Cardiac substructure segmentation and dosimetry using a novel hybrid MR/CT cardiac atlas.
Morris E, Ghanem A, Pantelic M, Walker E, and Glide-Hurst C. Cardiac substructure segmentation and dosimetry using a novel hybrid MR/CT cardiac atlas. Med Phys 2018; 45(6):e700-e701.
Purpose: While radiation dose to cardiac substructures has been linked to increased cardiac toxicities, these structures are not considered in treatment planning because they cannot be visualized on CT. We have developed a hybrid MR/CT cardiac segmentation atlas with the overarching goal of application to non-contrast CT sets, either for substructure sparing or cardiac radiosurgery. Methods: Eighteen left breast cancer patients underwent a T2-weighted MR and non-contrast CT at end exhalation. A radiation oncologist delineated 14 substructures (chambers, great vessels, atrial fibrillation (AFib) target, etc.) by employing MR/CT information via a cardiac-confined rigid registration. Contours for 15 patients were inputted into an intensity-based deformable registration atlas and applied to three validation patients. Atlas performance was evaluated via Dice similarity coefficient (DSC), mean distance to agreement (MDA), centroid displacement, and volume percent difference (VPD) between propagated and ground truth delineations. Dosimetric assessment of sensitive substructures was conducted via mean heart dose (MHD), left ventricular volume receiving 5 Gy (LV-V5), and left anterior descending artery mean and maximum doses (LADA (mean), LADA (max)). Results: Across three test cases, over half of all contours had VPD < 10% and centroid displacements <5 mm, with largest shifts in coronary arteries. Eight substructures had mean DSC > 0.75 and 11 had MDA < 3 mm (3-5 mm for coronary arteries), suggesting excellent atlas performance. Propagation of AFib contours were useful for localization, but the DSC of ∼0.4 suggests further manual edits are required. Atlas-generated contours were robust for dosimetric endpoints, with the LV-V5, MHD, and LADA (max) < 5% different from ground truth. In a case with unfavorable anatomy, the propagated contour yielded an LADA (mean) = 18.4 Gy and LADA (max) = 35.4 Gy, which was within 2.5% of ground truth. Conclusion: The hybrid MR/CT atlas provided reliable substructure segmentation for patients receiving non-contrast CT, suggesting potential widespread applicability for routine treatment planning. Future work is warranted to evaluate planning strategies for robust substructure sparing.