Source-linked AI summary
CONFLUX: A Latent Diffusion Model for 3D Chest-CT Synthesis with RL Post-Training
Max Van Puyvelde, Halil Ibrahim Gulluk, Wim Van Criekinge, Olivier Gevaert
TL;DR
Controllable 3D chest-CT synthesis must balance volumetric fidelity with faithful realization of requested clinical attributes. CONFLUX combines latent rectified-flow generation with reinforcement-learning post-training, achieving tri-planar FID 32.3 versus 74.6 for MAISI and recovering 47% of the base-to-real faithfulness gap.
Problem
Controllable chest-CT synthesis requires samples that are simultaneously high-fidelity, natively 3D, and faithful to structured clinical conditioning.
Method
CONFLUX uses a 3D latent rectified-flow model conditioned on radiological metadata, followed by GRPO post-training optimized for conditioning faithfulness.
Results
Tri-planar FID is 32.3 versus 74.6 for MAISI, while independent evaluation shows post-training recovers 47% of the base-to-real faithfulness gap.
Takeaways & Limitations
CONFLUX provides controllable natively 3D chest-CT synthesis and releases the model with a ∼200,000-volume metadata-conditioned synthetic dataset.
Abstract
from arXiv · showhide
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.
1 Introduction
CONFLUX addresses the challenge of generating high-fidelity, controllable 3D chest CT by combining latent rectified-flow synthesis with structured radiological conditioning. It further improves conditioning faithfulness through GRPO post-training and releases the model with a large synthetic dataset.
- Motivation: 3D chest CT synthesis requires samples that are simultaneously high-fidelity and controllable for specified clinical attributes.Chest CT is volumetric, high-resolution, and described by structured radiological metadata, making it an especially difficult target.
- Model: CONFLUX is a natively 3D latent rectified-flow model conditioned on abnormality findings, sex, age, and reconstruction kernel through adaptive layer normalization.Its architecture combines a 3D convolutional VAE with a single-stream rectified-flow transformer.
- Motivation: Flow matching optimizes aggregate realism but does not ensure that each generated volume realizes its requested findings.A realistic sample may therefore under-express or omit a requested abnormality.
- RL post-training: GRPO post-training fine-tunes the flow model with a faithfulness reward, improving the reliability of requested findings and validating the gain with an independent held-out judge.The authors describe this as the first GRPO post-training of a 3D medical flow model, to their knowledge.
- Release: ∼200,000-volume synthetic chest-CT dataset released with CONFLUX carries conditioning metadata spanning a wide variety of findings.The dataset is described as controllable and faithfulness-optimized.
2 Related Work
Prior work combines latent diffusion with diffusion-transformer architectures, rectified-flow objectives, and adaptive-normalization conditioning, including applications to 3D medical imaging. Reinforcement-learning methods further fine-tune diffusion or flow-matching generators with non-differentiable rewards, including GRPO-based approaches.
- Latent diffusion compresses volumes with an autoencoder and models diffusion or flow dynamics in the latent representation.
- Diffusion transformers trained with rectified-flow or flow-matching objectives and adaptive-normalization conditioning have become common in image generation.
- Conditional latent diffusion has been applied to brain MRI, whole-body CT, and chest CT; MAISI is the reference competitor for synthesis quality.
- Policy-gradient fine-tuning treats denoising as a multi-step decision process and optimizes diffusion models against non-differentiable rewards.
- GRPO replaces a learned value function with a group-relative advantage, while Flow-GRPO and DanceGRPO extend GRPO to flow-matching and visual generation settings.
3 Method
CONFLUX uses a three-stage pipeline that compresses 3D CT volumes into latents, generates them with a conditioned rectified-flow transformer, and post-trains the model for conditioning faithfulness. Its reinforcement-learning stage uses classifier-based rewards while constraining updates relative to the frozen pre-training model.
- Three-stage architecture: CONFLUX comprises a 3D VAE, a single-stream rectified-flow transformer, and a post-training stage targeting faithfulness to requested conditioning.The VAE provides the latent representation, the transformer generates within that space, and post-training improves attribute realization.
- Latent representation: The 3D convolutional VAE encodes each CT volume into a diagonal-Gaussian latent, downsampling by f=8 per spatial axis to C=16 channels before reconstruction.The decoder reconstructs the volume from the compact latent representation, reducing the cost of generation relative to voxel-space modeling.
- RL post-training: The post-training stage adapts Flow-GRPO by rewarding agreement between generated volumes and requested findings using a frozen classifier, while penalizing KL divergence from the frozen pre-RL policy.Action noise makes each sampling step probabilistic, and the reward is based on the negative weighted cross-entropy between predicted and requested conditioning.
4 Experiments
CONFLUX is evaluated on QC-filtered CT-RATE data using tri-planar distribution metrics and an independent image-space judge for conditioning faithfulness. It outperforms the reported generative baselines in synthesis quality, improves optimized finding control significantly, and releases a large metadata-paired synthetic dataset.
- Data and evaluation: The autoencoder trains on approximately 40,800 QC-passing volumes, while the flow model uses 18,417 one-scan-per-patient volumes resized to 216×176×200.Volumes are lung-box cropped, intensity-scaled to approximately [−1, 1], and filtered by lung fraction, left–right balance, voxel spacing, acquisition direction, and reconstruction kernel.
- Synthesis quality: 32.3 tri-planar FID leads MAISI at 74.6 and GenerateCT at 145.4, approaching the 22.6 VAE-reconstruction ceiling.On the resize-robust axial plane, CONFLUX also leads with 24.7 versus 55.6 for MAISI and 70.2 for GenerateCT.
- Synthesis quality: The model is 3.5× denser and covers 6× more of the real manifold than MAISI, while diversity remains near the real floor at 0.51 versus 0.41.All methods remain near the real diversity floor, where collapse would approach 1; reported FID and density/coverage gaps are statistically significant.
- Reward modeling: The frozen latent reward classifier reaches macro AUROC 0.793 over 18 findings on real validation data and outperforms CT-CLIP zero-shot on 17 of 18 findings.Its performance bounds the GRPO reward signal, while operating on the 8×-compressed latent indicates that the tokenizer preserves diagnostic information.
- Conditioning faithfulness: Both conditioning-faithfulness gains are significant, with p=0.042 and 0.014, and the improvement is specific to optimized findings rather than sex, age, or kernel.Faithfulness is judged by an independent image-space classifier on generated samples, avoiding evaluation with the classifier optimized by GRPO.
- Release: The release includes a approximately 200,000-volume synthetic chest-CT dataset with conditioning vectors spanning diverse findings, plus trained model checkpoints.The dataset conditions volumes according to real CT-RATE metadata proportions and is intended for cohort augmentation and conditional study design.
5 Conclusion
CONFLUX is a controllable, natively 3D latent rectified-flow model for chest CT that combines strong 3D quality with direct clinical-attribute control. RL post-training improves conditioning faithfulness, and the authors release the model alongside a ∼200,000-volume synthetic dataset.
- 5 Conclusion: CONFLUX is a controllable, natively 3D latent rectified-flow chest-CT model conditioned on structured radiological metadata.It provides direct control over clinical attributes while achieving strong quality relative to 3D baselines.
- 5 Conclusion: RL post-training raises conditioning faithfulness and is presented as the first GRPO post-training of a 3D medical flow model.
- 5 Conclusion: The authors release CONFLUX and a ∼200,000-volume synthetic chest-CT dataset.
A Training Parameters
Training uses a three-stage configuration with cached latent moments and frozen reward-classifier and reference-policy components. The post-training SDE uses the FLUX time-shifted Euler grid, while per-channel latent standard deviations range from 1.27 to 2.48.
- Training configuration: Training is organized into three stages.The configuration is described explicitly as covering three stages.
- Training configuration: Latent moments are cached during training.
- Post-training: The post-training SDE uses the FLUX time-shifted Euler grid.
- Post-training: The reward classifier and reference policy remain frozen.
- Latent representation: Per-channel latent standard deviations range from 1.27 to 2.48.
B Per-finding reward-classifier accuracy
This section reports per-finding AUROC on the CT-RATE real-validation set across all 18 findings, comparing the latent classifier with CT-CLIP variants and CT-Net.
- B Per-finding reward-classifier accuracy: Per-finding AUROC is evaluated on CT-RATE real validation for all 18 findings.The findings are sorted by the latent classifier’s AUROC, with the best method per row highlighted.
- B Per-finding reward-classifier accuracy: The comparison includes the latent classifier, CT-CLIP variants, and CT-Net [7].The table identifies the best-performing method separately for each finding.
C Additional samples and conditioning
CONFLUX generates diverse synthetic chest-CT volumes with coherent coronal and axial views. The accompanying conditioning metadata spans sex, age, reconstruction kernel, and varied combinations of abnormalities, including patients with no conditioned abnormalities.
- Additional samples: Forty randomly sampled volumes show diverse body habitus, anatomy, and abnormalities with cross-plane 3D coherence.Each volume is presented through matching coronal and axial mid-slices in a lung window.
- Conditioning metadata: Table 6 records per-volume conditioning across sex, age, kernel, and multiple abnormality combinations, including a case with none.Examples include arterial wall calcification, lymphadenopathy, lung nodule, pulmonary fibrotic sequela, and mosaic attenuation pattern.
- Additional samples: Two further sets of 40 randomly sampled volumes extend the examples with paired coronal and axial lung-window views.Figures 4 and 5 provide conditioning metadata in Tables 7 and 8.
- Conditioning metadata: Tables 7 and 8 show broad conditioning combinations spanning demographic attributes, reconstruction kernels, and cardiopulmonary abnormalities.The listed cases include both no-abnormality examples and combinations involving emphysema, lung nodules, lymphadenopathy, consolidation, and pleural or pericardial effusion.