SAGE: An Expert-Annotated South Asian GI Endoscopy Dataset for Multimodal Learning and Hallucination Analysis

The article introduces the SAGE dataset, a South Asian GI endoscopy dataset designed for multimodal learning and hallucination analysis, comprising 1,300 images, captions, 18 labels, and 14,726 question-answer pairs. It highlights significant performance drops in existing models when evaluated on this dataset, with task-specific classifiers experiencing an average performance decline of 58%, and contemporary large multimodal models showing reduced GREEN scores for anatomical landmark and abnormality detection. This dataset is crucial for developing AI tools that are geographically inclusive and can improve diagnostic accuracy in underrepresented regions.