[论文解读] Abnormality Detection and Localization in Chest X-Rays using Deep Convolutional Neural Networks

Mohammad Tariqul Islam, Abdul Aowal+2 more|arXiv (Cornell University)|May 27, 2017

COVID-19 diagnosis using AI参考文献 25被引用 115

一句话总结

本文评估多种深度卷积网络在胸部X线异常上的表现，使用 Indiana、JSRT 和 Shenzhen 数据集，结果显示集成 DCN 模型优于基于规则的方法，并引入基于遮挡的定位用于心脏增大和肺水肿的定位。

ABSTRACT

Chest X-Rays (CXRs) are widely used for diagnosing abnormalities in the heart and lung area. Automatically detecting these abnormalities with high accuracy could greatly enhance real world diagnosis processes. Lack of standard publicly available dataset and benchmark studies, however, makes it difficult to compare various detection methods. In order to overcome these difficulties, we have used a publicly available Indiana CXR, JSRT and Shenzhen dataset and studied the performance of known deep convolutional network (DCN) architectures on different abnormalities. We find that the same DCN architecture doesn't perform well across all abnormalities. Shallow features or earlier layers consistently provide higher detection accuracy compared to deep features. We have also found ensemble models to improve classification significantly compared to single model. Combining these insight, we report the highest accuracy on chest X-Ray abnormality detection on these datasets. We find that for cardiomegaly detection, the deep learning method improves the accuracy by a staggering 17 percentage point compared to rule based methods. We applied the techniques to the problem of tuberculosis detection on a different dataset and achieved the highest accuracy. Our localization experiments using these trained classifiers show that for spatially spread out abnormalities like cardiomegaly and pulmonary edema, the network can localize the abnormalities successfully most of the time. One remarkable result of the cardiomegaly localization is that the heart and its surrounding region is most responsible for cardiomegaly detection, in contrast to the rule based models where the ratio of heart and lung area is used as the measure. We believe that through deep learning based classification and localization, we will discover many more interesting features in medical image diagnosis that are not considered traditionally.

研究动机与目标

• 评估不同 DCN 架构（AlexNet、VGG、ResNet）在公开数据集上的胸部X线异常检测性能。
• 评估使用集成模型相对于单一 DCN 对心脏增大等异常的收益。
• 开发并应用基于遮挡敏感性的定位方法以识别对分类决策有贡献的区域。
• 将 DCN 基于心脏增大检测与基于规则的特征进行比较并报告改进。
• 证明对其他数据集（结核病检测）具有泛化能力。

提出的方法

• 在胸部X线数据集上对预训练的 DCN（AlexNet、VGG-16/19、ResNet-50/101/152）进行微调，使用 Adam 优化器（lr=0.001）。
• 从选定层提取特征（如 res4f、res4b22、res4b35；pool5），并训练用于异常检测的二分类器。
• 通过在多个 DCN 的模型概率上进行简单线性平均来探索集成预测。
• 使用 40×40 的遮挡模板进行遮挡敏感性分析，通过在 CXRs 上滑动遮挡物并跟踪病理概率的下降来生成定位概率图。
• 在平衡的训练/测试划分和多次随机划分下，使用准确率、AUC、灵敏度、特异性进行评估以提高鲁棒性。
• 将该方法应用于深圳 TB 数据集以证明迁移能力。

实验结果

研究问题

• RQ1 不同 DCN 架构在公开数据集上对心脏增大和其他胸部X线异常的表现如何？
• RQ2 DCN 的集成是否相较单一模型提升检测准确性和 AUC？
• RQ3 基于遮挡的定位能否可靠定位胸部X线中的异常，如心脏增大和肺水肿？
• RQ4 DCN 基于心脏增大检测与基于规则的方法相比如何，且该方法是否对深圳数据集的结核病检测具有泛化性？

主要发现

模型	准确率 (%)	AUC	灵敏度 (%)	特异性 (%)
AlexNet	86.00	0.92	86.00	86.00
VGG-16	86.00	0.87	96.00	76.00
VGG-19	92.00	0.94	92.00	92.00
ResNet-50	87.00	0.93	94.00	80.00
ResNet-101	92.00	0.92	88.00	96.00
ResNet-152	90.00	0.91	92.00	88.00

• 集成 DCN 模型在心脏增大检测的准确率最高可达 93.0%，AUC 最高可达 0.9728，优于单一模型。
• 单一模型的性能随异常而异；对于心脏增大，VGG-19 的准确率高达 92% 且 AUC 为 0.94，而 ResNet-152 在随机划分上的平均表现较强。
• 较浅层的特征在心脏增大检测中常常优于深层特征；来自 ResNet-152 早期层（如 res4b35）的特征提供更好准确性。
• Dropout 的放弃通常有助于更深的模型，但可能降低深层网络的性能；六个 DCN 的集成在心脏增大任务上提供最佳整体性能（准确率 93%，AUC 0.97）。
• 基于遮挡的定位将心脏区域作为心脏增大检测的主要区域、将肺部区域作为肺水肿的定位区域；定位结果与临床期望区域一致，尽管对水肿结果有时超出解剖范围。
• 在 Shenzhen TB 数据集上，集成模型达到最高准确率（90.0%）和 AUC（0.94）。