基于Faster R-CNN的图像检索手腕骨折分类研究  

Research on Classification of Wrist Fracture Based on Faster R-CNN Image Retrieval

作　　者：杨锋 陈雷[1] 邢蒙蒙 YANG Feng;CHEN Lei;XING Mengmeng(Department of Asset Equipment,Affiliated Hospital of Shandong University of Traditional Chinese Medicine,Jinan Shandong 250013,China;Equipment Division,Department of Medical Engineering,China Rehabilitation Research Center,Beijing 100071,China)

机构地区：[1]山东中医药大学附属医院资产设备处,山东济南250013 [2]中国康复研究中心设备处医工科,北京100071

出　　处：《中国医疗设备》2023年第2期1-6,24,共7页China Medical Devices

基　　金：国家自然科学基金(81973981)。

摘　　要：目的为克服手腕X射线图像病灶区域排列复杂容易造成骨科医生漏诊误诊及诊断效率低的问题,提出一种更快速的基于区域的卷积神经网络(Faster Region-Convolutional Neural Network,Faster R-CNN)的医学图像检索手腕分类算法。方法首先利用限制对比度自适应直方图均衡化算法对手腕样本数据进行预处理,然后利用Faster R-CNN快速定位手腕图像的感兴趣区域,并提取其方向梯度直方图特征、Haralick纹理特征以及深度特征,最后利用卷积神经网络将提取到的多种特征进行有效融合后,送入本文改进的图像检索诊断模型完成对手腕图像的分类任务。结果本文提出的手腕图像检测模型分类的曲线下面积均值为0.893,诊断的准确率优于对比实验结果,较之前的研究方法提高了约5%。结论本文提出的Faster R-CNN的图像检索手腕骨折分类研究具有一定的有效性和鲁棒性。Objective In order to overcome the complicated arrangement of focal areas in X-ray images of wrist,which is easy to cause missed diagnosis,misdiagnosis and low diagnosis efficiency by orthopedic doctors,to propose a wrist classification algorithm for medical image retrieval bases on faster region-convolutional neural network(Faster R-CNN).Methods Firstly,the limited contrast adaptive histogram equalization algorithm was used to preprocess the wrist sample data,and then Faster R-CNN was used to locate the region of interest in the wrist image and extract its directional gradient histogram features,Haralick texture features and depth features.After using convolutional neural network to effectively fuse various features extracted,they were fed into the improved image retrieval diagnosis model in this paper to complete the classification task of wrist images.Results The average area under curve of the wrist image detection model proposed in this paper was 0.893,and the diagnostic accuracy was better than the results of the comparative experiment,which was about 5%higher than the previous research methods.Conclusion This paper proves that the proposed Faster R-CNN image retrieval for wrist fracture classification is effective and robust.

关 键 词：Faster R-CNN 图像检索 传统特征 深度特征 手腕骨折 

分 类 号：R687.3[医药卫生—骨科学] R197.3[医药卫生—外科学]