Image registration consists of aligning two (potentially 3D) images such that their contents match. In a multi-modal setting, the data are imaged through different sensors and the style and contents of the images might not correspond across modalities (e.g. CT/X-Ray/CAT scans), but still represent the same object. The goal is to recover the transformation that aligns the images semantically. Recent methods  have shown great success at efficiently and accurately aligning images of different modalities. The nature of the algorithms used (FFT, linear algebra on matrices & tensors) makes the full pipeline perfectly parallel. As such, GPUs are well suited for this task, but do require a high amount of memory to handle 3D data. We discovered some improvements to the methods listed above that improve the speed and decrease the search space of the sought-after transformations, which helps transition the method to big 3-D biomedical images. We also got access to an expansive dataset consisting of 3D objects, such as e.g. images of 16000x16000x80 (with 2 color channels). We would like to expand this research by testing multiple global optimization methods and assessing the convergence rate to the right set of parameters.
 Öfverstedt, Johan, Joakim Lindblad, and Nataša Sladoje. "Fast computation of mutual information in the frequency domain with applications to global multimodal image alignment." arXiv preprint arXiv:2106.14699 (2021).
 Öfverstedt, Johan, Joakim Lindblad, and Nataša Sladoje. "Cross-Sim-NGF: FFT-Based Global Rigid Multimodal Alignment of Image Volumes using Normalized Gradient Fields." arXiv preprint arXiv:2110.10156 (2021).