SU‐E‐J‐210: Thoracic Atlas Segmentation Model Based On Consensus Contours

Abstract

Purpose:Generation of a thoracic segmentation atlas based on consensus contours obtained through input from individual expertsMethod and Materials:A cohort of 30 patients treated for lung disease was selected for this study. All patients had 4D‐CT studies with image segmentation based on respiration‐averaged images. For each subject, three experts delineated the eighteen agreed‐upon structures using tools available in Eclipse ver.10.0. Consensus structures were generated using CERR(4.0β2). The EM‐STAPLE method was chosen to generate the consensus contours based on published reports which state that iterative methods yield better results over analytical methods in medical imaging. Once the consensus contours were generated they were re‐imported in Eclipse for further processing and evaluation. For each structure, each expert's contour was scored against the consensus using two scoring indices: Dice similarity coefficient, DSC, and Dice‐Jaccard coefficient, DJC. Visual inspection was always performed to detect large segmentation errorsResults:The respiration‐averaged CT set proved to be adequate for all structures included in the thoracic atlas. As expected, anatomic variability between patients was larger than the variability due to the level of expertise. Across the entire cohort the smallest structure segmented was the brachial plexus with a 9.4cc (+/− 3.5) volume, while the chest wall was the largest (2933cc +/‐560). The DSC and DJC coefficients were smallest for Left Brachial Plexus 0.894 (+/−0.138) and 0.842 (+/− 0.138), respectively, while the largest were 0.992(+/−0.004) and 0.990(+/− 0.004), observed for the right lung.Conclusions:Consensus contours generated using expert input were analyzed using comparison indices and visual evaluation. With appropriate atlas selection (e.g. patient gender, arm position) a robust thoracic atlas segmentation tool can be set up and may be used in conjunction with an automatic segmentation algorithm.Project supported by Varian Medical System