Shadow and Light: Digitally Reconstructed Radiographs for Disease Classification

Area of Science:

• Medical Imaging
• Artificial Intelligence in Radiology
• Dataset Development

Background:

• Large-scale datasets are crucial for training robust medical AI models.
• Existing chest X-ray datasets may have limitations in view diversity and associated detailed reports.
• Synthetic data generation offers a controllable method to augment real-world medical imaging data.

Purpose of the Study:

• Introduce DRR-RATE, a novel, large-scale synthetic chest X-ray dataset.
• Facilitate research in multimodal AI applications using paired CT, X-ray, text, and labels.
• Evaluate the performance of AI models on this synthetic dataset.

Main Methods:

• Generated 50,188 frontal Digitally Reconstructed Radiographs (DRRs) from the CT-RATE dataset.
• Paired each DRR with radiology text reports and binary labels for 18 pathology classes.
• Utilized DRR generation for controllable inclusion of various imaging views.

Main Results:

• CheXnet achieved sufficient to high AUC scores for six common pathologies when trained/tested on DRR-RATE.
• CheXnet demonstrated accurate out-of-distribution pathology identification when trained on CheXpert.
• Generated DRR images effectively capture pathology features from CT scans.

Conclusions:

• DRR-RATE is a valuable resource for multimodal AI research in medical imaging.
• Synthetic DRRs can effectively represent pathologies for AI model training and validation.
• The dataset supports research into novel AI applications leveraging diverse imaging modalities and reports.