多尺度特征融合注意力新冠肺炎病灶分割网络
COVID-19 lesion segmentation network based on multi-scale feature fusion attention
林洁沁 1黄新2
作者信息
- 1. 桂林电子科技大学电子工程与自动化学院,广西桂林 541004
- 2. 桂林电子科技大学电子工程与自动化学院,广西桂林 541004;广西自动检测技术与仪器重点实验室,广西桂林 541004
- 折叠
摘要
新冠病毒传染性极强,尽早的诊断和治疗是减少疫情造成损失的关键因素.为辅助医生诊断新冠病情,高效、准确地从肺部CT切片中分割新冠病灶,提出了一种改进的编码器-解码器深度神经网络——多尺度融合注意力网络MSANet(Multi-scale Attention Network),以图像分割效果较为出色的U-Net网络为基础,通过全局池化层和设置空洞卷积的采样率,增大网络感受野,捕获多尺度信息,实现对大目标的有效分割;使用通道注意力与空间注意力,在空间维度上建模,有效提取图像深层特征.测试结果表明,改进后的算法与U-Net网络相比,分割的平均交并比提升了 1.46％,类别平均像素准确率提升了 0.8％,准确率提升了 1.17％.
Abstract
Because of COVID-19's highly infectious,early diagnosis and treatment are the key factors to reduce the losses caused by the epidemic.In order to assist doctors in the diagnosis of COVID-19 and efficiently segment CO-VID-19 lesions from lung CT slices,an improved encoder-decoder deep neural network based on the U-Net with ex-cellent image segmentation effect,Multi-scale Attention Network(MSANet)is proposed.By using a global pooling layer and setting a sampling rate for void convolution,the network receptive field is increased,and multiscale informa-tion is captured to achieve effective segmentation of large objects.MSANet uses channel attention and spatial attention to model in the spatial dimension to effectively extract deep image features.The test results show that compared with U-Net network,the improved algorithm improves the mean intersection over union of segmentation by 1.46％,improves the mean pixel accuracy of category by 0.8％,and improves the accuracy by 1.17％.
关键词
图像处理/特征提取/卷积块注意力模块/空洞空间卷积池化金字塔/U-Net结构/多尺度特征融合Key words
image processing/feature extraction/convolutional block attention module/atrous spatial pyramid poo-ling/U-Net structure/multi-scale feature fusion引用本文复制引用
出版年
2024
NETL
NSTL
万方数据