Aim: This study aimed on improvement of a Coronary Sinus Veins (CS) reconstruction using the proprietary cardiac computer tomography (CT) protocol. Specifically, increasing of performance in terms of Dice score and Hausdorff Distance and also improving overall performance to reduce computation time. These cardiac structures are crucial for pre-procedural CRT assessment and successful LV lead implantation but are often poorly visible. The goal of the study was to accurately identify suitable target structures and finally improve the efficiency of CRT within our noninvasive ECG mapping technology.
Methods: The method used in this study relies on an ensemble of state-of-the-art neural networks (NN) for medical image segmentation, precisely: SwinUnet, UNet++ and SegResNet. It utilizes a large dataset and introduces a novel approach to training time loss function, combining loss function based on Dice, Common Entropy and Hausdorff Distance loss functions. Our research team extended an early collected a high-quality dataset consisting of two sets of cardiac CT studies, in particular: 130 cases of patients selected for CRT procedure (CRT subset) and 127 cases of patients selected for primary pulmonary vein isolation surgery of the left atrium (Extra subset). Each DICOM image was manually segmented by radiologist and further independently validated by two cardiologists. To test the method, 39 cases were randomly selected from the CRT subset, while the other 91+127 cases were used to train neural networks during a 5-fold cross-validation procedure.
Results: The results of the study indicate that the ensemble neural network achieved a quality of 0.81 of the dice score, 88.59~mm in terms of Hausdorff Distance and 0.85~mm in terms of mean surface distance on the test subset. This represents a significant improvement over the early results obtained.
Conclusion: The study improved an overall quality of automatic CS anatomy reconstruction and also extended a high-quality dataset.