Automatic Delineation of Cardiac Substructures Using A Region-Based Fully Convolutional Network
Joseph Harms,Yang Lei,Sibo Tian,Neal S. McCall,Kristin A. Higgins,Jeffrey D. Bradley,Walter J. Curran,Tian Liu,Xiaofeng Yang
DOI: https://doi.org/10.1002/mp.14810
IF: 4.506
2021-01-01
Medical Physics
Abstract:Purpose: Radiation dose to specific cardiac substructures, such as the atria and ventricles, has been linked to post-treatment toxicity and has shown to be more predictive of these toxicities than dose to the whole heart. A deep learning-based algorithm for automatic generation of these contours is proposed to aid in either retrospective or prospective dosimetric studies to better understand the relationship between radiation dose and toxicities. Methods: The proposed method uses a mask-scoring regional convolutional neural network (RCNN) which consists of five major subnetworks: backbone, regional proposal network (RPN), RCNN head, mask head, and mask-scoring head. Multiscale feature maps are learned from computed tomography (CT) via the backbone network. The RPN utilizes these feature maps to detect the location and region-of-interest (ROI) of all substructures, and the final three subnetworks work in series to extract structural information from these ROIs. The network is trained using 55 patient CT datasets, with 22 patients having contrast scans. Threefold cross validation (CV) is used for evaluation on 45 datasets, and a separate cohort of 10 patients are used for holdout evaluation. The proposed method is compared to a 3D UNet. Results: The proposed method produces contours that are qualitatively similar to the ground truth contours. Quantitatively, the proposed method achieved average Dice score coefficients (DSCs) for the whole heart, chambers, great vessels, coronary arteries, the valves of the heart of 0.96, 0.94, 0.93, 0.66, and 0.77 respectively, outperforming the 3D UNet, which achieved DSCs of 0.92, 0.87, 0.88, 0.48, and 0.59 for the corresponding substructure groups. Mean surface distances (MSDs) between substructures segmented by the proposed method and the ground truth were <2 mm except for the left anterior descending coronary artery and the mitral and tricuspid valves, and <5 mm for all substructures. When dividing results into noncontrast and contrast datasets, the model performed statistically significantly better in terms of DSC, MSD, centroid mean distance (CMD), and volume difference for the chambers and whole heart with contrast. Notably, the presence of contrast did not statistically significantly affect coronary artery segmentation DSC or MSD. After network training, all substructures and the whole heart can be segmented on new datasets in less than 5 s. Conclusions: A deep learning network was trained for automatic delineation of cardiac substructures based on CT alone. The proposed method can be used as a tool to investigate the relationship between cardiac substructure dose and treatment toxicities. (C) 2021 American Association of Physicists in Medicine