TGME-Net: Text-Guided Multi-Scale Enhancement Network for Semi-Supervised Medical Image Segmentation

Pseudo labels quality is pivotal to the performance of semi-supervised medical image segmentation. Recent studies have explored the use of vision language models (VLM) to improve the quality of pseudo labels by enhancing their ability to describe lesion regions. Nevertheless, existing methods often exhibit limitations such as excessive reliance on textual modalities, performance plateaus, and insufficient integration of multi-grained information. To address these issues, we propose a Text-Guided Multi-scale Enhancement (TGME) network, which enhances pseudo labels through the accumulation of multi-scale cross-modal supervision. Specifically, TGME establishes cross-modal semantic alignments between multi-scale image patches and text tokens, thereby preserving both fine-grained and coarse-grained informative features. Additionally, we introduce a novel learning paradigm named Conflict Feature Supervised Learning (CFSL) to achieve spatial perception of text-described regions within feature. Extensive experiments on three datasets demonstrate the effectiveness of the proposed method.