Balanced Convolutional Neural Networks for Pneumoconiosis Detection.

Area of Science:

• Occupational Medicine
• Medical Imaging
• Artificial Intelligence

Background:

• Pneumoconiosis is a prevalent and costly occupational disease in China.
• Current diagnosis relies heavily on radiologist experience, hindering large-scale, rapid detection.
• Machine learning, particularly artificial neural networks (ANNs), shows promise for computer-aided detection (CAD).

Purpose of the Study:

• To develop and validate a deep learning model for accurate pneumoconiosis detection from radiographs.
• To address challenges of imbalanced datasets and lack of interpretability in existing ANN models.
• To provide a reliable diagnostic reference for surgeons through result visualization.

Main Methods:

• Established a comprehensive dataset of pneumoconiosis radiographs with positive and negative samples.
• Compared deep convolutional neural network (CNN) approaches for pneumoconiosis detection.
• Employed balanced training strategies to improve model recall.
• Utilized visualization techniques to explain model predictions.

Main Results:

• Achieved a high diagnostic accuracy of 88.6% for pneumoconiosis detection.
• Demonstrated the effectiveness of balanced training in improving detection performance.
• Successfully visualized suspected opacities, aiding in diagnostic interpretation.

Conclusions:

• Deep convolutional diagnosis approaches, with balanced training, offer a viable solution for pneumoconiosis detection.
• The developed model provides high accuracy and interpretability, facilitating clinical adoption.
• This AI-driven approach can assist surgeons in diagnosing pneumoconiosis more efficiently and reliably.