基于改进U-Net的新冠肺炎图像分割方法

A Novel Coronary Pneumonia Image Segmentation Method based on Improved U-Net

近两年,新冠肺炎在全球暴发,给人类带来了极其严重的生命安全隐患,为尽可能地提高医生的诊断效率,研究新冠肺炎图像的病灶分割方法是极具价值的。利用U-Net网络模型作为基础网络,在第一、二层拼接前,引入改进的通道注意力,强化重要信息;在第三、四层拼接前,引入空间注意力,提取空间信息,对目标区域的细节信息进行补充;最后使用混合损失函数,加快网络收敛速度,避免样本不均衡。对比基础网络,改进后的网络模型既能够分割出较大的目标区域,也能够分割出较小的目标区域,更好地避免出现梯度消失问题,捕捉特征更加充分,有效提高分割性能和网络的可靠性。

In the past two years,the outbreak of new coronary pneumonia in the world has brought extremely serious security risks to human beings.In order to improve the efficiency of doctors in diagnosis as much as possible,it is of great value to study the lesion segmentation method of new coronary pneumonia images.This paper uses the U-Net network model as the basic network to process the pneumonia images,before the first and second layer splicing,the improved channel attention is introduced to strengthen important information;at the same time,before the third and fourth layers are spliced,the spatial attention is introduced to extract the spatial information,and the detailed information of the target area is supplemented.Finally,the mixed loss function is used to accelerate the convergence speed of the network and avoid the sample imbalance.Compared with the basic network,the improved network model can segment both large target areasand small target areas,and avoid the problem of gradient disappearance,and effectively improve the segmentation performance and the reliability of the network.