Transfer Learning Approaches for Neuroimaging Analysis: A Scoping Review.

Area of Science:

• Artificial Intelligence in Medicine
• Neuroimaging and Machine Learning

Background:

• Deep learning shows moderate success in medical image diagnosis, particularly neuroimaging.
• Transfer learning effectively addresses limited annotated data by transferring knowledge from source to target domains.
• Various transfer learning approaches yield diverse performance outcomes in clinical problem diagnosis, detection, and classification.

Purpose of the Study:

• To review transfer learning approaches, their design attributes, and applications in neuroimaging.
• To identify prevalent research areas and common methodologies within transfer learning for neuroimaging.

Main Methods:

• Systematic review of two major literature databases.
• Inclusion of relevant studies based on predefined criteria.
• Analysis of 50 selected studies focusing on transfer learning in neuroimaging.

Main Results:

• Over half of the reviewed studies focused on transfer learning for Alzheimer's disease, followed by brain mapping and tumor detection.
• ImageNet was the most common source dataset, indicating a preference for pre-trained models.
• Magnetic Resonance Imaging (MRI) was the predominant imaging modality.
• Transfer learning consistently outperformed non-transfer learning methods in diagnosis, classification, and segmentation.
• Fine-tuning all layers or freezing convolutional layers with fine-tuning fully-connected layers showed superior accuracy.

Conclusions:

• Transfer learning significantly enhances diagnostic, classification, and segmentation performance in neuroimaging across various diseases.
• Specific fine-tuning strategies offer superior accuracy and efficiency.
• These advanced transfer learning methods require reduced computational resources and time, making them highly practical.