Bayesian convolutional neural network based MRI brain extraction on nonhuman primates.

Area of Science:

• Neuroimaging
• Computational Neuroscience
• Medical Image Analysis

Background:

• Brain extraction (skull stripping) is critical for neuroimaging analysis.
• Existing automated methods struggle with nonhuman primate MRI data.
• Accurate brain extraction is essential for reliable neuroscience research.

Purpose of the Study:

• To develop a fully-automated pipeline for accurate brain extraction in nonhuman primates.
• To improve upon existing methods that are unsatisfactory for nonhuman primate neuroimaging.
• To provide a robust tool for neuroscience research involving nonhuman primates.

Main Methods:

• A novel pipeline combining a deep Bayesian convolutional neural network (CNN), Bayesian SegNet, and fully connected 3D conditional random field (CRF).
• Bayesian SegNet provides probabilistic, high-resolution segmentation and measures model uncertainty using Monte Carlo dropout.
• Fully connected 3D CRF refines segmentation results within the entire 3D brain volume context.

Main Results:

• The proposed method achieved a mean Dice coefficient of 0.985 and a mean average symmetric surface distance of 0.220 mm on a dataset of 100 nonhuman primates.
• Outperformed six popular brain extraction packages and three established deep learning methods, verified by statistical significance (p < 10^-4).
• Model uncertainty was quantified (mean value of 0.116) and shown to correlate with training set size and label consistency.

Conclusions:

• The developed pipeline offers superior performance for nonhuman primate brain extraction compared to existing methods.
• The probabilistic nature of the method allows for uncertainty estimation, crucial for evaluating segmentation reliability.
• The findings advance automated neuroimaging analysis for nonhuman primates, supporting broader neuroscience investigations.