Deep learning in medical image super resolution: a review

Super-resolution (SR) reconstruction is a hot topic in medical image processing. SR implies reconstructing corresponding high-resolution (HR) images from observed low-resolution (LR) images or image sequences. In recent years, significant breakthroughs in SR based on deep learning have been made, and many advanced results have been achieved. However, there is a lack of review literature that summarizes the field’s current state and provides an outlook on future developments. Therefore, we provide a comprehensive summary of the literature on medical image SR (MedSR) based on deep learning since 2018 in five aspects: (1) The SR problem of medical images is described, and the methods of image degradation are summarized. (2) We divide the existing studies into three categories: two-dimensional image SR (2DISR), three-dimensional image SR (3DISR), and video SR (VSR). Each category is subdivided. We analyze the network structure and method characteristics of typical methods. (3) Existing SR reconstruction quality evaluation metrics are presented in detail. (4) The application of MedSR methods based on deep learning is discussed. (5) We discuss the challenges of this phase and point out valuable research directions.