Toward Improved Cardiac Sparing: Development and Validation of a Hybrid MR/CT Cardiac Segmentation Atlas

Abstract

Radiation dose to the heart and cardiac substructures has been linked to toxic effects including coronary artery disease and cardiomyopathy. As cardiac substructures are poorly visualized on treatment planning computed tomography (CT), we employed magnetic resonance imaging (MRI)’s superb soft tissue contrast to generate a hybrid MRI/CT atlas, with the ultimate goal of enabling cardiac substructure dose assessment in CT-based planning. A local rigid registration encompassing the heart was performed between the end-expiratory (EE) phase of a non-contrast 4DCT and EE axial T2-weighted MR for a cohort of 11 left breast cancer patients enrolled in an IRB study. Contour delineations using hybrid MR/CT information of 8 patients were made for 13 cardiac substructures namely: heart, right and left atria/ventricles (RA, LA, RV & LV), right (RCA), left main (LM) and left anterior descending (LAD) coronary arteries; ascending aorta (AA), superior/inferior vena cavae (SVC/IVC), and pulmonary artery/veins (PA/PV). Radiation oncologist/Radiologist ground-truth delineations were inputted into an intensity-based deformable registration atlas. Atlas performance was evaluated for 3 test cases via: Dice similarity coefficient (DSC), mean distance to agreement (MDA), centroid displacement, volume percent difference (VPD), and qualitative consensus scoring on a 5-point scale (1 = clinically unacceptable; 2, 3, and 4 = major, moderate, and minor modifications required, respectively; 5 = clinically acceptable). Eight out of 13 contours had a mean DSC >0.6 across 3 test cases, with the heart having the highest DSC (VPD <5%), and the poorest performance in the coronary arteries and pulmonary veins. One-third of all contours had mean VPD <10%. The greatest centroid displacements occurred in the superoinferior direction. Consensus scores for the heart, LV, LA, and AA required minor modifications, typically at the inferior aspects. RA, RV, PA, and SVC/IVC scored between 3-4, requiring moderate modifications. Major modifications were necessary for the PV, RCA, and LAD (scores = 1.7-2) although propagated contours were deemed useful for localization. LM yielded consistent values of 1, suggesting inadequate segmentation.Abstract 1075; TableStructureDSCMDA (mm)Heart0.91 ± 0.012.17 ± 0.14R/L Atria0.77 ± 0.032.73 ± 0.41R/L Ventricles0.73 ± 0.094.18 ± 1.20AA0.72 ± 0.012.72 ± 0.22SVC0.64 ± 0.082.55 ± 0.49IVC0.64 ± 0.104.06 ± 1.97PA0.72 ± 0.062.36 ± 0.67PV0.34 ± 0.014.55 ± 0.75Coronary Arteries (3)0.08 ± 0.076.79 ± 3.30 Open table in a new tab We are the first to demonstrate the application of a hybrid MR/CT atlas for cardiac segmentation, offering strong potential for the assessment of dose to these critical substructures. Atlas performance was clinically acceptable for most substructures. However, further improvement is needed for smaller or more complex structures (e.g., coronary arteries, pulmonary veins) that may require more sophisticated segmentation techniques, such as deep learning.