Accuracy of an articulated head-and-neck motion model using deep learning-based instance segmentation of skeletal bones in CT scans for image registration in radiotherapy

Abstract

Knowing about anatomical deformations in patient images is crucial for adaptive image-guided radiation therapy. Biomechanical models ensure biofidelity in deformable image registration, but manual contouring limits their clinical use. We investigate the application of automatically generated contours for a biomechanical registration model in head and neck cancer treatment. For that, we automatically generate individual bone segmentations on planning CT scans examining a custom-trained nnU-Net model and the ready-trained TotalSegmentator model. Both sets of segmentations are evaluated using DICE, Hausdorff Distance and surface DICE. We investigate their impact on the build-up of the biomechanical articulated skeleton model by deviations in joint positioning and CT-CT registration accuracy using target registration error (TRE). The custom-trained model achieves 1.51 ± 0.26 mm TRE, with no significant difference in registration accuracy. While the TotalSegmentator does not provide all structures needed for the complete biomechanical model build-up. Overall, deep learning–based automatic bone segmentation can replace manual contouring in this model, matching its performance.

Keywords

• Biomechanical modelling
• biofidelity
• image registration
• head and neck cancer
• medical image segmentation