The deep learning model achieved an average Dice score of 0.917 for automated 3D segmentation of multiple cardiac substructures in CT images.
Does a deep learning model based on X2-Net architecture accurately segment cardiac substructures in CT imaging?
A novel deep learning model provides highly accurate, fully automated 3D segmentation of multiple cardiac substructures on CT imaging, facilitating rapid quantitative analysis.
Absolute Event Rate: 0% vs 0%
Abstract Background Cardiac CT (CCT) provides high-resolution data on cardiac structures, essential for personalized medicine. However, manual 3D segmentation is time-consuming, labor-intensive, and prone to variability, making it impractical in clinical settings. Purpose This study aims to develop a deep-learning model for precise, efficient, and automated 3D segmentation of different cardiac substructures in CCT images. Methods Three distinct datasets (1559 CT scan) with varying structural characteristics were used to develop a deep learning segmentation model based on the X2-Net architecture within the nnUZoo framework. Furthermore, a human-in-the-loop approach was implemented to refine both the model and the dataset, aiming to create a unified dataset encompassing all cardiac substructures for model development. Experts visually inspected the segmentations, selected the highest-quality segmentation, and made necessary modifications. The model incorporated extensive data augmentation to improve segmentation accuracy across various cardiac substructures. The segmented cardiac substructures included the blood cavities of the Left and Right Ventricles (LV, RV), the Left and Right Atria (LA, RA), and the LA Appendage. Vascular structures encompassed the aorta, superior and inferior vena cava, pulmonary arteries, pulmonary veins, and coronary arteries. Additionally, tissue segmentation included the left ventricular myocardium (LVM), as well as peri- and epicardial fat. The model's performance was evaluated using the Dice similarity coefficient (DSC) for each segmented cardiac substructure. Results A dice average of 0.917 achieved for all cardiac substructures. The blood cavities of the Left and Right Ventricles (LV, RV) achieved Dice scores of 0.991 and 0.978, respectively, while the Left and Right Atria (LA, RA) scored 0.981 and 0.961. The LA Appendage yielded a Dice score of 0.958. Among the vascular structures, the aorta achieved 0.990, the superior and inferior vena cava scored 0.789 and 0.904, respectively, and the pulmonary arteries and veins obtained 0.967 and 0.934. The coronary arteries were segmented with a Dice score of 0.747. For tissue segmentation, the Left Ventricular Myocardium (LVM) achieved a Dice score of 0.971, while peri- and epi-cardial fat were segmented with scores of 0.860 and 0.817, respectively. These results demonstrate the model's ability to achieve accurate and consistent segmentation across various cardiac substructures. Conclusion We developed a deep learning model for fully automated and precise segmentation of cardiac substructures in 3D CT images. It enables rapid, consistent segmentation and extraction of quantitative metrics. The model is openly accessible, supporting future research in cardiac imaging.
Kazaj et al. (Sat,) reported a other. The deep learning model achieved an average Dice score of 0.917 for automated 3D segmentation of multiple cardiac substructures in CT images.