Medical image restoration (MedIR) seeks to handle a wide range of imaging modalities and degradation types with a single universal model. According to a new paper posted on Hugging Face, existing approaches have concentrated mainly on modeling the heterogeneity between tasks — such as differing data distributions and degradation patterns — while largely overlooking the homogeneity inherent in medical images.
The authors argue that widely shared anatomical structures, both within and across modalities, can be exploited to ease model training and improve generalization. To that end, they propose UniH3, a framework that unifies hierarchical homogeneity and heterogeneity for all-in-one medical image restoration.
UniH3 introduces a Hierarchical Homogeneity Memory (H2M) module that progressively distills intra- and inter-task homogeneity priors from high-quality images during training, then adaptively retrieves the priors most relevant to each input. These retrieved priors are injected into the restoration pipeline through an efficient Homogeneity-Guided Attention (HGA) mechanism.
To address heterogeneity, the framework adds a Hierarchical Heterogeneity Balancer (H2B) that mitigates both inter- and intra-task conflicts during optimization, supporting balanced multi-task learning.
Experiments on two large-scale benchmarks, MedIR-2D-500K and MedIR-3D-3K, show that UniH3 reaches state-of-the-art performance on both all-in-one and single-task medical image restoration, according to the paper. The authors say they hope the work establishes a strong benchmark and advances general-purpose medical image restoration models. Code is available on GitHub.