A Physiological Variability Inspired Spatial Regularization for Joint Rigid-Deformable Abdominal MR Image Registration

Isabella Poles, Eleonora D'Arnese, Darvin Yi, Marco D. Santambrogio

Abstract: Deep Learning deformable Image Registration (DLIR) has exhibited valuable results for accurately analyzing different patients’ medical images. However, the most common DLIR approaches produce deformations assuming high-quality rigid/affine pre-registered images and global regularization regardless of image content and admissible motion nature. To address these shortcomings, we present MANGO, a spatial regularization strategy to perform jointly rigid and deformable abdominal MRI DLIR. In particular, MANGO fixes translational and rotational deformation components resulting from a suboptimal rigid pre-registration. Furthermore, physiologically generated Deformation Vector Fields guarantee DL model predictions that capture the patients’ physiological heterogeneity as much as possible. Compared to state-of-the-art iterative and DLIR methods, our solution leads up to 10% DSC improvements on inter-patient abdominal MR images and proves to preserve (i.e., no statistical difference) the same accuracy when removing the pre-registration step.