Segmenting Based on UNETR Network and 3D Reconstruction of Interventricular Septal-Free Wall Structure

Background: Percutaneous intramuscular septal radiofrequency ablation (PIMSRA) provi des an innovative alternative to septal myectomy (SM) and alcohol septal ablation (ASA). The precise segmentation of the interventricular septal is of paramount importance for the successful execution of the PIMSRA procedure. The study is based on the feasibility of applying the neural network-based method of UNETR model to the automatic segmentation of ventricular septal free wall structure in cardiac CT images of patients with HOCM and the three-dimensional reconstruction method based on visualization toolkit (VTK). Materials and Methods: The MedSAM-2, UNET and UNETR models were used to automatically segment the interventricular septal wall structure from 700 cardiac CT images of 23 patients. The image annotation tool was Labelme software, and the ratio of training set, test set, as well as validation set was 6:2:2. The higher Dice coefficient of the segmentation model was chosen to address the images, and the segmentation results were uesd to reconstruct in 3D with the metod of moving cubes. Results: Via training the UNETR segmentation model, when the parameter of batch and epoch were 8 and 32, respectively, the Dice coefficient of the interventricular septal-free wall structure test set is 0.89, which is higher than the Dice coefficient of 0.81 of the UNET model and 0.83 of the MedSAM-2 model. The model of UNETR was chosen to achieve the better segmentation results, and VTK three-dimensional reconstruction was performed based on the better segmentation results which is more closer to the real structure of heart. Conclusion: The results show that the segmentation method is feasible, and the three-dimensional reconstruction of the interventricular septal free wall structure by VTK based on the segmentation results is also feasible.