Enhance the Image: Super Resolution using Artificial Intelligence in MRI

Abstract: This chapter provides an overview of deep learning techniques for improving the spatial resolution of MRI, ranging from convolutional neural networks, generative adversarial networks, to more advanced models including transformers, diffusion models, and implicit neural representations. Our exploration extends beyond the methodologies to scrutinize the impact of super-resolved images on clinical and neuroscientific assessments. We also cover various practical topics such as network architectures, image evaluation metrics, network loss functions, and training data specifics--including downsampling methods for simulating lowresolution images and dataset selection. Finally, we discuss existing challenges and potential future directions regarding the feasibility and reliability of deep learning-based MRI superresolution, with the aim to facilitate its wider adoption to benefit various clinical and neuroscientific applications. Keywords: Single-image super-resolution, deep learning, convolutional neural network, generative adversarial network, transformer, diffusion model, implicit neural representation, loss function, transfer learning, uncertainty estimation. Introduction MRI with higher spatial resolution provides more detailed insights into the structure and function of living human bodies non-invasively, which is highly desirable for accurate clinical diagnosis and image analysis. The spatial resolution of MRI is characterized by in-plane and through-plane resolutions (Figure 1). On the other hand, the through-plane resolution, also referred to as slice thickness, is determined differently for 2D and 3D imaging. In 2D imaging, the slice thickness is defined by the full width at half maximum (FWHM) of the slice-selection radiofrequency (RF) pulse profile. In 3D imaging, the slice-selection direction is encoded by another phase encoding gradient. Consequently, the through-plane resolution is determined similarly to the in-plane resolution by the maximal extent of the k-space along slice-selection direction as in Eq. 1. The in-plane resolution is dictated by the k-space coverage, and a larger k-space coverage brings higher spatial resolution (a). The slice thickness is determined by the slice-selective RF pulse for 2D imaging, and by k-space extent along sliceselection direction for 3D imaging (b).