The main aim of radiotherapy is to maximise dose to the tumour volume without harming the surrounding normal tissue. However, this is complicated by internal organ motion, typically managed with treatment volume margins, which can increase dose to healthy tissue. Monitoring the tumor's position during treatment allows for margin reduction, achieving the treatment goal. In prostate radiotherapy, limited soft tissue contrast in CBCT images hampers IGRT capabilities, often necessitating the implantation of gold fiducial markers to track prostate motion.
The aim of this project was to develop a workflow for the real-time, automatic tracking of fiducial markers in prostate radiotherapy patients. This was to be used as a means of tracking both interfraction and intrafraction variations in marker position.
A deep learning object detector was trained on the YOLOv4 detection network using CBCT projection images of prostate patients implanted with gold fiducial markers. Using this detector, a GUI was then developed that can be coupled with the Elekta XVI software to perform real time fiducial marker tracking as projection images are acquired during a patient CBCT. The performance of this workflow was assessed based on the ability to determine marker positions during a CBCT of a phantom with markers inserted.
Preliminary tests of the marker tracking workflow demonstrate the ability to successfully determine marker positions in real-time. Errors introduced in the marker positions were detected using the workflow and an alert was sent to the user to halt the treatment as expected.
The result of this study demonstrates the feasibility of using a deep learning based automated tool for the assessment of prostate motion in real time without the need for additional hardware and with minimal computation power. The ability to track tumour movements during treatments could pave the way for the determination of a site-specific CTV-PTV margin.