基于改进AlexNet的手腕骨图像成熟等级识别  

Recognition of maturity level of wrist bone image based on improved AlexNet

作　　者：丁维龙[1] 李涛 丁潇 余鋆 毛科技[1] DING Weilong;LI Tao;DING Xiao;YU Yun;MAO Keji(College of Computer Science&Technology,Zhejiang University of Technology,Hangzhou 310023,China)

机构地区：[1]浙江工业大学计算机科学与技术学院,浙江杭州310023

出　　处：《浙江工业大学学报》2021年第6期614-622,共9页Journal of Zhejiang University of Technology

基　　金：浙江省重点研发项目(2018C01082);浙江省公益技术研究计划/工业资助项目(LGF21F020015,LGG20F020018)。

摘　　要：CHN法是目前适用于我国评估骨龄的方法之一,其评估骨龄最关键的步骤是对手腕骨图像的成熟等级评定。传统的方法是由医学专家人工阅片,不仅工作量大、耗时长,评定的准确性还受到人的主观因素干扰。为了提高骨骼等级识别的准确率,提出了一种基于改进AlexNet的手腕骨图像等级识别方法,将优化的空间变换网络加入到AlexNet网络结构中,对特征图进行旋转、平移和缩放等变换操作以获取更有辨识度的特征信息;采用Maxout激活函数作为网络中卷积层的激活函数,训练手腕骨图像成熟等级识别模型。实验结果表明:相比于原始AlexNet网络与其他几个常见的卷积神经网络,改进的AlexNet网络提高了网络模型对头状骨、钩骨、掌骨Ⅰ、远节指骨Ⅰ和中节指骨Ⅴ等成熟等级识别的准确率,分别达到了88.39%,85.35%,79.69%,79.41%和81.29%。该方法可以为基于深度学习的骨龄评估方法提供新的技术参考。The CHN method is currently one of the methods suitable for assessing bone age in China.The most critical step in assessing bone age is to assess the maturity level of the wrist bone image.The traditional method is that medical experts read the film manually,which is not only requires a lot of work and time-consuming,but the accuracy of the assessment is also interfered by human subjective factors.In order to improve the accuracy of bone grade recognition,a new method of wrist bone grade recognition based on improved AlexNet was proposed.The optimized spatial transformation network is added to the AlexNet network structure,and transformation operations such as rotation,translation,and scaling on the feature map are taken to obtain more recognizable feature information.The Maxout activation function is used in the convolutional layer in the neural network to train the recognition maturity level model of wrist bone image.The experimental results show that compared with the original AlexNet network and several other common convolutional neural networks,the improved AlexNet network improves the performance of the network model on the head bone,hook bone,metacarpal bone I,distal phalanx I,and middle phalanx V.Maturity level recognition accuracy ratesreach 88.39%,85.35%,79.69%,79.41%,and 81.29%respectively.The method can provide a new technical reference for bone age assessment methods based on deep learning.

关 键 词：深度学习 骨龄评估 注意力机制 等级识别 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]