Research On Medical Image Segmentation Based On Convolutional Neural Network And Attention Mechanism

Medical images reflect the internal structure or function of the human body,which are important basis for assisting doctors in disease diagnosis,preoperative analysis,and postoperative evaluation.The analysis of a large number of medical images requires enormous human and material resources,causing certain pressure on the medical system.With the rapid development of computer technology and medical imaging technology,the use of computer-assisted diagnosis by doctors has become a breakthrough in alleviating medical pressure,and medical image segmentation tasks have attracted much attention.With the development and maturity of deep learning,convolutional neural networks have been widely applied in the field of medical image segmentation and have achi eved excellent performance.Different from natural image segmentation tasks,medical image segmentation tasks have many specific problems due to its own characteristics,such as limited Receptive field,weak generalization ability,lack of high-quality annotation data,etc.This paper focuses on the problems of limited Receptive field and weak generalization ability in medical image segmentation,and proposes a series of effective theories and methods.The main innovative work can be summarized as follows:(1)Aiming at the problem of limited Receptive field in medical image segmentation tasks,this paper studies the segmentation task of malignant thyroid nodules,and proposes a feature fusion network based on super convolution kernel and attention mechanism.This network uses super large convolution kernels,combined with deep convolution and structural reparameterization,to significantly expand the effective Receptive field of the network.At the same time,super large convolution kernels are highly sensitive to features such as shape and contour,which can fully retain the complete shape information of malignant thyroid nodules in the image.Then,in order to reduce the influence of confusing background and fuzzy boundary on the focus extraction and further enhance the feature expression of the focus area,this paper designs an enhanced Transformer module,which uses the embedded large convolution kernel and the multi head self attention mechanism to model the correlation between pixels,providing more abun dant Semantic information for the subsequent feature fusion stage.Finally,the multi-scale feature fusion module uses cross fusion strategy and two different attention mechanisms at the same time to fuse the feature information of two branches to obtain a feature map with complete shape information and rich Semantic information.A series of experiments conducted on multiple datasets have shown that the large convolutional kernel and attention mechanism feature fusion network have excellent performance in medical image segmentation tasks.(2)To address the issues of weak generalization ability and insufficient training data,this paper proposes a medical image segmentation network based on style enhancement and attention mechanism.First,the networkexpands the training data by introducing a bidirectional style enhancement module,which adjusts the gray distribution of the image through different nonlinear functions,and enhances the original input data into two categories: source similar and source different.Then,based on the research of large convolution kernel,this paper also proposes a double regularization large convolution kernel module.In replacing the single batch normalization operation of the large convolution kernel module with an independent double batch normalization layer,the characteristics of images with similar sources and those with different sources are aggregated respectively to preserve the feature distribution information of images with different styles.In addition,this article uses Transformer self attention to model pixel relationships in enhanced images,and introduces global spatial attention for auxiliary training to compensate for the lack of structural and positional information.This method can simulate segmentation tasks in different scenes using the distribution characteristics of different styles of data,save the domain distribution features of images through two independent batch normalization operations,and break away from the style dependency of a single source data.This study conducted experiments on multiple datasets,and the experimental results showed that compared with traditional image segmentation algorithms,this method has better segmentation performance and stronger generalization ability.