Convolutional Neural Net Learning Can Achieve Production-Level Brain Segmentation in Structural Magnetic Resonance...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Neuroscience

Background:

• Intra-cranial segmentation from magnetic resonance images (MRIs) is crucial for brain image analysis but is often a time-consuming bottleneck.
• Deep learning, particularly convolutional neural networks (CNNs), shows significant promise in medical imaging applications.

Purpose of the Study:

• To present a method for high-quality, high-throughput intra-cranial segmentation of whole head MRIs.
• To achieve "production-level" segmentation suitable for routine use in processing pipelines.

Main Methods:

• Utilized a deep learning approach with convolutional neural networks.
• Trained and tested the algorithm on an extremely large archive of structural brain MRIs.
• Focused on dataset size, variety, and quality of ground truth, alongside appropriate neural network architecture.

Main Results:

• Achieved consistently high-quality and high-throughput intra-cranial segmentation.
• Demonstrated uniform performance across diverse national imaging cohorts.
• Obtained Dice metric scores exceeding those of other recent deep learning brain extraction methods.

Conclusions:

• The developed method provides production-level intra-cranial segmentation suitable for routine brain image analysis.
• Large and varied datasets play a crucial role in achieving high performance in brain segmentation.
• Further algorithm development may have diminishing returns beyond a certain capability threshold.