Sensitivity analysis of self-supervised variational manifold learning based accelerated dynamic upper-airway collapse MRI

Motivation: This study aims to quantitatively validate a physics-guided stopping criterion for variational manifold learning applied to upper-airway sleep MRI. Goal(s): We retrospectively analyze reconstruction sensitivity using a reference dynamic dataset and varied regularization constraints. Approach: Given a target time resolution, healthy volunteer performing the Muller maneuver was imaged with 2D gradient-echo-based variable-density spirals, generating a reference image from a mild total-variation-regularized sparse-SENSE reconstruction. Retrospective undersampling was then applied to assess reconstruction sensitivity to various regularizers quantitatively. Results: Proposed criterion effectively stabilizes adaptive learning by guiding it with validation loss, establishing a robust quantitative basis for stopping criteria in dynamic MRI reconstruction. Impact: Enables robust sleep MRI methods for patient specific imaging diagnostics and therapeutic planning.