Deep Learning-Based Prediction of Gadolinium-Enhanced MRI from Non-Contrast MRI and MR Fingerprinting

Motivation: Gadolinium-enhanced MRI is crucial for brain tumor diagnosis, but drawbacks associated with gadolinium-based contrast agents (GBCAs) drive interest in developing alternative imaging methods. Goal(s): Our goal is to develop a model capable of predicting gadolinium-enhanced MR images that closely resemble those obtained with contrast injection, in terms of tumor enhancement overlap and overall appearance. Approach: We developed a deep-learning based model that can be trained with pre-contrast MRF maps, synthetic contrasts, and/or clinical images and predict gadolinium-enhanced MR images. Results: The models obtained competitive enhancing tumor overlap values (0.599, 0.603, 0.622) and quantitative metrics (average SSIM=0.944, CC=0.941, and MAE=0.261). Impact: The clinical use of a deep learning model to obtain gadolinium-enhanced MR images would replace the need for GBCAs, thus eliminating associated problems including longer scan times, higher costs, and certain patient risks.