Acceleration and validation of optical flow based deformable registration for image-guided radiotherapy

Area of Science:

• Medical Imaging
• Image Registration
• Computational Anatomy

Background:

• Accurate image registration is crucial for radiotherapy planning and monitoring.
• Deformable registration methods are computationally intensive, limiting their clinical application.
• Graphics Processing Units (GPUs) offer potential for accelerating complex image processing tasks.

Purpose of the Study:

• To evaluate two GPU-accelerated optical flow-based deformable registration methods.
• To assess the accuracy and speed of these methods on 4DCT thorax and CBCT head and neck datasets.
• To determine the feasibility of online registration using GPU acceleration.

Main Methods:

• Implementation of Horn & Schunck and Cornelius & Kanade optical flow algorithms on a GPU.
• Testing on a 4DCT thorax dataset with 41 landmarks per phase.
• Testing on CBCT and planning CT datasets for head and neck cancer patients with 6 landmarks per dataset.
• Comparison of rigid and deformable registration accuracy and computation time.

Main Results:

• Deformable registration reduced average landmark error in 4DCT from 3.5 mm to 1.1 mm.
• In CBCT to CBCT registration, deformable registration achieved 1.6 mm error vs. 1.8 mm for rigid.
• In CBCT to CT registration, deformable registration achieved 1.8 mm error vs. 2.2 mm for rigid.
• GPU acceleration improved Horn & Schunck method speed by 48x, enabling 37s 4DCT and 64s head and neck registration.

Conclusions:

• GPU-accelerated deformable registration provides acceptable accuracy for clinical use, comparable to image slice thickness limitations.
• The speed improvements allow for potential online CBCT registration.
• Further validation with soft tissue landmarks is planned to fully assess deformable registration performance.