DeepRelaxo: A Generalizable Self-supervised Method for Brain R2* Mapping

Motivation: R2* is a marker to monitor the progression of neurodegenerative diseases. Existing algorithms face challenges with varying acquisition parameters, Rician noise levels, and undersampling. Extended scanning times hinder R2* implementation in clinical settings. Goal(s): To introduce a framework that achieves parameter generalizability and noise robustness for R2* mapping without the requirement of in-vivo dataset . Approach: We propose DeepRelaxo, a cascaded framework of transformer followed by a 3D-UNet. Results: DeepRelaxo is trained on a synthetic dataset with varying number of echoes, echo times, spacing, and rician noise levels. It significantly outperforms NLLS in shorter echoes, particularly in 2-echo R2* reconstructions with high noise. Impact: Self-supervised training removes the need for real-world scans and allows DeepRelaxo adaptable across diverse clinical sites. Improved shorter echo reconstructions facilitates significantly scan time reduction.