RatLesNetv2

RatLesNetv2: Fully Convolutional Neural Network for Rodent Brain Lesion Segmentation

RatLesNetv2 performs automated segmentation of rodent brain lesions in magnetic resonance (MR) imaging using a fully convolutional neural network with an autoencoder-like architecture and residual blocks, implemented in Python and PyTorch. It supports end-to-end training directly on 3D T2-weighted MR images without preprocessing.

Key Features:

• Fully Convolutional Architecture: Autoencoder-like ConvNet with residual blocks to improve optimization and segmentation accuracy.
• 3D End-to-End Training: Processes 3D T2-weighted rat brain MRI volumes without preprocessing.
• Quantitative Performance: Achieves higher Dice coefficients, lower Hausdorff distance, and more compact segmentations with fewer holes compared to other ConvNets.
• Large-Scale Evaluation: Validated on 916 T2-weighted rat brain MRI scans from 671 rats across nine lesion stages in focal cerebral ischemia studies.

Scientific Applications:

• Focal Cerebral Ischemia Research: Enables automated lesion segmentation in rodent brain MR images to support neurological studies and drug development.
• Reproducible Quantification: Exceeds inter-rater agreement in manual segmentation, improving precision and reproducibility.

Methodology:

RatLesNetv2 employs a fully convolutional neural network with residual blocks in an autoencoder-like configuration, trained end-to-end on 3D T2-weighted rat brain MRI volumes. Performance was assessed using Dice coefficient and Hausdorff distance metrics on a dataset of 916 scans spanning nine lesion stages.