Image registration for patient positioning in external beam radiotherapy

This project aims at development of image registration techniques using electronic portal imaging device (EPID) imaging for automated on-line patient positioning in radiation therapy. The project consists of two parts. In part one, we developed an iterative subregion match (ISM) algorithm and a variable step-size grid search (VSGS) algorithm. The ISM was applied to intramodal registrations involving EPID images. User intervention is limited to selection of rectangular subregions in the reference image. Phantom experiments show that the ISM algorithm has the accuracy of 0.2 mm and 0.2°. The registration takes typically 8 sec on a 2 GHz PC. The VSGS was applied to intermodal registrations involving digitally reconstructed radiographs (DRRs) and EPID images. It was combined with a multi-resolution strategy to increase computation speed without adversely affecting robustness. Measurements indicated that VSGS has an accuracy of 0.5 mm and 0.5°, and requires about 8 seconds on a 2 GHz PC. Validation was carried out using phantom and clinical imaging data from a video camera-based EPID and amorphous silicon flat-panel system with 6 MV megavoltage x-rays.; In part two, the concept of registration quality evaluator (RQE) was introduced. The RQE consists of a decision criterion to determine the acceptability of registration results. It is generated using initial training data generated from a single pair of reference and target images to which known shifts are applied and similarity measures are computed. The RQE was applied to patient positioning in cranial radiotherapy using an anthropomorphic phantom. Two types of RQEs were constructed---one for the EPID/EPID registration and the other for the DRR/EPID registration. Using receiver operating characteristic (ROC) analysis, mutual information was proved to be an appropriate similarity measure for both types of RQEs. We also established a methodology in the estimation of the performance of a RQE by calculating the sensitivity and the specificity and their confidence intervals. These estimations were confirmed using phantom imaging with radio-opaque fiducial markers. RQE combined with a fast and accurate registration algorithm offers the potential of a robust automated patient positioning system with reduced user intervention.