Aim: This study is focused on improvement of a automatic segmentation of epi-endocardial ventricular anatomy using the proprietary cardiac computer tomography (CT) protocol intended for the reconstruction of CS veins. The main challenge of this research is RV endocardial surface which was not previously automatically reconstructed using CT images. RV and LV anatomical structures are crucial for generating a relevant noninvasive electrical activation map. The goal of the study was to accurately identify these target structures to improve the efficiency of CRT within our noninvasive ECG mapping technology.
Methods: The method used in this study is based on a state-of-the-art neural networks (NN) for medical image segmentation namely SwinUnet, UNet++ and SegResNet and large dataset. The dataset consisting two subsets of cardiac CT studies: 130 cases of patients selected for CRT procedure and 200 selected for cardiac interventions due to ventricular rhythm disturbances. Each CT image was manually segmented by a radiologist and further independently validated by two cardiologists. Three classes were introduced: ventricular epi-endocardial surface, LV and RV cavities. To test the method, 39 cases were randomly selected while the other 291 cases were used to train NNs. Results: The results of the study indicate that the proposed model achieved a quality of 0.82 in terms of the dice score and 0.95 mm in terms of mean surface distance. These results are recognized as being of suitable quality sufficient for further noninvasive electrical activation map generation as well as for other cardio-anatomical analyses such as high-precision ventricular volume estimation.
Conclusion: The study successfully achieved a satisfactory quality of ventricular anatomy reconstruction, also leading to the collection of a high-quality dataset. Furthermore, the study confirmed the efficiency of the developed proprietary standardized cardiac CT protocol for the reconstruction of ventricular anatomy and its further use in clinical practice.