A new AI model can generate synthetic 3D chest CT scans conditioned on specific clinical findings — and it ships with a public dataset of roughly 200,000 examples.
CONFLUX is a latent diffusion model built around a 3D variational autoencoder and a rectified-flow transformer. Researchers condition generation on 18 abnormality types plus patient sex, age, and scan reconstruction kernel. On the standard tri-planar Frechet distance benchmark, it scores 32.3 against 74.6 for MAISI, the previous leading volumetric baseline — a meaningful gap. The team then added a reinforcement-learning post-training stage using group-relative policy optimization, rewarding the model when a classifier could correctly recover the requested findings from generated scans. Judged by a separate, independent classifier, that stage removed 47% of the reliability shortfall compared to real scans.
The shortage of labeled medical imaging data is one of the harder bottlenecks in clinical AI — hospitals are reluctant to share scans, and annotating them is slow and expensive. A model that synthesizes high-fidelity, clinically conditioned volumes on demand could let researchers train diagnostic systems without touching a single real patient file. The RL fine-tuning step is the less obvious contribution: it treats clinical attribute fidelity as an optimizable objective rather than a training-data problem.
Synthetic medical data has a trust problem — reviewers and regulators are skeptical of models trained on it, and rightly so, since the gap between synthetic and real reliability has historically been wide. Cutting that gap by 47% is progress, but it also means 53% of the shortfall remains.