Hurdles to Artificial Intelligence Deployment: Noise in Schemas and "Gold" Labels

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Data Science

Background:

• Clinicians express concerns regarding the safety and robustness of artificial intelligence (AI) in medical imaging, despite AI performance reports.
• Underreported sources of noise in imaging AI include variations in labeling schema definitions and noise within the labeling process itself.

Purpose of the Study:

• To investigate the impact of labeling schema variations and labeling process noise on medical imaging AI.
• To quantify the extent of label inconsistency and noise in publicly available datasets and radiologist annotations.

Main Methods:

• Compared schema overlap between two public datasets and a third-party vendor.
• Analyzed individual radiologist annotations from the CheXpert test set to quantify labeling noise.
• Evaluated label agreement across different classes to identify sources of unreliability.

Main Results:

• Low agreement (<50%) was found between labeling schemas of different datasets and vendors.
• Label noise varied by class; high agreement (>90%) for pneumothorax and medical devices, but low agreement for pneumonia and consolidation.
• The reliability of 'ground truth' labels for low-agreement classes was questionable, indicating dependence on the annotating radiologist group.

Conclusions:

• Noise in labeling schemas and gold standard annotations is prevalent in medical imaging classification.
• This noise negatively impacts the downstream clinical deployment and trustworthiness of AI tools.
• Potential solutions involving task design, annotation methods, and model training are discussed to enhance clinical AI trust.