摘要
【目的】肺结节图像具有相似度高和关联度高等特点,但传统图像哈希方法不能完整表达图像内容和语义信息导致检索的精度下降,因此,探讨一种基于深度哈希学习的肺结节CT相似图像检索方法。【方法】采用LIDC-IDRI公开数据集,首先,通过构造加入注意力机制的卷积神经网络与双向长短期记忆网络提取肺结节图像中带有权重信息的图像区域特征与区域间上下文相关信息,并将两种网络提取的深度特征进行融合,通过全连接层过渡到哈希层,实现哈希码的有效映射;其次,采用分级检索策略,利用本文的深度网络预测待查询图像的标注信息以获取对应的类库,在类内检索得到一组具有相似哈希码的候选对象构成候选池,然后根据池内图像高层语义特征进行相似度排序获取相似的肺结节图像列表。【结果】通过对公开数据集LIDC-IDRI进行实验分析,本文所提方法的平均检索精度提高到91.00%;与其他模型相比,准确率、召回率均有明显提升。【结论】本文构建了一种基于深度哈希网络的肺结节CT相似图像检索方法,该方法对肺结节病灶检索性能优于传统方法,可为临床医学诊断提供一定的参考价值。
【Objective】Pulmonary nodule images have the characteristics of high similarity and high correlation,but traditional hashing algorithms cannot fully express the image content and semantic features,which leads to a decrease in retrieval accuracy.Therefore,this paper proposes a method for retrieving similar images of pulmonary nodules based on deep hashing network.【Methods】Using a public data sets LDC-IDRI to expose.Firstly,the attention mechanism was add⁃ed to the convolutional neural network(CNN)and the bidirectional long-short-term memory network(BiLSTM)to obtain the regional features and inter-regional context-related information of lung nodule images with weight information,and the features extracted by the two kinds of network were fused,and then the effective mapping of hash codes was achieved through the full connection layer to the hash layer.Secondly,a hierarchical retrieval strategy was used.The annotation in⁃formation of the image to be queried was predicted by using the deep hashing network to obtain the corresponding class li⁃brary,and a set of candidates was retrieved with similar hash codes within the class.Then,similarity ranking was per⁃formed based on the high-level semantic features of the images in the set of candidates for final search results.【Results】Through experimental analysis of a public data set,LIDC-IDRI,the average retrieval accuracy of the proposed method was increased to 91.00%.Compared with other models,the accuracy and recall rate have been significantly improved.【Conclusions】In this study,a similar image retrieval for pulmonary nodules based on deep hashing network was proposed.The retrieval performance of this method for pulmonary nodule lesions was better than that of traditional methods,which could provide a reference value for clinical diagnosis.
作者
郝瑞
秦亚雪
甄俊平
强彦
HAO Rui;QIN Ya-xue;ZHEN Jun-ping;QIANG Yan(School of Information,Shanxi University of Finance&Economics,Taiyuan 030006,China;School of Management Science&Engineering,Shanxi University of Finance&Economics,Taiyuan 030006,China;Department of Medical Imaging,The Second Hospital of Shanxi Medical University,Taiyuan 030001,China;College of Information and Com-puter,Taiyuan University of Technology,Taiyuan 030600,China)
出处
《中山大学学报（医学科学版）》
CAS
CSCD
北大核心
2022年第4期667-674,共8页
Journal of Sun Yat-Sen University:Medical Sciences
基金
国家自然科学基金(61373100,61872261)
中国博士后科学基金(186544)
山西省自然科学基金(201901D111319)。
关键词
肺结节
图像检索
卷积神经网络
双向长短期记忆网络
注意力机制
深度哈希
计算机断层扫描
pulmonary nodule
image retrieval
convolutional neural network
bidirectional long-short-term memo⁃ry network
attention mechanism
deep hash
computed tomography
作者简介
通信作者:郝瑞,博士,副教授,研究方向:肺癌影像智能诊断,E-mail:sxtytutu@163.com。
