UniH^3: Unifying Hierarchical Homogeneity and Heterogeneity for All-in-One Medical Image Restoration
While most all-in-one medical image restoration methods focus on handling heterogeneous data distributions and degradation types, they often neglect the anatomical consistencies shared across modalities. The proposed UniH^3 framework addresses this balance by introducing a Hierarchical Homogeneity Memory to extract and inject shared structural priors, coupled with a balancer module to mitigate optimization conflicts across varied tasks. Benchmarked on MedIR-2D-500K and MedIR-3D-3K, the unified model achieves consistent restoration performance across diverse single-task and multi-task scenarios.
Why it's worth reading
By leveraging shared anatomical priors alongside task conflict balancing, UniH^3 demonstrates how an all-in-one model can resolve cross-modality tensions in medical image restoration.