Generative Adversarial Network (GAN)-based High-Fidelity Diffusion Tensor Imaging in Brain Cancer Patients

Motivation: Acquiring more than 30 diffusion weighted images (DWIs), required for precise estimation of DTI metrics leads to longer scan time, hindering its clinical application. Goal(s): We aim to develop and validate a deep-learning (DL) network for high-fidelity DTI in brain cancer patients. Approach: A Generative Adversarial Network (GAN) was trained using 32-direction DWI to predict 6-direction high-quality DWI volumes for accurate estimation of DTI metrics. Results: The predicted DWI and fractional anisotropy (FA) maps showed higher structural similarity and peak signal-to-noise ratio (PSNR) with the ground-truth than the input images. Impact: Our study demonstrates the potential of DL to reduce DTI scan time for brain cancer patients by enabling accurate estimation of DTI metrics from only 6-direction DWI volumes; and enhancing their image quality and signal-to-noise ratio.