Segmentation of 3D Non-Contrast CT Brain Images Using Transformer Neural Networks

We propose a method for semantic segmentation of 3D non-contrast computed tomography brain images of acute ischemic stroke using transformer neural networks. To improve the segmentation quality of lesion areas, the pre-processing methods were implemented. The 3D Swin UNETR model is employed for segmentation, which is based on the attention mechanism. The sum of DICE loss and Focal loss are used to train the model, and DICE score as well as sensitivity and precision is utilized to evaluate the quality of model's predictions. The model was trained and tested using cross-validation on real images of patients at the International Tomography Center SB RAS. Research and comparison of the performance of the model and its analogues was carried out. The proposed algorithm demonstrates 30% greater DICE metric in comparison with the analogous 3D U-Net model. The main feature of the 3D Swin UNETR model is the increase in false positives and the decrease in false negatives compared to 3D U-Net.