NHP

NHP applies transfer learning with a U-Net convolutional neural network to perform brain extraction of non-human primate MRI, improving preprocessing accuracy for NHP neuroimaging studies.

Key Features:

• Transfer Learning Framework: U-Net is pretrained on large human imaging datasets and fine-tuned on NHP data to mitigate limited sample sizes.
• Multi-Site Generalizability: Training incorporates datasets from multiple research sites within the Primate Data-Exchange (PRIME-DE) consortium to enhance robustness across heterogeneous acquisition parameters.
• Superior Performance: The transfer-learning model converges faster and achieves higher accuracy than a U-Net trained exclusively on NHP data and outperforms AFNI, FSL, and FreeSurfer brain extraction routines across PRIME-DE samples.
• Computational Efficiency: Processing time is reduced to approximately 20 seconds to 10 minutes per scan.
• Cross-Species Adaptability: The model can be updated for scans from macaques, chimpanzees, marmosets, and other mammals such as pigs.

Scientific Applications:

• Neuroimaging Research: Enables high-quality brain extraction for downstream analyses in non-human primate MRI studies.
• Comparative Neuroscience: Supports multi-species analyses across primate models and other mammals for comparative studies.
• Data Standardization: Facilitates standardized preprocessing across multi-site neuroimaging projects.

Methodology:

The approach uses transfer learning: U-Net is pretrained on extensive human imaging data, then fine-tuned with NHP-specific data from PRIME-DE, with iterative updating using additional datasets to improve generalizability.