一种改进的双通道多尺度医学图像分割网络模型

An Improved Dual-Channel Multi-Scale Medical Image Segmentation Network Model

近年来,深度学习在医学图像处理中的应用越来越广泛。UNet作为医学图像分割的主要方法,虽然具备一定的优势,但在细节处理及精确度上仍不够理想。为了优化分割效果,本研究提出一种改进的UNet分割模型。该模型在UNet的网络架构基础上,使用残差连接代替编码器和解码器之间的跳跃连接;同时将编码器中卷积操作更改为双通道多尺度卷积,并在编码器中加入挤压激励块。在Kvasir-SEG、CVC-ClinicDB、ISIC2018三个数据集上的实验结果表明,该模型与其它几种UNet改进算法相比,分割精度有所提高。

In recent years,deep learning has been increasingly used in medical image processing.As the main method of medical image segmentation,UNet has certain advantages,but it is still not ideal in detail pro‐cessing and accuracy.In order to optimize the segmentation effect,this study proposes an improved UNET seg‐mentation model.Based on the network architecture of UNet,the model uses residual connection to replace the jump connection between the encoder and the decoder.At the same time,the convolution operation in the encod‐er is changed to dual channel multi-scale convolution,and a squeeze excitation block is added to the encoder.The experimental results of Kvasir-Seg,CVC-ClinicDB and ISIC2018 data sets show that comparing with other UNet improved algorithms,the segmentation accuracy of this model is improved.