MedicDeepLabv3+

MedicDeepLabv3+ performs automated segmentation of cerebral hemispheres in rat magnetic resonance imaging (MRI) volumes to support analysis of ischemic lesions.

Key Features:

• CNN architecture: A convolutional neural network based on DeepLabv3+ with an enhanced decoder that integrates spatial attention layers and additional skip connections.
• Automatic segmentation: Segments cerebral hemispheres and contralateral hemisphere masks in rat MRI scans without requiring preprocessing steps such as bias-field correction or registration to a template.
• Speed: Processes each scan in under one second.
• Adaptive resource utilization: Operates with adjustable GPU memory requirements to accommodate different computational environments.
• Segmentation accuracy: Reported average Dice coefficient of 0.952 for the brain region and 0.944 for the contralateral hemisphere.
• Robustness: Maintains satisfactory performance under constrained GPU memory and limited training data.

Scientific Applications:

• Rodent neuroimaging: Automates brain mask and contralateral hemisphere extraction to reduce manual workload in rat neuroimaging studies.
• Ischemic lesion analysis: Facilitates detailed analysis of ischemic lesions across different lesion stages in rat MRI volumes.

Methodology:

MedicDeepLabv3+ was optimized alongside six other state-of-the-art CNNs (DeepLabv3+, UNet, HighRes3DNet, V-Net, VoxResNet, and Demon) using a diverse training set from 11 cohorts that encompassed MR volumes acquired at various ischemic lesion stages. Its performance was evaluated against other methods, including rodent MRI skull-stripping approaches RATS and RBET, on a dataset of 655 rat brain MRI volumes.