Deep learning models for radiography body-part classification and chest radiograph projection/orientation...

Area of Science:

• Radiology
• Medical Imaging
• Artificial Intelligence

Background:

• Large-scale radiographic datasets frequently contain labeling errors, compromising automated image analysis.
• Accurate metadata is crucial for reliable clinical interpretation and deep learning research in medical imaging.

Purpose of the Study:

• To develop and externally validate deep learning models for automated body part categorization and chest radiograph projection/rotation identification.
• To enhance quality control and data integrity in large, multi-institutional radiographic databases.

Main Methods:

• Retrospective collection of over 860,000 radiographs from multiple institutions and public repositories.
• Development of two models: Xp-Bodypart-Checker (seven categories) and CXp-Projection-Rotation-Checker (projection and rotation).
• External validation on diverse datasets, assessing performance using accuracy and area under the receiver operating characteristic curve (AUC).

Main Results:

• Xp-Bodypart-Checker achieved AUC values of 1.00 for most categories and 0.99 for 'Incomplete Chest'.
• CXp-Projection-Rotation-Checker demonstrated AUC values of 1.00 across all projection and rotation classifications.
• Both models showed high performance in classifying radiographs and identifying projection/rotation errors.

Conclusions:

• The developed deep learning models effectively verify image labels in large radiographic databases.
• These tools significantly improve quality control and data integrity across multiple institutions.
• Enhanced data reliability supports both clinical workflows and advancements in deep learning research.