An anatomically-informed correspondence initialisation method to improve learning-based registration for radiotherapy

TL;DR

This work tackles interpatient CT non-rigid registration in radiotherapy by introducing CorrTPS, an anatomically informed initialisation that uses predicted dense surface correspondences to build a 3D thin-plate spline (TPS) deformation prior to a second registration step. The approach is evaluated on 31 head-and-neck CT scans against two baselines, Voxelmorph (DL-based) and NiftyReg (iterative), showing that CorrTPS reduces mean distance-to-agreement by up to $1.8$ mm for structures included in the TPS and $0.6$ mm for structures not included, while preserving a substantial speed advantage (about $5$ s vs $72$ s). CorrTPS brings the DL-based registration closer to traditional iterative performance for included structures, though gains for non-included structures are more variable and depend on control-point distribution. The method demonstrates potential for faster, more accurate radiotherapy registration and could be extended with auto-segmentation and broader structure inclusion to further improve robustness and automation in clinical workflows.

Abstract

We propose an anatomically-informed initialisation method for interpatient CT non-rigid registration (NRR), using a learning-based model to estimate correspondences between organ structures. A thin plate spline (TPS) deformation, set up using the correspondence predictions, is used to initialise the scans before a second NRR step. We compare two established NRR methods for the second step: a B-spline iterative optimisation-based algorithm and a deep learning-based approach. Registration performance is evaluated with and without the initialisation by assessing the similarity of propagated structures. Our proposed initialisation improved the registration performance of the learning-based method to more closely match the traditional iterative algorithm, with the mean distance-to-agreement reduced by 1.8mm for structures included in the TPS and 0.6mm for structures not included, while maintaining a substantial speed advantage (5 vs. 72 seconds).

Figures (2)

• Figure 1: Correspondence of each structure between the fixed and moving datasets, represented by matching colours. Different colour maps are used for each structure and identified correspondences are only shown in the cranio-caudal direction for clarity. The resulting deformation vectors (OARs in orange, bony anatomy and envelope in blue) are the input for the TPS to set up the non-rigid initialisation.
• Figure 2: Mean distance-to-agreement results for each registration pipeline. The structures in a) were included in CorrTPS, i.e., their surface correspondence was used to set up the TPS, whereas the structures in b) were not. Significant deviations between methods with and without CorrTPS, as identified with a Wilcoxon signed-rank test, are indicated. Black asterisks show an improvement using CorrTPS, whereas the red show performance reductions.