Deep Learning-Based Contrast Transformation of High Resolution 3D Gradient Echo Images Trained Using Low Resolution 2D Turbo Spin Echo Images

Abstract

Acquiring multiple clinically relevant MR image contrasts from a single scan is an emerging trend in MR imaging to reduce the total scan time of an exam. A deep learning-based approach that takes this idea one step further is BoneMRI, which generates synthetic CT images from MR images. These images are generated using a 3D RF-spoiled T1‐weighted multiple Gradient-Echo sequence (GE), but the sequence is often combined with a 2D T1-weighted Turbo Spin Echo (TSE), because the latter provides better T1 contrast from a clinical perspective. To reduce the total scan time, we investigate a deep learningbased approach to generate synthetic TSE images from the GE images. We propose a training approach, called the High-to-low approach, to keep the high through-plane resolution of the GE images, while still applying the contrast transformation using only the lower resolution TSE as target data. Additionally, we implement a network architecture, called HighResNet, and redesign it for the synthesis of TSE images from GE images. The proposed approach and network do not necessarily need to be used together and both were validated against a more often used approach and network, respectively. Experiments using scans of the cervical spine showed that the High-to-low approach was capable of keeping the higher through-plane resolution of the GE images, while also achieving significantly higher image similarity to the lower resolution ground-truth TSE after downsampling. The experiments also demonstrated that HighResNet synthesized high resolution synthetic TSE images with fewer artifacts than the often used U-Net. The results show that a neural network is capable of learning an MR contrast transformation between a higher resolution input image and lower resolution output image without sacrificing performance or image resolution.