Text this: Multimodal LLMs for retinal disease diagnosis via OCT: few-shot versus single-shot learning