Leveraging deep learning to enable real-time beam-view image-guided radiotherapy

Abstract

Real-time image-guided radiotherapy (IGRT) is essential for minimising the clinical impact of patient motion during radiotherapy. Existing IGRT systems often rely on additional imaging dose or specialised, expensive technology to enable continuous intrafraction monitoring. Beam-view imaging offers a promising alternative, enabling real-time tumour monitoring directly in the treatment beam without additional imaging dose and with broad compatibility on standard linear accelerators (linacs). However, its clinical use is limited by treatment-beam occlusions and poor megavoltage (MV) image quality.

This thesis aims to develop, implement, and evaluate a deep learning-enabled beam-view IGRT framework that is safe, accurate, and compatible with standard linacs. Novel deep learning approaches were leveraged to overcome the aforementioned challenges facing beam-view imaging.

Three research objectives are addressed: (i) develop deep learning-enabled beam-view target tracking approaches for abdominopelvic and thoracic treatment sites; (ii) experimentally evaluate real-time beam-view marker tracking using an anthropomorphic pelvic phantom, and develop associated workflows and patient-specific quality assurance procedures to facilitate safe clinical deployment for prostate cancer radiotherapy; and (iii) clinically implement and evaluate real-time beam-view marker tracking during prostate stereotactic body radiotherapy (SBRT).

Real-time beam-view IGRT has been developed and investigated, with clinical feasibility demonstrated for prostate SBRT. Key implementation barriers are addressed, establishing a foundation for broader clinical adoption of real-time beam-view IGRT on standard linacs.