Transformer-Based Anomaly Detection for Neurodegenerative Screening in MRI Images.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Neuroscience

Background:

• Automatic anomaly detection in medical images is crucial for diagnosing neurodegenerative diseases.
• Alzheimer's disease diagnosis can be improved with advanced imaging analysis tools.

Purpose of the Study:

• To develop and evaluate a Transformer-based anomaly detection model for brain MRI.
• To enable unsupervised detection of atypical structural patterns in neuroimaging.

Main Methods:

• A Vision Transformer encoder combined with a memory bank module was utilized.
• The model employed a one-class learning approach, trained exclusively on normal brain images.
• A 3D to 2D preprocessing step was implemented to adapt volumetric data.

Main Results:

• The model demonstrated a strong capability in characterizing normal brain structures.
• Reliable predictions for anomaly detection were generated, confirming model viability.
• The approach successfully flagged atypical structural patterns in neuroimaging data.

Conclusions:

• Transformer-based architectures are viable for unsupervised anomaly detection in neuroimaging.
• This method provides a foundation for clinical applications in medical imaging analysis.
• The model shows promise for assisted diagnosis of neurodegenerative conditions.