Safeguarding AI in Medical Imaging: Post-Hoc Out-of-Distribution Detection with Normalizing Flows.

Area of Science:

• Artificial Intelligence in Medicine
• Medical Imaging Analysis
• Machine Learning for Healthcare

Background:

• Out-of-distribution (OOD) detection is vital for AI medical imaging reliability.
• Current OOD methods require model retraining, limiting clinical adoption.
• Clinical diagnostic errors risk is high with undetected OOD data.

Purpose of the Study:

• To develop a post-hoc OOD detection method for AI medical imaging.
• To enable seamless integration with pre-trained models without weight alteration.
• To enhance clinical reliability and reduce diagnostic errors.

Main Methods:

• Proposed a novel post-hoc normalizing flow-based approach.
• Integrated the method with existing pre-trained models seamlessly.
• Evaluated on the in-house MedOOD dataset and MedMNIST benchmark.

Main Results:

• Achieved 84.61% AUROC on MedOOD, outperforming ViM and MDS.
• Reached 93.8% AUROC on MedMNIST, surpassing ViM and ReAct.
• Demonstrated superior performance and practical integration capabilities.

Conclusions:

• The proposed post-hoc method is a practical safeguard for clinical imaging.
• Effective OOD detection enhances AI diagnostic accuracy and safety.
• Seamless integration facilitates adoption in regulated clinical environments.