UCLA

Onboard imaging (OBI) is widely used in radiotherapy for alignment, delivery verification, and adaption. However, its acquisition is subject to configuration constraints and pulsing control, compromising view optimality and decision quality. This study combines the complementary advantages of inference power from Generative Radiance Fields (GRAF) representation and robustness from iterative reconstruction targeting instance; and alleviates their respective limitations in volume reconstruction and efficiency. Initialize with a pre-trained GRAF representation, an online instance learning module refines the GRAF using the instantaneous projection from an arbitrary angle acquired during treatment. A view from an alternative angle is generated to provide clinically preferable triggers for real-time setup review or beam adaptation decisions. To support offline delivery verification, dense synthesized projections are fed into an iterative reconstruction module to yield the full volumetric rendering. Qualitative and quantitative evaluations were performed for 2D projection query-generation and 3D volume reconstruction. Vessel bifurcation landmarks were used as surrogates for tumor targets, and their localization accuracy was analyzed to reflect motion-tracking efficacy. Detailed characterization of dependency on query angle, number of inputs, and operational conditions was reported. Our assessment indicates that paired query with instance learning enhances model's adaptation to target subjects in reducing nMSE by 17% for synthesized projections and 33% for volumetric rendering, and improving localization accuracy by 47% in bifurcation landmark analysis. This study shows the promise of using single or pair real-time projection for view toggling and volume tracking, demonstrating its utility for onboard monitoring, adaptation, and retrospective reconstruction in radiotherapy. We are working on further improving absolute accuracy and reconstruction resolution.