1.1. Pneumoconiosis Diagnosis
Pneumoconiosis is a disease caused by long-term inhalation of mineral dust [
1]. Its retention in the lungs during occupational activities, mainly characterized as diffuse fibrosis of lung tissue, is the most serious and common occupational disease in China. The high prevalence and costly treatment of pneumoconiosis bring huge economic losses to society. According to the national occupational disease report, by the end of 2018, more than 970,000 cases of occupational diseases were reported in China, and more than 870,000 cases of pneumoconiosis were included, accounting for about 90% of all occupational disease cases. Since 2010, the number of new pneumoconiosis cases reported each year has exceeded 20,000 cases. According to relevant surveys, the average annual medical cost per pneumoconiosis case in China is 19.05 thousand yuan, and other indirect costs are 45.79 thousand yuan on average. Simplifying an average survival period after diagnosis as 32 years, the average economic burden caused by pneumoconiosis for each patient is 2.075 million yuan without taking inflation into account [
2,
3].
Though pneumoconiosis is prevalent and costly, many cases have confirmed that the earlier that pneumoconiosis is diagnosed and treated, the better treatment could be. The main cause of death in cases of pneumoconiosis lies in a variety of complications that emerge in the late-developed stage, of which respiratory complications account for 51.8% and cardiovascular disease complications account for 19.9%. Early diagnosis and treatment of pneumoconiosis will largely inhibit the development of complications, which is of great importance for treatment.
In China, the diagnosis of pneumoconiosis based on chest X-ray radiographs is still manual in clinical practice, rather than computer-aided and automatic diagnosis, which creates two drawbacks. First, the accuracy rate is not high enough. Manual radiograph reading requires high diagnostic skills, and the variation in diagnosis of pneumoconiosis caused by inconsistency of professional level and experience can be as high as 75.6%. Second, stability is not good enough. When workload is high, physicians may overlook subtle lesions due to fatigue, some of these being small pulmonary nodules and subtle calcified spots. Therefore, in order to improve the accuracy and stability of pneumoconiosis diagnosis, there are two major bottlenecks that need to be addressed, and an automatic and data-driven pneumoconiosis diagnosis system will make early and accurate diagnosis possible.
1.2. Data-Driven Methods and Deep Learning
In the literature on the computer analysis scheme of chest radiographs in the twentieth century, three main areas are distinguished by Ginneken et al. [
4]: (1) general processing techniques, (2) algorithms for segmentation and (3) analysis for a particular application. However, these methods emphasize utilization of imaging processing techniques, rather than data mining and pattern recognition. Electronic health records provide massive image data and rich patient information, especially chest radiographs and graphic details, which make data-driven methods possible.
Recent research has shown that a data-driven automatic diagnostic system can be simplified to a framework in which image features and texture patterns of chest radiographs are first extracted, followed by a data-driven classifier based on a machine learning algorithm. Yu et al. first enhanced opacity details on images by applying a multi-scale difference filter bank algorithm, and then they calculated histogram features and co-occurrence matrices as artificially encoded information [
5]. Zhu et al. utilized 22 wavelet-based energy texture features. Then, they applied a support vector machine (SVM) to distinguish between normal and abnormal samples and reached an AUC of 0.974 ± 0.018 and accuracy of 0.929 ± 0.018 [
6]. In addition, Zhu et al. compared the classification ability of decision tree (DT) and support vector machine (SVM) with four different kernels for pneumoconiosis diagnosis, and they finally reached the conclusion that the AUCs of DT and SVM were 0.88 and 0.95, respectively. Furthermore, among all tested SVM kernels, polynomial kernel has performed best [
6]. Researchers [
7] also have utilized three-stage artificial neural network (ANN) for hierarchical classification, while four extracted features are still calculated in fixed paradigm, including gray-level histogram, gray-level difference histogram, gray-level co-occurrence matrix (GLCOM) feature map and gray-level run-length matrix (GLRLM) feature map in each ROI, which is still not end-to-end.
In 1998, inspired by individual neurons in the primary visual cortex of cats, Yann LeCun et al. proposed LeNet [
8], the first modern convolutional neural network, to classify handwriting digits, which provides an end-to-end differentiable model for image classification. Convolutional neural networks (CNNs) differ from other neural network models that have convolutional operations as the main character. In 2012, AlexNet [
9], the latest CNN at that time, outperformed the second-place system by 12% in the ImageNet image classification competition. Since then, CNNs have been widely studied and largely improved. By now, researchers have proposed ZFNet [
10], VGGNet [
11], GoogleNet [
12], ResNet [
13], DenseNet [
14], EfficentNet [
15] and many other deep convolutional structures, which are called deep learning models. CNNs provide end-to-end solutions for image feature extraction and outperform traditional benchmarks in nearly all image recognition tasks, for instance, image classification, semantic segmentation, image retrieval and object detection.
Recently, data-driven deep learning has made achievements in assisting physicians with lung disease diagnosis. Cai et al. applied texture analysis in pneumonia diagnosis [
16] and achieved an accuracy of 0.793 on diagnosing 29 images. Kermany et al. established a deep learning framework for pneumonia diagnosis and applied it to CT and X-ray datasets [
17]. They demonstrated that performance of diagnosis based on deep learning is comparable to that of human experts. Specifically in silicosis diagnosis, Wang XH et al. investigated the powerful capability of deep learning and demonstrated that the performance of Inception-V3 is better than that of two certified radiologists [
18]. However, detection performance and balance between accuracy and recall still have much room for improvement. Moreover, lack of interpretation impedes CNNs from playing a part in clinical application.
In summary, contributions of this paper include the following:
- (1)
We have proposed a pneumoconiosis radiograph dataset based on electronic health records provided by Chongqing CDC, China, which is a full image dataset under privacy protection guidelines. The URL is
https://cloud.tsinghua.edu.cn/f/d8324c25dbb744b183df/ (accessed on 14 August 2021)
- (2)
We have established two data-driven deep learning models based on ResNet and DenseNet, respectively. A brief comparison and discussion has been conducted on their performance. We rebalance weights of positive and negative samples, which trade off well between accuracy and recall.
- (3)
We have explained diagnosis results by interpreting feature maps and visualizing suspected opacities on pneumoconiosis radiographs, which could provide a solid diagnostic reference for surgeons.