BRAX, Brazilian labeled chest x-ray dataset

Area of Science:

• Radiology and Medical Imaging
• Artificial Intelligence in Healthcare
• Machine Learning for Medical Diagnosis

Background:

• Chest radiography interpretation requires specialized expertise.
• Machine learning (ML) offers potential for automated medical image analysis.
• Development of large, labeled datasets is crucial for training and validating ML models in radiology.

Purpose of the Study:

• To introduce the Brazilian labeled chest x-ray dataset (BRAX).
• To provide a resource for researchers validating ML models for chest radiograph analysis.
• To facilitate the advancement of automated diagnostic tools in medical imaging.

Main Methods:

• Compiled a dataset of 24,959 chest radiography studies (40,967 images) from a Brazilian hospital.
• Images were verified by trained radiologists and de-identified for patient privacy.
• Utilized Natural Language Processing (NLP) to extract 14 distinct labels from free-text radiology reports in Brazilian Portuguese.

Main Results:

• Created BRAX, a large-scale, automatically labeled chest x-ray dataset.
• Dataset includes 40,967 verified chest radiograph images.
• 14 labels derived from clinical reports are available for ML model training and validation.

Conclusions:

• The BRAX dataset serves as a valuable resource for the ML research community.
• Facilitates the development and validation of AI tools for chest x-ray interpretation.
• Supports advancements in automated medical image analysis and diagnostic support systems.