Acceleration of T2* mapping on an MR Linac using a self-supervised convolutional neural network.

Motivation: T2* mapping could inform biologically-adaptive MR-guided radiotherapy, but requires improvement in processing time and precision for clinical implementation. Goal(s): To accelerate intravoxel field inhomogeneity correction and generation of T2* maps. Approach: We developed a physics-informed self-supervised convolutional neural network for whole volume T2* mapping of complex multi-echo data from an MR Linac. Bias in T2* estimation is accounted for by calculating the additional signal decay from 3D derivatives of the field inhomogeneity map. Results: Our model generates T2* parameter maps 30% faster than an existing time-efficient algorithm. Resulting T2* maps are less affected by noise compared to the reference. Impact: Our AI-based algorithm is a step towards integration of whole volume T2* mapping for hypoxia assessment into clinical MR-guided radiotherapy workflows. It could enable real-time mapping of dynamic changes, for example during an oxygen challenge and enable biologically adaptive radiotherapy.