深度哈希算法在心脏超声图像检索中的应用

Application of deep hash algorithm in cardiac ultrasound image retrieval

针对传统图像特征提取方法提取图像特征不够细致以及矢量量化检索的时间和内存消耗大等问题,结合B型心脏超声图像的先验知识,提出了基于深度哈希检索网络的心脏超声图像检索系统。首先,通过深度神经网络提取图像特征,生成哈希特征编码;然后,通过预先标注好的心脏超声图像利用坐标下降法直接对编码进行离散优化,对网络权重进行训练调整;最后,将现有的超声图像通过训练好的检索网络生成超声图像哈希编码数据库,对之后新采集的超声图像比较检索网络生成的哈希编码与数据库中的编码的汉明距离以找到以往类似的超声图像和病例,实现辅助诊疗的功能。在大规模真实采集的数据集上,对不同模型实验和测试表明,该方法相比传统检索方法有更高的准确率,检索精度可以达到0.9966,比尺度不变特征变换(SIFT)等方法可提高50%以上。该基于深度哈希的心脏超声图像检索能较好地应用于多种场景下的心脏超声图像检索,实现高精度的心脏超声辅助诊疗。

For the fact that the extracted features are not detailed enough and vector quantiatition retrieval costs large time and memory,combined with the prior knowledge of B-mode echocardiogram,a deep hash retrieval network-based echocardiogram retrieval system was proposed.Firstly,image features were extracted by deep neural network to generate hash feature coding;then,the pre-labeled echocardiographic images were dicretely optimized by coordinate descent method to train and adjust the weight of the network;finally,the existing ultrasound images were trained to generate the hash coding database of the ultrasound images through the trained retrieval network,and then the newly acquired ultrasound images were processed.By comparing the Hamming distance between the Hash coding generated by the retrieval network and the coding in the database,the similar ultrasound images and cases in the past could be found for function of assistant diagnosis and treatment.The extensive real-collected data sets,as well as different model experiments and tests show that this method has a higher accuracy than traditional retrieval methods,and the retrieval accuracy can reach 0.9966,more than 50%compared with Scale-Invariant Feature Transform(SIFT)and other methods.The deep hash based echocardiographic image retrieval can be applied to echocardiographic image retrieval in a variety of scenarios,and achieves high-precision echocardiography-assisted diagnosis and treatment.