Benchmarking Hybrid CNN-Transformer Versus Pure Transformer Architectures for Accelerated Hyperpolarized 129Xe MRI Reconstruction

Ramtin Babaeipour; Matthew S Fox; Grace Parraga; Alexei Ouriadov

doi:10.1002/jmri.70314

Benchmarking Hybrid CNN-Transformer Versus Pure Transformer Architectures for Accelerated Hyperpolarized ¹²⁹Xe MRI Reconstruction

J Magn Reson Imaging. 2026 Mar 27. doi: 10.1002/jmri.70314. Online ahead of print.

Authors

Ramtin Babaeipour¹, Matthew S Fox^{2

3

4}, Grace Parraga^{1

3

4

5}, Alexei Ouriadov^{1

2

3

4}

Affiliations

¹ School of Biomedical Engineering, Faculty of Engineering, The University of Western Ontario, London, Ontario, Canada.
² Department of Physics and Astronomy, The University of Western Ontario, London, Ontario, Canada.
³ Lawson Research Institute, London, Ontario, Canada.
⁴ Department of Medical Biophysics, The University of Western Ontario, London, Ontario, Canada.
⁵ Robarts Research Institute, The University of Western Ontario, London, Ontario, Canada.

PMID: 41891379
DOI: 10.1002/jmri.70314

Abstract

Background: Hyperpolarized ¹²⁹Xe MRI faces technical challenges including low signal-to-noise ratio and breath-hold constraints. Current literature focuses on proprietary deep learning methods or image-domain enhancements.

Purpose: To present a comprehensive evaluation of transformer and hybrid CNN-transformer architectures integrating dual-domain (k-space and image) processing for HP ¹²⁹Xe MRI reconstruction.

Study type: Retrospective.

Population: Two hundred five participants (22 healthy [male and female, 18-85 years], 26 COPD [male and female, 50-85 years], 90 asthma [male and female, 18-70 years], 67 long-COVID [male and female, 18-70 years]) yielding 1640 2D slices. Dataset split: 80% training (1312 slices), 10% validation (164 slices), 10% test (164 slices).

Field strength/sequence: 3 T; 3D fast gradient-recalled echo.

Assessment: Five architectures were compared: KTMR (hybrid transformer-CNN), KIKI-net (pure CNN), ReconFormer, SwinMR, and MR-IPT (pure transformer) at acceleration factors of 3, 7, and 10. Performance was assessed using peak signal-to-noise ratio (PSNR), structural similarity index measure (SSIM), and normalized mean squared error (NMSE). Ventilation defect percentage (VDP) agreement with semi-automated analysis was evaluated.

Statistical tests: Friedman test with post hoc Dunn's test and Benjamini-Hochberg correction for multiple comparisons. Significance level: p < 0.05.

Results: At 10-fold acceleration, KTMR produced PSNR of 36.4 ± 2.8 dB and SSIM of 0.88 ± 0.12, significantly outperforming KIKI-net (32.5 ± 3.4 dB, 0.81 ± 0.12), ReconFormer (29.7 ± 2.6 dB, 0.76 ± 0.12), SwinMR (30.5 ± 2.8 dB, 0.76 ± 0.09), and MR-IPT (28.8 ± 2.4 dB, 0.74 ± 0.11). VDP measurements showed mean bias of 1.94% at 3-fold, 2.12% at 7-fold, and 2.69% at 10-fold acceleration.

Data conclusion: KTMR demonstrated superior performance for HP ¹²⁹Xe MRI reconstruction at high acceleration factors.

Evidence level: 3.

Technical efficacy: Stage 1.

Keywords: COPD; MRI reconstruction; Vision Transformers; asthma; deep learning; hyperpolarized 129Xe MRI; long‐COVID; lung imaging; medical imaging; pulmonary imaging.

Plain language summary

Hyperpolarized ¹²⁹Xe MRI is a specialized lung imaging technique that allows doctors to visualize how air moves through the lungs, helping diagnose conditions like asthma, COPD, and long‐COVID. However, the scan requires breath‐holding and uses expensive gas. This study tested five artificial intelligence methods to reconstruct high‐quality lung images from faster, reduced data acquisitions. Using data from 205 patients, it was found that a hybrid approach combining convolutional neural networks and transformers produced the best image quality, even at very high acceleration. This could reduce breath‐hold times from 10 to 2–3 s, making the scan more accessible for vulnerable patients.

Grants and funding

R9245A04/Natural Sciences and Engineering Research Council of Canada