Room: Exhibit Hall | Forum 3
Purpose: To develop a robust method to automatically delineate cardiac substructures and to evaluate its performance as compared to manual contours for treatment plans from the Radiotherapy Comparative Effectiveness (RADCOMP) randomized trial.
Methods: We developed a method to delineate the whole heart and 8 cardiac substructures using affine transformation based on landmark comparison between an individual patientâ€™s CT and a pre-constructed cardiac atlas. To test the performance of the auto contouring method, we conducted a leave-one-out cross validation with the 30 atlas models using 6 anatomical landmarks. In addition, we obtained ten manually-contoured breast radiotherapy patients enrolled in the RADCOMP trial for clinical studies. We selected one atlas model after transformation out of the atlas library which best matched those of a given RadComp patient from the developed method. The cardiac contours from our method were compared with the manual contours using Dice Similarity Coefficient (DSC) for the whole heart, atria, and ventricles, and Average Surface Distance (ASD) for LAD.
Results: Leave-one-out cross validation demonstrated an average DSC of 86, 65, 63, 78, and 70% for for the whole heart, LA, RA, LV, and RV, respectively. The comparison of cardiac substructures between auto and manual contours showed strong agreement with the average DSC of 85, 60, 74, 74, and 63% for the whole heart, LA, RA, LV, and RV, respectively, and with the average ASD of 0.61 cm for LAD. It took less than 10 minutes to finish auto contouring the whole heart and the 5 cardiac substructures per patient, whereas it is estimated to take about 30 minutes per patient when traced manually by trained staff.
Conclusion: Automated contouring of cardiac structures is technically feasible and demonstrates strong reproducibility with manual contours, while substantially reducing contouring time. Our goals are to continue to refine this algorithm with additional datasets.
Funding Support, Disclosures, and Conflict of Interest: This work was funded by the intramural program of the National Institutes of Health, National Cancer Institute, Division of Cancer Epidemiology and Genetics.
Segmentation, Heart, Radiation Risk