Volumetric white matter tract segmentation with nested self-supervised learning using sequential pretext tasks.

Area of Science:

• Neuroimaging
• Computational Neuroscience
• Medical Image Analysis

Background:

• White matter (WM) tract segmentation using diffusion MRI (dMRI) is crucial for analyzing brain development, function, and disease.
• Deep learning significantly enhances WM tract segmentation accuracy.
• Acquiring large datasets of manual WM tract delineations for training deep learning models is challenging and often infeasible.

Purpose of the Study:

• To develop a deep learning-based WM tract segmentation method that performs effectively with limited annotated training data.
• To leverage abundant unannotated dMRI data through self-supervised learning to overcome data scarcity.

Main Methods:

• Explored self-supervised learning by pretraining deep networks on pretext tasks using unannotated dMRI data.
• Designed two pretext tasks: predicting fiber streamline density maps and mimicking registration-based segmentation.
• Developed a nested self-supervised learning strategy combining both pretext tasks before fine-tuning on limited annotated data.

Main Results:

• The proposed self-supervised learning approach significantly improved WM tract segmentation performance.
• The method demonstrated effectiveness even when only a small number of annotated scans were available.
• Experiments on the Human Connectome Project dataset validated the approach's utility.

Conclusions:

• Self-supervised learning offers a viable solution for deep learning-based WM tract segmentation in low-annotation scenarios.
• Pretraining with pretext tasks effectively transfers knowledge from unannotated dMRI data.
• This approach enhances the accessibility and applicability of advanced dMRI analysis techniques.