采用轻量级姿态估计网络的脊柱侧弯筛查方法

PDF下载HTML

Title:

Scoliosis screening method using lightweight pose estimation network

作者:

魏旋旋¹, 黄子健¹, 曹乐², 杨皓¹, 方宇¹

1. 上海工程技术大学 机械与汽车工程学院, 上海 201620;
2. 上海工程技术大学 电子电气工程学院, 上海 201620

Author(s):

WEI Xuanxuan¹, HUANG Zijian¹, CAO Le², YANG Hao¹, FANG Yu¹

1. School of Mechanical and Automotive Engineering, Shanghai University of Engineering Science, Shanghai 201620, China;
2. School of Electronic and Electrical Engineering, Shanghai University of Engineering Science, Shanghai 201620, China

关键词:

脊柱侧弯; 姿态估计; 轻量化; 反池化; 反卷积; 热图回归; 分类器; 筛查系统

Keywords:

scoliosis; pose estimation; lightweight; depooling; deconvolution; heatmap regression; classifier; screening system

分类号:

TP391.4；TP183

DOI:

10.11992/tis.202203038

摘要:

脊柱侧弯是一种复杂的脊柱三维畸形，如不及时矫正将对身体健康产生严重影响。通过拍摄X光片或测量人体表面形貌的方法可以对脊柱侧弯进行筛查，但现有方法存在成本高、效率低且不适用于所有人群等缺点。本文提出了一种采用轻量级姿态估计网络的脊柱侧弯筛查方法，首先，将MobileNetV3的前13层作为轻量级人体姿态估计网络的编码器，经过坐标解码得到关键点的二维坐标。其次，利用各关节点的坐标计算人体姿态的空间特征；最后，用3个SVM（support vector machine）二分类器对脊柱侧弯进行详细分级，并将训练好的姿态估计和脊柱侧弯筛查模型移植到嵌入式平台。实验结果显示，该系统可以对4种不同程度的侧弯进行筛查，准确率分别为93.0%、81.7%、81.3%、86.6%。该方法的提出为脊柱侧弯筛查工作提供了一种便捷解决方案，易于在全民健康普测工作中进行推广。

Abstract:

Scoliosis is an abnormal deformity of the spine, and if left untreated, may eventually lead to severe health problems. X-rays and surface appearance measurements of the human body can be used to screen scoliosis; however, the existing methods pose several limitations, including high cost, low efficiency, and limited suitability for all patients. Herein, a scoliosis screening method is proposed based on a lightweight human pose estimation network. First, using the first 13 layers of MobileNetV3 as the encoder of the lightweight human pose estimation network, the two-dimensional coordinates of coilas can be obtained by coordinate decoding. Second, the coordinates of each coila are used to calculate the spatial features of the human body posture. Finally, scoliosis is graded in detail with three support vector machine binary classifiers, and the trained posture estimation and scoliosis screening models are transplanted into the embedded platform. The experimental results reveal that the system can screen four different degrees of scoliosis, with accuracy rates of 93.0%, 81.7%, 81.3%, and 86.6%. This proposed method provides a convenient solution for scoliosis screening and can be easily disseminated within national health survey work.