Research map / Vision-language and multimodal / Cross-modal learning

DuSSS: Dual Semantic Similarity-Supervised Vision-Language Model for Semi-Supervised Medical Image Segmentation

AAAI 2025 · Vision-language and multimodal › Cross-modal learning · approach: Contrastive vision-language alignment · 10 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 DuSSS, a VLM-based semi-supervised medical image segmentation method that uses dual (cross- and intra-modal) contrastive learning with uncertainty-based semantic similarity supervision to produce text-guided masks that refine teacher pseudo-labels.

Problem

Semi-supervised medical image segmentation suffers from low-quality pseudo-labels, and existing vision-language models suffer from cross-modal uncertainty where one image/text corresponds to multiple texts/images.

Approach

Dual Contrastive Learning (cross-modal and intra-modal) with a Semantic Similarity-Supervision strategy that modulates cosine similarity by a distribution-based (2-Wasserstein) uncertainty level, plus a text-guided mask from the pretrained VLM that merges with teacher pseudo-labels for consistency regularization.

Key results

• Achieves 82.52% Dice and 75.87% mIoU on QaTa-COV19 with 50% labels, versus 81.24% Dice for MGCA.
• Achieves 74.61% Dice on BM-Seg (50% labels), slightly below MGCA's 75.17% Dice, and 78.03% Dice on MoNuSeg, the best reported.
• Ablation on 50% labels shows the full model (SSS+DCL+text-guided loss) reaches 82.52% Dice versus 79.64% for the teacher-student baseline on QaTa-COV19.

Measured effects reported in the paper

Surprising findings

• On BM-Seg with 50% labels, MGCA reports higher Dice (75.17%) than DuSSS (74.61%), which the paper does not discuss in detail.

Limitations

• The method relies on a pretrained VLM and text annotations, which are not available for all datasets.
• Gains over MGCA on BM-Seg at 50% labels are not observed in the main table.
• Theoretical analysis is limited to a short proof that the uncertainty term scales with the uncertainty level.

Experimental setup

Datasets: QaTa-COV19 BM-Seg MoNuSeg

Benchmarks: QaTa-COV19 BM-Seg MoNuSeg

Metrics: Dice mIoU

Baselines: U-Net CLIP ViLT MT CCT BCP MC-Net SS-Net UCMT LViT CMITM ASG

Builds on: CLIP Mean Teacher InfoNCE LViT MGCA

Closest papers by meaning

• SemiDAViL: Semi-supervised Domain Adaptation with Vision-Language Guidance for Semantic Segmentation (CVPR 2025 · similarity 0.87)
• RankMatch: Exploring the Better Consistency Regularization for Semi-supervised Semantic Segmentation (CVPR 2024 · similarity 0.86)
• SGTC: Semantic-Guided Triplet Co-training for Sparsely Annotated Semi-Supervised Medical Image Segmentation (AAAI 2025 · similarity 0.86)
• Steady Progress Beats Stagnation: Mutual Aid of Foundation and Conventional Models in Mixed Domain Semi-Supervised Medical Image Segmentation (CVPR 2025 · similarity 0.86)
• Improving Semi-Supervised Semantic Segmentation with Sliced-Wasserstein Feature Alignment and Uniformity (CVPR 2025 · similarity 0.85)
• DuPL: Dual Student with Trustworthy Progressive Learning for Robust Weakly Supervised Semantic Segmentation (CVPR 2024 · similarity 0.85)