DeepDECOMPOSE: A Deep Learning based framework for solving DECOMPOSE QSM

We propose a deep learning (DL) approach to accelerate and improve the accuracy of the parameter fitting problem of DECOMPOSE-QSM. The approach allows a triple complex exponential model to be fitted in <1s on CPUs and <20ms on a GPU for a 256x256 image, vs. 5+ min for the original solver. The DL solver can be implemented with either fixed echo times or adaptive to a range of echo times and number of echoes. Trained with various additive noise levels, the DL-solver performs more robustly compared to the conventional optimization-based solver when the signal has a very low SNR.