ABSTRACT Purpose We introduce DeepRelaxo, a fast and generalizable deep learning method for estimating brain R2* maps from multi‐echo gradient echo (ME‐GRE) acquisitions with arbitrary echo configurations, including shortened echo trains for accelerated scans. Methods DeepRelaxo is a cascaded two‐stage self‐supervised network comprising: (1) a voxel‐wise Transformer‐MLP for initial R2* estimation, and (2) a patch‐based 3D U‐Net for denoising. Both stages are trained entirely on synthetic ME‐GRE data simulated at 3 T with a varied number of echoes, echo times, and noise levels. We evaluate on simulated and in vivo brain datasets, comparing it against conventional non‐linear least squares (NLLS) and the standalone Transformer‐MLP. Experiments assess robustness under increased noise and shortened TEs. Results In simulations, DeepRelaxo consistently outperforms NLLS and Transformer‐MLP, particularly in accelerated conditions. For example, with 4× scan time reduction at low SNR (= 10), DeepRelaxo improves SSIM by 13.5% and reduces RMSE by 76% compared with baseline methods. In in vivo 3 T and 7 T data, DeepRelaxo produces consistent R2* values in deep gray matter and preserves anatomical detail, even with only two short echoes. Conclusion DeepRelaxo effectively models ME‐GRE decay, leveraging temporal and spatial context to deliver accurate, robust, and computationally efficient R2* mapping. It enables reliable reconstruction under accelerated protocols, making it suitable for time‐sensitive workflows.
Prima et al. (Sun,) studied this question.