Dynamic Pseudo-Labeling via Large Language Models for Robust Medical Image Segmentation

In the field of medical image segmentation, the high cost of annotated data and the poor generalization performance of the model have always been problems that need to be solved urgently. In this study, we innovatively propose a cross-modal collaboration framework based on large language model (LLM), MedSegLLM, which aims to overcome the two major difficulties of weakly supervised learning and domain adaptation at the same time, and open up an efficient and interpretable new path for medical image analysis: 1) Cross-modal semantic alignment network: Through a unique mapping strategy, medical image features and accurate text descriptions generated by LLM are projected into the same semantic space, cleverly bridging the understanding gap between vision and language, and achieving deep semantic integration across modalities. 2) Dynamic pseudolabel generator: deeply excavate the powerful logical reasoning potential of LLM, intelligently generate high-credibility segmentation masks, provide an effective way to break the dilemma of a lack of annotated data, and the generation process has adaptive optimization capabilities. 3) Adaptive feature fusion module: According to the image differences generated by different medical devices, it intelligently integrates multi-source feature information, comprehensively improves the generalization and adaptability of models across devices and scenarios, and ensures the stability of clinical applications. MedSegLLM has been experimentally proven to perform well on BraTS (Brain Tumor Segmentation Benchmark), ISIC (Dermoscopy Image Dataset) and proprietary CT datasets. In particular, under the severe conditions of low annotation data rate (only $10 \%$), the Dice coefficient is still as high as 0.91, which is significantly higher than that of traditional strongly supervised U-Net (0.85) and advanced TransUNet (0.88) ($\mathbf{p}\lt \mathbf{0 . 0 1}$). The cross-device generalization error of the framework is reduced by $32 \%$ compared with the existing schemes, which fully demonstrates its adaptability and robustness in complex medical environments(The model can maintain stable and high - accuracy segmentation performance in complex conditions without significant drops due to environmental changes or data - quality issues). Even in low-resource hospitals (less than $10 \%$ of labeled data) and emergency diagnosis scenarios, it can still maintain high-precision output, which is of great clinical practical value. It will provide a strong boost for the inclusive development of global medical imaging technology, and is expected to accelerate the popularization of precision medicine in medical institutions at all levels.