Researchers have developed CONFLUX, a novel latent diffusion model designed for synthesizing 3D chest CT scans with specific clinical attributes. The model utilizes a 3D variational autoencoder for compression and a rectified-flow transformer for generation, conditioned on detailed radiological metadata. An additional reinforcement learning post-training stage, using group-relative policy optimization, significantly enhances the model's ability to reliably generate requested clinical findings, reducing a substantial portion of the shortfall compared to real scans. The project also includes the release of a large dataset of synthetic chest CTs and an interactive demo. AI
IMPACT Enables more precise and controllable synthesis of medical imaging data for research and training purposes.
RANK_REASON Publication of a research paper detailing a new generative model for medical imaging.
Read on Hugging Face Daily Papers →
- 3D chest CT
- Max Van Puyvelde
- 3D Chest-CT Synthesis
- 3D variational autoencoder
- GenerateCT
- Group Relative Policy Optimization
- Hugging Face
- Latent diffusion model
- MAISI
- Rectified Flow transformer
- RL Post-Training
AI-generated summary · Google Gemini · from 3 sources. How we write summaries →
