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

MAPSeg: Unified Unsupervised Domain Adaptation for Heterogeneous Medical Image Segmentation Based on 3D Masked Autoencoding and Pseudo-Labeling

CVPR 2024 · Label-efficient and robust learning › Domain adaptation and generalization · approach: Cross-modality medical segmentation adaptation · 35 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 MAPSeg, a unified UDA framework combining 3D multi-scale masked autoencoder pretraining, masked pseudo-labeling with a teacher-student EMA, and global-local collaboration, applied to centralized, federated, and test-time UDA for volumetric medical segmentation.

Problem

Domain shifts (cross-modality, cross-site, cross-sequence, cross-age) degrade segmentation of volumetric medical images, and existing UDA methods often need co-located data, synchronous training, or target labels for model selection.

Approach

Combine 3D multi-scale masked autoencoding pretraining with masked pseudo-labeling (teacher-student EMA on unmasked target images, student on masked images) and global-local feature fusion, so one framework covers centralized, federated, and test-time UDA without target labels for model selection.

Key results

• Centralized UDA on cardiac CT-to-MRI reaches 80.3 average Dice, versus 74.8 for SDUDA and 75.5 for FSUDA-V2 (about 5.7 points above the prior best).
• Centralized brain MRI cross-sequence average Dice of 77.7 vs. 68.2 for DAR-UNet; cross-site 73.1 vs. 62.7 for HRDA; cross-age 75.2 vs. 64.7 for MIC.
• Fed-MAPSeg achieves 69.9 Dice on cross-sequence vs. 28.4 for DualAdapt and 27.6 for FAT; test-time UDA drops target Dice by under 3%.

Measured effects reported in the paper

Surprising findings

• Applying masked pseudo-labeling alone yields only a minor gain (39.5 Dice in the cross-sequence ablation), and adding MAE pretraining is what raises it to 75.3.

Limitations

• Test-time adaptation causes a noticeable Dice drop on the source domain (up to 5.8 points), indicating forgetting.
• Performance is sensitive to local patch size (4 or 16 degrades results significantly).
• Brain MRI evaluation uses a small private dataset, and small subcortical structures show large class imbalance.

Experimental setup

Datasets: BCP50 infant brain MRI ECHO5 MCRIB10 MMWHS 2017 cardiac CT-MRI

Benchmarks: private infant brain MRI segmentation (cross-sequence, cross-site, cross-age) MMWHS 2017 cardiac CT to MRI segmentation

Metrics: Dice

Baselines: AdvEnt DAFormer HRDA MIC DAR-UNet FAT DualAdapt PnP-AdaNet SIFA-V1 SIFA-V2 MPSCL MA-UDA

Builds on: masked autoencoder mean teacher DeepLabV3 SIFA MIC

Closest papers by meaning

• A Unified Degradation-Robust Approach to SSL and UDA for 3D Medical Images (AAAI 2025 · similarity 0.89)
• UniPro: Unified Multi-Mode Medical Image Segmentation from 2D Images to 3D Volumes via Propagation (NeurIPS 2026 · similarity 0.87)
• Constructing and Exploring Intermediate Domains in Mixed Domain Semi-supervised Medical Image Segmentation (CVPR 2024 · similarity 0.87)
• One-to-Multiple: A Progressive Style Transfer Unsupervised Domain-Adaptive Framework for Kidney Tumor Segmentation (NeurIPS 2024 · similarity 0.87)
• Unleashing the Potential of SAM for Medical Adaptation via Hierarchical Decoding (CVPR 2024 · similarity 0.86)
• Annotation Ambiguity Aware Semi-Supervised Medical Image Segmentation (CVPR 2025 · similarity 0.86)