Transfer Learning in Magnetic Resonance Brain Imaging: A Systematic Review

Area of Science:

• Machine Learning in Medical Imaging
• Neuroimaging Analysis
• Artificial Intelligence in Healthcare

Background:

• Transfer learning enhances machine learning model generalization across different magnetic resonance imaging (MRI) protocols and scanners.
• It enables the reuse of models trained on related tasks for improved performance in specific applications.
• This review focuses on transfer learning applications within MR brain imaging.

Purpose of the Study:

• To identify research trends, knowledge gaps, and common strategies in transfer learning for MR brain imaging.
• To categorize existing transfer learning approaches applied to brain MRI.
• To highlight areas for future research, including privacy and handling of unlabeled data.

Main Methods:

• Systematic literature search of transfer learning applications in MR brain imaging.
• Screening of 433 studies for relevance and extraction of key information (task, application, labels, methods).
• Detailed examination of brain MRI-specific methods and general medical imaging challenges (privacy, unseen domains, unlabeled data).

Main Results:

• 129 articles were identified applying transfer learning to MR brain imaging.
• Dementia-related classification and brain tumor segmentation were the most common applications.
• Convolutional neural networks (CNNs) were the predominant machine learning technique; few studies addressed MRI-specific issues or data privacy.

Conclusions:

• There is a growing interest in transfer learning for brain MRI analysis.
• Public datasets and pre-trained CNNs have driven the popularity of applications like Alzheimer's diagnostics and tumor segmentation.
• Further research is needed on the interpretability and comparative analysis of transfer learning strategies in this domain.