Purpose: Deformable image registration (DIR) aims to find a geometric transformation between corresponding image data and brings them into a common coordinate frame. DIR accuracy can be improved using a hybrid image similarity metric which matches both image intensities and the location of known structures. This work evaluates the application of the point-to-distance map (PD) hybrid similarity metric for DIR when the segmented structures have been created using a deep learning segmentation method.
Methods: The approach was evaluated using a publicly available head and neck dataset. Brainstem, mandible, larynx, and parotid glands were segmented automatically using a 3D UNet model. Pairs of images with segmented structures were registered using a hybrid similarity metric, which penalizes the distance between structure boundaries. Cubic B-splines were used for the registration transform and sum of squared difference for the intensity similarity.
Results: Accuracy of the deformable registration was quantified using the Dice Similarity Coefficient (DSC) and 95% Hausdorff Distance (95% HD), with and without the PD metric. Results demonstrate much improved overlap between the fixed and warped contours once the PD metric is applied. For example, with a grid spacing of 40 mm, mean DSC for the larynx increases from 0.72 to 0.8 with the PD metric. The computational overhead associated with adding the PD metric is minimal.
Conclusion: A PD hybrid similarity metric was found to improve DIR accuracy when the segmented structures were created by deep learning. This means that a significant improvement in terms of registration accuracy can be achieved without resorting to manual contouring.
Funding Support, Disclosures, and Conflict of Interest: This material is based upon work supported by the National Science Foundation under Grant Nos. 1553436, 1642345 and 1642380 and the National Institutes of Health under NCI R01CA229178.