Deformable Image Registration for Image-Guided Adaptive Radiation Therapy (IGART) Based on Massive Parallelism and Real-Time Scheduling

This thesis introduces a novel solution for accelerating deformable image registration (DIR), a critical tool in clinical applications like medical imaging and radiotherapy that ensures accurate image alignment and analysis. In collaboration with the German Cancer Research Center (DKFZ), we propose CLAIRE-ROP, a Rapid Overlapped Partitioning-based approach designed to address the challenges associated with multi-GPU implementations, such as high communication times and increased complexity, which have hindered their use in clinical settings. Our framework incorporates a unique partitioning scheme that allows for dynamic adjustment of the number and size of partitions, enabling real-time DIR that operates within milliseconds. This is the first method to employ partitioning for accurately registering large misalignments in medical images. Moreover, our approach extends beyond medical imaging, offering a versatile solution for various image registration tasks. Our results show that our method achieves the fastest registration times on the DIR-Lab dataset among all published methods for 4DCT, maintaining or even surpassing the accuracy of existing techniques, including well-known ones like deformable ANTs. Notably, our method registers images from the largest openly available lung dataset (512 × 512 × 136) in under 0.5 seconds with a Dice score of 0.991. Our code is available on GitHub.