Probabilistic imaging
Reconstructing images, and understanding their uncertainty.
PET image reconstruction with generative flow models, bringing prior and posterior modelling to medical imaging. Published at IEEE ISBI 2026.
Problem
Medical image reconstruction is not only about a sharp image; it is also about whether the estimate can be trusted. Framing PET reconstruction as Bayesian inference makes it possible to study the posterior distribution and its uncertainty, not just the reconstruction.
Key decisions
- Probabilistic formulation
- Built a deep probabilistic imaging pipeline around the prior and posterior distributions of tracer activity.
- Learned flow priors
- Studied normalizing flows such as RealNVP and Glow for medical image reconstruction.
- Scalable experiments
- Trained in parallel on multiple GPUs on a Slurm cluster to support experiments on reconstruction quality and uncertainty quantification.
Outcome
The paper Bayesian PET Reconstruction with Learned Normalizing Flow Priors was published at IEEE ISBI 2026 (8–11 April 2026). Authors: Hengjia Ran, Yuxuan Lin, Huafeng Liu, Bo Zhao; I am the second author.
Related notes (Chinese)