一种改进的Unet医学影像分割模型

An Improved Unet Medical Image Segmentation Model

高准确率和高精确度的医学图像分割是业界研究的一大难题。Unet是优秀的近年来被广泛应用于医学影像分割的神经网络模型,在具体应用时,样本量少、前后背景不均衡等实际问题极大地影响了模型的分割准确率和分割精度。针对上述问题,提出了一个改进的Unet医学影像分割模型。首先改进了Unet下采样模块,在应用传统卷积的同时引入空洞卷积分支,既扩大感受野又保留位置等空间特征;其次,结合医学影像分割常用的损失函数引入通道注意力,让模型在正负样本不均衡时更加关注少样本;最后,在finding-lungs-in-ct数据集上进行了实验,模型的iou准确率超过96%,与传统Unet相比,分割性能得到了较大的提升。

High-accuracy and high-precision medical image segmentation is a major research problem in the in⁃dustry.Unet is an excellent neural network model that has been widely used in medical image segmentation in recent years.In specific applications,practical problems such as small sample size and unbalanced positive and negative samples greatly affect the segmentation accuracy and segmentation precision of the model.Aiming at these problems,an improved Unet medical image segmentation model is proposed.Firstly,the Unet downsampling module is improved,and the hole convolution branch is introduced while applying traditional convolution,which not only expands the receptive field but also preserves spatial features such as position.Secondly,combined with the loss func⁃tion commonly used in medical image segmentation,channel attention is introduced,so that the model pays more at⁃tention to the few samples when the positive and negative samples are unbalanced.Finally,experiments are carried out on the finding-lungs-in-ct dataset,and the iou accuracy of the model exceeds 96%,compared with the traditional Unet,the segmentation performance has been greatly improved.