ADVANCING LUNG AND LIVER MRI: FROM COMPRESSED SENSING TO DEEP LEARNING

Abstract

Magnetic resonance imaging (MRI) has advanced with the development of accelerated imaging techniques, including parallel imaging, compressed sensing (CS), and more recently, deep learning. These innovations have enhanced MRI speed and image quality, allowing for broader clinical applications. However, adapting these methods to specific organs remains challenging, specially for organs affected by respiratory and cardiac motion, such as the lungs and liver. Lung MRI, for example, suffers from low proton density, limiting the signal-to-noise ratio (SNR) in traditional proton based MR imaging. Hyperpolarized 129Xe imaging provides an effective alternative, enhancing lung visualization but necessitating breath-holds that can be difficult for patients. CS has shown promise in reducing MRI scan durations. In this work, a thorough investigation of its application in hyperpolarized 129Xe ventilation MRI was conducted in Chapter 3. The resultant image quality was found to be sensitive to the chosen Gaussian undersampling pattern, even when the undersampling percentage remained constant. This sensitivity to undersampling patterns, along with parameters like regularization factors, complicates CS optimization and highlights the need for methods with greater flexibility and reduced reliance on manual tuning of reconstruction settings, such as deep learning techniques. In a similar problem, ultra-high-field (7T) MRI holds great potential for achieving higher SNR and improved image quality compared to standard 1.5T and 3T scanners, but image resolution remains limited by breath-hold requirements for motion-sensitive organs. To address this a free-breathing 7T abdominal imaging protocol, employing a pseudo-spiral sampling pattern with CS reconstruction, was developed in Chapter 4. This approach demonstrated high image quality while addressing the challenge of patient discomfort associated with breath-holding and reducing motion artifacts through retrospective respiratory gating. However, manually tuning CS to optimize hyperparameters and maximize reconstruction performance can be time-consuming. This suggests that using a deep learning technique with increased freedom could benefit the reconstruction process. Recently, unrolled CS techniques have leveraged the power of deep learning algorithms to overcome the challenges associated with conventional CS. The integration of deep learning into CS facilitates the automatic optimization of underlying parameters, enhancing the quality of reconstructed images. To address the challenges of conventional CS mentioned earlier, the Cascade of Independent Recurrent Inference Machine (CIRIM) was considered a viable alternative in this study. First, unlike CS, which relies on a predetermined l1 constraint to promote image sparsity, CIRIM iteratively enhances image quality through learned priors. Furthermore, previous studies have shown CIRIM's ability to outperform conventional CS and other state-of-the art deep learning (DL) reconstruction techniques. It also demonstrates promising robustness to variations in undersampling patterns, making it an excellent candidate for this study. CIRIM demonstrated its ability to adapt across varying Gaussian undersampling patterns, successfully reconstructing both hyperpolarzied 129Xe lung MRI as discussed in Chapter 5 and 7T breath-hold abdominal data in Chapter 6, outperforming CS in terms of image quality and robustness. Furthermore, CIRIM's adaptability was assessed by training on Cartesian breath-hold abdominal data undersampled with a Gaussian pattern and applying it to prospectively undersampled, free-breathing pseudo-spiral data in Chapter 6. By successfully reconstructing these images, the method demonstrated a particular capacity to generalize across Cartesian and pseudospiral sampling. In conclusion, to achieve high-quality imaging at low scan durations, a robust reconstruction technique capable of handling high acceleration factors is necessary. This study advances the state of the art in MRI by demonstrating that combining the advantages of deep learning reconstruction and non-Cartesian acquisition can address key challenges of conventional CS, including its sensitivity to sampling patterns and the need for manual hyperparameter tuning, while enabling robust, high-quality reconstructions even under suboptimal undersampling conditions. Specifically, this work applies these advancements to 129Xe lung MRI, addressing its sensitivity to Cartesian undersampling, and to free-breathing 7T liver MRI, overcoming challenges associated with breath-hold requirements, motion artifacts, and scan efficiency. Future work should build on these advancements by evaluating alternative non-Cartesian sampling methods, such as radial sampling, to further optimize imaging for motionsensitive organs. By tailoring sampling strategies and improving reconstruction robustness, these techniques can enhance imaging resolution, increase scan efficiency, and address patient comfort through tolerable scan times, paving the way for broader clinical and research applications.