Shortening of Analysis time for T2wsup Synthetic Diffusion imaging (T2wsup-dMRI) with Deep Neural Network (DNN)

Motivation: T2wsup-dMRI combining Triangle-pattern sampling in (TE, b) space with the nonlinear least square (LSQ) fitting provides lower errors (RE and CV) for the quantitative maps but is time-consuming. Goal(s): To shorten the analysis time of LSQ fitting by applying a Deep Neural Network (DNN) while minimizing errors. Approach: The errors between DNN and 2dLSQ fitting for several sampling patterns were compared using digital phantom and MRI data by simulation. Results: The DNN outperformed the 2dLSQ fitting in terms of time (~1/500) while maintaining comparable errors. The caveat for the DNN is that the learning is required depending on the data pattern and SNR. Impact: The analysis time for the 2d LSQ fitting with a Triangle-pattern in (TE, b) space was significantly shortened by the DNN while maintaining comparable REs and CVs. Although further optimization is required, clinical application is promising.