Physics-guided self-supervised learning for retrospective T1 and T2 mapping from conventional weighted brain MRI

Quantitative MRI directly measures tissue physical parameters, but has limited clinical adoption due to additional scan time and specialized sequence requirements. Supervised deep learning methods were developed to estimate relaxation maps from conventional weighted images. However, paired weighted images and quantitative maps required for training are hard to obtain. In this work, a physics-guided self-supervised learning approach was developed to estimate T1 and T2 maps from conventional weighted images. Using the Bloch equations to decode the estimated maps back to weighted images and enforcing similarity in the image space, the approach realized label-free training and provided maps comparable to references.