Leveraging large language and vision models for knowledge extraction from large-scale image–text colonoscopy records

Abstract

The development of artificial intelligence systems for colonoscopy analysis often necessitates expert-annotated image datasets. However, limitations in dataset size and diversity impede model performance and generalization. Image–text colonoscopy records from routine clinical practice, comprising millions of images and text reports, serve as a valuable data source, although annotating them is labour intensive. Here we leverage recent advancements in large language and vision models and propose EndoKED, a data mining paradigm for deep knowledge extraction and distillation. EndoKED automates the transformation of raw colonoscopy records into image datasets with pixel-level annotation. We apply EndoKED to multicentre datasets of raw colonoscopy records (~1 million images), showing its superior performance in detecting polyps at the report and image levels, as well as annotating polyps at the pixel level. The state-of-the-art performance and generalization ability of polyp segmentation models are achieved through EndoKED pretraining. Furthermore, the EndoKED vision backbone enables data-efficient learning for optical biopsy, achieving expert-level performance in internal, external and prospective validation datasets.