Preview

PDF, English - main document Download (72MB) | Lizenz: Rights reserved - Free Access

Abstract

Image-guided adaptive radiation therapy (IGART) significantly enhances modern radiotherapy by adapting to patient movements and anatomical changes in real-time. However, the high computational demands of the technology prevent its deployment on standard computing systems. Furthermore, the requirement for recalculating doses in real-time, which allows for immediate treatment adjustments in response to these variations, further complicates its implementation. Given these substantial computational challenges, with optimization processes constituting the majority of execution time, this study focuses on the image registration (IR) task within the IGART application. IR, which relies on iterative optimization, is a time-intensive process that is essential for accurate image alignment and analysis. Due to the deformable nature of anatomical changes during treatment, deformable image registration (DIR) is crucial for addressing variations in the shape and size of internal organs between the initial and adaptive planning images. This study specifically targets the registration of three-dimensional computed tomography (CT) lung scans, which are among the most deformable and complex medical datasets due to their elasticity and the influence of respiratory motion. However, our method is adaptable to images from various regions of interest and can extend beyond the medical imaging domain. Accelerating DIR is essential for efficient and timely medical procedures given its high computational demands. While multi-GPU implementations offer significant potential for DIR acceleration, challenges such as high communication overhead and increased complexity have made such configurations impractical for real-time applications. To overcome these challenges, this study proposes a novel solution named CLAIRE-ROP (constrained large deformation diffeomorphic image registration - rapid overlapped partitioning-based). This framework leverages multi-GPU systems to accelerate DIR without increasing programming complexity, aiming to enable rapid and precise registration, thus facilitating real-time adaptation in IGART. Our approach employs a partitioning scheme that divides lung images into multiple partitions, thereby enabling the individual registration of each partition without compromising accuracy. Our method has successfully registered images from the largest publicly available lung dataset (512×512×136) in under 0.5 seconds, achieving a Dice score of 0.991. This work represents a significant advancement in lung image registration techniques, facilitating more efficient DIR in clinical applications. Currently, our achieved registration time on the DIR-Lab dataset is the fastest among all published DIR methods for 4DCT, encompassing both deep learning and optimization-based approaches. Notably, our results demonstrate that our approach maintains competitive registration accuracy.