Research map / Label-efficient and robust learning / Domain adaptation and generalization

Why Invariance is Not Enough for Biomedical Domain Generalization and How to Fix It

NeurIPS 2026 · Label-efficient and robust learning › Domain adaptation and generalization · approach: Cross-modality medical segmentation adaptation · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes MaskGen, which concatenates raw image intensities with frozen anatomix foundation-model features and applies view masking during training to force use of both stable and unstable inputs for domain-generalized 3D biomedical segmentation.

Problem

Domain generalization for 3D biomedical image segmentation under clinical shifts (modality, sequence, disease stage, site, phase) without domain labels, test-time adaptation, or complex architectures.

Approach

Combine in-domain image intensities with domain-invariant features from a frozen 3D biomedical foundation model (anatomix) via channel concatenation, and randomly mask the image or representation channels during training so the predictor cannot shortcut onto the unstable input; theory argues stationarity then forces use of stable features.

Key results

• All-data: MaskGen achieves average rank 1.5 across six datasets, with Dice 67.5 on AMOS and 47.8 on CoW versus 57.4 and 24.6 for ERM.
• Few-shot: MaskGen gains of 15.6, 6.6, and 11.0 Dice points over the second-best method on AMOS, CoW, and Prostate.
• Without test-time adaptation, MaskGen obtains best average rank 1.3 in the TTA comparison, outperforming entropy minimization and SFB/Alpha Tuning which use test-domain data.

Surprising findings

• Training on stable anatomix representations alone gives inconsistent performance, and naive concatenation without masking often fails to improve over image-only training.
• Generic channel dropout on the image-only model underperforms targeted view masking, and fine-tuning the foundation model does worse than keeping it frozen.

Limitations

• Adds a forward pass through a frozen foundation model, increasing training overhead by 26% on BraTS.
• Theoretical analysis relies on simplifying assumptions, including that masking either removes all stable channels or the image and a condition on gradient non-conflict verified only empirically.
• Hyperparameter (mask rate) tuned on validation per dataset in main results; untuned r=0.5 variant is reported separately.

Experimental setup

Datasets: AMOS BraTS CoW (TopCoW) HVSMR PanDG Prostate multi-site MRI

Benchmarks: AMOS BraTS CoW HVSMR PanDG Prostate

Metrics: Dice score average rank

Baselines: ERM GIN GIN+IPA MixUp CutOut DSBN MixStyle RSC nnUNet EntMin SFB Alpha Tuning

Builds on: anatomix ERM GIN IPA channel dropout

Closest papers by meaning

• Domain Generalized Medical Landmark Detection via Robust Boundary-Aware Pre-Training (AAAI 2025 · similarity 0.88)
• Generative Medical Segmentation (AAAI 2025 · similarity 0.87)
• MaskTwins: Dual-form Complementary Masking for Domain-Adaptive Image Segmentation (ICML 2025 · similarity 0.87)
• LangDAug: Langevin Data Augmentation for Multi-Source Domain Generalization in Medical Image Segmentation (ICML 2025 · similarity 0.87)
• Learning General-purpose Biomedical Volume Representations using Randomized Synthesis (ICLR 2025 · similarity 0.87)
• Learning Generalized Medical Image Segmentation from Decoupled Feature Queries (AAAI 2024 · similarity 0.87)