Exploring the performance and robustness of Natural Domain Vision Language models for Segmenting Medical Image