Abstract / Summary
Abstract Background : Deformable image registration is an essential technique in image-guided adaptive radiotherapy. It provides voxelwise estimates of the anatomical deformations from medical image volumes acquired at different time points. However, the accuracy of the estimated motion strongly depends on image contrast, presence of noise and artifacts, and the overall image quality. Purpose : This study introduces a novel similarity metric based on the geometry of Sobolev spaces. The proposed metric provides control over the contribution of high-order image features, allowing the metric to adapt to the characteristics of the images. It unifies several established similarity metrics and enables efficient numerical computation for integration into existing registration methods. Methods : The metric was integrated into a deformable image registration algorithm and evaluated on three experiments relevant for adaptive radiotherapy: thoracic and abdominal CT registration and pre- to post-operative brain MR image registration. Registration accuracy was assessed using alignment of anatomical landmarks. Results : The proposed similarity metric achieved rapid and accurate image alignment in all three scenarios. Increasing the contribution from higher-order image features improved the estimated deformation fields by a reduction of the landmark error. In particular, increased weighting of these features improved the alignment of liver vessels and the brain cavity following tumor resection. Conclusions : The proposed metric provides the flexibility to control the relative contribution of image features for deformable image registration. It unifies several established similarity metrics while maintaining computational efficiency and accuracy standards that are compatible with image-guided radiotherapy.