Item

Abstract

Aiming to tackle data deficiency in 9-Tesla Magnetic Resonance Image(MRI) anatomic images of human brain, which fits an adequate amount for deep neural network training, we applied generative neural networks to produce super-resolution 3D images based on extensive amount of 3T data. Such synthetic data own two main attributes to provide training model with essential features included in 9-Tesla images, ultra-high spatial resolution and the distinguishable contrast, thus a supervised neural network would gain better prediction accuracy benefiting from such realistic data augmentation. Additionally, such augmentation scheme avoids offending privacy from real patients as well as expensive scanning, especially when it comes to such data-driven neural network jobs. Moreover, high quality MR images better resolved contours of tissues and are helpful for follow-up data analysis, e.g. image registration, segmentation, etc., which employed advantage of the prevailing convolutional neural networks.