Recognition And Dynamic Tracking Of Retinal Features In Laser Photocoagulation

Diabetic retinopathy is a typical complication of diabetes.Effective laser treatment are the keys to delaying deterioration of diabetic retinopathy.Fundus laser photocoagulation surgeryrobot can assist doctors in performing operations,reducing doctor's labor intensity and improving treatment efficiency.Visual recognition and tracking system is the core of this surgical robot,which can effectively assist doctors in analysis of making preoperative plan and intraoperative tracking execution.At present,visual recognition and tracking of fundus laser surgery still faces many challenges: Firstly,it is difficult to define safe area of laser operation.Secondly,it is hard to identify small abnormal blood vessels and bleeding points in fundus.Thirdly,central point of fundus macula which is the key target of laser treatment is difficult to be located.Finally,dynamic tracking of fundus images in photocoagulation surgery is the key challenge of the operation.The research focuses on feature recognition and dynamic tracking during operation in fundus images.The main work and innovation are as follows:Aiming at dividing safe zone before laser surgery,a pre-segmentation strategy for fundus surgery is proposed.Based on Gamma transform to optimize image.And the fundus segmentation effect of edge recognition operators such as Canny is studied.Pre-segmentation binary image of safe zone contour with complete and clear edges is obtained,which provides a basic reference for more detailed segmentation.Aiming at identification and segmentation of abnormal fundus blood vessel,a method of blood vessel segmentation based on U-Net network is studied.Aiming at problems of professionally annotated data set being too small and insufficient training data,this paper proposes a segmentation training model based on optimization and data enhancement.The data set is increased by optimization methods such as image segmentation and overlap-tile strategy.Finally obtain a fundus blood vessel segmenter with high-precision and high-accuracy.Aiming macula identification and location,an identification and location method based on improved YOLO model is proposed.Based on OIA data set,images with lesions and macular spots are recollected to make a new data set.The training set is enhanced by scaling transformation,mosaic enhancement,and color space adjustment.The algorithm adaptability is improved by setting automatic learning anchor frame.The calculation performance is improved by setting adaptive scaling function.The loss function is improved by using GIOU Loss.And finally obtain a recognition and positioning model for macula with high flexibility,high adaptability,fast calculation speed,and good accuracy.A dynamic fundus tracking strategy with optic disc as the target is proposed to provide a basis for servo control of surgical robots.Fundus video of some relevant motions are simulated.Anchor frame is determined by comparison of optic disc template.This paper researches on different tracking effects of particle filter and MOSSE filter.And it researches on a new model based on dual-channel fusion of HOG features and color histograms.It completes some relevant experiments and finally obtains a very robust optical disc tracker against changes in illumination,motion blur and size changes.