Evaluation of two registration techniques applied to lung adaptive radiotherapy

Abstract

During the course of a radiotherapy treatment, volumes of interest might change and the doses delivered to patients may not match the planned dose any more. The delivered dose distributions can be directly estimated from the Cone Beam Computed Tomography (CBCT), images acquired at each session. For this purpose, it is first necessary to perform a registration between the computed tomography (CT) planning images and the CBCT images to extract the contours of the structures of interest in order to calculate dose volume histograms. This study evaluates registration methods adapted to these imaging modalities in the case of lung stereotactic treatments. CBCT images acquired on the linear accelerator (Varian 2100 CD) are registered on planning CT acquisitions (GE Lightspeed RT 16). Phase-shift can be seen as the iterative solution of an optimization problem in which a similarity criterion is maximized. Digital processing has also been implemented to change the resolution of the two images. The goal of this preliminary study is to establish the best suited registration algorithm suited. Two methods were compared: diffeomorphic demons algorithm and B splines warping algorithm, both from the ITK open source library. The selection of these algorithms was based on their diffeomorphism property, required for dose matrix applications. Their evaluation is based on qualitative and quantitative methods. The evaluation between the two registration algorithms was based on ten patients. CBCT were performed at each session with fixed acquisition parameters. The results show the importance of the choice of interpolation technique. A tri-linear interpolation is not relevant for tumors located in the lower lobes where the respiratory cycle is more pronounced, while the two algorithms provide satisfactory results in terms of accuracy for tumors located in the upper lobes. Optimization of CT-CBCT registration was performed in this study. This registration will then be applied to matrices of dose distribution.