Unsupervised Brain MRI Anomaly Detection via Inter-Realization Channels.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Neurology

Background:

• Accurate anomaly detection in brain MRI is vital for early neurological disorder diagnosis.
• High heterogeneity of brain abnormalities and limited annotated data pose significant challenges.
• Traditional methods often require extensive normal data, limiting adaptability.

Purpose of the Study:

• To introduce MadIRC, an unsupervised anomaly detection framework for brain MRI.
• To develop a robust nominal model without reliance on labeled data.
• To evaluate the generalizability of MadIRC across different medical imaging modalities.

Main Methods:

• MadIRC leverages Inter-Realization Channels (IRC) to build a nominal model.
• The framework operates in an unsupervised manner, requiring no labeled data.
• Evaluated on brain MRI, liver CT, and retinal images.

Main Results:

• MadIRC achieved a localization AUROC of 0.96 on brain MRI.
• Outperformed state-of-the-art supervised anomaly detection methods.
• Demonstrated generalizability across liver CT and retinal images.

Conclusions:

• MadIRC offers a scalable, label-free solution for brain MRI anomaly detection.
• The framework shows promise for integration into clinical workflows.
• Provides a robust approach for identifying neurological abnormalities.