Research On Multi-scale Feature Pyramid Network In Medical Image Segmentatio

Segmentation of pathological regions,human organs and regions of interest in medical image data sets is a crucial step in computer-aided clinical diagnosis.In recent years,with the rapid development of deep learning,convolutional neural network is rapidly active in the field of medical image processing.Convolutional neural network can effectively extract features from medical images.Many neural network models have been widely used in medical image segmentation.Among them,the codec structure represented by U-Net has made great achievements in the field of medical image segmentation,but there is a huge challenge when segmenting fuzzy edge regions and organs with low contrast in complex organs.Based on multi-scale strategy,pyramid mechanism and other technologies,this thesis studies two improved network models,aiming to effectively achieve accurate segmentation from organs with complex lesions,fuzzy edges and low contrast.(1)A hybrid attention network model for medical image segmentation is proposed.Firstly,an edge attention module is inserted into the coding part of the network to enhance the extraction of edge contour features.Secondly,a multi-scale location attention module designed based on the multi-scale strategy and location attention mechanism is embedded at the skip connection of the network,which can effectively capture the features of interest from the context information.Finally,at the bottom of the network,this thesis introduces a scale feature pyramid attention module which can extract adjacent feature information from different receptive field feature maps.(2)Combining multi-scale strategy and feature pyramid mechanism,a multiscale feature pyramid fusion network architecture for medical image segmentation is proposed.Firstly,a multi-scale attention module designed on the basis of multi-scale strategy and residual connection is embedded in the skip connection of the network,so that the network can extract different levels of contextual detail features.Secondly,the stacked feature pyramid module is nested in the deepest part of the network to process features of different scales at the same time,so that the network can better focus on the target area.Finally,the feature perception module,a sub module of the attention module of the stacked feature pyramid,is introduced to adaptively increase the weight of the features of interest while suppressing the influence of the background area.The above two networks were validated and evaluated on the data sets of cell tissue,lung,liver and skin lesions.The experimental results show that the proposed two network models can effectively deal with the problems of fuzzy edges and low contrast.