The adventitious irradiation of cardiac substructures during radiation therapy can lead to a variety of cardiac complications including pericarditis, myocardial fibrosis, and coronary artery disease. However, cardiac substructures are not routinely delineated due to limited image quality of treatment planning CTs and time constraints. The purpose of the current project was to develop an automatic segmentation method for performing cardiac substructure dose calculations on a large number of patients as needed for epidemiological studies or clinical trials. We used a most-similar atlas selection algorithm and 3D deformation combined with 30 detailed cardiac atlases. We cross-validated our method within the atlas library by evaluating geometric comparison metrics and by comparing cardiac doses for simulated breast radiotherapy between manual and automatic contours. More recently, machine learning techniques have been explored for improved accuracy, but the methods still rely on the quality of patient images so that the performance for radiotherapy planning CT is uncertain. As a preliminary effort with deep learning segmentation, our team applied a 2D U-Net method with a 50-patient training set to predict the whole heart outer contour. While we believe these deep learning-based results can be improved with further research, in parallel our team has continued research on the cardiac atlas-based method.
