结合区域检测和注意力机制的胸片自动定位与识别  

Automatic thoracic disease localization and recognition by combining region proposal network and attention mechanism

作　　者：朱伟 张帅 辛晓燕[2] 李文飞[1] 王骏[1,3] 张建[1] 王炜[1] Zhu Wei;Zhang Shuai;Xin Xiaoyan;Li Wenfei;Wang Jun;Zhang Jian;Wang Wei(School of Physics,Nanjing University,Nanjing,210023,China;Department of Radiology,Nanjing Drum Tower Hospital The Affiliated Hospital of Nanjing University Medical School,Nanjing,210008,China;State Key Laboratory for Novel Software Technology of Nanjing University,Nanjing University,Nanjing,210023,China)

机构地区：[1]南京大学物理学院,南京210023 [2]南京大学附属鼓楼医院放射科,南京210008 [3]南京大学计算机软件新技术国家重点实验室,南京210023

出　　处：《南京大学学报（自然科学版）》2020年第4期591-600,共10页Journal of Nanjing University（Natural Science）

基　　金：国家自然科学基金(11774158,31671026,11774157,11574132)。

摘　　要：胸部X光片(以下简称胸片)是胸部相关疾病的常用诊断手段,具有辐射量低、速度快、价格低廉等优点,但样本数量巨大,所以开发基于人工智能的、对胸片进行自动识别、分类以及定位的系统具有重大的应用价值.由于胸片拍摄设备不同、胸片质量参差不齐、涉及疾病众多,尤其是缺乏标注框数据集等问题,将深度学习用于胸片的疾病检测和定位仍是一项具有挑战性的任务.为此构建了胸片标注框数据集Chest‐box,该数据集中包含3952张阳性胸片和9960个标注框.基于此数据集,提出并训练了一个区域检测网络模型,用于提取胸片中所有可能的病变区域,即图像处理领域中的感兴趣区域.以区域检测网络提取的感兴趣区域为注意力信息,进一步发展了DenseNet卷积网络和注意力机制相结合的方法,通过融合原始胸片和感兴趣区域的特征,使模型更专注于感兴趣区域,再对疾病进行识别和定位.在ChestX‐ray14数据集上的测试表明,该网络模型相比之前的工作,具有极佳的分类性能,并能提供更好的疾病定位信息.Chest X‐rays(CXR)are commonly used in diagnosing thorax diseases since they have low radiation,quick and inexpensive.The volume of CXR samples is large.It is of practical importance to develop automatic systems for recognizing,classifying and localizing diseases from CXR images.However,it is challenging to develop systems with deep learning,due to the difference in X‐ray devices,quality of images,larger number of relevant diseases,and particularly,the lack of bounding box dataset.In this work,we built a bounding box dataset,named Chest‐box,which contains 3952 positive CXR images and 9960 bounding boxes.We then developed a region proposal model to extract the regions of interest(ROI).Using the ROIs as attention information,we further employed the DenseNet121 network with attention mechanism to recognize the images and localize relevant diseases.The network uses both the whole image as well as the ROIs to extract features,and is better at focusing on the ROIs than previous models.The testing on the ChestX‐ray14 dataset and comparison with previous works show that our approach has the state‐of‐the‐art performance in both classification and localization of diseases.

关 键 词：胸片 深度学习 卷积神经网络 标注框数据集 区域检测网络 注意力机制网络 

分 类 号：TP391.[自动化与计算机技术—计算机应用技术]