Open Access

ARTICLE

Digital Radiography-Based Pneumoconiosis Diagnosis via Vision Transformer Networks

Qingpeng Wei^1,#, Wenai Song^1,#, Lizhen Fu¹, Yi Lei², Qing Wang^2,*

1 School of Software, North University of China, Taiyuan, 030051, China
2 Pharmacovigilance Research Center for Information Technology and Data Science, Cross-Strait Tsinghua Research Institute, Xiamen, 361009, China

* Corresponding Author: Qing Wang. Email:
# These authors contributed equally to this work

Journal on Artificial Intelligence 2025, 7, 39-53. https://doi.org/10.32604/jai.2025.063188

Received 08 January 2025; Accepted 17 March 2025; Issue published 23 April 2025

Abstract

Pneumoconiosis, a prevalent occupational lung disease characterized by fibrosis and impaired lung function, necessitates early and accurate diagnosis to prevent further progression and ensure timely clinical intervention. This study investigates the potential application of the Vision Transformer (ViT) deep learning model for automated pneumoconiosis classification using digital radiography (DR) images. We utilized digital X-ray images from 934 suspected pneumoconiosis patients. A U-Net model was applied for lung segmentation, followed by Canny edge detection to divide the lungs into six anatomical regions. The segmented images were augmented and used to train the ViT model. Model component evaluations were conducted to assess the impact of lung segmentation and data augmentation. The ViT model achieved an accuracy of 78.8% and a specificity of 89.2% in pneumoconiosis classification. Furthermore, the region-based classification method utilizing detailed lung segmentations substantially improved diagnostic precision and was closely aligned with established clinical evaluation standards in pneumoconiosis assessment. This study demonstrates the clinical effectiveness of the ViT model in pneumoconiosis classification and highlights the importance of detailed lung region segmentation for structured clinical assessment. The findings strongly suggest that deep learning approaches, especially region-specific methods, can significantly enhance diagnostic accuracy, providing clinicians with a systematic, reliable framework for medical image analysis.

---

Keywords

---