Author: Jin Sung Kim, Chanwoong Lee, Young Hun Yoon 👨🔬
Affiliation: Department of Radiation Oncology, Yonsei Cancer Center, Heavy Ion Therapy Research Institute, Yonsei University College of Medicine 🌍
Purpose: Chest contrast-enhanced CT (CECT) serves as a valuable tool for cardiac imaging, but its lack of detailed anatomical visualization limits its utility in segmentation tasks. While CECT offers superior anatomical detail, its acquisition is often limited due to challenges such as allergic reactions and renal impairments. To address these limitations, we propose a mask-guided framework that utilizes organ masks derived from non-CECT (NCECT) to generate sCECT, enabling deep learning-based auto-segmentation (DLAS) of cardiac substructures in CECT-limited circumstances.
Methods: Organ and tissue masks were derived from 269 contrast-enhanced CT (CECT) cases in the lung image database consortium (LIDC) dataset using TotalSegmentator (TS) and used to train a conditional generative adversarial network (cGAN) for generating synthetic CECT (sCECT). A total of 71 paired CECT and non-contrast CT (NCECT) cases from our institution were prepared. Interim sCECT was generated using organ masks from NCECT, and the final sCECT combined skeletal structures from NCECT with organs from the interim sCECT. Ground truth for evaluation was established from actual CECT using TS to segment cardiac substructures, including atria, ventricles, and myocardium. Using this ground truth, segmentation models were trained on NCECT, CECT, and sCECT to evaluate cardiac substructure segmentation performance on CECT. Statistical significance was assessed with the wilcoxon rank-sum test.
Results: DSC for NCECT was 83.5%, 62.6%, 69.5%, 77.4% and 79.3% for the left atrium, right atrium, left ventricle, right ventricle, and myocardium, respectively. For sCECT, DSC values were 91.2%, 72.0%, 82.3%, 91.2% and 86.3% for the same substructures. There were statistically significant differences observed in the left atrium, left ventricle, right ventricle, and myocardium.
Conclusion: The mask-guided framework effectively generates sCECT with improved visualization of cardiac structures, addressing the challenges of contrast agent and limited dataset. The approach shows promise for enhancing cardiac imaging with significant improvements in key metrics.