Department of Automation, Tsinghua University, Beijing 100084, China.
Chongqing Center for Disease Control and Prevention, Department of Occupational Health and Radiation Health, Chongqing 400042, China.
Int J Environ Res Public Health. 2021 Aug 28;18(17):9091. doi: 10.3390/ijerph18179091.
Pneumoconiosis remains one of the most common and harmful occupational diseases in China, leading to huge economic losses to society with its high prevalence and costly treatment. Diagnosis of pneumoconiosis still strongly depends on the experience of radiologists, which affects rapid detection on large populations. Recent research focuses on computer-aided detection based on machine learning. These have achieved high accuracy, among which artificial neural network (ANN) shows excellent performance. However, due to imbalanced samples and lack of interpretability, wide utilization in clinical practice meets difficulty. To address these problems, we first establish a pneumoconiosis radiograph dataset, including both positive and negative samples. Second, deep convolutional diagnosis approaches are compared in pneumoconiosis detection, and a balanced training is adopted to promote recall. Comprehensive experiments conducted on this dataset demonstrate high accuracy (88.6%). Third, we explain diagnosis results by visualizing suspected opacities on pneumoconiosis radiographs, which could provide solid diagnostic reference for surgeons.
尘肺病仍然是中国最常见和最具危害性的职业病之一，其高发病率和昂贵的治疗费用给社会带来了巨大的经济损失。尘肺病的诊断仍然强烈依赖放射科医生的经验，这影响了对大量人群的快速检测。最近的研究集中在基于机器学习的计算机辅助检测上。这些方法已经达到了很高的准确率，其中人工神经网络（ANN）表现出了优异的性能。然而，由于样本不平衡和缺乏可解释性，在临床实践中的广泛应用遇到了困难。为了解决这些问题，我们首先建立了一个包含阳性和阴性样本的尘肺病 X 光片数据集。其次，我们比较了在尘肺病检测中使用的深度卷积诊断方法，并采用平衡训练来提高召回率。在这个数据集上进行的综合实验表明，该方法具有很高的准确率（88.6%）。第三，我们通过对尘肺病 X 光片上的可疑混浊进行可视化来解释诊断结果，这为外科医生提供了可靠的诊断参考。
