Cautious Artificial Intelligence Improves Outcomes and Trust by Flagging Outlier Cases.

Area of Science:

• Medical image analysis
• Artificial intelligence in healthcare
• Machine learning for diagnostics

Background:

• AI diagnostic models typically use curated data, limiting performance on outlier images in clinical settings.
• Outlier images (rare diseases, different setups) can cause AI errors, posing risks in clinical practice.
• Clinicians cannot be expected to manually identify and discount AI results for outlier cases.

Purpose of the Study:

• To develop a method for training cautious AI that automatically flags outlier medical images.
• To ensure AI models can reliably identify and flag unusual cases encountered in real-world clinical data.
• To improve the safety and trustworthiness of AI in medical image diagnosis.

Main Methods:

• Introduced ClassClust, a method using supervised contrastive learning to form tight clusters of training images.
• ClassClust identifies outliers during testing by detecting images that fall outside these learned clusters.
• Evaluated ClassClust against three other methods on diverse datasets (pathology, dermatology, radiology), holding out specific image types as outliers.

Main Results:

• ClassClust demonstrated consistently higher outlier detection performance (AUC ROC) than competing methods.
• The method also achieved higher average accuracy on non-outlier images.
• Visualizations from ClassClust provided more informative insights into the image regions used for AI decisions.

Conclusions:

• AI models can effectively flag outlier cases without compromising accuracy on non-outlier cases.
• The ability to detect outliers is crucial for clinical AI deployment, complementing classification accuracy.
• ClassClust offers a promising approach for developing more robust and reliable AI diagnostic tools.