Intracranial blood vessel segmentation plays an essential role in the diagnosis and surgical planning of cerebrovascular diseases. Recently, deep convolutional neural networks have shown increasingly outstanding performance in image classification and also in the field of image segmentation. However, cerebrovascular segmentation is a challenging task as it requires the processing of more information compared to natural images. In this paper, we propose a novel network for intracranial vessel segmentation in computed tomography angiography, which is termed as global channel attention network (GCA-Net). GCA-Net combines a four-branch at the shallow feature that captures global context information efficiently that focuses on preserving more feature details. To achieve this, we formulate an UpSampling Module (USM) by introducing the channel attention mechanism when aggregating high-level features and shallow features, leading to learning the global feature information better. This novel design is developed into different branches to learn feature information at different levels. Furthermore, we introduce Atrous Spatial Pyramid Pooling (ASPP) for capturing more details in feature maps with different resolutions. Comprehensive experimental results demonstrate the superiority of our proposed method, whereby it can achieve a dice coefficient score of 96.51% and a Mean IoU score of 92.73%, outperforming the state-of-the-art methods.
Keywords: Atrous spatial pyramid pooling; Global channel attention network; Shallow features; Vessel segmentation.
Copyright © 2020 Elsevier Ltd. All rights reserved.