Multi-Image Super Resolution in Multi-Contrast MRI

Abstract

Acquisition of high-resolution magnetic resonance images (MRI) under distinct contrasts can enhance diagnostic information required in clinical diagnosis. Yet, acquiring high-resolution images might be impractical due to increased noise, prolonged scan durations and hardware costs. In such situations, an alternative solution can be the synthesis of high-resolution images from low-resolution images. Common methods perform super resolution of a single image. However, in multi-contrast MRI, the images of a single contrast might not contain sufficient prior information required for a successful deblurring. To enhance the required prior information, complementary prior information available in other contrasts can be used. Here, a multi-contrast MRI super resolution method is proposed to simultaneously deblur the images of multiple distinct contrasts. The proposed method relies on generative adversarial networks that can produce as realistic images as possible by better recovering high-frequency details. Qualitative and quantitative evaluations on a multi-contrast MRI dataset demonstrated that the proposed method outperforms the alternative single image MRI super resolution method.