Medical images rarely look the same twice. A retinal scan captured on one clinic’s fundus camera can differ dramatically, in color balance, contrast, and vessel appearance, from a seemingly identical scan taken on another machine in another hospital, or even another room. For the artificial intelligence models that radiologists and ophthalmologists increasingly rely on to outline blood vessels, optic discs, and organs at the pixel level, this seemingly cosmetic variation poses a serious technical problem known as domain shift: models trained on data from one source tend to falter when confronted with images from another. A new study published in the journal Cognitive Computation offers a fresh approach to this stubborn issue, presenting a neural network architecture designed to learn from many medical datasets at once without being overwhelmed by their differences.
The work, led by Xu Han and Chaobin Wang, with contributions from Junni Huang, Meijun Sun, Jinchang Ren, and corresponding author Zheng Wang, introduces a framework called MDCL-UNet — short for Multi-Domain Collaborative Learning UNet. Its central insight is deceptively simple: instead of trying to erase the stylistic fingerprints that different scanners and clinics leave on medical images, the model explicitly separates those fingerprints from the underlying anatomical content, and then cleverly reuses both streams of information when it makes its final segmentation prediction.
Most existing segmentation networks, including the now-classic UNet and its many descendants, are trained and evaluated on a single dataset. Under those conditions they can achieve impressive accuracy. But in real clinical settings, training data are scarce, expensive to annotate at the pixel level, and often siloed behind privacy restrictions, so the same model must ideally work across data from multiple sources. When researchers test single-dataset models on images from a different medical center, performance frequently collapses. Prior attempts to solve this have taken two main routes: domain adaptation, which adjusts a model to a specific target domain using unlabeled target images, and domain generalization, which trains models to ignore domain-specific style cues altogether so they perform well on unseen sources.
The team behind MDCL-UNet took issue with a third, more recent strategy known as Cross-Dataset Collaborative Learning, or CDCL, which trains a single unified model on multiple labeled datasets simultaneously. While elegant in principle, this approach shares a hidden weakness: it lumps all features together in shared convolutional layers without distinguishing what is truly domain-specific from what is domain-invariant. When the gap between datasets is large, those shared layers are asked to do too much, and training becomes unstable. The Tianjin University-led team set out to build a framework that would acknowledge this distinction from the ground up.
At the heart of MDCL-UNet is a two-branched encoder. One branch is dedicated to capturing domain-specific features — the idiosyncratic visual style of each dataset, such as intensity distributions or scanner-related artifacts. The other branch extracts domain-invariant features: the semantic information about anatomy that remains consistent regardless of which machine produced the image. To keep the style branch well-behaved across datasets, the researchers introduced a module called Domain Style Instance Normalization, or DSIN. Instead of using batch normalization, which computes statistics across an entire mini-batch and becomes unreliable when batches are small or contain images from mixed domains, DSIN applies instance normalization on a per-domain, per-sample basis. This choice, the team found, dramatically reduces training instability, particularly in scenarios where batch sizes are as small as two images, as was the case in some of their retinal experiments.
Keeping the two branches truly separate is enforced through two complementary mechanisms. The first is a domain adversarial classifier, an N-way discriminator attached to the domain-invariant branch. Its job is to guess which of the training datasets a given feature vector came from. Meanwhile, the domain-invariant branch tries to fool it, using a gradient reversal layer borrowed from the classic DANN architecture to flip the direction of gradient updates during backpropagation. When the discriminator can no longer tell domains apart, the features it is looking at have, by definition, shed their domain identity. The second mechanism is a Maximum Mean Discrepancy, or MMD, loss — a staple of domain adaptation research that measures the distance between two probability distributions in a reproducing kernel Hilbert space. Here it is repurposed to push the outputs of the two encoder branches toward orthogonality, minimizing the coupling between style and content representations.
What truly sets MDCL-UNet apart, however, is its refusal to throw away the style features. Conventional domain generalization methods treat domain-specific information as noise to be suppressed. The Tianjin group reasoned that this is wasteful: style cues carry genuine information about how a given image’s appearance relates to the anatomy within it. Their solution is a Domain Fusion Attention Module, or DFAM, inspired in part by well-known attention designs such as Squeeze-and-Excitation networks and the Convolutional Block Attention Module. DFAM applies channel attention to the domain-specific features — since style information tends to be encoded along the channel dimension — and mixed channel-plus-spatial attention to the domain-invariant features, which the authors argue contain both fully domain-invariant signals and partially invariant ones. The fused representation is then passed to a UNet-style decoder, giving the network access to the full spectrum of information extracted from every training domain.
The researchers evaluated MDCL-UNet on three distinct segmentation tasks of increasing domain difficulty: retinal vessel segmentation using the DRIVE, STARE, and CHASE_DB1 datasets; optic disc and cup segmentation across four fundus image datasets; and abdominal multi-organ segmentation using two CT volumes datasets, BTCV and TCIA, annotated for eight organs including the liver, spleen, pancreas, and kidneys. Performance was measured using the Dice coefficient, a standard overlap metric, and the 95th-percentile Hausdorff distance, or HD95, which quantifies the worst-case boundary error in millimeters or pixels.
The results were consistent across all three tasks. On retinal vessel segmentation, MDCL-UNet improved the Dice score by 1.35 points over a standard UNet while cutting the HD95 distance by nearly three units, and outperformed the existing CDCL approach by roughly 2.86 units on HD95 without requiring any special training strategy. The gains widened as domain shift increased. In optic disc and cup segmentation, where the four source datasets diverge most dramatically, MDCL-UNet achieved a Dice score of 91.95 on average, a 2.61-point improvement over UNet and a striking 12.51-point improvement over CDCL with domain adversarial training. Its HD95 dropped by more than 15 units compared to the best competing collaborative method. Notably, the framework even edged out a recent unsupervised domain adaptation method, DDF-UDA, despite operating under a far more demanding setting that requires no access to unlabeled target-domain images.
An interesting and clinically relevant finding emerged from these comparisons: the domain adversarial training strategy that previous work had championed turned out to be unreliable. While it helped in retinal vessel segmentation, it actively degraded performance in optic disc/cup and abdominal organ segmentation, sometimes misleading the model rather than helping it. The authors attribute MDCL-UNet’s stability to its structural approach — explicit two-branch decoupling — rather than reliance on a single training trick that may or may not suit a given dataset.
Ablation experiments on the fundus task confirmed that each of the three novel components earns its place. Removing the domain adversarial classifier cost 1.60 Dice points. Removing the MMD-based decoupling loss cost 3.72 points and made training markedly less stable. Most strikingly, removing DFAM — and thereby discarding domain-specific features as domain generalization methods conventionally do — cost 4.19 Dice points, underscoring the team’s argument that style information is a resource to be harvested, not discarded. The authors also showed that DSIN’s instance normalization maintains robust performance across batch sizes ranging from 2 to 8, while batch-normalization-based competitors such as CDCL exhibited noticeable training fluctuations under small-batch conditions.
One further practical advantage deserves mention: unlike many multi-branch architectures whose parameter counts balloon with the number of training domains, MDCL-UNet’s size remains essentially fixed regardless of how many datasets are added. This makes it far more portable for real-world deployment, where a hospital might want to pool data from a handful of partner institutions one year and a dozen the next. Training curves published with the study show consistently smoother convergence for MDCL-UNet compared to single-branch baselines, particularly on the most domain-diverse fundus task.
The work arrives at a moment of intense interest in generalizable medical AI, as foundation models like the Segment Anything Model are being adapted for clinical use, and as Mamba-style state space models begin to compete with convolutional and Transformer architectures. By combining ideas from adversarial domain adaptation, distribution alignment, and style-aware attention within a single supervised framework, MDCL-UNet offers a pragmatic middle path: it does not chase adaptation to one target domain, nor does it pretend style differences do not exist. Instead, it treats every training domain as a source of complementary information.
The authors, whose work was supported by the National Natural Science Foundation of China, note that future efforts will push the framework toward entirely unseen domains and toward three-dimensional volumetric and multi-modality segmentation, where domain gaps are often even more pronounced. The source code is slated for release on GitHub, opening the door for other research groups to build on the approach. For a field where a model’s usefulness can hinge on whether it was trained on the same scanner brand as the images it will see in the clinic, architectures that gracefully absorb rather than ignore such variation may prove to be a meaningful step toward AI tools that work everywhere medicine is practiced, not just where their training data happened to come from.
Subject of Research: Multi-domain collaborative learning for medical image segmentation using domain feature disentanglement (MDCL-UNet)
Subject of Research: Technology and Engineering
Article Title: MDCL-UNet: A Multi-Domain Collaborative Learning Method for Medical Image Segmentation
Article References: Han, X., Wang, C., Huang, J., Sun, M., Ren, J., & Wang, Z. (2026). MDCL-UNet: A Multi-Domain Collaborative Learning Method for Medical Image Segmentation. Cognitive Computation, 18(1), Article 89. https://doi.org/10.1007/s12559-026-10628-0
Image Credits: AI Generated
DOI: 10.1007/s12559-026-10628-0
Keywords: medical image segmentation, multi-domain collaborative learning, domain shift, domain feature disentanglement, domain adversarial classifier, instance normalization, attention module, Dice score, HD95, retinal vessel segmentation, optic disc and cup segmentation, abdominal multi-organ segmentation
Cite Scienmag News
APA MLA Chicago
Blake Davidson. (September 4, 2026). Multi-domain collaborative learning improves medical image segmentation via MDCL-UNet. Scienmag. https://scienmag.com/multi-domain-collaborative-learning-improves-medical-image-segmentation-via-mdcl-unet/
Blake Davidson. “Multi-domain collaborative learning improves medical image segmentation via MDCL-UNet.” Scienmag, 4 September 2026, https://scienmag.com/multi-domain-collaborative-learning-improves-medical-image-segmentation-via-mdcl-unet/. Accessed 4 September 2026.
Blake Davidson. “Multi-domain collaborative learning improves medical image segmentation via MDCL-UNet.” Scienmag. September 4, 2026. https://scienmag.com/multi-domain-collaborative-learning-improves-medical-image-segmentation-via-mdcl-unet/
Copy citation Download RIS
Tags: AI for blood vessel and organ segmentationcross-domain neural network modelscross-source retinal scansdeep learning for medical imagingdomain shift in medical imaginghandling scanner and clinic differences in medical imageshandling variability in medical imagesimproving generalization of AI models in healthcareimproving robustness of medical image modelsMDCL-UNet architecturemedical dataset variabilitymedical image segmentationmulti-domain collaborative learningmulti-hospital medical image datasetsmulti-source medical image segmentationmulti-source medical image trainingneural networks for medical image analysisneural networks for multi-center medical dataretinal scan image analysisrobust AI models for medical diagnostics
