Nadu, India
Received:
21
May
2025
Accepted:
24
June
2025
Published online:
13
July
2025
High-fidelity low-dose X-ray imaging is essential in contemporary diagnostic radiology, offering a balance between minimizing radiation exposure and maintaining diagnostic accuracy. However, reducing the radiation dose significantly exacerbates challenges such as noise, photon starvation, motion artifacts, patient variability, and ring artifacts, each of which severely degrades image quality and diagnostic reliability. To address these persistent limitations, we propose MSACNet: An Advanced Multi-Scale Artifact Correction Network with Motion-Adaptive Attention for high-fidelity low-dose X-ray imaging. MSACNet is a modular deep learning framework designed to tackle the multifaceted challenges of low-dose acquisition through the integration of four synergistic components. First, the Multi-Scale Contrast-Adaptive Enhancement Module (MSCAEM) improves image contrast across scales while preserving anatomical structures and suppressing noise and ring artifacts. This significantly enhances feature visibility in soft tissues and fine structural regions. Second, the CSWin-X Transformer is augmented with the Motion-Adaptive Dual-Stream Attention (MADSA) module, which dynamically incorporates spatial and temporal attention to address motion-induced distortions, offering a computationally efficient alternative to non-rigid registration. Third, to handle variability across patients, we introduce MetaNorm + + , a meta-learned normalization strategy that adapts to distributional shifts in anatomical patterns. By learning normalization parameters conditioned on patient-specific features, it significantly improves generalization and reduces overfitting in diverse clinical settings. Finally, the Global-Selective Attention Transformer Block (GSAT Block) is employed to model both global context and local detail, enhancing artifact correction and structural consistency. This combination ensures that the network captures both the broader anatomical layout and fine-scale details essential for clinical interpretation. Comprehensive evaluations on a low-dose X-ray dataset demonstrate the superior performance of MSACNet over existing methods. Quantitatively, the model achieves a PSNR of 43.7 ± 0.2 dB, SSIM of 94.5 ± 0.3, LPIPS of 0.055 ± 0.003, MAE of 0.028 ± 0.002, RMSE of 0.070 ± 0.004, FSIM of 93.3 ± 0.3, and a NIQE of 3.10 ± 0.09%. These results confirm MSACNet’s ability to restore image fidelity while preserving clinical relevance under challenging low-dose conditions.
Copyright comment Springer Nature or its licensor (e.g. a society or other partner) holds exclusive rights to this article under a publishing agreement with the author(s) or other rightsholder(s); author self-archiving of the accepted manuscript version of this article is solely governed by the terms of such publishing agreement and applicable law.
© The Author(s), under exclusive licence to Società Italiana di Fisica and Springer-Verlag GmbH Germany, part of Springer Nature 2025
Springer Nature or its licensor (e.g. a society or other partner) holds exclusive rights to this article under a publishing agreement with the author(s) or other rightsholder(s); author self-archiving of the accepted manuscript version of this article is solely governed by the terms of such publishing agreement and applicable law.
