This 1.3B-Parameter AI Generates Chest X-Rays So Real, Doctors Can't Tell the Difference
Radiographs so realistic, clinical experts can't tell them from real ones.
A team of researchers from multiple institutions (including Fabio De Sousa Ribeiro, Emma Stanley, and Ben Glocker) has introduced the first billion-parameter generative foundation model for chest radiography, trained from scratch. Called a Rectified Flow Transformer, the model boasts 1.3B parameters and was trained on 1.6 trillion tokens derived from a curated dataset of 1.2 million chest radiographs with expert-guided metadata. This scale makes it the largest specialist generative model for medical imaging to date.
The model enables controllable synthesis and editing of chest X-rays across multiple dimensions: demographic subgroups (age, sex), acquisition settings (views like PA or lateral), and a dozen pathologies. Crucially, the generated images are clinically indistinguishable from real radiographs, according to expert evaluation. This capability promises to diversify training datasets for diagnostic AI, reduce biases against underrepresented populations, and provide a robust stress-test for existing diagnostic models—addressing a key barrier to real-world deployment of medical AI.
- 1.3 billion parameters trained on 1.2 million chest radiographs for 1.6 trillion tokens.
- Supports controllable generation across demographic subgroups, acquisition views, and 12 pathologies.
- Generated radiographs are indistinguishable from real ones to clinical experts.
- Aims to improve generalisation of diagnostic AI across patient subpopulations and hospital settings.
Why It Matters
Synthetic yet realistic chest X-rays can debias medical AI and stress-test diagnostic models before clinical deployment.