The Research On Ischemic Cerebrovascular Segmentation Algorithm For DSA Images

Because of its high mortality,disability and recurrence rate,cerebrovascular diseases have surpassed heart disease and cancer(malignant tumors),ranking first among the three diseases that pose the greatest threat to health.Cerebrovascular diseases mainly include ischemic cerebrovascular diseases and hemorrhagic cerebrovascular diseases,among which ischemic cerebrovascular diseases account for about 80%of cerebrovascular diseases.At present,the main treatment of ischemic cerebrovascular diseases is interventional therapy.An important problem in interventional surgery is that doctors and patients are exposed to highintensity X-rays during surgery,which greatly damages doctors’ health.Robotic interventional surgery system can avoid iatrogenic radiation risks and improve surgical accuracy.Therefore,in order to strengthen doctors’health protection and improve the accuracy of interventional surgery,The research of interventional surgical robot is particularly important.The interventional surgical robot system can avoid iatrogenic radiation risk and improve surgical accuracy,and the ischemic cerebrovascular segmentation technology based on digital subtraction angiography(DSA)is an indispensable part of the interventional surgical robot.The high resolution DSA imaging segmentation can greatly reduce the intensity of the intraoperative doctor’s film reading.Ischemic cerebrovascular segmentation technology can help physicians quickly locate the responsible vessels,accurately locate the focal location,and provide a clear vascular path for clinicians to accurately manipulate the catheter guide wire to move through human blood vessels.With the rapid development of convolutional neural networks,more and more researchers associate artificial intelligence technology and convolutional neural networks with cerebrovascular segmentation tasks in medical images.Therefore,this paper will focus on image segmentation technology based on convolutional neural network,study ischemic cerebrovascular segmentation algorithm on DSA images,and build a cerebrovascular segmentation model for DSA image.(1)We establish a new dataset of cerebrovascular segmentation on DSA images.Different from the publicly available vascular segmentation datasets such as DRIVE and CHASEDB1,there is no publicly available dataset for DSA images in the field of vascular segmentation.Therefore,in order to study the cerebrovascular segmentation algorithm targeting the DSA image,we first established a complete dataset of cerebrovascular data on DSA images.In addition,in order to obtain pixel-level annotation information,we explored an annotation method for our DSA dataset,and completed the annotation work of the dataset through image processing software,manual filling and secondary review by the clinician.(2)We propose a method of generating annotation information for blood vessel boundary.In the conventional vascular dataset,including our DSA dataset,there is no labeling information for the vascular boundary,only the labeling information for the main vessel.Our proposed method is based on the idea of regression,by calculating the gradient of ground truth map in the dataset and special mathematical transformation,to convert the discrete label into the continuous label in the form of thermal map,so that obtain the label of blood vessel boundaries.(3)We propose a CNN-based Two-branch Boundary Enhancement Network(TBENet)for automatic segmentation of cerebrovascular in DSA images.The TBENet is inspired by U-Net and designed as an encoderdecoder architecture.Given that the cerebrovascular size and shape changes widely and that the boundary of cerebrovascular of various sizes has varying degrees of blurring,we propose a boundary branch to segment the boundary of cerebrovascular and integrate it with the results of the main segmentation branch to achieve better segmentation performance.Moreover,in order to improve information fusion between two branches,we redesigned the skip connection for this network structure,which is a Main and Boundary branches Fusion Module(MBFM).Through the ablation experiment,it is proved that our Two-branch structure and MBFM are very effective in improving the segmentation performance.The TBENet was evaluated on our DSA dataset and has performed very well.Meanwhile,we tested our TBENet on the public vessel segmentation benchmark DRIVE,and the results show that our TBENet can be extended to diverse vessel segmentation tasks.