Patient Positioning Using a Fast Robust 2D-3D Registration Strategy With a Registration Quality Evaluator

Abstract

Implement 2D-3D registration on linacs with electronic portal imaging devices (EPIDs) to yield accuracy similar to 3D-3D registration using linac-based conebeam CT (CBCT) for registrations based on bony anatomy. Our specific methodology overcomes 1) long computation times due to multiple digitally reconstructed radiograph (DRR) generation requirement, and 2) potential positioning errors due to local minima trapping. It offers high precision 3D patient positioning without requiring onboard CBCT on each treatment linac. 2D-3D registration uses multiple 2D projection images to achieve 3D positioning by computing the best alignment based on bony anatomy of the 2D portal images with DRRs at incrementally different poses. Challenge 1) can be overcome with specialized code that assigns DRR generation to the graphics processing unit (GPU) rather than CPU but this requires dedicated software/hardware for each EPID. Challenge 2) can be overcome by requiring human oversight to check the validity of the registration but this limits the automation level needed for large patient throughput. We present a 2D-3D registration methodology that can be implemented either on 1) a single workstation equipped with a NVIDIA GeForce 8800 GPU and a DRR server software, which is networked to multiple standard EPID viewing stations via TCP/IP, hence minimizing hardware requirements, or 2) each viewing workstation containing the GPU and DRR software. A registration quality evaluator (RQE) is used to avoid local optima associated with 2D-3D registrations. RQE is an algorithm based pattern classifier that identifies local optima trapping of an optimization, which can lead to incorrect patient positioning. We implemented a 2D-3D registration using two strategies: 1) GPU on a server and a remote client PC connected through intranet, which achieved an accuracy of 1 mm and 1° for phantom and clinical imaging with computation times <30 sec; 2) both the server and the client in a single workstation, in which computation time <10 sec. The use of RQE eliminated any local optima trapping. RQE training yielded a sensitivity and a specificity of 0.9804 (0.8955–0.9995) and 0.9388 (0.8313–0.9872), respectively, at 95% confidence interval. Using test dataset from phantom imaging, the sensitivity and the specificity of RQE were 0.939 and 0.937, respectively. Currently CBCT generally has insufficient image quality for accurate deformable registration based on soft tissue, and 3D-3D registrations are largely driven by bony anatomy. Our 2D-3D registration provides accuracy comparable to 3D-3D registration using CBCT for registrations based on bony anatomy. The proposed proof-of-concept system offers a simple and inexpensive solution for those radiotherapy patients requiring precise 3D patient positioning based on bony anatomy without invasive fiducial markers that can be implemented on any linac with EPID imaging.