Classification Comparison of Models Trained on Real Versus Synthetic Chest X-Ray Data

Dawood Rehman's 9.5 KB spreadsheet compares classification performance for models trained on real and synthetic medical imaging datasets. The data, last updated in April 2026, contrasts results from CheXpert, MIMIC-CXR, ChestX-ray14, and a synthetic dataset using ResNet50 and Swin-T encoders. Bold formatting indicates where synthetic data led to improved model performance.

Use Cases

• Benchmarking model performance based on training data source (real vs. synthetic) mentioned in the description
• Analyzing fairness and bias mitigation in medical AI models based on the platform tags
• Comparing the effectiveness of different neural network encoders (ResNet50 vs. Swin-T) for chest X-ray classification
• Studying the impact of synthetic data on reducing spurious shortcuts in model training as suggested by platform tags

Strengths

• Directly compares performance across four distinct datasets (CheXpert, MIMIC-CXR, ChestX-ray14, Synthetic)
• Explicitly marks performance improvements with bold formatting for easy visual analysis
• File size is 9.5 KB, ensuring quick download and processing
• Released under a permissive CC-BY-4.0 license for broad reuse

Limitations

• Row count is unknown, which may limit suitability assessment for large-scale analysis
• Column-level documentation is absent; field semantics must be inferred after download
• The 9.5 KB size suggests a very limited scope, likely containing summary metrics rather than raw data

Provenance

Source

Dawood Rehman via figshare

Collection Method

Likely contains experimental results from model training and evaluation comparisons.

Time Range

Dataset creation or coverage timeframe is not specified.

Freshness

Last updated 2026-04-03 17:45:18; freshness should be verified

Geography

Spatial coverage is not specified.