Research On Tiny Bleeding Lesions Detection In Endoscope Images Based On Deep Learning

As a non-invasive modern medical imaging technology,Wireless Capsule Endoscopy?WCE?has been widely used in gastrointestinal?GI?disease examination.However,the massive data generated has placed a heavy burden on the doctor.The use of computers to assist doctors in intelligent diagnosis is an urgent need.Bleeding are common symptoms of many GI diseases,and tiny bleeding lesions are particularly important for early detection of diseases such as early gastric cancer.Therefore,it is of great theoretical and practical significance to construct a detection method for bleeding lesions in endoscopic images,especially for tiny lesions.A detection model based on deep convolutional neural network is proposed for the detection of tiny bleeding lesions.First,a basic feature extraction network is constructed to extract features from the image automatically through a deep convolutional neural network.Then using the multi-scale region proposal network to select candidate regions of the bleeding target from the feature pyramid.Pre-screening the possible regions of the target can not only highlight the target features,but also help to reduce the fusion of invalid features.And next,by constructing a top-down feature fusion network,the fusion of the shallow features and the deep features of the target is realized.By fusing the different levels of features,the problem of the loss of the features of the tiny bleeding lesions is solved,which leads to the problem that the tiny lesions cannot be effectively detected.Thereby achieving accurate detection of tiny bleeding lesions.Model training and testing were carried out in clinical datasets.The experimental results show that the multi-scale feature fusion bleeding detection method based on deep convolutional neural network is superior to the existing methods in many indicators such as sensitivity,accuracy and F₁ score on image datasets containing tiny bleeding lesions.Among them,the sensitivity achieved a high sensitivity of 98.9%and F₁ score of 93.5%,which was 31.69%higher than the existing method,and the overall performance F₁ score increased by 22.12%.