Research map / Label-efficient and robust learning / Domain adaptation and generalization
LangDAug: Langevin Data Augmentation for Multi-Source Domain Generalization in Medical Image Segmentation
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 LangDAug, which trains pairwise energy-based models between source domains via contrastive divergence and uses Langevin-dynamics iterates as augmented samples for multi-source domain generalization in 2D medical image segmentation.
Problem
Segmentation models for medical images generalize poorly to unseen domains (sites, scanners, protocols); the paper targets multi-source domain generalization for 2D fundus (OC/OD) and prostate MRI segmentation.
Approach
Train energy-based models for each ordered pair of source domains with contrastive divergence, run Langevin dynamics from one domain toward another, and use the intermediate Langevin iterates (keeping original labels) as augmentation data within ERM; theory links this to a gradient/Hessian regularizer and a Rademacher bound depending on intrinsic dimension for GLMs.
Key results
- On retinal fundus leave-one-out, LangDAug reaches average IoU 78.84 and DSC 87.61, the best among compared methods (TriD 77.92 / 85.95, RAM 77.40 / 85.39).
- On prostate MRI, LangDAug averages ASD 0.81 mm and DSC 89.16%, versus TriD at 0.90 mm and 87.68%.
- Added on top of domain-randomization methods, LangDAug improves TriD by +3.05 IoU and +2.21 DSC on fundus and +2.90 DSC on prostate; RAM and FedDG also improve.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use LangDAug instead of TriD | Retinal Fundus Segmentation · Retinal Fundus Segmentation (average) | IoU: 77.92 -> 78.84 | Helped |
| Use LangDAug instead of TriD | Retinal Fundus Segmentation · Retinal Fundus Segmentation (average) | DSC: 85.95 -> 87.61 | Helped |
| Use LangDAug instead of TriD | Prostate MRI Segmentation · Prostate MRI Segmentation (average) | ASD: 0.90 -> 0.81 | Helped |
| Use LangDAug instead of TriD | Prostate MRI Segmentation · Prostate MRI Segmentation (average) | DSC: 87.68 -> 89.16 | Helped |
| Add LangDAug to domain-randomization baselines | FedDG, RAM, and TriD · Retinal Fundus Segmentation (average) | mIoU: +1.38%, +1.86%, and +3.05% | Helped |
| Add LangDAug to domain-randomization baselines | FedDG, RAM, and TriD · Retinal Fundus Segmentation (average) | mDSC: +1.47%, +2.04%, and +2.21% | Helped |
| Add LangDAug to domain-randomization baselines | FedDG, RAM, and TriD · Prostate MRI Segmentation (average) | DSC: +1.40%, +1.67%, and +2.90% | Helped |
| Vary the number of Langevin steps K (ablation) | LangDAug · Retinal Fundus Segmentation | mIoU | No clear effect |
| Vary the Langevin step size β (ablation) | LangDAug · Retinal Fundus Segmentation | mIoU | Mixed |
| Vary the number of EBM convolutional blocks (ablation) | LangDAug · Retinal Fundus Segmentation | mIoU | No clear effect |
| Vary the number of stored samples per Langevin chain (ablation) | LangDAug · Retinal Fundus Segmentation | mIoU | Mixed |
Surprising findings
- Adding LangDAug roughly doubles training time and memory (e.g., TriD 5.53 to 7.49 GPU hours), a notable cost for the gain.
Limitations
- Training time and memory increase substantially because of the many augmented samples.
- The number of required EBMs scales as the number of source domains, grows quadratically with pairs.
- Only 2D slices are used; 3D volumes are processed as slices.
Experimental setup
Datasets: Retinal Fundus Segmentation (RFS) Prostate MRI Segmentation (SAML prostate dataset)
Benchmarks: Retinal Fundus Segmentation (RFS) Prostate MRI Segmentation
Metrics: IoU DSC ASD
Baselines: CORAL RSC SagNet SWAD Hutchinson Fish Fishr RandConv MixStyle FedDG RAM TriD
Builds on: Hutchinson Fishr RAM TriD FedDG
Closest papers by meaning
- Learning Generalized Medical Image Segmentation from Decoupled Feature Queries (AAAI 2024 · similarity 0.88)
- Why Invariance is Not Enough for Biomedical Domain Generalization and How to Fix It (NeurIPS 2026 · similarity 0.87)
- Steady Progress Beats Stagnation: Mutual Aid of Foundation and Conventional Models in Mixed Domain Semi-Supervised Medical Image Segmentation (CVPR 2025 · similarity 0.86)
- S2S2: Semantic Stacking for Robust Semantic Segmentation in Medical Imaging (AAAI 2025 · similarity 0.86)
- Generative Medical Segmentation (AAAI 2025 · similarity 0.86)
- Constructing and Exploring Intermediate Domains in Mixed Domain Semi-supervised Medical Image Segmentation (CVPR 2024 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.