Towards reconstructing experimental sparse-view X-ray CT data with diffusion models
This work addresses the problem of translating synthetic-trained diffusion models to real-world medical imaging data for researchers, highlighting incremental improvements in handling domain and forward model mismatches.
The study tackled the challenge of applying diffusion models to reconstruct experimental sparse-view X-ray CT data, revealing that domain shift causes model collapse and hallucinations, while forward model mismatch leads to artifacts, but these issues can be mitigated with diverse priors and annealed likelihood schedules.
Diffusion-based image generators are promising priors for ill-posed inverse problems like sparse-view X-ray Computed Tomography (CT). As most studies consider synthetic data, it is not clear whether training data mismatch (``domain shift'') or forward model mismatch complicate their successful application to experimental data. We measured CT data from a physical phantom resembling the synthetic Shepp-Logan phantom and trained diffusion priors on synthetic image data sets with different degrees of domain shift towards it. Then, we employed the priors in a Decomposed Diffusion Sampling scheme on sparse-view CT data sets with increasing difficulty leading to the experimental data. Our results reveal that domain shift plays a nuanced role: while severe mismatch causes model collapse and hallucinations, diverse priors outperform well-matched but narrow priors. Forward model mismatch pulls the image samples away from the prior manifold, which causes artifacts but can be mitigated with annealed likelihood schedules that also increase computational efficiency. Overall, we demonstrate that performance gains do not immediately translate from synthetic to experimental data, and future development must validate against real-world benchmarks.