Rethinking domain generalization in medical image segmentation: One image as one domain

Domain shifts in medical image segmentation, particularly when data comes from different centers, pose significant challenges. Intra-center variability, such as differences in scanner models or imaging protocols, can cause domain shifts as large as, or even larger than, those between centers. To address this, we propose the “one image as one domain” (OIOD) hypothesis, which treats each image as a unique domain, enabling flexible and robust domain generalization. Based on this hypothesis, we develop a unified disentanglement-based domain generalization (UniDDG) framework, which simultaneously handles both multi-source and single-source domain generalization without requiring explicit domain labels. This approach simplifies training with a fixed architecture, independent of the number of source domains, reducing complexity and enhancing scalability. We decouple each input image into content representation and style code, then exchange and combine these within the batch for segmentation, reconstruction, and further disentanglement. By maintaining distinct style codes for each image, our model ensures thorough decoupling of content representations and style codes, improving domain invariance of the content representations. Additionally, we enhance generalization with expansion mask attention (EMA) for boundary preservation and style augmentation (SA) to simulate diverse image styles, improving robustness to domain shifts. Extensive experiments show that our method achieves Dice scores of 84.43% and 88.91% for multi-source to single-center and single-center generalization in optic disc and optic cup segmentation, respectively, and 86.96% and 88.56% for prostate segmentation, outperforming current state-of-the-art domain generalization methods, offering superior performance and adaptability across clinical settings.

💡 Research Summary

This paper tackles the persistent problem of domain shift in medical image segmentation by fundamentally redefining what constitutes a “domain.” The authors observe that intra‑center variability—caused by different scanner models, acquisition protocols, and parameter settings—can be as large as, or larger than, inter‑center differences. Consequently, grouping images by center or scanner leads to an inaccurate domain taxonomy. To overcome this, they propose the “One Image as One Domain” (OIOD) hypothesis, which treats every individual image as its own domain, thereby eliminating the need for explicit domain labels and the labor‑intensive process of domain recognition.

Building on OIOD, the authors introduce UniDDG (Unified Disentanglement‑based Domain Generalization), a framework that works for both single‑source and multi‑source scenarios with a fixed architecture and constant computational cost. UniDDG first encodes each input image into two latent spaces: a content code that captures anatomy‑related structures, and a style code that captures scanner‑specific characteristics such as contrast, illumination, and noise. The key operation is a cross‑batch recombination: within each training batch, content codes are paired with style codes from other images, producing mixed content‑style pairs. The model is then trained to (1) segment using the content code, (2) reconstruct the original image from the mixed pair, and (3) enforce disentanglement through reconstruction loss and a style‑consistency loss. By keeping a distinct style code for every image, the framework ensures that content representations become truly domain‑invariant.

Two complementary enhancements are added. Expansion Mask Attention (EMA) expands the ground‑truth mask by a small margin and uses this enlarged region as an attention map during reconstruction, which protects boundary pixels and improves segmentation of low‑contrast structures. Style Augmentation (SA) randomly generates synthetic style codes during training, effectively simulating a wide variety of unseen acquisition styles and further strengthening the model’s robustness to domain shifts.

The method is evaluated on two clinically relevant tasks: optic disc/cup segmentation from fundus images and prostate segmentation from MRI. Experiments cover two generalization settings: (a) multi‑source → single‑center and (b) single‑center → multi‑source. UniDDG achieves Dice scores of 84.43 % (optic disc/cup, multi‑source → single) and 88.91 % (optic disc/cup, single‑center → multi), as well as 86.96 % and 88.56 % for prostate segmentation in the respective settings. These results consistently surpass state‑of‑the‑art domain generalization methods such as BigAug, DCA‑Net, CDDSA, and others, often by 2–5 % absolute Dice improvement.

Ablation studies reveal that (i) removing the cross‑batch content‑style exchange degrades performance dramatically, (ii) omitting EMA leads to poorer boundary accuracy, and (iii) excluding SA reduces robustness to unseen styles. The authors also emphasize that UniDDG’s architecture does not grow with the number of source domains, unlike many GAN‑based multi‑domain approaches that require a separate encoder/discriminator per domain. This fixed‑complexity property makes the method practical for real‑world deployment where new centers may appear without any retraining of the network architecture.

Limitations include increased GPU memory consumption proportional to batch size, especially for 3‑D volumetric data, and the reliance on a deterministic style encoder that may need further refinement for extremely heterogeneous datasets. Future work is suggested to explore memory‑efficient style representations and unsupervised style extraction techniques.

In summary, by redefining domains at the image level, disentangling content and style, and reinforcing boundary preservation and style diversity through EMA and SA, the paper presents a scalable, label‑free, and highly effective solution to domain generalization in medical image segmentation, offering strong performance across diverse clinical settings.