Research On Medical Image Segmentation Method Based On Deep Learnin

Medical image segmentation is a critical and challenging component of computer-aided diagnosis(CAD)systems,which can help medical professionals to understand and grasp the condition more intuitively and provide a sufficient basis for diagnosis,treatment and prognosis.With the development of deep learning,image processing techniques have been successfully applied to medical image segmentation,with the goal of using neural networks to learn discriminative image features from defined pixel-level objective functions to segment some important and interested regions in medical images.Among them,U-Net and its variants have become the gold standard in the field of medical image segmentation.However,medical images are affected by their own properties,such as category imbalance and noise problems,and obtaining ideal segmentation results still faces great challenges.In addition,the training and learning of deep neural network models for medical images still need further exploration.Previous methods for medical images tend to have low pixel point dependency,deficiencies in the acquisition of important discriminative features,loss of global information and utilization of multi-scale features also need to be further improved.Therefore,this paper focuses on some key problems in the field of medical image segmentation and combines the U-Net framework to conduct an in-depth study.First,in order to solve the problem of difficult extraction of discriminative features caused by the imbalance of medical images’ own categories and strong noise interference,i.e.,the problem of dilution of important information in network inference,we propose a feedback fusion network structure based on attention flow guidance,which fully extracts discriminative,potentially valuable information of features in different layers,resists the influence of feature dilution,and uses the attention mechanism to effectively removes the strong interference of noise and similar categories,makes feedback of high-level semantic information in each level and efficiently combines low-level spatial information to the backbone network,and promotes the extraction of discriminative features in the cross-level corresponding features aggregation with each other to ensure the accuracy of segmentation.Secondly,we also argue that for different scales and unknown segmentation targets in medical images,the multi-scale feature fusion in previous work fusing feature maps with different perceptual fields and size sizes by simple convolution and direct summation is suboptimal.In addition,global information is not secured in the feature inference process,which leads to an unrealistic and incomplete segmentation.Therefore,we propose a global pyramid attention module to solve the above problems,and adopt a new "activation-response" approach to integrate multi-scale information and model global contextual relationships,which greatly facilitates the capture of targets of different sizes and locations in medical images and ensures the integrity of segmentation.