Surgical AI training requires human-defined ontology and verified data

ddonoho · x · 2026-08-18

This post discusses the challenges in teaching AI to understand surgical videos. Beyond recognizing tools, understanding actions and anatomy requires teaching AI a specific language. The training of surgical tool detection models relies on thousands of human-verified annotations. Due to the "garbage in, garbage out" principle, mislabeled examples lead to learned errors. Thus, human oversight remains essential: humans define the ontology, verify data, and determine if the model's output is meaningful in the real world.

Original post →

More from Research

Research channel →