Research And Application Of Medical Image Segmentation Technology Based On Deep Learning

Thanks to the development of human imaging technology,increasingly more diseases can be diagnosed with the help of a variety of imaging equipment.After we get the scanned image,some target tissues or lesion areas in the image usually need to be segmented for further research or application.In specific application scenario such as cardiac magnetic resonance imaging,scanners need to geometrically plan the related cardiac planes,including four-chamber,three-chamber,two-chamber and short-axis planes.The manual planning method have complicated steps and it is time-consuming.So the scanning efficiency is low.Recently,the advances in artificial intelligence technology have pushed computers to replace amounts of manual operations in the medical field.This thesis mainly studies the segmentation accuracy of new model designing methods on cardiac pre-scanned data in computer vision,and then explores the application of the segmentation technology in cardiac magnetic resonance imaging.This thesis proposes a series of medical image segmentation models based on weighted feature fusion.Firstly,a separate attention mechanism is introduced in the model.Specifically,the method is to divide the input data into multiple paths,to apply self-attention weights to the adjacent data paths,and finally to fuse the weighted values to form a basic convolutional block.This process contains multiple parallel data paths,which increase the width of the network,so the feature extraction ability of the model can be improved.Secondly,a bidirectional feature pyramid network containing jump connections,top-to-bottom path and the reverse for medical image segmentation is applied.This network can fully interact with feature maps of different scales.After that,a new activation function named Mish is also introduced,and the advantages of this function over other activation functions is also proved.Finally,in light of the fact that obtaining medical image annotations is difficult,a semi-supervised learning method is added in model training.We demonstrate the results of this method through experiments in this study.In the application part of this thesis,the technology of medical image segmentation is applied for cardiac planes planning,and the newly proposed method of automatic cardiac planes planning is used.Firstly,we segment the pre-scanned image through the abovementioned medical image segmentation model to obtain the segmentation results of various cardiac substructures.After we process the segmentation results,the key points used to plan the related planes can be retrieved.According to the coordinates of these points,the final desired planes can be obtained.We compare the planning results of the proposed method with other methods,and evidence that our proposed method outperforms others.The automatic planning method in this thesis replaces part of operation in the process of manual planning,which reduces the operation difficulty of the scanner and improves the efficiency of cardiac magnetic resonance imaging.And finally,the application value of this study is proved.