Motivation: Clinical abdominal MRI scans often yield anisotropic data due to technical constraints, resulting in varying resolutions across spatial dimensions, which limits diagnostic accuracy and volumetric analysis. Goal(s): To create a simultaneous multi-plane self-supervised learning method for realistic generation of isotropic MRI volume from anisotropic data. Approach: Our model uses 3D generator with two conditional discriminators (coronal and axial planes) and initially applies single-plane super-resolution to create paired data. We evaluated its performance on an abdominal MRI dataset prone to motion artifacts from respiration and peristalsis. Results: Our model outperforms state-of-the-art methods both quantitatively using distribution-based performance metrices and semi-quantitatively through radiologist evaluations. Impact: Our model introduces simultaneous multi-plane self-supervised learning method to restore isotropic MRI from anisotropic data. The isotropic volumes improve volumetric analysis and 3D reconstructions, showing strong potential to enhance clinical diagnostics. It adapts to various contrast types and acquisition methods.
Benisty et al. (Tue,) studied this question.