BRAIN LESION DETECTION USING A ROBUST VARIATIONAL AUTOENCODER AND TRANSFER LEARNING

Area of Science:

• Medical Imaging
• Machine Learning
• Neurology

Background:

• Automated brain lesion detection from multi-spectral MRI aids clinicians but supervised methods require extensive manual data and lack generalizability.
• Unsupervised models like autoencoders eliminate the need for manual delineations but struggle with pre-trained model adaptation to new datasets with different parameters, demographics, or preprocessing.
• Clinical datasets often contain anomalies, posing challenges for unsupervised learning due to outlier impact on model performance.

Purpose of the Study:

• To address the challenges of unsupervised brain lesion detection, particularly concerning dataset variability and outliers.
• To develop a robust and adaptable unsupervised learning framework for accurate brain lesion detection.

Main Methods:

• A robust variational autoencoder (VAE) model utilizing β-divergence was employed to handle datasets with outliers.
• A transfer-learning approach was implemented to enable model adaptation across datasets with differing characteristics.
• The combined approach was tested on MRI datasets for brain lesion detection.

Main Results:

• The proposed method demonstrated improved accuracy in brain lesion detection.
• The robust VAE effectively handled data containing anomalies and outliers.
• Transfer learning facilitated successful adaptation of models to new datasets.

Conclusions:

• Adapting robust statistical models and transfer learning within a VAE framework significantly enhances unsupervised brain lesion detection accuracy.
• The developed method offers a more reliable solution for automated lesion detection in diverse clinical settings.
• This approach overcomes key limitations of existing unsupervised methods, paving the way for broader clinical application.