Encrypted login | home

Program Information

A Systematic Analysis of Rigid Image Registration Using Patient CTs and Simulated Setup Images with a Unique Gold Standard Registration

no image available
J Kieselmann

J Kieselmann*, A Rosselet , S Scheib , S Thieme-Marti , Varian Medical Systems, Imaging Laboratory GmbH Baden

Presentations

SU-E-J-118 (Sunday, July 12, 2015) 3:00 PM - 6:00 PM Room: Exhibit Hall


Purpose:To quantify the reliability of fully automated image registration (IR) of planning CTs with 2D MV/kV projections or kV-CBCTs in terms of its failure rate, and to investigate the influence of algorithmic and external parameters on registration quality.

Methods:A gold standard registration is necessary to evaluate the failure rate. We develop a novel framework to obtain unique gold standards for MV imaging: we simulate setup images from 50 patient CTs (5 anatomic sites), using HU mapping, simple raytracing, a semi-empirical scatter model and noise addition. Comparing simulated and acquired images yields excellent agreement. We introduce translations (≤2cm) and rotations (≤2°) with respect to the planned isocenter between planning CTs and setup images according to reported setup-errors in the literature, followed by using the clinically available IR software from Varian Medical Systems (Palo Alto). Application of this simulation framework to 2D/3D kV imaging is work in progress.

Results:We investigate the dependence of the registration failure rate on the field size of MV projection images (square, centered at planned isocenter, detected shift >2mm from known shift). The results suggest a minimum field size of 10x10 cm² to achieve reasonably small failure rates (<10%). We define an optimized parameter set for 2D/3D MV IR, which outperforms the default one of the IR software with respect to the failure rate (<3%), while not significantly (<50%) increasing computation time.

Conclusion:The developed framework demonstrates the possibility of systematic tests of automated IR for arbitrarily large patient datasets. To our current knowledge this is a novel strategy. We are able to define optimized default parameter sets for registration algorithms and to estimate the influence of parameters, such as imaging dose, on the quality of IR. The improvements made for MV IR in this work look promising. Similar studies for kV based IR are work in progress.


Contact Email: