Vision-language foundation models have transformed medical image segmentation over the past three years. These models pair large image encoders with text prompts, so a single model can segment many anatomical structures, lesion types, and imaging modalities through natural language. This survey reviews vision-language foundation models designed for medical image segmentation. We describe the technical background from contrastive vision-language pretraining to the Segment Anything Model and its medical variants. We propose a three-part taxonomy that covers text-prompt guided models, large language model embedded architectures, and hybrid frameworks. We examine adaptation strategies such as full fine-tuning, Low-Rank Adaptation, adapters, and prompt engineering. We organize the literature by modality and cover computed tomography, magnetic resonance imaging, pathology, chest radiography, and ultrasound. We discuss clinical uses such as organ segmentation, tumor delineation, and radiotherapy planning. We summarize evaluation metrics and benchmark datasets. We identify four open challenges: prompt dependence, mask hallucination, slow volumetric inference, and limited annotated data. We close with a research roadmap for trustworthy deployment, multimodal pretraining, and clinical integration.