Medical Image Based On Deep Convolutional Neural Network Research On Segmentation Methods

Developing appropriate algorithms to better interpret images is a fundamental problem in the field of medical imaging analysis.Recent advances in machine learning,especially deep convolutional neural networks(DCNN),have demonstrated significant improvements in speed and accuracy for many medical image analysis tasks,such as image alignment,anatomical structures,tissue segmentation,and computer-aided diagnosis.Despite the advances in image analysis,these problems remain challenging due to the limited amount of labeled data and the large anatomical differences between patients.The feature engineering methods used in traditional algorithms are manual feature extraction methods based on expert a priori knowledge,which are demanding and generally robust for researchers and signal analysis equipment.The popular neural network-based modulation recognition in recent years can effectively simplify the process and improve the robustness of the algorithm,but there are still some problems.In addition,the complexity of traditional method models and some deep learning models is high,which makes it difficult to meet the demand of lightweight deployment in practical applications.In order to solve the above problems,two improved U-Net models based on deep learning are proposed and applied to de-segment intestinal polyps and fundus vessels,respectively,and deployed in a deep learning inference framework to solve the problem of insufficient applicability and robustness in traditional algorithms and achieve end-to-end training and inference,the main research content of this paper is as follows1.Segmentation of intestinal polyps based on gated attentionThe proposed approach based on gated attention will use a novel attention fusion mechanism to better fuse channel attention and spatial attention,and also to capture feature information of different layers more effectively.This attention unit takes reference from GRU and effectively combines the spatial attention mechanism and the channel attention mechanism in terms of computational flow.Similarly,the improved U-Net network has the ability to selectively memorize the effective feature layers and selectively forget the invalid features of some layers because the gate mechanism of GRU is retained.Finally,it is experimentally shown that the proposed method achieves better segmentation results.2.Fundus vascular segmentation method based on dense junctionThe densely connected based method mainly optimizes the U-Net network model and employs a densely connected U-shaped network that can better capture small blood vessels in the fundus retina.The proposed method is compared with the underlying U-Net segmentation model,which efficiently fuses coder and decoder features at the same stage and across stages by more dense jump connections.Subsequently,the FPN network model approach is referenced to fuse features from different scales to provide higher segmentation accuracy.