3D U-Net Improves Automatic Brain Extraction for Isotropic Rat Brain Magnetic Resonance Imaging Data.

Area of Science:

• Neuroimaging
• Medical Image Analysis
• Computational Neuroscience

Background:

• Manual brain extraction in rodent MRI is labor-intensive and limits high-resolution data analysis.
• Previous 2D deep learning models for brain extraction do not leverage 3D spatial context.
• Accurate brain segmentation is crucial for quantitative analysis in rodent neuroimaging studies.

Purpose of the Study:

• To develop and validate a 3D U-Net framework for automated rodent brain extraction from MRI data.
• To compare the performance of the 3D U-Net model against existing rodent brain extraction tools.
• To provide a robust and efficient tool for pre-processing high-resolution 3D rodent brain MRI data.

Main Methods:

• A 3D U-Net deep learning architecture was implemented, replacing 2D convolutions with 3D counterparts.
• The model was trained and validated on a rat brain MRI dataset (CAMRI) with isotropic resolution (T2-weighted and T2*-weighted images).
• Performance was evaluated using metrics including Dice, Jaccard, center-of-mass distance, Hausdorff distance, and sensitivity, and compared to other tools.

Main Results:

• The 3D U-Net model demonstrated superior performance across all evaluated metrics compared to 2D U-Net and other existing methods.
• The model showed reliability under varying noise levels and identified optimal training sample sizes.
• Qualitative metrics confirmed the improved accuracy of the 3D U-Net for rodent brain boundary identification.

Conclusions:

• The proposed 3D U-Net framework offers an accurate and automated solution for rodent brain extraction in MRI.
• This deep learning approach streamlines pre-processing, reduces human bias, and benefits the rodent MRI research community.
• The developed source code is publicly available to facilitate broader adoption and research advancement.