计算机工程 ›› 2020, Vol. 46 ›› Issue (7): 306-311,320.doi: 10.19678/j.issn.1000-3428.0055204

• 开发研究与工程应用 • 上一篇    下一篇

基于深度学习的胸部常见病变诊断方法

张驰名1, 王庆凤1, 刘志勤1, 黄俊1, 陈波1, 付婕1, 周莹2   

  1. 1. 西南科技大学 计算机科学与技术学院, 四川 绵阳 621010;
    2. 绵阳市中心医院 放射科, 四川 绵阳 621010
  • 收稿日期:2019-06-14 修回日期:2019-07-18 发布日期:2019-08-09
  • 作者简介:张驰名(1994-),男,硕士研究生,主研方向为人工智能、医学图像分析、机器学习;王庆凤,博士研究生;刘志勤(通信作者),教授;黄俊,博士研究生;陈波,教授、博士;付婕,讲师、硕士;周莹,主治医师。
  • 基金项目:
    四川省科技计划项目(2019JDRC0119);四川省军民融合研究院开放基金(2017SCII0220,2017SCII0219)。

Diagnostic Method of Frequently Occurring Chest Diseases Based on Deep Learning

ZHANG Chiming1, WANG Qingfeng1, LIU Zhiqin1, HUANG Jun1, CHEN Bo1, FU Jie1, ZHOU Ying2   

  1. 1. School of Computer Science and Technology, Southwest University of Science and Technology, Mianyang, Sichuan 621010, China;
    2. Radiology Department, Mianyang Central Hospital, Mianyang, Sichuan 621010, China
  • Received:2019-06-14 Revised:2019-07-18 Published:2019-08-09

摘要: 胸透X射线广泛应用于多种胸部常见病变的筛查任务,由于不同类型的胸科疾病在病理形态、大小、位置等方面往往具有多样性以及较大的差异性,且疾病样本具有比例不平衡等问题,导致难以通过深度学习技术来检测并定位胸部疾病区域。针对该问题,提出一种基于深度学习的胸部疾病诊断算法。通过压缩激励模块实现自适应特征重标定,以提高网络的细粒度分类能力。采用全局最大-平均池化层增强网络病理特征的空间映射能力,使用焦点损失函数降低简单易分类样本的权重,使得模型在训练时更专注易错分样本的学习。在此基础上,通过梯度加权类激活映射实现弱监督病变区域的可视化定位,为网络预测结果提供相应的视觉解释。在ChestX-Ray14官方数据划分标准下进行训练与评估,结果表明,该算法对14种常见胸部疾病的诊断效果较好,平均AUC值达到0.83。

关键词: 卷积神经网络, 医学图像分类, 计算机辅助诊断, 胸部X射线, 胸部病变诊断

Abstract: Chest X-ray is commonly used in the examination of multiple types of frequently occurring chest diseases.However,there is high difference and diversity of chest diseases in pathological morphology,size and location,and the ratio of disease samples is imbalanced.So it is challenging to detect and locate chest diseases by deep learning.To address the above problems,a diagnostic algorithm for chest diseases is proposed.Firstly,the adaptive feature recalibration is implemented through the squeeze-excitation module to improve the fine-grained classification ability of the network.Secondly,the spatial mapping ability of the pathological features of the network is enhanced by the global max-average pooling layer.Then the focus loss function is used to reduce the weight of easily classified samples,so that the model can focus more on the learning of easily misclassified samples in training.Finally,the visualized location of weakly supervised lesion areas is implemented through the Gradient-weighted Class Activation Mapping(GCAM),providing corresponding visual interpretation of network prediction results.Training and evaluation results on the official data division criteria of ChestX-Ray14 show that the proposed algorithm has excellent performance in the diagnosis of 14 frequently occurring chest diseases with an average AUC of 0.83.

Key words: Convolutional Neural Networks(CNN), medical image classification, Computer Aided Diagnosis(CAD), chest X-ray, diagnosis of chest diseases

中图分类号: