Improving Performance, Robustness, and Fairness of Radiographic AI Models with Finely-Controllable Synthetic Data.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Deep Learning

Background:

• Clinical deep learning models struggle with performance and fairness across diverse patient populations.
• Synthetic data generation offers a solution to limited dataset scale and diversity.

Purpose of the Study:

• Introduce RoentGen-v2, a text-to-image diffusion model for generating chest radiographs with controlled demographic attributes.
• Develop and evaluate a novel training strategy using synthetic data for improved downstream disease classification.

Main Methods:

• Utilized RoentGen-v2 to create a large, demographically balanced synthetic dataset (>565,000 images).
• Implemented a two-stage training: supervised pretraining on synthetic data, followed by fine-tuning on real data.
• Evaluated performance, generalization, and fairness on over 137,000 real chest radiographs from five institutions.

Main Results:

• Synthetic pretraining increased downstream model accuracy by 6.5%, significantly outperforming naive data combination (2.7%).
• Reduced the underdiagnosis fairness gap by 19.3% across sex, age, and race/ethnicity subgroups.
• Demonstrated improved generalization to out-of-distribution settings and label efficiency.

Conclusions:

• Demographically controllable synthetic imaging data can advance equitable and generalizable medical AI.
• The proposed data-centric approach with multi-stage training reduces reliance on extensive annotated real data.
• Open-sourced code, models, and dataset to facilitate further research.