Journal: Nature communications
This study presents a novel computer-aided diagnosis (CAD) model that uses narrow-band imaging endocytoscopy (EC-NBI) to classify colorectal lesions. These include non-neoplastic lesions, adenomas, and invasive cancers.
The model employs a multi-stage pre-training strategy combined with supervised deep clustering, outperforming existing supervised methods in a multi-center retrospective cohort.
It shows higher diagnostic accuracy than endoscopists in human-machine competitions and improves clinician performance when used as an assistive tool.
The approach enhances diagnostic accuracy and consistency, especially in distinguishing superficial from deep submucosal invasive cancers.
This has potential implications for early colorectal cancer screening, pending further validation.