Evaluating normative representation learning in generative AI for robust anomaly detection in brain imaging.

Area of Science:

• Medical Imaging Analysis
• Artificial Intelligence in Healthcare
• Computational Pathology

Background:

• Normative representation learning in AI synthesizes typical anatomical patterns from healthy medical scans.
• Generative AI leverages normative patterns to detect and correct anomalies in pathological data, reducing reliance on expert labeling.
• Traditional anomaly detection often overlooks the significance of normative learning in AI model performance.

Purpose of the Study:

• To introduce novel metrics for evaluating normative learning capabilities in generative AI models.
• To assess the performance of various generative AI frameworks, including diffusion models, in detecting brain pathologies.
• To compare AI-driven anomaly detection metrics with expert evaluations through a multi-reader study.

Main Methods:

• Development and application of new metrics to evaluate normative learning in AI.
• Testing generative AI models against diverse brain pathologies.
• Conducting a large-scale multi-reader study to validate metrics against expert assessments.

Main Results:

• AI models with strong normative learning capabilities demonstrate versatility in detecting a wide spectrum of unseen medical conditions.
• Novel metrics effectively evaluate the normative learning facet of AI models.
• Expert evaluations align with the performance indicated by the new metrics.

Conclusions:

• Proficiency in normative learning enhances generative AI's ability to detect diverse medical anomalies.
• The developed metrics provide a robust framework for assessing AI models in medical anomaly detection.
• Generative AI shows significant promise for automated anomaly detection in medical imaging.