Max Van Puyvelde, H. Ibrahim Gulluk, Wim Van Criekinge, Olivier Gevaert
4 min
Abstract
Controllable generative models of 3D medical images can synthesize volumes with specified clinical attributes, but this demands samples that are simultaneously high-fidelity, natively 3D, and faithful to the requested conditioning. We present CONFLUX, a latent diffusion model for chest computed tomography (CT): a 3D variational autoencoder compresses each volume, and a rectified-flow transformer generates in the latent space. Generation is conditioned on structured radiological metadata (18 abnormality findings, sex, age, and reconstruction kernel) through adaptive layer normalization. The model leads strong volumetric baselines on tri-planar Frechet distance (FID 32.3 vs. 74.6 for MAISI) while exposing direct control over clinical attributes. To strengthen that control we add an online reinforcement-learning post-training stage (group-relative policy optimization) that rewards how reliably a classifier recovers the requested findings from each generated volume. Judged by a separate, independent classifier, post-training removes 47% of the shortfall relative to real-scan reliability. We release the model and a ~200k synthetic chest-CT dataset with conditioning metadata spanning a wide variety of clinical findings.
Sam: But doesn't this whole system depend on how good the judge is? If the judge misses a subtle finding, the model never learns it's doing something wrong.
Alex: That is the central limitation, and the authors are direct about it. The model's accuracy is strictly bounded by the judge's accuracy. If the judge can't reliably detect a subtle feature, the reward signal becomes unreliable, and the model won't learn that detail. It's only as capable as its teacher.
Sam: So did this second stage actually make a measurable difference?
Alex: It did. By adding it, they were able to close nearly half the gap between their model's initial performance and the reliability of real-world medical scans. In a field where the margin for error is very small, that's a meaningful improvement.
Sam: And they're releasing the model alongside a large dataset of synthetic scans—which could be a real resource for researchers who can't access real patient data.
Alex: That's one of the more practically significant parts of the work. The hope is that other researchers can use these synthetic scans to train their own diagnostic tools, without needing to navigate the legal and ethical complexities of working with real patient records. The authors also suggest that future versions might pair these images with automatically generated diagnostic reports—so the system could produce both the scan and a written description of what it shows.
Sam: It sounds like a careful, incremental step in a field where accuracy is non-negotiable.
Alex: That's exactly the right framing. The value isn't in making images that look impressive—it's in making images that are genuinely useful. And this work suggests that reinforcement learning, with the right judge, is a credible path toward that goal. Thanks for listening to ResearchPod.