Room: Exhibit Hall | Forum 6
Purpose: To quantitatively assess segmentation uncertainty in an automatic atlas-based framework, developed primarily for delineation of the heart and cardiac substructures in CT imaging of radiotherapy patients. Extending this concept to develop an iterative atlas selection tool to reduce segmentation errors and support the accurate and consistent segmentation of cardiac structures and coronary vessels.
Methods: Two retrospective datasets, comprising 15 and 20 independently contoured CT images of breast cancer patients from Denmark and Australia, respectively, were automatically segmented using the other dataset as an atlas set. Patient-specific models of segmentation uncertainty mapped across the surface of the heart are generated using the propagated atlas labels. For each patient image, this model of spatial variation on the heart surface is used to quantify the relative spatial discordance of each atlas, providing a metric used in a procedure for iterative atlas selection. To generate segmentations of coronary arteries a splining and tube extrusion method was developed. Performance of the segmentation algorithm was assessed using the Dice Similarity Coefficient (DSC) and mean absolute surface to surface distance (MASD) measured between the manual delineation and automatic segmentation.
Results: When aggregated across testing datasets, the highest level of segmentation uncertainty is at the caudal limit of the heart, where registration inaccuracies are more common due to the similar intensity of nearby tissues. The implementation of the iterative atlas selection algorithm as part of the segmentation pipeline resulted in substantial reduction in segmentation errors, with a corresponding improvement in segmentation accuracy.
Conclusion: Mapping segmentation uncertainty over the surface of the heart serves as a useful quantitative tool for assessing variation in atlas quality. The developed iterative atlas selection procedure successfully removes errors induced by inaccurate registration. For planned analysis of large, retrospective datasets this is vital to improve the consistency of automatic segmentations.
Not Applicable / None Entered.