基于VGG-Net的X射线全脊柱冠状面图像分割方法
Segmentation method of X-ray whole spine coronal image based on VGG-Net
申学泉 1张勇 2张润杰 2石琼芳 3宋宇锋 2张权1
作者信息
- 1. 中北大学生物医学成像与影像大数据山西省重点实验室 太原 030051
- 2. 山西医科大学第二医院 太原 030001
- 3. 太原市杏花岭区医疗集团中心医院 太原 030002
- 折叠
摘要
在计算机辅助脊柱图像分析和疾病诊断应用中,从X射线脊柱图像中自动分割脊柱和椎骨是一个关键且具有挑战性的问题.为进一步提升脊柱图像分割精度,提出一种基于VGG-Net改进的模型.首先,将VGG16网络去掉了后面的全连接层,用作U-Net的特征提取网络;其次,为了增强图像的细节信息,在特征提取网络引入小波分解模块;最后,在上采样网络中设计了一种逐像素相减的自空间注意力模块(SUB-SSAM)机制,进一步提高网络模型识别关键特征的能力.实验结果表明,改进后的模型相较于原VGG-Net模型在平均交并比(mIoU)上提高了2.39%、召回率(recall)提高了0.96%、准确率(accura-cy)提高了1.31%,训练的该网络模型可以定位到每一块椎骨,准确分割椎体区域.
Abstract
Automatic segmentation of the spine and vertebrae from X-ray spine images is a crucial and challenging task in computer-aided spine image analysis and disease diagnosis applications.To further improve the accuracy of spine image segmentation,this paper proposes an improved model based on VGG-Net.Firstly,the VGG16 network is modified by removing the fully connected layers and used as the feature extraction network for U-Net.Secondly,to enhance the detail information of the images,a wavelet decomposition module is introduced into the feature extraction network.Finally,a self-subtracted spatial self-attention module(SUB-SSAM)mechanism is designed in the upsampling network to enhance the network's ability to identify key features.Experimental results show that the improved model achieves a 2.39%improvement in mean intersection over union(mIoU),a 0.96%improvement in recall,and a 1.31%improvement in accuracy compared to the original VGG-Net model.The trained network model can accurately locate each vertebra and segment the vertebral area.
关键词
图像分割/U-Net/VGG-Net/小波分解/SUB-SSAMKey words
image segmentation/U-Net/VGG-Net/wavelet decomposition/SUB-SSAM引用本文复制引用
基金项目
山西省应用基础研究计划项目(201901D111153)
生物医学成像与影像大数据山西省重点实验室开放研究基金()
出版年
2024
NETL
NSTL
万方数据