Medical imaging data are heterogeneous, spanning different modalities, anatomies, and acquisition conditions. To handle this variety, researchers often train domain-specialized diffusion models, or "experts." But having a set of experts is not enough: during sampling, their predictions must be combined in some way, and the choice of combination strategy is not settled by specialization alone.
An arXiv preprint (2610.16082) addresses this by learning a global router that decides how to weight or select expert predictions. The abstract emphasizes that expert specialization alone does not determine the combination, and frames the router as "regime-aware," suggesting it adapts to different data regimes. The full details of the router architecture and training procedure are not available in the truncated abstract.
This line of work could make multi-domain medical image generation more practical by replacing hand-crafted combination rules with a learned, data-driven policy. However, because the abstract is cut off, the preprint's experimental results and comparisons remain unknown from this source alone.