UATTA-QSM: Uncertainty-Aware Test Time Adaptation for Improved Quantitative Susceptibility Mapping

This work addresses the problem of deriving improved quantitative susceptibility mapping (QSM) from magnetic resonance (MR) acquisitions. Deep learning based models that map measured MR local phase field to QSM maps, have recently demonstrated satisfactory reconstruction quality. However, they suffer from poor generalization on acquisitions that deviate from training settings. To address this, this work proposes a patient-specific, test-time adaptation method for modifying pre-trained deep learning models, conditioned on individual subject data. The proposed method of uncertainty-aware test-time adaptation (UATTA-QSM), for the first-time introduces, lipschitz constraints on local phase field measurements and QSM reconstructions to jointly reduce uncertainty and adapt model weights during inference time with functional regularization. This framework was evaluated on multiple adaptation scenarios like changing field strength, limited data training and different architectures. On unseen datasets, the proposed method of UATTA-QSM consistently demonstrates performance improvement on all metrics of interest.