In this talk, Dr. Lin will review contemporary MRI reconstruction methods, and propose a stable network for MRI reconstruction. MRI is an essential imaging tool for diagnosis and prognosis, but its application has largely been restricted because it needs a long scanning time. During the past two decades, accelerated MRI acquisition methods have been hotly discussed. Most recent MRI acceleration methods acquire the subsampled k-space, which exploits advanced image reconstruction algorithms and parallel imaging to reconstruct the image. A priori information is used in reconstruction methods to eliminate image artifacts. Several variant algorithms, such as iterative reconstructions, compressed sensing and non-linear optimization methods, have been reported in the literature.

Iterative reconstructions generally deliver excellent image qualities but can entail heavy computation burdens and substantial costs in architecture developments. Recent developments in parallel computing architectures using graphic processing units (GPUs) have shortened the computation times, leading to near-real-time performance. However, maintaining heterogeneous computing architectures incurs significant hardware/software complexities. Moreover, the design of sparse transforms depends on users' heuristics, although the experiences in one research area can rarely be translated to medical imaging. Sparse transforms usually accompany subtle artifacts but radiologists can perceive them. For example, stair-casing, oil-painting, or over-smoothing artifacts have hampered recent development in compressed sensing.

The recent machine learning (ML) frameworks have taken an empirical approach and have rerouted multi-layers of neurons, yielding a linear mapping between the distorted image and the result. Although many general-purpose ML frameworks calculate results faster than optimization methods, the disadvantages of these methods include numerical instabilities, inexplicable results, dependence on large data, and lack of support for complex values. In the literature, the complex-domain back-propagation (CDBP) was used for reducing the ringing artifacts of constrained MRI reconstruction, but these earlier methods have vanished in recent ML models. Some image domain learning methods, such as U-net and ResNet, are restricted to magnitude-only imagery, while many recent attempts to reroute complex data, such as AUTOMAP and hybrid technique, have reinvented the CDBP in an unsystematic way. Given the fundamentals of MR physics, CDBP requires the network for MRI reconstruction to be redesigned.

Transfer learning might be used but the performance of transfer learning has proved inferior and, given the hierarchy of hospitals, small hospitals cannot offer training models with the same size of data that large ones can. Recently, researchers in ML for MRI have explored the unrolled version of iterative reconstructions, thereby achieving an image quality comparable with that of compressed sensing. Regarding the practical barriers against current ML frameworks, many factors in clinical settings can degrade performance. For instance, the coil sensitivities and the geometry of patients can change between patients and organs. These logistic problems (the difficulties of acquiring sufficient data) can diminish the efficacy of general ML, and future research is necessary.

Dr. Jyh-Miin Lin, MD, MSc, PhD

• Education

2012–2016 Ph.D. Radiology, University of Cambridge

2005–2007 M.Sc. Department of Electrical Engineering, National Taiwan University

1999–2005 M.D. Department of Medicine, National Taiwan University

• Academic Appointments

2019–2020 Senior Research Associate CEA Grenoble

2018–2019 Research Associate University College London

2016–2018 Postdoc Department of Electrical Engineering, National Taiwan University

2011–2012 Post-doctoral (MD) research assistant Duke University Medical Center, NC, USA