Self-Supervised, Deep Learning Denoising of GRAPPATINI T2-Maps in Magnetic Resonance Imaging

Magnetic Resonance Imaging (MRI) is a widely used medical imaging modality, offering non-invasive and non-ionizing diagnostic capabilities. However, conventional MRI relies on contrast-weighted images influenced by various physical parameters and scanner characteristics, making their interpretation still highly qualitative and subjective. This approach limits analysis only to relative changes in intensities or morphological features and no absolute comparisons can be performed. Quantitative MRI (qMRI) is shifting this perspective by enabling the generation of parameter-specific maps, such as relaxation times T1 and T2 maps, capable of providing absolute and quantitative information, enabling longitudinal studies, inter-subject comparisons, and the establishment of normative ranges. Specifically, T2 relaxation times have been shown to be highly sensitive to physiological and pathological changes, such as intra- and extra-cellular water accumulation and myelin loss. Despite these advantages, gold standard T2 mapping techniques are still constrained by long acquisition times, impacting patient comfort and increasing the susceptibility to motion artifacts. GRAPPATINI, an accelerated technique, address these limitations by leveraging the orthogonality of two previously established methods in accelerating MRI, GRAPPA and MARTINI. However, this acceleration comes at the cost of increased noise level and reduced signal-to-noise ratio (SNR) in the reconstructed T2 maps. Consequently, there is an urgent need to improve the final quality of T2 maps reconstructed using GRAPPATINI. This thesis proposes two novel strategies to denoise GRAPPATINI T2 maps using a self-supervised machine learning framework to train a deep learning model, with the aim of bypassing the need for large datasets typically required for supervised learning approaches. The first strategy, operating in the k-space domain, was found to be ineffective, while the second, implemented in the image-space domain, demonstrated exceptional performance on the test set. Furthermore, the image-space strategy showed remarkable generalizability considering 7T brain and knee datasets acquired with different resolutions, field strengths, anatomies, and orientations compared to those used during the training process. Statistical validation was performed through scan-rescan analyses, with results confirming its ability to preserve unbiased and reproducible T2 values. Reproducibility and generalizability are the key milestones of this work, enabling potential integration of the strategy into scanner reconstruction pipelines without retraining or modifications, independently from the anatomy or the acquisition settings. Eventually, achieving superior denoising performance compared to traditional methods, this work represents a significant advancement toward the potential clinical adoption of GRAPPATINI, both at 3T and 7T, where the combination of high-resolution and high-quality imaging can drive innovations in both research and clinical practice.