Enhanced Automatic Segmentation of Epi-Endocardial Ventricular Anatomy Using Multi-Architecture Neural Network Ensemble

The modern development of cardiology as a branch of medicine is strongly linked to the improvement of neural networks for automatic heart segmentation.Our approach is based on three state-of-the-art neural networks for medical image segmentation: SwinUnet, UNet++ and SegRes-Net.The dataset consists of two subsets: 133 patients selected for cardiac resynchronization therapy (CRT subset) and 127 candidates recruited for heart interventions due to ventricular arrhythmias (Extra subset).Each computed tomography scan was manually segmented by a radiologist and independently validated by two cardiologists.Three classes were introduced: ventricular myocardium (wall), left, and right cavities.To test the method, 39 of the 133 cases in the CRT subset were randomly selected.The remaining patients in the CRT subset and all cases in the extra subset were used for training.The results of the study indicate that the model achieved a quality of 0.80 in terms of Dice score and 1.2 mm in terms of mean surface distance.The quality of automatic ventricular segmentation obtained demonstrates that this solution can be applied in real clinical practice.