UniH3 achieves state-of-the-art all-in-one medical image restoration

All-in-one medical image restoration, or MedIR, aims to handle many different restoration tasks, across different imaging modalities and different kinds of degradation, using a single universal model rather than one model per task. According to the paper's authors, existing approaches to this problem mostly model the differences between tasks, such as distinct data distributions and degradation types, and largely ignore something medical images have in common: anatomical structures that repeat within and across modalities. The authors argue that this shared structure, or homogeneity, can be used to make training easier and to improve how well a model generalizes, instead of treating every task as unrelated to the others.

To act on that idea, Zhiwen Yang and co-authors propose UniH3, short for Unifying Hierarchical Homogeneity and Heterogeneity, built around two components. The first, a Hierarchical Homogeneity Memory (H2M) module, is trained to progressively distill homogeneity priors, patterns shared within a single task and across different tasks, from high-quality images during training. At inference time it adaptively retrieves whichever priors are most relevant to the specific input, and these retrieved priors are injected into the restoration pipeline through a Homogeneity-Guided Attention (HGA) mechanism the authors describe as efficient.

The second component addresses the opposite problem: how tasks conflict rather than what they share. A Hierarchical Heterogeneity Balancer (H2B) mitigates both inter-task and intra-task conflicts that arise during optimization when a single model trains on many restoration tasks at once, keeping multi-task training balanced instead of letting some tasks dominate others.

The authors report testing UniH3 through what they call extensive experiments on two large-scale benchmarks, MedIR-2D-500K and MedIR-3D-3K, and say it achieves state-of-the-art performance both in the all-in-one, multi-task setting and on single restoration tasks on their own. The abstract does not give specific performance numbers, such as PSNR or SSIM scores, for UniH3 or for the methods it is compared against, and it does not explain what the '500K' and '3K' in the two benchmark names refer to. The authors say they hope the work establishes a strong benchmark and helps advance general-purpose medical image restoration models, and they have released the code for UniH3 on GitHub at https://github.com/Yaziwel/UniH3.

Key facts

UniH3, short for Unifying Hierarchical Homogeneity and Heterogeneity, is a new framework for all-in-one medical image restoration: handling many tasks, modalities and degradation types with one universal model instead of a model per task.
Its Hierarchical Homogeneity Memory (H2M) module distills shared anatomical priors from high-quality images during training and retrieves the most relevant ones for a given input, injecting them into restoration through a Homogeneity-Guided Attention (HGA) mechanism.
A separate Hierarchical Heterogeneity Balancer (H2B) mitigates conflicts between and within tasks during training so multi-task learning stays balanced.
On two benchmarks, MedIR-2D-500K and MedIR-3D-3K, the authors report UniH3 reaching state-of-the-art results in both all-in-one and single-task restoration, though the abstract gives no numerical scores to check the claim against.
Zhiwen Yang and co-authors have released UniH3's code on GitHub at github.com/Yaziwel/UniH3.

Why it matters

Medical imaging restoration usually needs a different model for each task, modality and type of degradation: denoising an MRI scan is treated as unrelated to removing artifacts from a CT scan, for instance. UniH3's central argument is that this is wasteful, because medical images share real anatomical structure within and across modalities, and that shared structure can be used to train a single model that generalizes better instead of training many narrow ones. Modeling both what tasks share, homogeneity, and what makes them different, heterogeneity, rather than only the differences as prior methods do, is the paper's core conceptual contribution.

Who it affects

The work targets researchers and engineers who build medical image restoration systems across modalities such as CT and MRI and across different kinds of degradation, and who currently have to maintain a separate model for each task. A working, open-source implementation of UniH3 gives them a concrete architecture, including the H2M memory module and the H2B balancer, to test directly on their own restoration tasks and datasets.

How to use it

UniH3's code is public on GitHub at github.com/Yaziwel/UniH3. That lets researchers inspect the H2M, HGA and H2B components directly and adapt or retrain the framework on their own modalities and degradation types, rather than reimplementing the architecture from the paper's description alone.

How solid is it

The state-of-the-art claim rests on what the authors call extensive experiments across two large-scale benchmarks, MedIR-2D-500K and MedIR-3D-3K, covering both the all-in-one, multi-task setting and single-task restoration on its own. The abstract gives no numerical performance figures, such as PSNR or SSIM scores, for UniH3 or for the methods it is compared against, so the size of any improvement cannot be checked from this text. It also does not state a publication venue, peer-review status or date, and it does not explain what the '500K' and '3K' in the two benchmark names actually count, whether that is image pairs, scans or cases.

Risks and caveats

'State-of-the-art' here is the authors' own characterization in the abstract, not a number a reader can verify from this text, so it is worth treating as a claim pending the full paper and code rather than a settled result. Anyone considering the method for an actual restoration pipeline would need to check the underlying benchmarks and validation protocol in the full paper before treating the state-of-the-art claim as settled, since none of that numeric detail appears in the abstract itself.