Statusvortrag im Promotionsver­fahren von Peng Liu

Abstract: 

Nonrigid registration plays a central role in both medical diagnostics and surgical navigation. Recent advances in deep learning have shown promise in automating this challenging task by learning deformation patterns directly from data. However, registration accuracy often remains limited due to the noisy, incomplete, and heterogeneous nature of intraoperative data, specifically, volumetric data from OCT or sparse 3D liver surface point clouds reconstructed from laparoscopic videos. To address these issues, this study investigates a range of nonrigid registration strategies tailored to different clinical scenarios. We first discuss general registration frameworks, including complete-to-partial and volume-to-surface registration, designed to handle sparse, noisy, and cross-modal data for both diagnostics and surgical guidance. Next, we present an online adaptive methodology that integrates expert knowledge, enabling real-time, human-in-the-loop correction to enhance clinical accuracy and reliability. Finally, we briefly discuss some work in 2D–3D registration, a rapidly evolving area with significant relevance to liver navigation.