Deep Learning Based Intracranial Vascular Segmentation And 3D Reconstruction

Globally,the mortality rate of stroke is lower than that of heart disease,which has become the second leading cause of death.About 15 million people suffer from ischemic stroke every year in the world,and intracranial atherosclerosis is an important factor causing stroke.Medical tomography technology is an important means to reconstruct the three-dimensional internal information of the human body.Magnetic resonance imaging is widely used in the diagnosis of intracranial arteries.It has the characteristics of high resolution and multisequence,and can clearly distinguish the structure of different tissues,providing a variety of sequence options for clinical diagnosis.Cerebrovascular labeling in magnetic resonance images should be widely used in clinical medicine,which is the first step to further obtain vascular information.Studies have found that the geometry of intracranial arteries is closely related to the formation of atherosclerosis,and it is difficult to accurately measure the geometry of arteries through traditional image sequences.Therefore,the reconstruction of three-dimensional vascular model is crucial for further vascular analysis.Traditional vascular segmentation algorithms based on boundary or region require manual marking,which is a huge workload and lack of sufficient verification.With the development of computing science,deep neural network has been widely combined with medical image processing,and full convolutional neural network has made great achievements in the field of medical image segmentation.In this paper,deep learning technology is applied to the segmentation of intracranial blood vessels.Tomography imaging related technology is used to reconstruct the information of the blood vessels on the fault through the backprojection algorithm,and on this basis,3D reconstruction is carried out.The main work contents of this paper are as follows:Inspired by Residual network and Dense network,Res-Unet and Dense-Unet networks were proposed to segment vascular images.Res-Unet introduces the feature of residual connection,which can effectively extract the characteristics of deep network and improve the training efficiency.Based on the residual module in the ResNet-v2,a residual module with shared weight is proposed.In this module,the same convolution kernel is used for two conv layers,which can effectively reduce network parameters.Density modules in Denset were used to replace the conventional layer in U-Net,and Dense-Unet network was proposed to segment blood vessels.The experimental data came from Provincial People’s Hospital.The vascular projection images of different angles were obtained through the maximum intensity projection,and the radiologists made auxiliary markers to form the data set of vascular projection images.Due to the limited amount of data in the dataset,the dataset is enhanced by random clipping.This paper compares the training efficiency and segmentation effect of U-net Res-Unet and Dense-Unet respectively.The dense-Unet has a better segmentation effect,.and the Dice coefficient reached 0.92428.Inspired by the back-projection reconstruction algorithm in computed tomography(CT)and based on the segmentation of the projected vascular images,the projected vascular segmentation images are substituted for the projection data obtained by the translation and rotation scanning method,and then the sectional reconstruction is performed by the filtering back-projection algorithm.The vascular information in the original magnetic resonance data was enhanced by the reconstructed tomography.By measuring the quality of the enhanced images,the reconstruction effects using different number of projection angles and filtering functions are compared.The experimental results show that the S-L filter function of 30 projection vessels iamges for back-projection has a good effect.Then,3D reconstruction of the enhanced data was carried out through the Marching Cube(MC)algorithm.Reconstruction effect of 3D vascular model was demonstrated through a specific craniocerebral magnetic resonance data,and the actual operation effect of artery segmentation,thickness measurement and artery segment Angle measurement of the model was also demonstrated.