Automatic Intracranial Aneurysm Segmentation Based on Spatial Information Fusion Feature from 3D-RA using U-Net

Abstract

Intracranial aneurysm is a severe disease that endangers human life and health, with an incidence of close to 6% in the population [1]-[3]. If there is no timely treatment, the intracranial aneurysm is likely to rupture, which will seriously endanger the patient's life. Accurate segmentation of intracranial aneurysms has essential clinical application value. However, it is challenging to achieve rapid and precise segmentation of intracranial aneurysms because of the complex structure of intracranial aneurysm, fuzzy boundary, and the overlap with normal brain tissue. Prompt and accurate segmentation can speed up the formulation of the treatment plan and improve the survival rate of patients. Currently, DSA (Digital subtraction angiography) is the gold standard for diagnosing intracranial aneurysms [4]. Using DSA data of intracranial aneurysm, segmentation has essential significance. In this paper, we proposed an automatic intracranial aneurysm segmentation method based on deep learning method, which used unreconstructed three-dimensional rotating angiography sequence to generate spatial information fusion feature images, and introduced three-dimensional spatial information for segmentation [5]. U-shaped deep neural network structure is used to achieve the pixel-level classification of images. We discussed the effect of aneurysm size on segmentation and compared the impact of aneurysm segmentation using conventional features and SIF features. During the training in traditional characteristics, the network used 2196 images for training and set aside 100 images for the test. In training with SIF features, the network used 1,749 positive and 1,741 negative images for training and reserved 100 images as test subjects. Finally, in the test image, the average Dice coefficient is 0.451, the maximum value can reach 0.883, and the minimum value is 0.022.