Purpose: To develop an automatic segmentation algorithm to contour heart substructures on a planning CT for estimating doses in support of research on cardiovascular morbidity subsequent to breast radiotherapy.
Methods: We collected DICOM-RT data for 100 breast cancer patients enrolled in the Radiotherapy Comparative Effectiveness clinical trial who were treated using photons. For each CT, a cardiology team provided a manual segmentation of the whole heart (WH), left/right atria, left/right ventricles and left anterior descending artery (LAD). A cardiac atlas library was developed for the automatic segmentation using 30 patients and the remaining 70 patients (40 left-side, 30 right-side) were used for performance testing. The input to the algorithm was the manual contouring of the WH which is drawn at time of treatment planning. The automatic segmentation consists of a selection of a most similar heart from the atlas library followed by a structure-guided B-spline transformation. The automatically contoured results were then compared with manual delineations to evaluate geometric and dosimetric similarity.
Results: The median dice similarity coefficient (DSC) for the four chambers ranged from 66% to 84% and the average surface distance (ASD) was ~4 mm. Performance for the LAD was worse with a mean DSC of 6% and median ASD of ~6 mm which can be explained by the difficulty segmenting such a small structure. The doses calculated using the manual and automatic delineations showed excellent correlation within expected variability of manual segmentation. For left-sided (right-sided) treatments, the mean dose difference for the chambers and LAD was <0.7 Gy (<0.4 Gy) and 1.8 Gy (0.4 Gy), respectively.
Conclusion: Cardiac structure doses based on our automatic segmentation agree with that of manual segmentation. When applied to large patient datasets our automatic segmentation method should facilitate the development of better prescriptive criteria for mitigating cardiovascular morbidity following breast radiotherapy.