Method Research On Medical Image Segmentation Based On Multi-path Attention Fusion

Medical images play a key role in diagnosis.Medical images have multiple modalities such as CT and X-ray.Different modalities are scanned by different equipment.Doctors make a diagnosis of the patient’s condition based on the medical images.The goal of a computer-aided diagnosis system is to accurately interpret medical images to assist medical diagnosis.One of the main research directions in the field of assisted medical care is medical image segmentation.Medical image segmentation is to outline the target from the medical image.Deep learning is the mainstream method in the field of medical image segmentation.However,existing segmentation algorithms still have unsolvable problems when segmenting images of skin,retinal blood vessels,lungs,etc.,such as the complicated distribution of blood vessels in retinal blood vessels,and the effects of light.The lesion has interference information,and the edge segmentation of the skin lesion is blurred.Existing medical image segmentation methods use simple jump connection encoding and decoding features,and cannot save encoding feature information,which makes the segmentation result poor.With a single encoding path,the encoder may suffer from feature entanglement and cannot obtain differentiated features.Even with the help of jump connections,the diversity and quality of features may still be limited.Not paying attention to the contextual information of the network,resulting in unclear target positioning.In order to solve these problems,the main research work of this paper is as follows:This thesis proposes a multi-scale aggregation network segmentation method based on dense connections.The dense connections used can ensure that the information of the encoding path is reused to obtain advanced encoding features;the residual connection acts on the decoding end,so that the gradient will not disappear or disappear when the gradient is returned.explosion.It is proposed that the multi-core pooling network module can obtain multi-scale context information,allowing the network to learn the surrounding information of the target,and solve the problem of mis-segmentation.A new aggregation method is also proposed,which combines the current low-resolution feature mapping with the high-resolution one.Compressed feature mapping is regarded as a timing relationship,and convolutional RNN and convolutional GRU are used to aggregate encoding and decoding features.A medical image segmentation method based on active conduction fusion of multi-stream features is proposed.The multi-stream encoder ensures that different encoding paths can learn differentiated features.The proposed attention-based data fusion model effectively cooperates with the multi-stream encoder.Inception Res-Atrous convolution block.Collecting relevant context information in the decoding stage,and gradually accumulating features from different paths,this method can establish meaningful connections between structural features and semantic features while maintaining a complete and flexible layout without the need for deep supervision.A large number of experiments on four medical image data sets show that this method can simultaneously obtain diverse and quality image features,which surpasses the current state-of-the-art segmentation methods.