Title: Non-Cartesian MR imaging in the deep learning era

Abstract: High-quality MRI reconstruction at highly accelerated acquisition rates (≥ 6) is challenging due to the ill-posed nature of the inverse problem. Traditional Compressed Sensing (CS) methods struggle to maintain quality at these acceleration factors, while deep learning approaches, though effective, face issues like overfitting and hallucinations when acquisition settings vary between training and testing. Unrolled neural networks improve image quality but require fully sampled k-space data for supervised training. Additionally, training such architecture depends on the forward operator (e.g., sampling pattern, coil configuration, FFT vs. NUFFT) involved in the acquisition process, thereby complicating clinical deployment if the acquisition setting changes from training to testing.

In this talk, I will discuss recent advancements in Plug-and-Play (PnP) approaches that address the limitations of CS and supervised learning. PnP algorithms replace handcrafted CS priors with powerful denoising deep neural networks (DNNs). However, existing PnP methods often yield suboptimal results due to instabilities in proximal gradient descent schemes and the lack of noiseless datasets for training robust denoisers. To overcome these issues, we propose a fully unsupervised preprocessing pipeline to generate clean, noiseless complex coi-combined MR images from multicoil data, enabling the training of high-performance denoising DNNs in 2D and 3D MR imaging. Additionally, we introduce an annealed Half-Quadratic Splitting (HQS) algorithm to address stability issues, leading to significant improvements over existing PnP algorithms. Combined with preconditioning techniques, our approach achieves state-of-the-art results in 2D MRI and unmatched results in 3D non-Cartesian highly accelerated (x10) anatomical MRI, offering a robust and efficient solution for high-quality MRI reconstruction.