[论文解读] Fast and accurate detection of Covid-19-related pneumonia from chest X-ray images with novel deep learning model

M. M. Ramadhan, A. Faza+11 more|arXiv (Cornell University)|May 10, 2020

COVID-19 diagnosis using AI被引用 8

一句话总结

本文提出CovIDNet，一种新型深度学习模型，可从胸部X光（CXR）图像中快速准确地检测与COVID-19相关的肺炎。该模型在GitHub和Kaggle的开源CXR数据集上进行训练，实现了98.44%的准确率、100%的COVID-19精确率以及96.97%的特异性，表明其在区分COVID-19与其他肺部疾病方面具有优异性能。

ABSTRACT

Background: Novel coronavirus disease has spread rapidly worldwide. As recent radiological literatures on Covid-19 related pneumonia is primarily focused on CT findings, the American College of Radiology (ACR) recommends using portable chest X-radiograph (CXR). A tool to assist for detection and monitoring of Covid-19 cases from CXR is highly required. Purpose: To develop a fully automatic framework to detect Covid-19 related pneumonia using CXR images and evaluate its performance. Materials and Methods: In this study, a novel deep learning model, named CovIDNet (Covid-19 Indonesia Neural-Network), was developed to extract visual features from chest x-ray images for the detection of Covid-19 related pneumonia. The model was trained and validated by chest x-rays datasets collected from several open source provided by GitHub and Kaggle. Results and Discussion: In the validation stage using open-source data, the accuracy to recognize Covid-19 and others classes reaches 98.44%, that is, 100% Covid-19 precision and 97% others precision. Discussion: The use of the model to classify Covid-19 and other pathologies might slightly decrease the accuracy. Although SoftMax was used to handle classification bias, this indicates the benefit of additional training upon the introduction of new set of data. Conclusion: The model has been tested and get 98.4% accuracy for open source datasets, the sensitivity and specificity are 100% and 96.97%, respectively.

研究动机与目标

• 开发一种完全自动化的深度学习框架，用于从便携式胸部X光（CXR）图像中检测与COVID-19相关的肺炎。
• 应对疫情期间对快速、可及诊断工具的临床需求，特别是考虑到CT诊断的局限性。
• 通过广泛可用的CXR影像，提升一线医护人员的诊断准确率和速度。
• 在包含多种病理状况的开源CXR数据集上评估模型性能。

提出的方法

• 开发了一种名为CovIDNet（Covid-19 Indonesia Neural-Network）的新型深度学习模型，用于从CXR图像中提取视觉特征。
• 使用从GitHub和Kaggle的开源仓库收集的胸部X光数据集对模型进行训练和验证。
• 采用为CXR图像分析优化的卷积神经网络架构，执行特征提取和分类。
• 在训练过程中使用SoftMax以缓解类别不平衡问题，提升分类公平性。
• 采用标准指标（包括准确率、精确率、敏感度和特异性）对模型进行评估。
• 在包含标记为COVID-19或其他肺部病理状况的图像的验证集上测试性能。

实验结果

研究问题

• RQ1 仅使用胸部X光图像，深度学习模型能否在区分与COVID-19相关的肺炎与其他肺部病变方面实现高准确率？
• RQ2 在包含多样化CXR数据集的情况下，该模型在检测COVID-19方面的敏感度和特异性表现如何？
• RQ3 训练数据中的类别不平衡在多大程度上影响模型的分类性能？
• RQ4 该模型在开源CXR集合中存在不同数据源和成像差异的情况下，是否具有良好的泛化能力？
• RQ5 使用SoftMax损失在减少对罕见疾病（如COVID-19）分类偏差方面起到了何种作用？

主要发现

• 该模型在开源数据集上对COVID-19与非COVID-19病例的分类准确率达到98.44%。
• 检测COVID-19的敏感度为100%，表明在验证集中无假阴性。
• 特异性为96.97%，意味着模型正确识别了96.97%的非COVID-19病例。
• COVID-19类别的精确率达到100%，表明所有阳性预测均正确。
• 对于“其他”病理类别，模型表现出97%的高精确率，表明其在非COVID疾病上的强泛化能力。
• 研究表明，当引入新数据时，可能需要进一步微调以维持高性能。