Clinical Scorecard: Utilizing Deep Learning for Classification of Body-Part Radiographs and Chest X-ray Orientation: Insights from a Multi-Institutional Study

At a Glance

Key Highlights

• Development and external validation of two deep-learning models: Xp-Bodypart-Checker for body part classification and CXp-Projection-Rotation-Checker for chest X-ray projection and rotation detection.
• Use of large, multi-institutional, and multinational datasets including institutional data and public datasets (MURA, CheXpert, PadChest) to enhance model generalizability.
• Automated detection of mislabeled or rotated radiographs addresses critical quality control challenges in large heterogeneous radiograph repositories.

Guideline-Based Recommendations

Diagnosis

• Utilize deep-learning models to classify radiographic body parts accurately before downstream analysis.
• Apply automated projection and rotation detection to chest radiographs to ensure correct image orientation.

Management

• Incorporate automated quality control steps in radiograph data preprocessing pipelines to detect and correct labeling errors.
• Use multi-institutional datasets for training to improve model robustness and generalizability.

Monitoring & Follow-up

• Regularly validate model performance across diverse clinical settings and datasets.
• Monitor for mislabeled or rotated images that may compromise deep-learning model training and diagnostic accuracy.

Risks

• Manual or semi-automated metadata entry errors can lead to mislabeled images affecting model reliability.
• Inclusion of rotated or incorrectly oriented images may degrade deep-learning analysis performance.

Patient & Prescribing Data

Adult patients undergoing radiographic imaging across multiple institutions and public datasets

Automated classification models improve data quality and reliability of radiograph-based deep-learning diagnostic tools, potentially enhancing clinical decision-making.

Clinical Best Practices

• Verify and correct DICOM metadata labels with expert radiologist review to establish accurate ground truth for model training.
• Use balanced datasets with equal distribution of image rotations to train rotation detection models effectively.
• Employ multi-institutional and multinational datasets to ensure model applicability across diverse populations and imaging protocols.
• Implement automated quality control to detect mislabeled or misoriented images prior to deep-learning model training and validation.

References

• CheXpert Dataset
• PadChest Dataset
• MURA Dataset
• STARD Guidelines