The Importance of Realistic Training Deformations for Respiratory Ct Registration

Deep learning enables fast deformable medical image registration but requires large training datasets, which are currently scarce. To overcome this, synthetic deformations can be generated to augment the training data. We proposed a method that incorporates prior knowledge of the physiological motion to generate more realistic deformations. Specifically, our method is developed on thoracic computed tomography scans to incorporate respiratory motion. We evaluated the effect of various synthetic deformation methods on deep learning-based registration performance, finding a better performance when trained on more realistic deformations, compared to when trained on random deformations. Generally, including realistic deformations, either real or synthetic, was found to be essential in achieving a good registration performance.