Training with Pooled Annotations from Multiple Surgeons Has No Effect on a Deep Learning Artificial Intelligence Model's Performance

Artificial intelligence (AI) has revolutionized image analysis. AI models learn from two inputs: data (videos/images) and accurate annotations (labels). Annotations require knowledgeable annotators (surgeons) and are time intensive, which has led to a paucity of annotated surgical videos. One solution includes pooling annotations from multiple annotators. Surgeons may vary in their annotations, which could train an inaccurate AI model. We studied model performance when trained with surgical videos annotated by one surgeon compared to training with videos annotated by multiple surgeons.
Source: Journal of the American College of Surgeons - Category: Surgery Authors: Tags: Surgical Education Source Type: research