Plug-and-Play Self-Supervised Denoising for Pulmonary Perfusion MRI

Changyu Sun; Yu Wang; Cody Thornburgh; Ai-Ling Lin; Kun Qing; John P Mugler 3rd; Talissa A Altes

doi:10.3390/bioengineering12070724

Plug-and-Play Self-Supervised Denoising for Pulmonary Perfusion MRI

Bioengineering (Basel). 2025 Jul 1;12(7):724. doi: 10.3390/bioengineering12070724.

Authors

Changyu Sun^{1

2}, Yu Wang¹, Cody Thornburgh², Ai-Ling Lin^{2

3

4}, Kun Qing⁵, John P Mugler 3rd⁶, Talissa A Altes²

Affiliations

¹ Department of Chemical and Biomedical Engineering, University of Missouri, Columbia, MO 65211, USA.
² Department of Radiology, University of Missouri, Columbia, MO 65211, USA.
³ Division of Biological Sciences, University of Missouri, Columbia, MO 65211, USA.
⁴ Institute for Data Science and Informatics, University of Missouri, Columbia, MO 65211, USA.
⁵ Department of Radiation Oncology, City of Hope National Medical Center, Duarte, CA 91010, USA.
⁶ Department of Radiology and Medical Imaging, University of Virginia, Charlottesville, VA 22908, USA.

Abstract

Pulmonary dynamic contrast-enhanced (DCE) MRI is clinically useful for assessing pulmonary perfusion, but its signal-to-noise ratio (SNR) is limited. A self-supervised learning network-based plug-and-play (PnP) denoising model was developed to improve the image quality of pulmonary perfusion MRI. A dataset of patients with suspected pulmonary diseases was used. Asymmetric pixel-shuffle downsampling blind-spot network (AP-BSN) training inputs were two-dimensional background-subtracted perfusion images without clean ground truth. The AP-BSN is incorporated into a PnP model (PnP-BSN) for balancing noise control and image fidelity. Model performance was evaluated by SNR, sharpness, and overall image quality from two radiologists. The fractal dimension and k-means segmentation of the pulmonary perfusion images were calculated for comparing denoising performance. The model was trained on 29 patients and tested on 8 patients. The performance of PnP-BSN was compared to denoising convolutional neural network (DnCNN) and a Gaussian filter. PnP-BSN showed the highest reader scores in terms of SNR, sharpness, and overall image quality as scored by two radiologists. The expert scoring results for DnCNN, Gaussian, and PnP-BSN were 2.25 ± 0.65, 2.44 ± 0.73, and 3.56 ± 0.73 for SNR; 2.62 ± 0.52, 2.62 ± 0.52, and 3.38 ± 0.64 for sharpness; and 2.16 ± 0.33, 2.34 ± 0.42, and 3.53 ± 0.51 for overall image quality (p < 0.05 for all). PnP-BSN outperformed DnCNN and a Gaussian filter for denoising pulmonary perfusion MRI, which led to improved quantitative fractal analysis.

Keywords: deep learning; denoising; plug-and-play; pulmonary perfusion MRI; self-supervised learning.