Segmentation of Intracranial Aneurysm Based on U-Net and BiConvGRU

Intracranial aneurysm (IA) cause a great risk to the health of patients. Digital subtraction angiography (DSA) is often used to diagnose IA. Early diagnosis and treatment of unruptured IA can effectively reduce the incidence of subarachnoid hemorrhage (SAH). In this paper, we proposed and evaluated a neural network structure for aneurysm segmentation to help doctors contour aneurysms from DSA sequences during aneurysm treatment. The network is based on the U-Net structure that often used in medical image segmentation. Bidirectional convolutional gated recurrent unit (BiConvGRU) module was added to the network. The module can captures the sequence changes between DSA images. In addition, the optical flow images corresponding to DSA images can be put into the network to extract the motion information of DSA. In this way, the network can obtain the spatial information, temporal information and motion information from DSA images. The experimental results showed that the dice coefficient score was 76.24% and the sensitivity was 92.82%.