SU‐E‐J‐92: Multi‐Atlas Based Prostate Segmentation Guided by Partial Active Shape Model for CT Image with Gold Markers

Abstract

Purpose: The gold markers implanted in prostate, which are used for IGRT, yield strong streak artifacts in planning CT image, which often causes a failure of automatic prostate segmentation. Our aim is to develop a novel method to utilize the partial active shape model to enable an automatic prostate segmentation on the CT image with strong streak metal artifacts. Methods: In the training stage, a groupwise surface correspondence was established via B‐spline deformable registration of binary structures. A joint active shape model was built for prostate, bladder and rectum together from 30 training CT images. In the segmentation stage, the gold markers on both atlas and target images were automatically detected. A mask was created to cover the region of gold markers and streak artifacts. By running the active shape model, the initial contours (from contours of atlas) was deformed to obtain the best fit with the edges of three pelvic organs in the target image, however the cost function is determined by matching only the parts of contours outside the mask based on gradient. The parts of contour inside the mask were purely estimated by the shape statistics. The segmentation was refined by a local profile based adjustment. Such segmentation was done 30 times using different training image as patient‐specific atlas, and the final Result was determined by a weighted majority voting algorithm. Results: The performance of the proposed method was evaluated for on independent dataset containing 5 CT images. The mean DSC overlap for prostate, bladder and rectum were 77.1%, 71.2% and 66.7%, respectively. The mean MAD for prostate, bladder and rectum were 3.1mm, 4.8mm and 4.6mm, respectively. Conclusion: The partial active shape model can very well estimate the contour passing the region of gold markers and streak artifacts from the rest part of contour and shape statistics.