Research On Fetoscopic Image Vascular Segmentation Based On Deep Learning

A fetoscope is an optical endoscope,which is often applied in fetoscopic laser photocoagulation to treat Twin-to-Twin Transfusion Syndrome(TTTS).In an operation,the clinician needs to observe the abnormal placental vessels through the endoscope,so as to guide the operation.Introducing an accurate placental vessel segmentation of fetoscopic images can help identify the abnormal vessels and effectively reduce the workload of manual reading in medical images.However,narrow Field of View(FOV),large variability in the shape,size,and location of vessels,as well as factors such as vessel highlight,pose difficulties in achieving accurate placental vessel segmentation algorithms.In order to solve the problem of placental vessel segmentation in fetoscopic images and provide assistance to doctors in treatment,this thesis has carried out research on fetoscopic image segmentation based on deep learning,including placental vessel segmentation algorithms based on dual-path encoder-decoder network and placental vessel segmentation algorithms based on vessel highlight detection.The specific research content and main contributions are as follows:1.A convolutional network named DPF-net with a dual-path structure is proposed.In this thesis,an improved encoder block and decoder block are designed.In each block,channel attention mechanism and continuous convolution structure are introduced to enable the network to better obtain multi-scale features with weights.In addition,the switch connection is introduced in the two paths of the network to strengthen the connection between the blocks.Through the vessel segmentation experiment on the open fetoscopic image dataset,it is confirmed that the network proposed in this thesis can accurately segment the placental vessels in fetoscopic images,and has achieved higher scores than the current mainstream medical image segmentation methods,raising the DSC,Io U and pixel accuracy by 5.8%,8.4% and 0.6%,respectively.2.Aiming at the problem that the highlight of placental vessels affects the segmentation effect in the fetoscopic image,a segmentation algorithm based on vessel highlight detection is proposed.The method is divided into three stages.First,the classification network is used to classify the fetoscopic image with or without vessel highlight.Then,the vessel highlight mask is generated using the proposed vessel highlight detection method based on the Frangi filter.Finally,the image is used as the additional input of the segmentation network to segment the placental vessels.Experimental results show that this method can improve the segmentation effect.DSC,Io U and pixel accuracy have increased by 2.1%,3.8% and 0.6% respectively.3.Based on the proposed methods,a fetoscopic image aided diagnosis system is designed and implemented.The system includes auxiliary diagnosis,case management,and electronic medical record functions for doctor users,as well as doctor management and model management functions for administrator users.Through the auxiliary diagnosis function,the segmentation methods proposed in this thesis and several other deep learning methods can be used to segment or enhance the fetoscopic image,assisting doctors in diagnosis.