Localised delineation uncertainty for iterative atlas selection in automatic cardiac segmentation

The heart is an important organ at risk during thoracic radiotherapy. Many studies have demonstrated a correlation between the mean heart dose and an increase in cardiovascular disease. Different treatments result in significant dose variation within the heart and individualised dose estimation increasingly requires more attention to delineation of various cardiac structures. Automatic segmentation tools are critical for consistent and accurate delineation of organs at risk in large, retrospective studies, however the challenge of ensuring a robust method must be addressed. In a multi-atlas based segmentation framework the uncertainty in delineation can be modelled over the surface of the heart. We extend this concept with an iterative atlas selection procedure designed to remove inconsistent atlas contours, in turn improving the reliability of the segmentation. Two independent datasets comprising 15 and 20 planning computed tomography (CT) images of Danish and Australian breast cancer patients, respectively, had the whole heart and left anterior descending coronary artery (LADCA) delineated. Using a cross-validation strategy, where each dataset is used as an atlas set to segment each image in the other, we assess segmentation performance qualitatively and quantitatively, using the dice similarity coefficient (DSC), mean surface-to-surface distance (MASD) and Hausdorff distance (HD). After using the iterative atlas selection procedure, every segmentation error was removed. For the whole heart, the resulting segmentation achieved a DSC, MASD and HD of mm, and mm.