Real-time radiotherapy error detection using transit beam image processing

Research Doctorate - Doctor of Philosophy (PhD)In modern radiotherapy treatments such as Modulated Radiation Therapy (IMRT) and Volumetric Modulated Arc Therapy (VMAT) are complex involving very sophisticated dynamic delivery techniques. Unfortunately with complex technology and techniques there is always some risk and errors can occur. The high dose gradients typically associated with these treatment techniques require rigorous verification of the radiation delivery, which is generally performed using pre-treatment verification. Therefore, a major gap in current radiation therapy safety for patient treatments is to absence of the actual treatment delivery verification to ensure that mistreatment do not occur. In this thesis, we have designed and developed a first real-time electronic portal imaging device (EPID)-based delivery verification system (Watchdog), which enables detection of gross treatment delivery errors before delivery of substantial radiation to the patient. This application addresses this important unmet need in radiation therapy safety. The system utilizes a comprehensive physics-based model to generate a series of predicted transit EPID image frames as a reference dataset and compares these to measured EPID frames acquired during treatment. The EPID image frames are acquired during treatment delivery at ~10 frames per second via a frame-grabber card and separate computer. When frames are received they are first synchronised to the predicted image data set. Currently for IMRT deliveries the multileaf collimator (MLC) aperture is segmented from the image and is compared to the MLC aperture in the predicted image set and the closest matching control point is found. For VMAT deliveries the gantry angle is used for synchronisation. The first 20 frames or ~2 s of delivery are ignored in the analysis as the beam dose-rate is rapidly increasing in this period making dosimetric assessment uncertain. The cumulative measured image (integrating all frames from the start of treatment) is then quantitatively compared to the predicted image (similarly integrated) in real-time using the chi comparison method with the criteria of 4%,4 mm. Comparisons are made in real-time with results reported at approximately 0.5 s intervals. For a pilot study of the system deployment in clinic, approximately 200 patient image data was corrected. This allows us to determine the appropriate action levels (threshold) using statistical process control (SPC) technique. In our investigation of system performance, errors are flagged when the results fall below the defined action level for approximately 3 seconds duration. In addition, we have developed new tools to verify the position and trajectory of dynamic MLCs and jaws from a linac machine using EPID. These tools were used to assist the error classification for estimating the source of detected errors by the system. Finally, we proposes the method for evaluating treatment quality to assist in continuous quality improvement using in-vivo EPID based dosimetry. The SPC was applied and a moving range control chart is constructed to monitor the individual patient treatment verification results. A process capability index is used to evaluate the treatment quality for three classes; linac machine, treatment techniques, and body-sites. The outcomes of this research include 1) development a first real-time EPID-based patient treatment verification, 2) determination of action levels for real-time error detection, 3) development of new tools for real-time collimation position and trajectory verification using EPID, 4) evaluating the long-term treatment quality for continue quality improvement