Uncertainty-Aware Deep Learning Based Deformable Registration

We introduce an uncertainty-aware deep learning deformable image registration solution for magnetic resonance imaging multi-channel data. In our proposed framework, the contributions of structural and microstructural data to the displacement field are weighted with spatially varying certainty maps. We produce certainty maps by employing a conditional variational autoencoder image registration network, which enables us to generate uncertainty maps in the deformation field itself. Our approach is quantitatively evaluated on pairwise registrations of 36 neonates to a standard structural and/or microstructural template, and compared with models trained on either single modality, or both modalities together. Our results show that by incorporating uncertainty while fusing the two modalities, we achieve superior alignment in cortical gray matter and white matter regions, while also achieving a good alignment of the white matter tracts. In addition, for each of our trained models, we show examples of average uncertainty maps calculated for 10 neonates scanned at 40 weeks post-menstrual age.