Optimization of Deep Learning Based Brain Extraction in MRI for Low Resource Environments.

Area of Science:

• Neuroimaging and Medical Image Analysis
• Artificial Intelligence in Healthcare
• Computational Neuroscience

Background:

• Brain extraction is crucial for neuro-imaging analysis but current methods struggle with pathological brains (e.g., tumors, MS).
• Deep learning (DL) offers advanced solutions but faces barriers in clinical translation due to high computational costs and hardware demands.

Purpose of the Study:

• To explore mathematical optimizations for making DL-based brain extraction methods more efficient.
• To enable the application of DL brain extraction in low-resource clinical environments.
• To evaluate the impact of optimizations on performance and resource utilization for pathological brain extraction.

Main Methods:

• Applied post-training mathematical optimizations and quantization to an existing DL brain extraction model.
• Focused on a model designed for pathologically-affected brains, independent of input imaging modality.
• Evaluated optimizations qualitatively and quantitatively, assessing segmentation performance and computational efficiency.

Main Results:

• Achieved substantial improvements in speedup, latency, throughput, and reduced memory usage.
• Maintained stable segmentation performance, validated by Dice Similarity Coefficient and Hausdorff Distance.
• Optimized models demonstrated comparable accuracy to the original model.

Conclusions:

• Post-training optimizations are effective in enhancing the efficiency of DL brain extraction methods.
• Optimized DL models can run on standard commercial-grade CPUs, overcoming hardware limitations.
• This approach facilitates the clinical translation of advanced neuro-imaging techniques in resource-constrained settings.