Breaking medical data sharing boundaries by using synthesized radiographs

Area of Science:

• Medical imaging analysis
• Artificial intelligence in healthcare
• Computer vision applications

Background:

• Computer vision (CV) holds significant potential for revolutionizing medical diagnostics.
• Current CV algorithms are hindered by small, incomplete, and heterogeneous training datasets.
• Data privacy concerns impede the sharing of medical data, further limiting algorithm development.

Purpose of the Study:

• To address limitations in medical CV training data using generative models (GMs).
• To produce high-resolution, privacy-preserving synthetic radiographs for AI training.
• To enhance the performance and applicability of CV diagnostic tools in medicine.

Main Methods:

• Utilized generative models (GMs) to synthesize high-resolution, anonymized radiographic images.
• Conducted blinded evaluations comparing synthetic and real radiographs with CV and radiology experts.
• Integrated pooled GM outputs to train CV algorithms on limited datasets.
• Incorporated federated learning strategies for collaborative GM training across institutions.

Main Results:

• Synthetic radiographs demonstrated high similarity to real images, validated by expert assessment.
• Combined GM outputs significantly improved CV algorithm performance on smaller datasets.
• Synthetic data integration compensated for underrepresented disease entities in training sets.
• Federated learning enabled data-scarce hospitals to contribute to and benefit from GM training.

Conclusions:

• Generative models offer a viable solution to data scarcity and privacy issues in medical CV.
• Synthetic data generated by GMs can effectively augment real datasets, improving AI diagnostic accuracy.
• Federated learning combined with GMs democratizes access to advanced AI training for all healthcare institutions.