A Learning-based Method for Quantifying the Fraction of Unsaturated Fatty Acid in Bone Marrow

Motivation: The level of unsaturated fatty acid in bone is an useful biomarker for assessing the health condition of bone. Goal(s): Predict the fraction of unsaturated fatty acid from a reduced number of gradient echo image data and the uncertainty level simultaneously. Approach: A Bayesian neural network is trained to refine the coarse parametric map fitted by a reduced number of images to the parametric map fitted by the full number of scans. Results: Experiments on knee data show that the model can achieve a relative error of less than 5% and the uncertainty prediction reflects the confidence. Impact: The feasibility of mapping the fraction of unsaturated fatty acid from a reduced number of imaging data in a learning-based way is validated.