Research map / Label-efficient and robust learning / Semi-supervised learning

Constructing and Exploring Intermediate Domains in Mixed Domain Semi-supervised Medical Image Segmentation

CVPR 2024 · Label-efficient and robust learning › Semi-supervised learning · approach: Semi-supervised medical segmentation · 45 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.

Introduces Mixed Domain Semi-supervised medical image Segmentation (MiDSS) and proposes UCP-based intermediate domains with symmetric guidance (SymGD) and training-progress-aware random amplitude MixUp (TP-RAM) on Fundus, Prostate, and M&Ms datasets.

Problem

Semi-supervised medical image segmentation with limited labels from one domain and unlabeled data from multiple domains, where domain shift degrades pseudo-label quality.

Approach

Build intermediate domains by applying unified bidirectional Copy-Paste between labeled and unlabeled images, use symmetric pseudo-label guidance merged from intermediate samples to supervise unlabeled data, and gradually inject low-frequency style transitions via training-progress-aware amplitude MixUp.

Key results

• On Prostate with 40 labels, the method reaches 87.98% average Dice versus 74.41% for FixMatch and 66.54% for UDA-VAE++ (the best baselines), a reported 13.57% Dice improvement.
• On Fundus with 20 labels, average Dice is 87.66% versus 83.39% for FixMatch and 83.05% for BCP.
• On M&Ms with 20 labels, average Dice is 84.31% versus 82.96% for FixMatch.

Measured effects reported in the paper

Surprising findings

• Some semi-supervised methods (e.g., UA-MT, CPS) perform worse than training with labeled data alone due to error accumulation under domain shift.
• Arbitrary random amplitude MixUp (RAM) without gradual progression can harm performance compared with the plain UCP+SymGD baseline.

Limitations

• Evaluated only on 2D slices from three public datasets with a single labeled domain setting.
• Relies on a fixed confidence threshold (tau=0.95) and hyperparameters (beta, lambda schedules) tuned on the benchmarks.
• Some dataset partition details are deferred to the supplementary material.

Experimental setup

Datasets: Fundus Prostate M&Ms

Benchmarks: Fundus Prostate M&Ms

Metrics: Dice Jaccard Hausdorff Distance Average Surface Distance

Baselines: U-Net UA-MT FDA SIFA FixMatch CPS CoraNet UDA-VAE++ SS-Net BCP CauSSL

Builds on: Mean Teacher CutMix BCP FixMatch amplitude MixUp

Closest papers by meaning

• Steady Progress Beats Stagnation: Mutual Aid of Foundation and Conventional Models in Mixed Domain Semi-Supervised Medical Image Segmentation (CVPR 2025 · similarity 0.91)
• MAPSeg: Unified Unsupervised Domain Adaptation for Heterogeneous Medical Image Segmentation Based on 3D Masked Autoencoding and Pseudo-Labeling (CVPR 2024 · similarity 0.87)
• Annotation Ambiguity Aware Semi-Supervised Medical Image Segmentation (CVPR 2025 · similarity 0.87)
• Semi-supervised TEE Segmentation via Interacting with SAM Equipped with Noise-Resilient Prompting (AAAI 2024 · similarity 0.87)
• SemiDAViL: Semi-supervised Domain Adaptation with Vision-Language Guidance for Semantic Segmentation (CVPR 2025 · similarity 0.87)
• Versatile Medical Image Segmentation Learned from Multi-Source Datasets via Model Self-Disambiguation (CVPR 2024 · similarity 0.86)