Evaluation of multi-atlas-based segmentation of CT scans in prostate cancer radiotherapy

In prostate cancer radiotherapy, accurate segmentation of prostate and organs at risk in planning CT and follow-up CBCT images is an essential part of the therapy planning and optimization. Automatic segmentation is challenging because of the poor constrast in soft tissues. Although atlas-based approaches may provide a priori structural information by propagating manual expert delineations to a new individual space, the interindividual variability and registration errors can introduce bias in the results. Multi-atlas approaches can partly overcome some of these difficulties by selecting the most similar atlases among a large data base but the definition of similarity measure between the available atlases and the query individual has still to be addressed. The purpose of this paper is the evaluation of different strategies to simultaneously segment prostate, bladder and rectum from CT images, by selecting the most similar atlases from a prebuilt 24 atlas subset. Three similarity measures were considered: cross-correlation (CC), sum of squared differences (SSD) and mutual information (MI). Experiments on atlas ranking, selection strategies and fusion decision rules were carried out. Propagation of labels using the diffeomorphic demons non-rigid registration were used and the results were compared with manual delineations. Results suggest that CC and SSD are the best predictors for selecting similar atlases and that a vote decision rule is better suited to cope with large variabilities.