Research On Medical Image Segmentation Method Based On Improved U-Net

Medical images can help clinicians better understand the disease and provide sufficient information for subsequent judgment in clinical diagnosis and treatment.Medical image segmentation separates lesions from complex medical images for further analysis.With the development of deep learning technology,many researchers have applied neural network to medical image segmentation and achieved good performances.However,there are still difficulties to obtain accurate segmentation results due to the common problems of medical images,such as fuzzy target boundaries,complex background,low contrast and variable target scales.In order to improve the performance of the segmentation network,many methods add complex modules,which increases the network complexity and affects the segmentation speed of the network.The accuracy and speed of medical image segmentation directly affect the efficiency and effect of medical diagnosis.Therefore,how to reduce the network complexity and improve the segmentation performance of the network is a problem worth exploring.This thesis focuses on the mentioned perspective and studies the U-Net structure that commonly used in medical image segmentation.First,according to the characteristics that the difference information can reflect the edge features of the target,a new self-adaptive difference convolution module is designed.The convolution operator of each channel can dynamically integrate the differential information with the feature information captured by the conventional convolution,and make full use of the differential feature information to improve the extraction ability of the convolution module for image features.Second,based on the generation process of feature map,a new convolution module is designed in this paper.The conversion method of feature information is improved by two convolution branches,and the multi-scale spatial information is fused in the convolution by self-calibration operation,which enriches the expression of features.Meanwhile,the conventional convolution is decomposed into asymmetric convolution to reduce the complexity of the module.Finally,an improved channel attention module algorithm based on discrete Fourier transform is proposed to retain both highfrequency and low-frequency information in the squeeze operation of channel attention and reduce the loss of feature information.Since the squeeze operation in the channel attention usually uses the global average pooling,and the results obtained by the global average pooling only represent the zero-frequency component of the discrete Fourier transform of the image,the information of the feature map is lost in the squeeze process.Therefore,the squeeze operation of the ECA channel attention module is improved by combining the idea of discrete Fourier transform,while retaining some low-frequency and high-frequency information and reducing the loss of feature information in the squeeze operation.In this work,the two designed modules are used to construct the U-Net network and verified on the DRIVE,STARE and CHASEDB1 datasets.The experimental results show that the proposed method has better feature extraction ability than U-Net,U-Net ++,Attention U-Net and Attention Channel U-Net.The AUC have achieved an improvement of 0.17% – 0.42%,0.69% – 1.16%,0.11% – 0.3% on the datasets while maintaining less number of learning parameters,and the segmentation speed is also speeded up.