Encrypted login | home

Program Information

Inferring 3D Lung Tumor Positions Based On Projected 2D Positions: A Comparison of Different Real-Time Tumor Tracking Methods

no image available
T Montanaro

T Montanaro1*, D Nguyen1 , P Keall1 , J Booth2 , V Caillet2 , T Eade2 , C Haddad2 , C Shieh1 , (1) University of Sydney, Camperdown, NSW, (2) Royal North Shore Hospital, St Leonards, NSW,


WE-F-605-3 (Wednesday, August 2, 2017) 1:45 PM - 3:45 PM Room: 605

Purpose: Most modern radiotherapy machines are built with a 2D kV imaging system. Utilising this imaging system would allow for a ready-made option in real-time tumor tracking. This work investigates and compares the accuracy of four existing 2D-3D inference methods using true 3D lung tumour motion for the first time.

Methods: Tumor motion data from 28 fractions (7 patients) of a lung SABR trial (NCT02514512) were used in this study. 3D tumour trajectories acquired from implanted Calypso electromagnetic beacons were used as the ground truth. The ground truth trajectories were then used in-silico to generate 2D positions projected on the kV detector. These 2D traces were then passed to the 2D-3D inference methods: interdimensional correlation, Gaussian probability density function (PDF), arbitrary-shape PDF, and Kalman filter with a respiratory prediction model. The inferred 3D positions were compared with the ground truth to determine the mean (±std) and 95th percentiles of the error distributions.

Results: The arbitrary-shape PDF, Gaussian PDF and interdimensional correlation methods had 3D mean errors of 0.82mm (±0.95), 0.88mm (±1.08) and 1.10mm (±0.90) while the 95th percentiles were 2.64mm, 2.96mm and 2.90mm respectively. The Kalman filter had the lowest overall error with a 3D mean error of 0.66mm (±0.73) and 95th percentile error of 2.05mm. For the Kalman filter and the arbitrary-shape PDF, directional errors had strong positive correlation with the directional motion magnitude.

Conclusion: Mean 3D errors around or smaller than 1mm can be expected from kV-based 2D-3D inference while 95th percentile errors were found to be between 2-3mm. Larger tracking errors were found for cases with larger tumour motion magnitude. These findings give valuable insight into the margins of safety needed when performing radiotherapy using real-time tracking by 2D-3D inference. The Kalman Filter, which incorporates a respiratory-based prediction model has led to the best tracking accuracy.

Contact Email: