Automated identification of the preclinical stage of coal workers' pneumoconiosis from digital chest radiography using three-stage cascaded deep learning model. (May 2023)

• Record Type: Journal Article
• Authors: Wang, Yan; Cui, Fengtao; Ding, Xinping; Yao, Yulong; Li, Guoming; Gui, Genjuan; Shen, Fuhai; Li, Bing
• Is Part Of: Biomedical signal processing and control. Volume 83(2023)
• Publication Date: 2023-05
• Subjects: Signal processing -- Periodicals; Biomedical engineering -- Periodicals; Signal Processing, Computer-Assisted -- Periodicals; Image Processing, Computer-Assisted -- Periodicals; Biomedical Engineering -- Periodicals; 610.28
• DOI: 10.1016/j.bspc.2023.104607
• Languages: English
• ISSNs: 1746-8094

Abstract

Highlights: The present study demonstrated the first attempt at the identification of preclinical coal workers' pneumoconiosis. The proposed three-cascaded approach can improve the diagnostic capabilities compared with classical convolutional neural networks. The created cascaded learning model can screen out the preclinical pneumoconiosis in a large number of coal workers' chest radiographs. Abstract: Objective: Coal workers' pneumoconiosis (CWP) is a broad and serious occupational disease caused by inhaling coal dust, which can cause permanent physical injury. There is no effective treatment currently. Screening of the preclinical stage of CWP can earlier detect the risk of CWP. However, it is laborious for radiologists to screen for CWP from a large number of chest radiographs. Missed diagnoses and misdiagnoses often occur. To address this problem, a novel three-stage cascaded learning model for screening the preclinical stage of CWP on digital chest radiography (DR) was proposed in this paper. Methods: 1447 DRs of coal workers from two hospitals were used in the study. At the first stage, a YOLOv2 network was trained for detecting lung boxes on DR images. At the second stage, six convolutional neural network (CNN) models were trained to identify the preclinical stage of CWP. At the third stage, an ensemble learning (EL) model based on the soft voting mechanism was implemented to integrate the outputs of the six CNN models. Results: Test results on the test set generated the area under the receiver characteristic operating curve (AUC), accuracy, sensitivity and specificity of 0.931, 84.7%, 75.0% and 95.7%, respectively. Conclusion: The proposed three-stage cascaded learning model could effectively screen coal workers at the preclinical stage. Significance: This is the first study for identifying the preclinical stage of pneumoconiosis on DR images, which could facilitate the secondary prevention of pneumoconiosis.