Machine learning accelerated on-line adaptive replanning

点击翻译按钮获取中文摘要

Abstract. The overall goal of this proposal is to develop and test a novel machine learning (ML) accelerated On-Line Adaptive Replanning (MOLAR) solution for magnetic resonance imaging (MRI) guided radiation therapy (RT) (MRgRT). During the multi-fraction RT process, the location, shape and size of tumors and normal organs vary significantly between the fractions. These interfraction variations are among the major factors that can limit the accuracy of RT targeting. The current standard practice of image-guided RT (IGRT), developed to address the interfraction variations based on cone-beam CT (CBCT), can only correct for translational errors, and thus does not fully account for interfraction changes. To address this issue, researchers recently introduced online adaptive replanning (OLAR) that generates a new plan based on the anatomy of the day and delivers the plan for the fraction. Currently, two main obstacles affect the success of OLAR: (1) the anatomy of the day cannot be delineated accurately based on CBCT, and (2) the time required to perform OLAR is long enough to render it impractical. One way to improve the delineation accuracy is to use MRI versus CT. MRI-guided OLAR is currently being introduced into the clinics to substantially improve RT targeting. However, the bottleneck is still the impractical length of time required to segment the anatomy of the day, which can exceed 30 minutes. Furthermore, available synthetic CT (sCT) generation methods are slow or inaccurate for MRI-guided OLAR. There is no method available to quickly and objective determine when OLAR is necessary. To address these issues, we plan to develop novel techniques in the MOLAR solution. We hypothesize that the MRI-based MOLAR solution will fully account for interfraction changes, thereby substantially improving tumor targeting during RT delivery and the effectiveness of RT. Specifically, we aim to (1) develop practical ML-based solutions to quickly determine the necessity of OLAR and to rapidly generate accurate synthetic CTs; (2) develop ML-based techniques to substantially accelerate segmentation for OLAR using a progressive three-step process; and (3) verify clinical practicality and effectiveness of MOLAR by retrospectively and prospectively applying the MOLAR on MRI sets to test its speed and effectiveness in accounting for interfraction variations. We will develop this novel MOLAR solution by forging unique collaborations between clinical physicists, radiation oncologists and industry developers via an established academic-industry partnership. The successful completion of this project will enable clinicians to routinely practice “image-plan-treat”, which is the optimal solution for MRgRT. This new paradigm will fully account for interfraction variations, improve tumor targeting, reduce normal tissue toxicity, and ultimately encourage clinicians to revise the current doses and/or dose fractionations to increase therapeutic gain, enhance patient quality of life, and/or substantially save on healthcare costs. Our proposed strategy represents a drastic departure from current practice. We firmly believe that this strategy is the future of RT delivery.