Motivation: At high-field strength, chemical-shift artifacts significantly impair image quality, especially for 3D radial sequences such as UTE or PETRA. Goal(s): To train deep-learning models to generate water-only and fat-only images from conventional single-echo UTE data, facilitating Dixon-like fat/water separation without additional acquisition time. Approach: We acquired UTE, water-excitation (WE), and fat-excitation (FE) UTE images of the knee and trained supervised models to predict water, fat, and chemical-shift-corrected images from the conventional UTE images. Results: Predicted water and fat images reflect anatomical details observed in the corresponding ground-truth images and a contrast was achieved free of chemical shift. Impact: We present a novel deep-learning Dixon-like fat/water separation and chemical-shift suppression for ultra-short echo time (UTE) MRI. Our method predicts water-only and fat-only images from conventional 7T UTE data, demonstrating a promising approach for high-field imaging applications.
Sommer et al. (Tue,) studied this question.