Segmentation of pelvic structures from planning CT based on a statistical shape model with a multiscale edge detector and geometrical likelihood measures.
Résumé
Accurate segmentation of the prostate and the organs at risk in CT images is a crucial step in prostate cancer radiotherapy planning. Because of the poor soft tissue CT contrast (prostate, bladder), an appropriate segmentation is challenging, even when this is manually performed by an expert. This paper introduces a Bayesian automatic segmentation method for prostate, rectum and bladder in planning CT. Firstly, a prior shape space for the organs is built with PCA decomposition from a population of manually delineated CT images. Then, for a given CT to be segmented, the most similar shape is selected by the associated probability which is set by a likelihood function. Finally, the local shape is deformed to adjust the particular local edges of each organ such that the most likely segmentation is produced. Experiments with real data from 30 patients treated for prostate cancer radiotherapy were performed under a leave-one out cross validation scheme. Results show that the method produces reliable segmentations (Averaged Dice = 0.91 for prostate, 0.94 for bladder, 0.89 for Rectum) and outperforms the best majority-vote multi-atlas based approach.