ABSTRACT
Magnetic resonance imaging (MRI) is versatile and has been extensively used in clinics for medical diagnosis and biomedical research. Multimodality MR images, which provide images with different contrast and diverse diagnostic information, have been proven highly effective in improving diagnostic performance. However, the acquisition of high-quality and multimodality MR images is often hindered by high cost, missing imaging protocols, or unacceptable long scanning time. Multimodality MRI synthesis, which synthesizes the missing target-modality image from any given source-modality image, is thus an attractive alternative. This chapter discusses the recent progress in multimodality MRI synthesis and presents an overview of various representative methods in this field. Specifically, this chapter elaborates on and discusses one patch-based conventional learning method, one fully supervised deep learning method with a convolutional neural network, and one semi-supervised deep learning method with a generative adversarial network. All these discussed methods are extensively evaluated on a 7T MRI prediction from the 3T MRI task. A brief review and discussion of possible future research are presented toward the end of this chapter. This chapter is aimed to promote a better understanding of the fundamentals of multimodality MRI synthesis and help feature researchers grasp the principle of designing different synthesis methods.
