Robust Multi-modal Registration of Cerebral Vasculature

Abstract

The assessment of neurological disorders benefits from a precise visualization of the vasculature and the surrounding brain tissue, leading to a more comprehensive understanding of the anatomical structures and pathological changes. While multi-modal registration of neuroimages has been extensively researched, existing methods face limitations regarding the alignment of angiographic and structural images and rely heavily on a good initial alignment. To address suboptimal initial alignments, manual or fiducial-based global alignment methods are commonly employed in clinical practice.

We propose a novel geometry-based approach to automate global alignment, leveraging deep image segmentation models to extract reference vasculature from structural MR images and subsequently directly align the vascular structures. We conducted a comprehensive evaluation of our method on both a clinical collection of DSA and MR images, and a large collection of publicly available data comprising angiographic time-of-flight MRA and five structural MR sequences. Our method was able to accurately align all 11,748 evaluated image pairs. Furthermore, we compared our method to three state-of-the-art image-based methods widely used for intracranial registration in the community. In our evaluation, we highlight the limitations of these methods that hinder their effectiveness in the context of vascular registration, a challenge our method successfully overcomes. Our method offers a solution that saves valuable time for medical experts by eliminating the need for manual pre-alignment, all without adding any additional prerequisites to the acquisition process. We share our models and evaluation code at our GitHub repository.