RUOK: Recuration of a Public Dataset Utilized to Optimize Knowledge for Multi-label Chest X-ray Disease Screening.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Public Health

Background:

• Rural Thai hospitals face radiologist shortages, impacting diagnostic capabilities.
• Existing chest X-ray datasets lack disease-specific labels, hindering AI model training for prevalent conditions.
• The RUOK system was developed to address these screening protocol challenges.

Purpose of the Study:

• To reclassify public chest X-ray datasets with labels for prevalent diseases in Thailand.
• To optimize a data-efficient image transformers model for disease screening using the enhanced dataset.
• To evaluate the model's performance in classifying common thoracic diseases.

Main Methods:

• Collaborated with expert radiologists to reclassify public chest X-ray (CXR) dataset labels into eight disease categories.
• Trained a data-efficient image transformers model using the reclassified dataset.
• Evaluated model performance using Area Under the Receiver Operating Characteristics (AUROC) on a test set.

Main Results:

• Achieved high AUROC scores for several disease classifications, including 93.06% for 'No finding' and 90.46% for 'Abnormal heart, great vessels, and mediastinum'.
• Demonstrated promising results for 'Suspected active tuberculosis' (89.21%) and 'COVID-19' (75.13%).
• AUROC scores ranged from 66.88% to 93.06% across the eight disease categories.

Conclusions:

• The reclassification of CXR datasets and subsequent model optimization significantly enhance AI capabilities for common disease screening.
• The developed model shows potential for improving diagnostic accuracy and healthcare outcomes in resource-limited settings with radiologist shortages.
• This approach offers a viable solution for advancing disease classification and supporting public health initiatives in Thailand and similar regions.