Experimental investigation of real‐time 3D beam's eye view image‐guided radiotherapy for prostate SBRT

Abstract

Background
Real‐time image‐guided radiotherapy (IGRT) is critical for accurate dose delivery during stereotactic body radiotherapy (SBRT). Beam's eye view (BEV) imaging offers the unique advantage of reporting motion in the most dosimetrically relevant frame of reference without additional imaging dose. However, its clinical use is limited by poor contrast‐to‐noise ratio and marker occlusion by treatment beam apertures. Deep learning enables fast identification of indistinct marker features even in low contrast images, facilitating real‐time BEV‐IGRT. To support integration with standard‐equipped linear accelerators, accurate 3D localization is essential—necessitating the development of a 3D BEV‐IGRT system.


Purpose
This study aimed to develop and experimentally evaluate a novel real‐time 3D BEV‐IGRT system for potential clinical implementation during prostate SBRT.


Methods
A real‐time 3D BEV‐IGRT system was developed by integrating a deep learning‐based 2D MV marker segmentation method with a 3D IGRT framework. Marker positions were segmented on MV images using a convolutional neural network (CNN) and used to predict 3D motion via a Gaussian maximum likelihood estimation method. A failure mode and effects analysis (FMEA) was performed by a multidisciplinary team. Mitigation strategies were implemented for high‐risk failure modes, and risk priority numbers (RPN) were recalculated.

Experimental system evaluation was guided by failure modes identified through the FMEA. An anthropomorphic pelvic phantom with three implanted gold markers was mounted on a 3D motion‐programmable platform. System performance was assessed under static and dynamic conditions, using treatment plans of increasing complexity, ranging from open fields to patient‐representative volumetric modulated arc therapy plans. Dynamic performance was evaluated using four patient‐derived prostate motion traces. Localization accuracy (mean error ± 1 SD) was assessed by comparing system‐reported positions to ground truth derived from known motion trajectories or static displacements. 5
th
/95
th
error percentiles were calculated. System latency was measured as the time delay between motion initiation and system‐reported displacement. Clinically acceptable accuracy was defined as within ± 2 mm in the superior‐inferior (SI), anterior‐posterior (AP) and left‐right (LR) directions, and latency ≤ 500 ms.



Results
Forty‐six failure modes were identified through the FMEA. High‐risk failure causes included algorithmic limitations, algorithmic errors, human error, and marker occlusion. Incorporation of mitigation strategies—including eligibility screening, staff training, and workflow formalization—resulted in an average RPN reduction of 43% across the top ten high‐risk failure modes. A risk‐informed quality assurance program was designed to support clinical implementation.

Overall 3D BEV‐IGRT system accuracy was 0.1 ± 0.7 mm (SI), ‐0.1 ± 0.8 mm (AP), and 0.1 ± 0.7 mm (LR). Accuracy remained within ± 2 mm in all directions across all individual tests. Overall 5
th
/95
th
percentile errors were [‐1.0, 1.3] mm (SI), [‐1.2, 0.9] mm (AP), and [‐0.9, 1.0] mm (LR). System latency was 300 ± 100 ms.



Conclusions
The 3D BEV‐IGRT system was experimentally validated, demonstrating clinically acceptable localization accuracy and latency, supporting its feasibility for clinical implementation. Integrated risk mitigation strategies effectively reduced workflow risk and promoted understanding of system vulnerabilities. Deployment is planned for a prostate SBRT clinical trial.