Semi-supervised deep learning of brain tissue segmentation.

Area of Science:

• Neuroimaging
• Computational Neuroscience
• Machine Learning

Background:

• Brain image segmentation is crucial for clinical applications and neuroscience research.
• Deep neural networks (DNNs) show promise but require extensive manual annotations.
• Manual annotation of 3D brain images is laborious and can lack consistency due to image variations.

Purpose of the Study:

• To propose a semi-supervised learning framework for DNN-based brain image segmentation.
• To reduce the reliance on large, expert-annotated datasets.
• To improve segmentation accuracy and stability using both labeled and unlabeled brain images.

Main Methods:

• Developed a semi-supervised learning framework leveraging image registration.
• Utilized a small set of labeled images (atlases) and a large set of unlabeled images.
• Generated pseudo-labels for unlabeled images via registration to train the DNN.

Main Results:

• The proposed method achieved superior segmentation results compared to existing methods.
• Demonstrated more stable segmentation performance across different datasets.
• Successfully applied to both human and marmoset brain image datasets.

Conclusions:

• The semi-supervised framework effectively enhances DNN-based brain image segmentation.
• Image registration-based pseudo-labeling is a viable strategy to utilize unlabeled data.
• The method offers a more efficient and robust approach for brain image analysis.