StRegA: Unsupervised anomaly detection in brain MRIs using a compact context-encoding variational autoencoder.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Machine Learning

Background:

• Neuroradiology relies on expert interpretation of brain MRIs, with machine learning (ML) assisting analysis.
• Current ML models require task-specific training and extensive manual annotations, challenging for diverse brain abnormalities.
• Unsupervised Anomaly Detection (UAD) offers a solution by learning normal data distributions to identify deviations without pathology-specific training.

Purpose of the Study:

• To develop a robust Unsupervised Anomaly Detection (UAD) pipeline for detecting anomalies in brain MRIs.
• To improve the detection of pathologies like brain tumors using a novel approach that does not require specific pathology training.
• To enhance the performance of UAD on clinical data compared to existing Variational Autoencoder (VAE) based methods.

Main Methods:

• Proposed a novel Unsupervised Anomaly Detection (UAD) pipeline named StRegA.
• StRegA integrates a compact context-encoding Variational Autoencoder (ceVAE) with pre- and post-processing steps.
• The pipeline was evaluated on the BraTS dataset for tumor detection and on artificially induced anomalies.

Main Results:

• The StRegA pipeline achieved a Dice score of 0.642 ± 0.101 for detecting tumors in T2w brain MRI images.
• Demonstrated superior performance compared to the best baseline method (Dice score 0.522 ± 0.135).
• Achieved a Dice score of 0.859 ± 0.112 for artificially induced anomalies, outperforming the baseline (0.783 ± 0.111).

Conclusions:

• The proposed StRegA pipeline offers a more robust approach to Unsupervised Anomaly Detection (UAD) in clinical brain MRI data.
• StRegA effectively detects anomalies such as brain tumors without prior specific pathology training.
• This method shows significant potential for assisting neuroradiologists in identifying abnormalities in brain scans.