arxiv:2607.02998

CONFLUX: A Latent Diusion Model for 3D Chest-CT Synthesis with RL Post-Training

Published on Jul 3

· Submitted by

Authors:

Abstract

A 3D latent diffusion model for chest CT generation that achieves high-fidelity results while enabling control over clinical attributes through metadata conditioning and reinforcement learning refinement.

Generated by Qwen/Qwen2.5-Coder-32B-Instruct

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.

View arXiv page View PDF Project page Add to collection

Community

mxvp

Paper submitter about 5 hours ago

CONFLUX is a conditional rectified-flow model for full-volume 3D chest-CT synthesis, with GRPO reinforcement-learning post-training that sharpens control over the requested clinical findings.

Flow matching trains the generator to match the data distribution but never checks that an individual sample realizes its requested attributes; the GRPO stage optimizes that agreement directly — to our knowledge, the first GRPO post-training of a 3D medical flow model.

Full 216×176×200 volumes from a 42-dimensional conditioning vector: 18 CT-RATE abnormality findings, sex, age, and reconstruction kernel.
Outperforms strong 3D baselines (MAISI, GenerateCT) on distribution-level quality.
GRPO post-training measurably improves how reliably the requested findings appear, verified by an independent held-out judge — recovering 47% of the gap to real scans.
Release of ~200,000 labeled synthetic chest CTs for cohort augmentation and controlled study design at a scale unavailable in real corpora.
Interactive demo for generating volumes from arbitrary clinical profiles.

📄 Paper: https://arxiv.org/abs/2607.02998
🤗 Model: https://huggingface.co/gevaertlab/conflux
🤗 Dataset (200k): https://huggingface.co/datasets/gevaertlab/conflux-chest-ct
🩻 Demo: https://huggingface.co/spaces/mxvp/conflux-chest-ct-demo

Feel free to reach out!