A Cramér–Rao Lower Bound-informed 3D Convolutional Neural Network for Multi-delay pCASL Estimation

Motivation: Weighted delay methods and mean squared error (MSE) fitting have limited capabilities in estimating ATT and CBF. The performance of deep learning models is constrained by the differences in orders of parameters within loss functions. Goal(s): To develop a novel loss function to overcome the limitations of MSE in estimating ATT and CBF. Approach: A CRLB-informed MSE was proposed. The accuracy of CBF and ATT was compared on a 3D CNN trained with MSE loss and CRLB-informed MSE loss. The performance of the proposed model was assessed on ASL with reduced PLDs and repetitions. Results: CRLB-informed MSE loss significantly improved CBF estimation. Impact: A CNN model with CRLB-informed MSE loss offered improved accuracy and robustness in CBF estimation and in handling outliers, which may provide an efficient method to quantify ASL images.