Deep Learning Experimental Design for Quantitative Parameter Mapping

Motivation: To maximise scanning time efficiency in diffusion MRI and quantitative MRI. Goal(s): To demonstrate a new deep learning approach for experimental design of acquisition protocols for MRI parameter mapping. Approach: We utilise TADRED (TAsk-DRiven Experiment Design) to simultaneously train two networks: one to optimise the acquisition protocol and one to optimise a neural network for parameter estimation. We demonstrate on three diffusion MRI applications - NODDI, VERDICT, and ADC mapping - and T1 inversion recovery. Code is available at https://github.com/sbb-gh/ED_MRI . Results: TADRED demonstrates superior or comparable performance in estimating model parameters compared to the Cramer-Rao lower bound (CRLB) baseline across all experiments. Impact: TADRED enables shorter, more efficient diffusion and quantitative MRI protocols. This can reduce scan time and costs, reduce motion artifacts, and enhance patient comfort.