ViTMARE - A Vision Transformer Pipeline for Anomaly Detection in 3D Brain MRI

Area of Science:

• Artificial Intelligence in Medical Imaging
• Neuroimaging Analysis
• Machine Learning for Anomaly Detection

Background:

• AI models in medical imaging struggle with dataset shifts and underrepresented patient groups.
• Detecting out-of-distribution scans is crucial for robust AI deployment in healthcare.
• Anomalies can stem from rare pathologies, atypical anatomy, or acquisition artifacts.

Purpose of the Study:

• To introduce ViTMARE, a volumetric anomaly detection pipeline for 3D brain MRI.
• To leverage Vision Transformer Masked AutoEncoders (ViTMAEs) for anomaly detection in volumetric data.
• To enhance the robustness and reliability of AI in medical imaging.

Main Methods:

• ViTMARE adapts ViTMAEs for volumetric data by processing axial slices as input channels.
• The model is fine-tuned on normal brain volumes.
• Inference involves multiple reconstructions, majority voting for anomaly masks, and morphological postprocessing.

Main Results:

• ViTMARE achieved a median Dice score of 0.793 on real images with synthetic anomalies.
• The pipeline demonstrated a median precision of 0.912 and a median recall of 0.748.
• Voting-based fusion combined with morphological postprocessing enabled robust voxel-level anomaly detection.

Conclusions:

• ViTMARE offers a reproducible pipeline for robust anomaly detection in 3D brain MRI.
• The study highlights the effectiveness of combining ViTMAEs with advanced postprocessing techniques.
• This approach significantly improves the ability to detect out-of-distribution scans in medical imaging.