Synopsis

Keywords: Hyperpolarized MR (Gas), Hyperpolarized MR (Gas)

The utility of a deep learning based reconstruction tool for improving the quality of hyperpolarized ¹²⁹Xe lung ventilation images was assessed. DL-reconstructed ¹²⁹Xe ventilation image quality and SNR was improved compared with conventionally reconstructed images. In a cohort of patients with asthma and/or COPD, a small bias towards increased ventilation defect percentage, and a bias towards decreased coefficient of variation, in DL-reconstructed vs. conventionally-reconstructed images, was observed. Initial feasibility of utilising this tool for reduced-cost ¹²⁹Xe ventilation imaging using natural-abundance xenon, and improved spatial resolution imaging with 129-enriched xenon, is demonstrated.

Introduction

Hyperpolarized ¹²⁹Xe images can be signal-to-noise limited, especially in patients with severe lung disease who struggle to fully inhale the gas dose. For this reason, most human ¹²⁹Xe imaging studies are currently performed with 129-enriched Xe (>85% ¹²⁹Xe), as it provides ~3-fold SNR benefits over natural-abundance Xe (26% ¹²⁹Xe) (1); yet, the enrichment process is expensive (~5–10-fold cost increase per litre). As such, methods to improve image quality and SNR are economically desirable.
A deep convolutional neural network (CNN)-based image reconstruction tool that acts on raw MR data and produces images with increased sharpness and reduced noise (2) has been recently commercialised, and applied to imaging of various organs (3, 4). The CNN was trained with supervised learning using pairs of low-noise, high resolution images and typical noisy lower resolution counterparts.
The purpose of this work was to assess the utility of this DL reconstruction to improve the quality of hyperpolarized ¹²⁹Xe lung ventilation images, and whether the fidelity of quantitative metrics were preserved.

Methods

Retrospective analysis: N=34 ¹²⁹Xe ventilation MRI datasets that were acquired from patients with asthma, COPD and combined asthma-COPD were randomly selected from a database of >100 clinical ¹²⁹Xe MRI examinations performed between 2020–2022. In all examinations, the xenon dose was a 50:50 mix of 129-enriched Xe (polarisation ~25% (5)) and N₂ of total volume 1L (or less depending on patient height (6)). Images were reconstructed using the DL-recon tool without re-training (2), with a denoising “level” (scalar factor: 0–1) of 0.25, 0.5, 0.75 and 1. For comparison, images were also reconstructed using the conventional manufacturer recon pipeline (i.e. Fermi filtering, FFT etc.). For DL-reconstructed images at a nominal denoising level of 0.75, quantitative metrics of lung ventilation, namely; ventilation defect percentage (VDP) (7) and coefficient of variation (CV) (8), were calculated for comparison with conventional images. The original mask of the lungs derived from the conventional ¹²⁹Xe images with same-session ¹H MR anatomical scans was used to process the DL-reconstructed images. Bland-Altman plots were produced to identify differences in each metric. Image quality was compared qualitatively between reconstructions, and the structural similarity index measure (SSIM) was evaluated for each pair of images at all denoising levels to quantify their similarity.
Prospective acquisition: to explore the effect of DL-recon on low SNR images, and the feasibility for low-dose natural-abundance Xe MRI, N=3 healthy volunteers were scanned with a 50:50 mix of natural-abundance Xe and N₂ in a 1L Tedlar bag. N=1 of these healthy volunteers was also scanned with the usual dose of 129-enriched Xe described above, but with increased spatial resolution.
For both retrospective, and prospective natural-abundance Xe studies, acquisition parameters were: 1.5 T GE Healthcare scanner (HDx/450w), T-R vest coil (CMRS), 3D SSFP acquisition with in-plane FOV between 36–48cm, matrix 100x100 (in-plane resolution 3.6–4.8cm₂), slice thickness 10mm, flip angle ~10°, bandwidth 31.25kHz. For the prospective enriched Xe acquisition, the slice thickness was halved (5mm) and flip angle decreased (~7°).

Results & Discussion

Example images from the retrospective DL reconstructions of images from subjects with asthma and asthma+COPD are shown in Figure 1 alongside the conventional reconstructions, and their difference image. In all cases, an increase in image SNR and sharpness of the lung boundaries were found, and qualitatively, the physiological ventilation distribution remained visually unchanged. VDP was increased for DL-reconstructed images when compared with conventional images; Bland-Altman analysis revealed a positive bias of 1.1% (Figure 2). Part of this bias appears to arise from the sharper edges of the lungs on DL-reconstructed ¹²⁹Xe images; i.e. the reduced VDP may be partially artefactual due to using the same lung cavity mask for both images, rather than adapting the mask for DL-reconstructed images.
An opposing trend was observed in CV; a negative bias (-12%) towards reduced CV for DL-reconstructed images. This implies a reduced local heterogeneity, or smoothing, corresponding visually to reduced noise granularity of the images, rather than a smoothing out of physiological variations in ventilation (see e.g. difference images and pixel signal histograms in Figure 1). Median SSIM values are reported in Table 1. As expected, SSIM decreased with increasing denoising level. When separately assessing the SSIM for the lungs+airway region and for the background only, it was found that the SSIM was relatively unchanged in the lungs+airway, but decreased sharply in the background. This is a promising indication that the main shapes / structural features of the image are unchanged by the DL reconstruction.
In Figures 3&4, the results of applying the DL reconstruction to natural-abundance Xe and high-resolution enriched Xe images are shown. Whilst the conventional reconstructions exhibit borderline SNR for clinical use, the DL-reconstructed images exhibit sufficient SNR and contrast for clinical evaluation.

Conclusion

DL-based image reconstruction was found to improve ¹²⁹Xe ventilation image quality and SNR. Further application of this tool on images acquired from patients with a range of lung pathologies is required to fully evaluate the physiological interpretation of the resulting images. This approach could enable routine low-cost ¹²⁹Xe ventilation imaging using natural-abundance xenon, and / or improved spatial resolution imaging with 129-enriched xenon.

Acknowledgements

No acknowledgement found.