Towards Real-Time Markerless Prostate IGRT During VMAT Treatment
A Mylonas1,2*, M Mueller1,2, P Keall1, J Booth3, D Nguyen1,2,3, (1) ACRF Image X Institute, Faculty of Medicine and Health, The University of Sydney, Sydney, NSW, AU, (2) School of Biomedical Engineering, University of Technology Sydney, Sydney, NSW, AU, (3) Northern Sydney Cancer Centre, Royal North Shore Hospital, Sydney, NSW, AU
Presentations
PO-GePV-M-157 (Sunday, 7/25/2021) [Eastern Time (GMT-4)]
Purpose: Image-guided radiation therapy (IGRT) and real-time IGRT for prostate cancer typically relies on markers to determine the target location in kilovoltage (kV) images. However, the placement of markers is an additional procedure that introduces costs and side effects. We present a novel deep learning method for markerless prostate IGRT.
Methods: A real-time tracking system using a conditional generative adversarial network (cGAN) was developed. A base model was trained on an imaging database of prostate cancer patients in the TROG 15.01 SPARK trial and was fine-tuned individually on unseen patients. The base model was trained using 17,000 kV images from 34 fractions of 17 patients. A patient-specific model was then trained for five different patients using 3,600 digitally reconstructed radiographs (DRR) per patient generated over 360 degrees. The DRRs were produced using the planning CTs and clinical target volume contours. The system was evaluated on 1,000 kV images/patient from two fractions. The implanted markers were used as the ground truth and were masked out for training and testing. The centroid geometric error was defined as the cGAN segmentation verses the ground truth and the Dice similarity coefficient (DSC) was calculated to gauge the similarity of the segmentations.
Results: On patient images, the mean error was 0.2±1.9mm and -0.2±1.5mm in the lateral/anterior-posterior and superior-inferior directions, respectively. The [2.5th, 97.5th] percentiles of the error were [-3.4, 3.9]mm in the lateral/anterior-posterior direction, and [-2.6, 2.2]mm in the superior-inferior direction. The mean DSC was 0.92±0.03.
Conclusion: The first algorithm for markerless prostate segmentation during an arc treatment was developed and evaluated on patient kV images. The results demonstrate that real-time markerless prostate IGRT is achievable to a high degree of accuracy using deep learning. This solution could enable more patients access to real-time IGRT without the costs, time and side effects for marker placement.
Funding Support, Disclosures, and Conflict of Interest: D T Nguyen is funded by an Early Career Research Fellowship from the Australian National Health and Medical Research Council (NHMRC) and the Cancer Institute of New South Wales. P J Keall is funded by a NHMRC Senior Principle Research Fellowship.
ePosters
Keywords
Image-guided Therapy, Segmentation
Taxonomy
IM/TH- image Segmentation: X-ray
Contact Email