Self-supervised deep learning generates quantitative T1/T2 maps from routine clinical brain MRI, invariant across scanners

A self-supervised physics-guided deep learning framework transformed clinical T1w, T2w, FLAIR MRI into quantitative T1, T2, PD maps in >600 test sessions, showing scanner invariance (inter-group CV ≤1.1%) and high voxel-wise reproducibility (T1/T2 correlation >0.82).

• Framework trained on 4121 scan sessions from four different 3T MRI systems over six years, capturing real-world clinical variability.
• Generated quantitative maps had white matter and gray matter values consistent with published literature ranges.
• Code and model weights are publicly available on GitHub.

Automated summary

RadPigeon summaries are original and for information only. They are not clinical advice.