A 3D Multi-domain U-Net Model for Intracranial Aneurysms Detecting

Intracranial aneurysm is a higher fatality rate disease, but it is often difficult to make early diagnosis of unruptured intracranial aneurysms in clinical practice. Therefore, this paper presents a 3D multi-domain U-Net model for detecting unruptured intracranial aneurysms based on 3D TOF-MRA images, to assist physicians in detecting unruptured intracranial aneurysms fast and accurately. First, 3D TOF-MRA images are divided into three different domains according to intracranial aneurysms size in our model. Second, the correlation of the three domains is learned through depth separable convolution to realize the detection of intracranial aneurysms. Additionally, the deep supervision and residual connection are added to further improve the performance of our model. In the evaluation of 208 3D TOF-MRA images, the sensitivity and specificity of our model respectively reach 85% and 81%, and the number of parameters and training time are reduced by 23% and 22%. Therefore, the proposed model can effectively help doctors detect intracranial aneurysms more quickly and accurately which has high clinical application value.