Research On Medical Image Segmentation Algorithm Based On Encoder-decoder Structure And Attention Mechanism

There are a large number of medical images in real life.It is too time-consuming and labor-intensive to rely on the human eye for analysis.Moreover,because of the differences among professional physicians,it will lead to misdiagnosis and missed diagnosis.The emergence of deep learning technology has effectively alleviated the above problems.However,the segmentation effect based on the depth model is not very ideal that the problems of low segmentation accuracy due to the unclear edges of the two-dimensional lesions and the difficulty of accurate identification of subtle branches in the threedimensional branched structure images.In order to solve the above problems,this paper mainly has the following two aspects:A method based on the edge attention helps segmentation(EHS)network combined with deep supervision is proposed.Aiming at the problem of low segmentation accuracy due to unclear lesion edges in medical images,this paper proposes to use shallow features with rich details to perform edge constraints to provide edge attention for deep networks.This method,taking the CT images of patients with pneumonia as example to segment the pneumonia lesions,use an improved multi-task encoder-decoder segmentation network to segment the lung images.The network shares a partial encoder to extract common features,and two decoders for different tasks are used to segment edges and lesions,respectively.Meanwhile,the edge segmentation network provides edge attention for the lesion segmentation network.The proposed model is validated on the COVID-UESTC and COVID-19-CT100 datasets and the experiments show that the EHS network achieves high Dice scores and m Io U values on two different datasets.An improved 3D U-Net network is proposed to segment 3D neuron images.Aiming at the problem that subtle branches in 3D branched structure images are not easy to be accurately identified,an attention module is used to pay attention to the target branches in the encoding part.This method improves the encoding module and skip connection part of the traditional 3D U-Net network.In the encoding module,the channel space attention module is extended to three-dimensional space for application,and the neuron feature images are scored from the two dimensions of channel and space in turn,and valuable features are selected to enhance their importance.Meantime,in the skip connection part,dilated convolution is used to expand the receptive field of the convolution kernel to resample the encoded part of the image to enhance the image features.Finally,the proposed model is verified on the SYNTANEI dataset and achieves good segmentation performance.