Fully automated whole brain segmentation from rat MRI scans with a convolutional neural network

Area of Science:

• Neuroimaging
• Artificial Intelligence
• Biomedical Engineering

Background:

• Automated whole brain delineation (WBD) is crucial for neuroimaging analysis.
• Existing WBD techniques lack accuracy and generalizability for preclinical MRI data with neuropathology.
• Conditions like organophosphate intoxication (OPI) and Alzheimer's Disease (AD) cause deformations challenging current WBD methods.

Purpose of the Study:

• To develop and evaluate a modified 2D U-Net framework for accurate and automated whole brain delineation (WBD) in rodent MRI scans.
• To assess the generalizability of the U-Net model across different rat models with neurological conditions.
• To overcome the limitations of manual segmentation and improve WBD efficiency.

Main Methods:

• A modified 2D U-Net with 27 convolutional layers, batch normalization, dropout, and data augmentation was employed.
• Training involved 125 T2-weighted 7.0T MRI scans from an OPI rat model.
• Testing and validation utilized 20 OPI and 120 AD rat model scans, with performance measured by Dice Coefficients (DC) and Hausdorff Distances (HD).

Main Results:

• The U-Net achieved high accuracy, with a median Dice Coefficient (DC) of 0.984 for OPI rats and a mean DC of 0.975 for AD rats.
• Hausdorff Distances (HD) were low, measuring 1.69 mm for OPI and 1.49 mm for AD rat models.
• The segmentation process was rapid, with a computational speed of 8 seconds per scan.

Conclusions:

• The modified 2D U-Net offers a fully automated, efficient, and generalizable solution for WBD in preclinical rodent MRI.
• The approach demonstrates robustness across different rat strains and longitudinal changes.
• This method significantly overcomes the limitations of manual segmentation for neuroimaging analysis in disease models.