基于深度判别性特征学习的医学影像分析方法  

Medical image analysis based on deep discriminative feature learning

作　　者：张丽[1] 闫鑫[1] 李敏[1] 姜丽丽[1] 徐虹霞[1] 张光 ZHANG Li;YAN Xin;LI Min(Jinan Central Hospital,Jinan 250013,China)

机构地区：[1]济南市中心医院,250013 [2]山东第一医科大学第一附属医院,250014

出　　处：《中国实用医药》2020年第10期196-198,共3页China Practical Medicine

基　　金：山东省软科学计划(项目编号:2018RKA01017);山东省软科学计划(项目编号:2017RKB14059)。

摘　　要：医学影像是临床诊断的重要工具,对医学影像进行自动分析能够提高医生的工作效率,减少由主观经验引起的诊断误差。特征提取是影像自动分析体系中的关键环节,对最后疾病诊断的精度具有重要的影响。然而,现有的方法忽略了特征的判别性信息。现有特征主要是利用病灶的低级特征,例如纹理、边缘等信息,虽然能够对病灶的主要特点进行表示,但是忽略了病灶之间的判别性,降低了识别性能。针对该问题,本文研究了一种基于深度判别性特征(DDC)学习的医学影像方法,提出了一种新的模型双路字典编码卷积神经网络来学习语义描述码和判别性描述码。引入图像区块(patch)划分层,生成大量的局部patch图像,从而为模型的训练提供充足的数据。引入字典编码层和分类层,加快特征学习速度,提高识别精度。提出的方法用在乳腺肿瘤识别,实验结果表明提出的方法在识别效果上优于传统方法。Medical image is an important tool for clinical diagnosis.Automatic analysis of medical images can improve the efficiency of doctors and reduce diagnostic errors caused by subjective experience.Feature extraction is a key step in the medical image analysis,and has an important impact on the accuracy of the final disease diagnosis.However,the discriminative information of the feature is ignored.In order to tackle this problem,this paper proposes a medical image analysis based on deep discriminative characteristics(DDC)learning.A novel two-way dictionary code convolutional neural network is proposed to learn discriminative code.The patch extraction layer is introduced to generate a large number of local patch images,thus providing sufficient data for the training of the model.Introducing dictionary code layers and classification layers to speed up feature learning and improve recognition performance.The proposed method is used for breast tumor recognition.Experimental Results show that the proposed method is superior to traditional Methods in recognition effect.

关 键 词：模式识别 医学影像分析 特征提取 深度学习 

分 类 号：R318[医药卫生—生物医学工程] TP391.41[医药卫生—基础医学] TP18[自动化与计算机技术—计算机应用技术]