Author: Freeman Jin, Paul J. Keall, Alistair MacDonald, Adam Mylonas, Chandrima Sengupta 👨🔬
Affiliation: Image X Institute, Faculty of Medicine and Health, University of Sydney, Image X Institute, Faculty of Medicine and Health, The University of Sydney, Image X Institute, School of Health Sciences, University of Sydney 🌍
Purpose: During radiation therapy, tumours in the prostate may move from the planned treatment position, leading to significant dose deviations above clinical tolerances Surveys have indicated the need for precise and cost-effective real-time motion management technologies accessible to all patients. All currently available solutions are either deployed on dedicated and expensive systems or require invasive marker implantation due to low contrast x-ray images. Marker implantation adds cost, time and requires additional resources. This research takes a major step towards clinical implementation of a marker-less prostate tracking method through the development of a real-time framework (RTF) software, followed by a large-scale investigation on multi-institutional data acquired from globally available radiation therapy systems.
Methods: The RTF had two steps: (1) A patient-specific marker-less prostate segmentation method using a conditional Generative Adversarial Network which was trained using 36,000 synthetic 2D images synthesized from the planning CT and contour data (2) An algorithm to convert the 2D points to 3D motion traces. An image simulator was created to stream x-ray images to the RTF simulating a treatment. Geometric accuracy of the RTF was assessed by comparing the mean and standard deviation of the differences between RTF-segmented position and the ground truth position for 32 fractions. The RTF latency was measured as the time difference between an input image and the segmented output position.
Results: The RTF displayed x-ray images, cGAN-predicted 2D contours and 3D tumour motion during simulated treatments. Averaged over all fractions, the RTF achieved mean ± standard deviation position estimation errors of 0.6 ± 1.9 mm in x and 0.2±1.9 mm in y on the imager plane. The system latency was <=250ms measured on a laptop computer.
Conclusion: The RTF detected prostate position during simulated real-time implementation, paving the way for experimental implementation under real treatment conditions on standard radiation therapy systems.