Aneurysm Identification in Cerebral Models with Multiview Convolutional Neural Network.

Mingsong Zhou,Xingce Wang,Zhongke Wu,Jose M. Pozo,Alejandro F. Frangi

DOI: https://doi.org/10.1007/978-3-030-46640-4_3

2019-01-01

Abstract:Stroke is the third most common cause of death and a major contributor to long-term disability worldwide. Severe stroke is most often caused by the rupture of a cerebral aneurysm, a weakened area in a blood vessel. The detection and quantification of cerebral aneurysms are essential for the prevention and treatment of aneurysmal rupture and cerebral infarction. Here, we propose a novel aneurysm detection method in a three-dimensional (3D) cerebrovascular model based on convolutional neural networks (CNNs). The multiview method is used to obtain a sequence of 2D images on the cerebral vessel branch model. The pretrained CNN is used with transfer learning to overcome the small training sample problem. The data augmentation strategy with rotation, mirroring and flipping helps improve the performance dramatically, particularly on our small datasets. The hyperparameter of the view number is determined in the task. We have applied the labeling task on 56 3D mesh models with aneurysms (positive) and 65 models without aneurysms (negative). The average accuracy of individual projected images is 87.86%, while that of the model is 93.4% with the best view number. The framework is highly effective with quick training efficiency that can be widely extended to detect other organ anomalies.