Anomaly detection in brain MRI: a comprehensive review.

Area of Science:

• Neuroimaging
• Artificial Intelligence
• Machine Learning

Background:

• Magnetic resonance imaging (MRI) is crucial for diagnosing neurological diseases but manual interpretation is time-consuming and variable.
• Large annotated datasets for supervised learning are scarce, hindering AI development in neuroimaging.
• Anomaly detection presents a scalable AI alternative by learning normal brain patterns to identify deviations.

Purpose of the Study:

• To review the landscape of anomaly detection techniques for brain MRI.
• To categorize deep learning approaches into reconstruction, generative, and self-supervised paradigms.
• To identify challenges and propose future research directions for robust AI in neuroimaging.

Main Methods:

• Comprehensive review of traditional statistics, classical machine learning, and deep learning methods for brain MRI anomaly detection.
• Categorization of deep learning methods based on their underlying principles.
• Analysis of current limitations and emerging strategies.

Main Results:

• Anomaly detection models normal brain anatomy and flags deviations, reducing reliance on expert labels.
• Deep learning paradigms include reconstruction, generative, and self-supervised approaches.
• Persistent challenges include high false positive rates, interpretability, and domain generalization.

Conclusions:

• Emerging strategies like hybrid learning and multimodal integration show promise for improving AI robustness in neuroimaging.
• Developing generalizable and interpretable AI systems is key for clinical integration.
• This review provides a guide for advancing reliable brain MRI anomaly detection.