Detection Of Cerebral Aneurysms Based On Convolutional Neural Network

Cerebral aneurysm is a typical condition of cerebrovascular disease.Cerebral aneurysm rupture often results in serious neurological sequelae and can be fatal.Clinically,3D time-of-flight MR angiography(3D TOF MRA)has been widely used as a screening method for cerebral aneurysms.However,the detection of unruptured cerebral aneurysms in unenhanced MRA is not always an easy task for radiologists by observing the images.The Convolutional Neural Network(CNN)has been widely used in the field of medical image analysis and has also shown good performance.The study of the use of the powerful recognition capabilities of convolutional neural networks to detect cerebral aneurysms is of great significance for the development of computer aided diagnosis or computer aided detection(CAD)systems.This paper presents three methods for detecting cerebral aneurysms using convolutional neural networks: 1)Positioning cerebral aneurysms directly in 3D TOF MRA data based on 3D U-Net;2)Using 2D CNN to identify the MIP(maximal intensity projection,MIP)image of artery to determine whether there is an aneurysm on the artery;3)Using 3D CNN to identify the 3D image of artery to determine whether there is an aneurysm on the artery.In this paper,the brain 3D TOF MRA data of 74 subjects,provided by XieHe Hospital,and in the course of the experiment,according to the aneurysm information marked by the doctor,different data processing methods were used to construct datasets for the three methods.In Method 1),We have established a positioning network based on 3D U-Net.The normalized 3D data is imported into the network,and we obtain a bounding box that may appear in the aneurysm.In Method 2)and 3),the artery is first extracted and then a volumes containing the artery are used as the dataset.In Method 2),a MIP image is generated from a volume containing the artery,with the MIP image as the input to the 2D CNN.In the difference,the input to the 3D CNN is a three-dimensional image of artery of this segment.In method 2)and 3),dataset is expanded using panning operations,at the same time,for the prediction results of the network,a voting mechanism is introduced to improve the accuracy of the classification.The results of the study showed that the effect of locating intracranial aneurysms directly in 3D TOF MRA data was not ideal,but it was able to locate most aneurysms.After the recognition area is concentrated in the artery,CNN can accurately identify whether there is an aneurysm on the artery.Data expansion can effectively improve the recognition performance of the CNN network,and the voting mechanism can also improve the accuracy of classification.Finally,the advantages and disadvantages of method 2)and 3_are compared.Method 2 can take up less computing resources,training and prediction are faster,and method 3 can get more accurate recognition results.