Beyond Respiratory Models: A Physics-enhanced Synthetic Data Generation Method for 2D-3D Deformable Registration

François Lecomte, Pablo Alvarez, Stéphane Cotin, Jean-Louis Dillenseger; Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, 2024, pp. 2413-2421

Deformable image registration is crucial in aligning medical images for various clinical applications yet enhancing its efficiency and robustness remains a challenge. Deep Learning methods have shown very promising results for addressing the registration process however acquiring sufficient and diverse data for training remains a hurdle. Synthetic data generation strategies have emerged as a solution yet existing methods often lack versatility and often do not represent well certain types of deformation. This work focuses on X-ray to CT 2D-3D deformable image registration for abdominal interventions where tissue deformation can arise from multiple sources. Due to the scarcity of real-world data for this task synthetic data generation is unavoidable. Unlike previous approaches relying on statistical models extracted from 4DCT images our method leverages a single 3D CT image and physically corrected randomized Displacement Vector Fields (DVF) to enable 2D-3D registration for a variety of clinical scenarios. We believe that our approach represents a significant step towards overcoming data scarcity challenges and enhancing the effectiveness of DL-based DIR in a variety of clinical settings.