DAVLMF-Seg: Vision-language model guided latent frequency-aware diffusion for semi-supervised medical image...

Area of Science:

• Medical image analysis
• Computer-aided diagnosis
• Machine learning for healthcare

Background:

• Medical image segmentation is crucial for diagnosis and treatment planning.
• Fully supervised methods require extensive annotated data.
• Semi-supervised learning (SSL) offers a solution by utilizing unlabeled data, but faces challenges in semantic modeling and domain adaptation.

Purpose of the Study:

• To propose DAVLMF-Seg, a domain-adaptive vision-language model guided frequency-aware SSL framework for medical image segmentation.
• To enhance semantic understanding and address domain gaps in medical imaging.
• To improve segmentation performance with limited labeled data.

Main Methods:

• Parameter-efficient adaptation to align medical images and textual descriptions in a shared latent space, providing semantic priors for pseudo-label refinement.
• A frequency-domain conditioned diffusion module for enhanced feature fusion and reduced decoding ambiguity.
• An uncertainty-aware regularization strategy for improved confidence calibration.

Main Results:

• Achieved 90.25% Dice on ACDC with 10% labels (+1.2%) and 90.48% Dice with 20% labels (+0.8%), reducing HD95 by 2.3 and 1.0, respectively.
• Attained 84.57% Dice on M&Ms (+2.0%) and 75.23% Dice on MyoPS (+1.1%), with HD95 reduced by 1.1 and 4.9, respectively.
• Demonstrated consistent improvements over state-of-the-art methods, particularly under limited supervision and cross-domain scenarios.

Conclusions:

• The proposed DAVLMF-Seg framework effectively improves medical image segmentation accuracy.
• The method shows significant advantages in scenarios with limited labeled data and cross-domain applications.
• The integration of vision-language models and frequency-aware techniques offers a promising direction for medical image analysis.