A comprehensive bedside chest radiography dataset with structured, itemized and graded radiologic reports

Area of Science:

• Radiology
• Artificial Intelligence
• Medical Informatics

Background:

• Automated analysis of bedside chest radiographs is hindered by a lack of large, annotated datasets.
• Standardized severity grading for radiographic findings is crucial for developing reliable AI models.

Purpose of the Study:

• To introduce TAIX-Ray, a comprehensive dataset of bedside chest radiographs with expert annotations.
• To facilitate the development of AI models for automated pathology detection and severity assessment in intensive care settings.

Main Methods:

• Collected 215,381 bedside chest radiographs from 47,724 ICU patients over 14 years.
• 134 radiologists provided structured reports assessing eight common pathological findings using a five-point severity scale.
• Dataset includes radiographs, structured reports, patient demographics, and temporal metadata.

Main Results:

• The TAIX-Ray dataset comprises 215,381 annotated chest radiographs.
• Structured reports cover cardiomegaly, pulmonary congestion, pleural effusion, pulmonary opacities, and atelectasis with severity grading.
• A baseline transformer model and implementation code are provided for reproducible research.

Conclusions:

• TAIX-Ray is a valuable resource for advancing AI in critical care radiology.
• The dataset enables research into automated detection and severity assessment of chest radiographic findings.
• Facilitates reproducible benchmarking and immediate research adoption.