Segmentation Of Brain Blood Vessels Based On Multi-Parametric,Multi-View Mra And Cascade Deep Learning

In recent years,neurosurgery navigation has been developing rapidly.Neurosurgery navigation provides the surgeons with the necessary tools to better visualize information about the patient’s brain anatomy and the positional relationships of these structures.Magnetic resonance angiography(MRA)has been used to achieve segmentation of brain blood vessels and visualization of vascular structures.The segmentation of brain blood vessels can assist neurosurgeons in surgical planning and treatment planning.Deep learning is a popular machine learning method which can automatically extract meaningful features from data,and many studies have shown that deep learning was successful and prospective in segmentation of brain blood vessels.However,most studies only considered the intracranial vessels,but ignored some extracranial vessels.multi-parametric multi-view MRA images can reflect cerebrovascular information from different views(axial,sagittal,coronal),which provide more robust cerebrovascular features for more accurate segmentation of brain blood vessels.In this study,we designed a cascaded deep learning networks to make full use of contextual information in multi-parametric multi-view MRA images to achieve accurate segmentation of brain blood vessels.The following three studies were performed in detail.(1)segmentation of brain blood vessels was performed with several traditional segmentation algorithms,of which the manual threshold selection method showed the best performance.The average Dice similarity coefficient was 73.3%,the average sensitivity was 64.2%,the average precision was 86.6%,and the average 95% Hausdorff distance was 4.6 mm.However,it is difficult for traditional algorithms to accurately segment extracranial vessels,low-contrast vessels,as well as small vessels.(2)2D Unet,a fully convolutional neural network,is used as the basic network.A cascade of UNets with 2 iterations of auto-context refinement,a cascade of UNets with 3 iterations of auto-context refinement,and multi-views cascade Unets were proposed.The multi-views cascade Unet showed the best segmentation performance,with the average Dice similarity coefficient of 90.3%,the average sensitivity of 87.5%,the precision of 93.2%,and the average 95% Hausdorff distance of 0.95 mm.These results showed the importance of comprehensive utilization of multi-views information and cascaded network.(3)A Nested Unet and Res Net are used as the basic network,and then the multi-parametric nested Res Unet,the multi-view and multi-parametric cascade 3D Unet,multi-view and multi-parametric cascade 3D Res Unet++ were proposed.The results showed that the multi-view and multiparametric cascade 3D Res Unet++ had the best segmentation performance,for which the average Dice similarity coefficient was 91.4%,the average sensitivity was 90.2%,the average precision was 92.7%,and average 95% Hausdorff distance was 0.58 mm.The above results show that the cascade network withmulti-views multi-parametric images can preliminarily solve the problem of segmentation of cranial scalp vessels,low-contrast vessels and small vessels.In summary,we proposed a multi-view spatial aggregation based cascaded segmentation framework to segmentation of brain blood vessels using three orthogonal 2D views(i.e.,axial,coronal,and sagittal)and multi-parametric of MRA images.The proposed method is shown to be able to provide the effective segmentation of three-dimensional cerebrovascular structures of patients to assist neurosurgical clinicians in surgical planning or treatment planning in the future.