Scaling Synthetic Brain Data Generation.

Area of Science:

• Neuroimaging
• Artificial Intelligence
• Computational Neuroscience

Background:

• Limited high-quality datasets hinder deep learning in neuroimaging.
• On-the-fly synthetic data generation is computationally intensive.
• Pre-generated data is inflexible and incurs high storage costs.

Purpose of the Study:

• Introduce Wirehead, a scalable in-memory pipeline for efficient on-the-fly synthetic data generation.
• Improve performance and reduce computational demands for deep learning in neuroimaging.
• Address data availability and storage challenges in the field.

Main Methods:

• Developed a scalable in-memory data pipeline (Wirehead).
• Decoupled data generation from training using parallel processes.
• Utilized MongoDB for efficient handling of large datasets.
• Evaluated Wirehead with SynthSeg for synthetic brain segmentation data generation.

Main Results:

• Achieved near-linear performance gains with parallel generators.
• Demonstrated a 15.7x throughput increase with 16 generators.
• Reduced model training time from 7 days to 9 hours with 20 generators.
• Successfully handled terabytes of data, minimizing storage costs.

Conclusions:

• Wirehead significantly accelerates deep learning experimentation cycles in neuroimaging.
• The pipeline offers a flexible, modular, and fault-tolerant solution.
• Enables more ambitious neuroimaging research through distributed deep learning.