Premium
WE‐A‐BRA‐01: The Hybrid Linac‐MR System for Real‐Time Tumour Tracking and Radiation Treatment
Author(s) -
Fallone BG
Publication year - 2012
Publication title -
medical physics
Language(s) - English
Resource type - Journals
SCImago Journal Rank - 1.473
H-Index - 180
eISSN - 2473-4209
pISSN - 0094-2405
DOI - 10.1118/1.4736053
Subject(s) - contouring , dosimetry , mean squared error , linear particle accelerator , nuclear medicine , computer science , image guided radiation therapy , tracking (education) , artificial intelligence , computer vision , medical imaging , mathematics , medicine , physics , optics , computer graphics (images) , statistics , psychology , pedagogy , beam (structure)
The development of a successful hybrid linac‐MR system is described that allows for both perpendicular and parallel radiation‐configurations to minimize perturbations in radiation dosimetry and to improve dosimetry due to magnetic‐field effects. Successful automatic contouring of tumours of MR images obtained during treatment at four frames a sec (as recommended for lung tracking) for various low magnetic‐field strengths is described, in addition to, our successful predictive tumour‐position algorithm based on patient‐specific (with and without initial weight (IW) for each patient and treatment fraction) feed‐forward 4 layered artificial neural networks (ANN) to compensate for delays in MLC leaf‐motions. The respective tracking and predictive performances of our algorithms are tested with a database of a large number of images for 29 patients obtained independently at very high frames, as well as, with in‐house motion MR phantoms that emulate the motion of any patient on the database. The automatic algorithm successfully contoured moving tumour from dynamic MR images obtained at 4 fps with Dice coefficients of >0.96 and >0.93, and tracked the tumour position with root‐mean‐squared‐errors (RMSE) of < 0.55 mm and <0.92 mm, for 0.5 and 0.2 T images, respectively. Mean RMSE values of 0.5 – 0.9 mm are achieved by our ANN predictor for MLC systems delays ranging from 120 – 520 ms for all the patients in the database. The advantage of using our patient‐specific ANN is shown by a 30 – 60 % decrease in mean RMSE values in motion prediction as compared to results achieved with a single ANN structure and randomly chosen IW. Our results successfully demonstrate the feasibility of using auto‐contouring in low field images and of using the intrafractional tumour‐motion auto‐tracking with our laboratory linac‐MR system Learning Objectives: 1. Understand the solutions to the mutual interferences associated with a linac and an MRI 2. Understand the development of MR autocontouring algorithms for low‐field MR images obtained at 4 fps 3. Understand the development of algorithm that predict the tumour positions from MR images obtained at 4 fps and thus compensate for the delay in MLC leaf motions to the tumour.