3D fetal head pose estimation from MRI navigators with equivariant networks

Motivation: Fetal brain abnormalities are identified by acquiring stacks of high-resolution 2D anatomical images along standard directions. However, unpredictable fetal motion and maternal abnormalities lead to double-oblique slices and coverage gaps, which pose challenges for radiological assessment. Goal(s): We aim to estimate fetal head pose from low-resolution 3D EPI navigators interleaved with 2D anatomical images to enable dynamic prescription of the imaging plane. Approach: We train a 3D rigid-equivariant network to rapidly estimate rigid head pose (translation and rotation). Results: We report accurate pose estimates on standard-alone EPI volumes, generalization to interleaved navigators, and robustness to large motion. Impact: Our method demonstrates the promise of fetal head pose estimation and opensnew possibilities for dynamic prescription of the imaging plane that tracks thefetal head in real time.