Research On Guidewire Segmentation And Tracking Algorithms In Vascular Interventional Surgical Robot

The morbidity and mortality of cardiovascular disease remain high in the world,and assisted intubation using vascular interventional surgery robots is an effective treatment method.Fast and accurate guidewire segmentation and tracking can help doctors better complete the intervention of medical devices,which is of great significance.However,the task of fast and accurate guidewire segmentation and tracking is challenging due to the low contrast of guidewire morphology and structure and the influence of uneven background noise in medical images.In this paper,on the basis of building an experimental platform for vascular interventional surgery robot,the image data acquisition of blood vessels and guidewire is completed,and a guidewire segmentation based on pixel intensity is realized.The focus is on the realization of a guidewire segmentation algorithm based on the deep learning MobileUNet neural network structure and a guidewire tracking algorithm based on the Gaussian Process(GP)model.The main work of the thesis is as follows:(1)Completed the construction of the vascular interventional surgery robot experimental platform and the acquisition of blood vessel and guidewire image data.Through the master-slave design of the robot,the intervention of the guidewire is simulated on the in vitro model to test the consistency of the movements of the master and slave ends.The angiography experiment was completed,the guidewire intervention path was planned through 3D reconstruction,and the guidewire image data acquisition was completed.(2)The research implements a guidewire segmentation algorithm based on pixel intensity.Using the collected guidewire image data,the guidewire feature enhancement and denoising are performed through image frame preprocessing,guidewire morphological feature extraction and multi-scale feature enhancement.The accuracy rate,recall rate,F1-score and segmentation speed of guidewire segmentation were analyzed.(3)A deep learning-based guidewire segmentation algorithm is proposed.By optimizing the Mobile-Net and U-Net network structures,the Mobile-UNet neural network structure is designed to better match the guidewire segmentation task,and the Dice＿loss and Binary＿crossentropy loss functions are optimized to reduce the loss of image training.By optimizing the network model training parameters to improve the segmentation accuracy and the speed of model calculation,segment the guidewire image on the test set,and obtain the comparison result of the guidewire segmentation experimental algorithm,which effectively improves the accuracy of the guidewire segmentation,PR curve value and segmentation speed and other indicators.(4)A guidewire tracking algorithm based on Gaussian process model is proposed.By establishing a Gaussian process model for the curvature and direction changes of the guidewire,and using Radon transform to extract the guidewire pixel features of the guidewire target,the extracted features are used as the prior information input by the GP,and the direction of the guidewire pixel points is analyzed.It is predicted that the center line of the guidewire is obtained and the tracking rate is improved,and the guidewire width is set to obtain the experimental results of the guidewire pixel extraction,which effectively improves the accuracy,sensitivity and Matthews correlation coefficient and other indicators.