Succesful Implementation of Atlas Segmentation for Cardiac Substructures

Abstract

Recent findings point toward the importance of cardiac substructures in thoracic radiotherapy. Incorporating an increased number of structures without adding considerable time to the clinical workflow call for automatic yet accurate segmentation. The aim of this work was to develop an atlas segmentation workflow including 14 cardiac substructures for contrast- and non-contrast-enhanced CT scans (CT+, CT-), and evaluate its performance relative to manual segmentations. One atlas was generated from manually segmented cardiac substructures in 20 CT+ scans, and another atlas for 20 CT- scans within a clinically available atlas software. The atlas performance was investigated in a subset of 10 scans/modality. In this subset, the modality-specific atlas was applied to each CT scan using a majority vote from the best three scans from the other 9 scans i.e. an overlap with at least two-thirds of the matching scans. The agreements between a reference observer and either the atlas or another observer were compared using the Dice Similarity Coefficient (DSC; significance: p≤0.05 using the Wilcoxon signed-rank test), and volumetric trends were assessed as the normalized relative volume difference (RVD). Ultimately, the time required to apply the atlas was compared to that of the manual segmentation. For both modalities, the atlas performance was overall satisfying for 9/14 structures: the aorta, chambers, heart, inferior and superior vena cava, and pulmonary artery, but the DSCs between the atlas and the reference observer were significantly lower (p<0.0001) vs. the DSCs between the reference observer and the other observer: structure averaged DSCs=0.52-0.91 vs. 0.86-0.97 for CT+; DSCs=0.39-0.90 vs. 0.70-0.97 for CT-. The atlas systematically over- and underestimated the reference observer for CT+ and CT-, respectively (RVD=0.02-0.40, and -0.02--0.73). Regardless of structure and modality, deviations between the atlas and the reference observer were typically observed in the coronary vessels, as well as in all boundary slices. On average, the atlas workflow required 2-3 min/scan and post-processing, including also the coronary vessels, added 15 min/scan in five scans for each modality. The average time required for manual segmentation was 1.5 h/scan. Our findings suggest that atlas segmentation is suitable for the nine major cardiac substructures for both contrast and non-contrast-enhanced CT scans. The time required to apply and post-process the atlas segmentations was considerably shorter compared to that of the manual segmentation (67% time saved), and the atlas could, therefore, also be used for the coronary vessels as an initial segmentation to minimize added load to the clinical workflow.