A multi-scale cross aggregation network for joint denoising and super-resolution of low-dose CT images

Introduction Low-dose computed tomography (LDCT) reduces radiation exposure but often introduces signal-dependent noise and spatially varying streak artifacts that obscure subtle anatomical structures. Existing restoration methods struggle to suppress these degradations while maintaining long-range anatomical consistency and recovering high-frequency details. To address these challenges, we propose MSCANet, a Multi-Scale Cross Aggregation Network for joint LDCT denoising and super-resolution. Methods MSCANet integrates three complementary components. The Token Fusion and Aggregation Block combines global structural context and local texture information through prototype-guided feature aggregation. Region-Aware Self-Attention models local contextual relationships with overlapping receptive fields, while the Multi-Scale Deformable Attention Module employs learnable sampling offsets to characterize irregular structures and spatially varying artifacts. We evaluated MSCANet on the public 3D-IRCADB and PANCREAS datasets at × 2 and × 4 scaling factors and compared it with nine representative image-restoration methods. An additional qualitative evaluation was conducted on a self-collected clinical CT dataset comprising 2,544 slices from 17 patients. Reconstruction quality was assessed using peak signal-to-noise ratio (PSNR), structural similarity index measure (SSIM), contrast-to-noise ratio (CNR), and Learned Perceptual Image Patch Similarity (LPIPS). Model efficiency was evaluated using the parameter count and floating-point operations (FLOPs). Results On the public datasets, MSCANet achieved the best PSNR, SSIM, and CNR among the compared methods at both scaling factors. At × 4 , it obtained PSNR/SSIM/CNR values of 31.9514 dB/0.8383/0.7592 on 3D-IRCADB and 32.1740 dB/0.8422/1.1692 on PANCREAS. These results corresponded to PSNR gains of 0.2320 and 0.0922 dB over the respective second-best methods. MSCANet also achieved the lowest LPIPS value of 0.0517 while requiring only 11.03 million parameters and 142.72 GFLOPs. Qualitative comparisons on the self-collected dataset across axial, coronal, and sagittal views showed clearer tissue boundaries and more continuous anatomical structures after restoration. Ablation studies further confirmed the complementary contributions of the proposed components. Conclusion MSCANet provides an effective balance between noise suppression, artifact reduction, anatomical structure preservation, and computational efficiency for joint LDCT denoising and super-resolution. The results support its potential for LDCT image restoration, while observer-based assessment and downstream clinical validation are still required to establish its diagnostic value.

Authors

Institutions

Publication Details

Journal
Frontiers in Medicine
Published
2026-11-17
DOI
https://doi.org/10.3389/fmed.2026.1961857
Primary Topic
Image and Signal Denoising Methods
Type
article
Field-Weighted Citation Impact
0.00
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
article

A multi-scale cross aggregation network for joint denoising and super-resolution of low-dose CT images

Qizhi Fu, Pengju Si, Haokun Zhang, Huan Wang et al.
2 links
Frontiers in Medicine
Image and Signal Denoising Methods
article

A multi-scale cross aggregation network for joint denoising and super-resolution of low-dose CT images

Qizhi Fu, Pengju Si, Haokun Zhang, Huan Wang, Min Wang
article en

Abstract

Introduction Low-dose computed tomography (LDCT) reduces radiation exposure but often introduces signal-dependent noise and spatially varying streak artifacts that obscure subtle anatomical structures. Existing restoration methods struggle to suppress these degradations while maintaining long-range anatomical consistency and recovering high-frequency details. To address these challenges, we propose MSCANet, a Multi-Scale Cross Aggregation Network for joint LDCT denoising and super-resolution. Methods MSCANet integrates three complementary components. The Token Fusion and Aggregation Block combines global structural context and local texture information through prototype-guided feature aggregation. Region-Aware Self-Attention models local contextual relationships with overlapping receptive fields, while the Multi-Scale Deformable Attention Module employs learnable sampling offsets to characterize irregular structures and spatially varying artifacts. We evaluated MSCANet on the public 3D-IRCADB and PANCREAS datasets at × 2 and × 4 scaling factors and compared it with nine representative image-restoration methods. An additional qualitative evaluation was conducted on a self-collected clinical CT dataset comprising 2,544 slices from 17 patients. Reconstruction quality was assessed using peak signal-to-noise ratio (PSNR), structural similarity index measure (SSIM), contrast-to-noise ratio (CNR), and Learned Perceptual Image Patch Similarity (LPIPS). Model efficiency was evaluated using the parameter count and floating-point operations (FLOPs). Results On the public datasets, MSCANet achieved the best PSNR, SSIM, and CNR among the compared methods at both scaling factors. At × 4 , it obtained PSNR/SSIM/CNR values of 31.9514 dB/0.8383/0.7592 on 3D-IRCADB and 32.1740 dB/0.8422/1.1692 on PANCREAS. These results corresponded to PSNR gains of 0.2320 and 0.0922 dB over the respective second-best methods. MSCANet also achieved the lowest LPIPS value of 0.0517 while requiring only 11.03 million parameters and 142.72 GFLOPs. Qualitative comparisons on the self-collected dataset across axial, coronal, and sagittal views showed clearer tissue boundaries and more continuous anatomical structures after restoration. Ablation studies further confirmed the complementary contributions of the proposed components. Conclusion MSCANet provides an effective balance between noise suppression, artifact reduction, anatomical structure preservation, and computational efficiency for joint LDCT denoising and super-resolution. The results support its potential for LDCT image restoration, while observer-based assessment and downstream clinical validation are still required to establish its diagnostic value.

Frontiers in MedicineVol. 13
Henan University of Science and Technology (CN), First Affiliated Hospital of Henan University of Science and Technology (CN)
Good health and well-being
Image and Signal Denoising Methods
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon