Development of Artificial Intelligence (AI) based algorithms to classify the Pneumoconioses

项目摘要 尘肺病是一种主要的职业性肺部疾病。接触工种工人的健康筛查计划 石棉、煤和二氧化硅以及黑肺福利补偿方案要求使用 国际劳工组织(劳工组织)对尘肺病X线片进行分类的指南。NIOSH有 制定了一项认证计划，以标准化分类。尽管使用了认证的B读卡器 筛选和补偿计划、获得认证的B级读卡器的数量较少、读卡器间和读卡器内的可变性、 潜在的财务利益冲突仍然是重要的挑战。迫切需要一种 提高尘肺分类的客观性和一致性。人工智能(AI)- 以此为基础的模型已被证明对其他肺部疾病有价值，如肺部病变、浮肿和肺炎。我们的 这项研究旨在开发基于人工智能的模型，以帮助对尘肺的X光片进行分类 根据劳工组织的指导方针。AIM 1将管理一套新的专家分类胸部X光照片，并 无尘肺用于人工智能模型的训练。目标2将开发机器学习方法，包括预 训练卷积神经网络(CNN)方法及与手工挑选相结合的混合CNN方法 实质异常和胸膜异常与正常X线片的鉴别特征。目标3将 根据劳工组织分类指南，将尘肺病X光片按四大类进一步分类 指小阴影；肺的受累区域和形状；三个大小的阴影和三个胸膜亚型 异常现象。在这一目标中，深度学习算法包括贝叶斯深度学习和分类学习 剩余注意力学习(CRAL)算法将被开发用于不确定性估计和更高的预测 在多类多标签分类问题中的准确性。 我们的项目将是美国第一个开发人工智能算法来对尘肺进行分类的研究 关于劳工组织的指导方针。在算法开发中将特别注意以下分类 具有估计不确定度的临界性X光照片(即丰富，0/1 vs.1/0)。中开发的人工智能算法 这项研究将使用一套新的X光片进行测试，期望根据以下因素对尘肺进行分类 国际劳工组织的指南具有很高的准确性，尤其适用于早期尘肺病患者。该项目 与NIOSH研究实践(R2P)方法保持一致，因为建议的算法的结果将是 与NIOSH分享，分发给B读者。这项研究中开发的计算机辅助算法将 提供客观和一致的分类，以帮助解决少数人的问题 认证的B级阅读器、阅读器间和阅读器内的可变性以及潜在的财务利益冲突。

Project Summary Pneumoconiosis is a major occupational lung disease. Medical screening programs of workers exposed to asbestos, coal and silica and the compensation program for Black Lung Benefits require the use of the International Labor Organization (ILO) guidelines to classify radiographs for pneumoconiosis. NIOSH has developed a certification program to standardize the classification. Despite the use of certified B readers in screening and compensation programs, the small number of certified B readers, inter- and intra-reader variability, and potential financial conflict of interest have remained important challenges. There is a pressing need for a system to improve the objective and consistent classification of the pneumoconioses. Artificial intelligence (AI)- based models have demonstrated value for other lung diseases such as lung lesions, edema, and pneumonia. Our study aims to develop AI-based models to assist in the classification of radiographs for the pneumoconioses according to the ILO guidelines. Aim 1 will curate a novel set of expert classified chest radiographs with and without pneumoconiosis for training AI models. Aim 2 will develop machine learned methods including pre- trained Convolutional Neural Network (CNN) methods and hybrid CNN methods combined with handpicked features to distinguish parenchymal abnormalities and pleural abnormalities from normal radiographs. Aim 3 will further classify pneumoconiosis radiographs based on the ILO classification guideline by four major categories of small opacities; affected zones of the lung and shape; three sizes of large opacities and three subtypes of pleural abnormalities. In this aim, Deep Learning (DL) algorithms including Bayesian deep learning and Category-wise residual attention learning (CRAL) algorithm will be developed for uncertainty estimation and higher prediction accuracy in multi-class and multi-label classification problem. Our project will be the first study in the US to develop AI algorithms to classify pneumoconiosis based on the ILO guidelines. Particular attention in algorithm development will be given to the classification of borderline radiographs (i.e. profusion, 0/1 vs. 1/0) with estimated uncertainties. The AI algorithms developed in this study will be tested using a new set of radiographs with expectation of classifying pneumoconiosis based on ILO guideline with high accuracy especially for individuals with an early stage of pneumoconiosis. The project aligns with the NIOSH Research to Practice (r2p) approach, as the results of the proposed algorithms will be shared with NIOSH for dissemination to B readers. Computer-aided algorithms developed in this study will provide an objective and consistent classification that will assist in addressing the problems of small number of certified B readers, inter- and intra- reader variability, and potential financial conflict of interest.