Mouse Brain Extractor: Brain segmentation of mouse MRI using global positional encoding and SwinUNETR.

Area of Science:

• Neuroimaging
• Machine Learning
• Biomedical Image Analysis

Background:

• Automating brain extraction in human MRI is advanced, but challenging for mouse models.
• Manual segmentation of mouse brains is labor-intensive and introduces variability.
• Accurate automated brain segmentation is crucial for studying mouse models of neurological disorders.

Purpose of the Study:

• To develop a robust deep learning method for automated mouse brain extraction from MRI.
• To improve segmentation accuracy and reduce manual correction needs.
• To address challenges posed by scale variance in heterogeneous mouse MRI datasets.

Main Methods:

• Adapted SwinUNETR architecture with Global Positional Encoding (GPE) for enhanced spatial information.
• Trained and tested the Mouse Brain Extractor on a diverse dataset of 223 mouse MRI scans (in vivo and ex vivo).
• Compared performance against seven conventional rodent brain extraction methods and state-of-the-art deep learning models (nnU-Net, SwinUNETR).

Main Results:

• Achieved high accuracy with average Dice scores of 0.98 and HD95 measures around 100 μm against manual segmentations.
• Significantly outperformed conventional methods in statistical analyses.
• Performed comparably to or better than nnU-Net and SwinUNETR on heterogeneous datasets.

Conclusions:

• Global Positional Encoding (GPE) effectively enhances contextual information for improved mouse brain segmentation.
• The Mouse Brain Extractor offers a competitive and accurate solution for automated brain extraction in diverse mouse MRI data.
• This method has the potential to accelerate research using mouse models of brain disorders by reducing manual segmentation efforts.