MU-Net

MU-Net performs simultaneous skull-stripping and region segmentation on mouse brain MRI using a convolutional neural network to support preclinical neuroimaging analysis.

Key Features:

• Simultaneous processing: Performs skull-stripping and anatomical region segmentation within a single framework.
• Convolutional neural network: Implements a CNN-based approach for automated image segmentation on mouse brain MRI.
• Network architecture: Employs a simplified architecture that includes skip connections and framing connections, with the simplest variant found most effective.
• Performance vs. multi-atlas methods: Demonstrates superior performance compared to state-of-the-art multi-atlas segmentation methods.
• Inference speed and preprocessing: Achieves an inference time of 0.35 seconds and does not require additional preprocessing steps.
• Validation dataset: Validated on 1,782 mouse brain MRI volumes including healthy and Huntington disease models.
• Quantitative accuracy: Reported Dice scores are 0.906 for the striatum, 0.937 for the cortex, and 0.978 for the brain mask.

Scientific Applications:

• Preclinical MRI studies: Automates skull-stripping and segmentation to reduce inter-rater and intra-rater variability in mouse neuroimaging.
• Large-scale and longitudinal studies: Enables efficient processing of large imaging datasets such as longitudinal or multi-center preclinical studies.

Methodology:

MU-Net implements a convolutional neural network that performs simultaneous skull-stripping and segmentation; the architecture includes skip connections and framing connections, was compared against multi-atlas segmentation methods, and was validated on 1,782 mouse brain MRI volumes with measured inference time of 0.35 seconds and reported Dice scores for target regions.