Deep Learning for Computer-aided Diagnosis of Pneumoconiosis

Background: The diagnosis of pneumoconiosis relies primarily on chest radiographs and exhibits signicant variability between physicians. Computer-aided diagnosis (CAD) can improve the accuracy and consistency of these diagnoses. However, CAD based on machine learning requires extensive human intervention and time-consuming training. As such, deep learning has become a popular tool for the development of CAD models. In this study, the clinical applicability of CAD based on deep learning was veried for pneumoconiosis patients. Methods: Chest radiographs were collected from 5424 occupational health examiners who met the inclusion criteria. The data were divided into training, validation, and test sets. The CAD algorithm was then trained and applied to processing of the validation set, while the test set was used to evaluate diagnostic ecacy. Three junior and three senior physicians provided independent diagnoses using images from the test set and a comprehensive diagnosis for comparison with the CAD results. A receiver operating characteristic (ROC) curve was used to evaluate the diagnostic eciency of the proposed CAD system. A McNemar test was used to evaluate diagnostic sensitivity and specicity for pneumoconiosis, both before and after the use of CAD. A kappa consistency test was used to evaluate the diagnostic consistency for both the algorithm and the clinicians. Results: ROC results suggested the proposed CAD model achieved high accuracy in the diagnosis of pneumoconiosis, with a kappa value of 0.90. The sensitivity, specicity, and kappa values for the junior doctors increased from 0.86 to 0.98, 0.68 to 0.86, and 0.54 to 0.84, respectively (p<0.05), when CAD was applied. However, metrics for the senior doctors were not signicantly different. Conclusion: DL-based CAD can improve the diagnostic sensitivity, specicity, and consistency of pneumoconiosis junior These results demonstrate that CAD can improve the consistency between diagnostic and performance approached level senior Specicity improved signicantly for and their diagnostic from moderate to good. Applying CAD also improved the sensitivity, specicity, and consistency for SD group, though the results were not statistically signicant.


Introduction
Pneumoconiosis is the most widely distributed and harmful occupational disease in China [1,2]. It is characterized by the diffuse brosis of lung tissue, caused by long-term inhalation of inorganic mineral dust and retention in the lungs during occupational activities. The cumulative number of cases has reached one million in china and it continues to increase at a rate of more than 20,000 new cases per year [2]. Dust inhaled through the respiratory tract can include silicon dioxide molecules, which cause pneumoconiosis alveolitis, focal lesions, nodular lesions, dusty brosis, massive brosis, and other pathological changes in the lungs. These lesions primarily manifest as small round structures with an irregular opacity, diffuse interstitial brosis of the lungs, and silicosis masses.
Conventional classi cation of pneumoconiosis is primarily based on the "International X-ray Classi cation of Pneumoconiosis" guidelines issued by the International Labor Organization in 2011 [3,4]. The current standard in China is the "Diagnosis of Occupational Pneumoconiosis" (GBZ 70-2015), Page 3/14 which is based on the correct interpretation of chest X-ray lms using the profusion of small opacity, lung zone distributions, and pleural plaques to diagnose and classify pneumoconiosis [5,6].
As the occurrence and progression of pneumoconiosis is a continuous process, the profusion of small opacity lesions in chest radiographs is also a sequential procedure. While professional training, comparisons of standard lm, and improvements to imaging equipment and technology can increase the diagnostic accuracy of occupational physicians, inter-clinician differences remain high due to the subjective nature of image interpretation. Prior studies under the same external conditions have suggested that inconsistent judgments based on the shape, size, and number of small opacity lesions are the primary cause of inconsistent pneumoconiosis classi cation [7]. As such, increasing the objectivity, accuracy, and consistency of the diagnostic process is highly important.
Computer-aided diagnosis (CAD) provides an objective methodology for improving the interpretation of medical images. The rapid development of computer technology and medical imaging equipment has enabled the effective combination of arti cial intelligence and image processing in recent years, which has led to improvements in detection and evaluation of disease severity [8]. Using computer-generated results as a reference, radiologists can draw more accurate conclusions for disease screenings and cancer risk assessments [9][10][11][12].
Several recent studies have investigated the application of CAD technology to the diagnosis of pneumoconiosis. New approaches have been developed for image preprocessing, characterization extraction, classi er selection, and optimization [13][14][15]. However, acquiring the annotated training data used in conventional machine learning models can be both time-and cost-prohibitive, as lesions must be labeled manually by clinical experts. Subjective error and bias can also become problematic in this process.
The rapid development of arti cial intelligence has led to broad applications of deep learning (DL) algorithms for medical image analysis. These models utilize a multi-layer network to facilitate the automated learning of implicit relationships within the data. The resulting characteristics are often more diverse and expressive, particularly in tumor imaging applications. DL can provide semi-supervised or unsupervised autonomous learning of a target image for classi cation tasks. It can also synthesize images with the same characteristics and imitate the independent learning and analysis capabilities of humans, thereby reducing the subjectivity of extracted features [16][17][18][19]. At present, there is no relevant report on the evaluation of pneumoconiosis imaging diagnoses based on deep learning computer-aided diagnostic technology.
Convolutional neural networks (CNNs) and deep residual network (DRNs), an extension of CNNs, are common DL algorithms used for image classi cation [20]. These models provide the advantages of simplicity, practicality, and generalizability. Alternative models have also been proposed with varying convolutional layer quantities, including ResNet18, ResNet50, and ResNet101, which includes ve network depths, a total of 100 convolutional layers, and a connection layer. These and other similar algorithms have been widely applied to image segmentation, detection, and recognition tasks [21,22]. This study focused on assessing the application value of DL-based CAD technology in diagnosing pneumoconiosis.

Materials And Methods
This study was approved by the ethics committee of the Occupational Safety and Health Research Center (No. 2018003). Informed consent was waived in this study due to its retrospective nature.

Patients
Chest radiographs were acquired from occupational health monitoring medical examiners from January to September 2017. A total of 6020 patient (all male) images were include in the dataset, aged between 22 to 67 years (mean age: 48.62±12.1 y), with a 2-38 years of dust exposure history. The quality and diagnosis of chest radiographs were based on GBZ70-2015. Inclusion criteria required DR image quality to meet second level lm standards. Exclusion criteria included signs of pneumonia, a tumor, tuberculosis, or other lung diseases that might affect the diagnosis. The number of control and study groups, as well as the stage criteria for study groups, are listed in Tables 1 and 2, respectively. The gold standard was produced by three experienced clinicians. Cases diagnosed as positive by two or more experts were classi ed as positive, and the remaining images were assumed to be negative. Pneumoconiosis in varying stages is shown in Figure 1.
To avoid the problem of over-tting (common in machine learning), fty images were selected using a random number table method to form positive and negative case groups. A total of 100 cases were used as the test set and the remaining positive and negative data were divided into training and veri cation sets using an 80/20 proportion. The training set was used for preliminary network training and the test set was used to evaluate the diagnostic e ciency of the CAD system.  Establishment of a CAD model for pneumoconiosis A DL algorithm based on DRN-Resnet101 was used to establish a CAD system for diagnosing pneumoconiosis. The included parameters were set as follows. The pre-training learning rate was 0.001 (using the SGD algorithm), which was decreased to 0.0001 after 6000 iterations. A total of 10000 epochs were used with a batch size of 64. All images were rst desensitized, and a U-Net was used to extract lung elds on both sides, by removing invalid areas and adjusting the image size to a single 256 × 256 pixel matrix [23]. Images were then converted to a JPG format and the training data were input to the model. The veri cation set was used to determine the effectiveness of the model and gradually improve the accuracy of model output through continuous iterative optimization.

Image reading
Six physicians were divided into junior doctor (JD) and senior doctor (SD) groups based on their diagnostic experience. The three doctors in the SD group each had more than 10 years of experience in diagnosing pneumoconiosis. Those in the JD group each had 3-5 years of experience. These participants conducted independent diagnoses on the test set and a comprehensive diagnosis with reference to the CAD results, following the GBZ70-2015 guidelines. The interval between the two diagnoses was 10 days. Chest radiograph images were interpreted using a PACS workstation and displayed on Jusha 5 M medical monitors.

Statistical analysis
The MedCalc 15.2.2 software package (MedCalc Software bvba, Ostend, Belgium; http://www.medcalc.org; 2015) was used for statistical analysis. A receiver operating characteristic (ROC) curve was used to evaluate the diagnostic e cacy of the CAD system for diagnosing pneumoconiosis [24]. Area-under-curve (AUC), sensitivity, and speci city values were also calculated. A McNemar test was used to evaluate diagnostic sensitivity and speci city for clinicians with and without the use of CAD software. A value of p<0.05 was considered to be statistically signi cant. A Kappa test was included to evaluate consistency between the CAD results, the physician diagnosis (with and without CAD), and the gold standard. Consistency was poor, fair, and good for K values of 0.01-0.39, 0.4-0.74, and 0.75-1, respectively.

CAD diagnostic e ciency for pneumoconiosis
The results of ROC analysis showed the AUC value for CAD-based pneumoconiosis was high (see Figure  2 and Table 3). Diagnostic sensitivity and speci city were also high and consistent with the gold standard. Classi cation results are shown in Tables 2 and 3, where it is evident that including CAD increased the diagnostic sensitivity of the JD group from 0.86 to 0.98 and the speci city from 0.68 to 0.86. These differences were statistically signi cant. Diagnostic sensitivity for the SD group increased from 0.94 to 0.98 and speci city increased from 0.90 to 0.94. This difference was not statistically signi cant. Diagnostic results with and without the use of CAD are provided in Tables 4 and 5.  Table 4 Diagnostic results with and without the use of CAD  Kappa test results showed that including CAD increased the diagnostic consistency between the JD group and the gold standard from moderate to good, as the kappa value increased from 0.54 to 0.84. The consistency of the SD group improved slightly, with the kappa value increasing from 0.84 to 0.92 (see Table 6). These results demonstrate that CAD can improve the consistency between diagnostic results and a gold standard, particularly for junior physicians whose performance approached the level of more senior physicians. Speci city improved signi cantly for the JD group (p<0.05) and their diagnostic consistency improved from moderate to good. Applying CAD also improved the sensitivity, speci city, and consistency for the SD group, though the results were not statistically signi cant.

Discussion
The diagnosis of pneumoconiosis has conventionally relied on manual interpretation, which can be subjective and inconsistent. This study made full use of the advantages offered by machine learning technology, developing a self-learning training model based on ResNet101 for the representation of pneumoconiosis in chest radiographs. ResNet101 can decompose a problem into multiple direct residual problems, using a residual vector coding scheme for image processing, without the need for additional network parameters or calculations. In this case, model training speed was increased, and classi cation accuracy was improved [20].
Pneumoconiosis is characterized in radiograph images by a diffuse distribution of low opacity objects of varying size. The shape, size, quantity, and distribution of these structures is di cult to accurately describe. As such, there is a 'semantic gap' between low-level image features and high-level medical terms [25]. This study used the entire lung eld on both sides of the image as the research object, including the diffuse structures and some surrounding tissue. There was no need to perform complex feature extraction on speci c object representations, thereby avoiding these semantic gaps. The algorithm extracted characteristic information for pneumoconiosis lesions and improved the accuracy of model classi cation. The collected data were used to train a CAD system, based on a ResNet101 model, for automated pneumoconiosis diagnosis.
Output results for 100 chest X-rays in the test group demonstrate the high diagnostic e ciency provided by the proposed CAD model. The AUC, sensitivity, speci city, and consistency were high when compared with a gold standard. In a future study, we will pursue various options for increasing the intuitive nature of predicted results [26]. For example, heat maps have been used in previous studies to visualize a portion of the test image [9,12]. Pixels with larger values indicate a greater contribution to the result, which could help physicians to better understand the conclusions of automated diagnoses.
Our results showed that thproposed CAD system improved the overall sensitivity, speci city, and consistency of pneumoconiosis diagnoses for physicians with varying levels of experience. Speci city increased for all three physicians in the JD group and sensitivity improved for two of the three. Sensitivity measures the probability of correctly diagnosing positive cases in a test group, while speci city measures the probability of correctly diagnosing negative cases. High sensitivity could identify patients with pneumoconiosis, providing for earlier treatment and an improved quality of life. High speci city could screen out patients who do not require further examination, thereby reducing caseloads. Junior physicians are generally more prone to misdiagnosis due to a lack of experience. However, the proposed CAD system effectively improved both sensitivity and speci city for the JD group, producing an accuracy that was comparable to more experienced clinicians. Although the senior physicians were more accurate before including CAD, the proposed system increased their sensitivity and speci city, though the difference was not statistically signi cant.
The presented study does have certain limitations. For example, although the proposed CAD system, based on a deep residual neural network (ResNet101), can achieve high diagnostic accuracy and consistency for independent diagnoses, pneumoconiosis requires a comprehensive diagnosis. It relies not only on chest X-rays, but also on professional history, epidemiology, and clinical manifestations. The inclusion of qualitative factors like these in computer-based decision making is a topic that requires further research. In addition, the number of patients included in this study was small and the results exhibited some deviation. In a future work, the amount of data will be increased for further analysis. Finally, this study is only a preliminary investigation of whether CAD can diagnose pneumoconiosis.
Research on different stages of pneumoconiosis should be conducted in the future.

Conclusion
In summary, the results presented in this study demonstrate that CAD can effectively improve the sensitivity, speci city, and consistency of pneumoconiosis diagnoses, particularly for junior physicians. As such, the proposed model could be a powerful new tool for reducing diagnostic subjectivity and interclinician variability.