Improved Upper-Airway Collapse Imaging in Obstructive Sleep Apnea Using Self-Supervised Spiral Variational Manifold Learning

To propose a self-supervised variational manifold learning framework for dynamic upper-airway MRI reconstruction in obstructive sleep apnea from prospectively under-sampled spiral data during natural sleep, without requiring fully sampled training data.

A CNN generator and low-dimensional latent vectors were jointly optimized directly from prospectively under-sampled spiral k-space measurements, learning a subject-specific temporal manifold without an explicit motion model. A validation loss guided adaptive learning rate provided principled early stopping without ground truth. Eleven concurrent axial slices from velo-pharynx to hypo-pharynx were reconstructed at 183 ms temporal resolution and $11\times$ acceleration across 8 OSA patients during natural sleep and 4 healthy volunteers performing the Müller maneuver. Reconstruction quality was compared against parallel imaging, compressed sensing, view sharing, and analysis-manifold reconstruction. Three expert raters independently assessed image quality using a four-point ordinal rubric.

The proposed method resolved multi-site airway collapse across 11 contiguous slices, including simultaneous visualization of complete concentric retro-palatal collapse, lateral collapse at the velo-pharyngeal to oro-pharyngeal transition, and partial tongue-base and oropharyngeal narrowing. Expert scores were superior across aliasing, spatial blurring, and temporal blurring categories ( $p<0.001$ , pairwise Mann-Whitney U tests). Reconstructed airway dynamics were corroborated by concurrent respiratory effort and SpO ${}_2$ signals, with up to 13 obstructive events identified within a single imaging run. Reconstruction artifacts occurred in 8 of 132 slices (6%) and did not compromise airway visualization.

Self-supervised variational manifold learning demonstrates the feasibility of robust dynamic upper-airway MRI during natural sleep, with potential to inform future OSA phenotyping and surgical-planning studies.