Research On Medical Image Segmentation Based On Scale-fusion Encoding And Decoding Network

Medical image segmentation is an important component in computer-aided medical diagnostic technology.In real medical scenarios,accurate and efficient segmentation of medical images plays an important role in assisting doctors in disease diagnosis,surgical planning and pathological analysis.Medical image possesses a relatively single structure and scattered features,which is susceptible to local noise and grayscale offset caused by the interference of equipment and environment during acquisition.Hence,the performance improvement of medical image segmentation model is a big challenge.With the development of deep learning,there are many algorithms related to medical image segmentation.Currently most algorithms still can not fully combine the characteristics of medical images themselves.And effectively utilizing the multimodal information,multi-scale features and spatial neighborhood relationships of the images is hard to realize at present.So it is difficult to improve the ideal segmentation effect and model robustness.Therefore,in view of the shortcomings of the existing methods,two medical image segmentation methods based on encoder-decoder structure are proposed,and both of them can significantly improve the segmentation accuracy and robustness in medical image segmentation.The main work of this paper is as follows:(1)An image segmentation method based on full-scale fusion and flow field attention is proposed.For the existing deep learning methods can not fully combine the characteristics of medical images to extract image features,and the robustness of the method is far away from expectation when the image is disturbed by noise and offset field.In this paper,starting from the design of the feature coding structure of CNN and the feature decoding structure of CNN respectively,the feature sensing field of the encoder is further expanded by adding the global feature information extracted by the convolutional MLP module to the encoder.Secondly,through the fullscale feature fusion module,the coarse-grained information and fine-grained information of each scale jump connection feature are effectively integrated.Finally,the decoder refines the image feature information through the proposed flow field attention decoding module layer by layer.The method combines semantic flow field and attention mechanism,which strengthens feature extraction while avoiding information redundancy and obtains predicted segmentation results.(2)An image segmentation method based on Transformer fusion with scale features is proposed.In view of the limitations of convolutional neural network extraction of image features in the previous method and the network’s inadequate use of multi-scale feature information,the method improves the feature coding structure and proposes the multi-scale interactive module.Firstly,the coding structure combining Transformer and CNN can effectively capture the local information and global feature information of the input image.Secondly,the multi-scale feature interaction module adaptively integrates the jumping connection features of each scale,thereby reducing the semantic differences between the features of each layer and improving the consistency of the extracted features.Finally,the decoder uses the flow field decoding module to extract the image feature information in the refinement coding stage to obtain the predicted segmentation result of the input image.Based on the encoder-decoder structure,the two medical image segmentation methods can make full use of the multimodal information and multi-scale feature information of the image and reduce the sampling loss by the semantic flow field.The experimental results of the two methods in this paper are on the four medical image datasets of Brain Web,MRBrain S,HVSMR and Choledoch.Compared with the existing methods,the segmentation of the organizational boundary acquired from the proposed method has more delicate details.And the segmentation accuracy and robustness of the network are greatly improved.The image segmentation method based on the fusion of Transformer and scale features improves the segmentation performance of the network and reduces the overall complexity of the network by introducing Transformer compared with the previous method based on full convolutional network.