A Deep Learning Framework for Real-Time Cancer Targeting in Radiation Therapy

Abstract

Growing evidence highlights the detrimental effects of underdosing tumours and overdosing organs at risk during high dose radiation therapy, emphasising the need for precision in treatment delivery. Tumour motion, due to normal physiological processes, is one factor that compromises accuracy. To address this, real-time motion management technologies have been developed to continuously monitor the tumour position. These technologies, though effective, are either prohibitively expensive or require the implantation of fiducial markers. These barriers were highlighted in a recent international survey which revealed that 71% of responding centres wish to implement real-time motion management for another treatment site but are limited by resources and capacity. This thesis presents the first large-scale proof-of-principle for x-ray-based markerless segmentation in globally available radiation therapy systems. It provides an important step towards making real-time motion management treatments accessible for all patients, eliminating the need for expensive, dedicated equipment or fiducial marker implantation. The proposed markerless approach relies solely on x-ray images acquired during treatment which in principle, covers the majority of linear accelerators, thus overcoming the resource and capacity barriers to real-time motion management. The first two studies investigate a deep learning framework for markerless segmentation of the prostate and pancreas head, demonstrating high accuracy across a large patient cohort. The third study explores using residual contrast agents from liver cancer chemotherapy as a surrogate for real-time motion monitoring, showing successful tracking in patients with liver cancer. A software application was also developed for ground truth data labelling. The thesis concludes with key findings and future directions for deep learning-based markerless tumour tracking.