Modified U-net with Different Attention Mechanisms for Acute Ischemic Stroke Segmentation using Non-Contrast CT

The instant diagnosis of acute ischemic stroke using non-contrast computed tomography brain scans is important for right decision upon a treatment. Artificial intelligence and deep learning tools can assist a radiology specialist in analysis and interpretation of CT images. This work aims at improving U-net model and testing it on real non-contrast CT images of acute ischemic stroke. In this work, we use the following attention modules to learn a better feature representation and for more accurate segmentation: Convolutional Block Attention Module on skip-connection stage, double attention gates on decoding stage, and Feature Pyramid Attention as bottleneck. Experiments were conducted using a combination of the Binary Cross-Entropy Loss and Dice Loss as the loss function, and separately with the Focal Tversky Loss. An anonymized sample of 500 patients with ischemic stroke was obtained from International Tomography Center SB RAS. After verification, 25 patients were used in our study. The application of the considered architecture in 2D ischemic stroke segmentation was quite successful.