DLMUSE: Robust Brain Segmentation in Seconds Using Deep Learning.

Area of Science:

• Medical Imaging and Radiology
• Artificial Intelligence in Medicine
• Neuroscience and Neuroimaging

Background:

• Accurate and efficient brain MRI segmentation is crucial for large-scale neuroimaging research.
• Current segmentation methods can be time-consuming and require significant human supervision.
• Deep learning offers potential for automating complex image analysis tasks in neuroscience.

Purpose of the Study:

• To introduce DLMUSE, an open-source deep learning model for fully automated brain MRI segmentation.
• To enable rapid segmentation for facilitating large-scale neuroimaging studies.
• To provide user-friendly tools for advanced segmentation methods.

Main Methods:

• Developed a deep learning model trained on 1900 diverse MRI scans with multi-atlas and human-supervised labels.
• Validated the model on 71,391 scans across 14 studies using Dice similarity and Pearson correlation.
• Assessed downstream predictive performance for brain age and Alzheimer disease classification.

Main Results:

• DLMUSE achieved high correlation (r=0.93-0.95) and agreement (Dice scores 0.84-0.89) with reference segmentations.
• Brain age prediction and Alzheimer disease classification performance were comparable to reference methods.
• DLMUSE segmentation speed was over 10,000 times faster than the reference method (3.5 seconds vs. 14 hours).

Conclusions:

• DLMUSE enables rapid brain MRI segmentation with performance comparable to state-of-the-art methods.
• The open-source tools and web interface facilitate large-scale neuroimaging research.
• DLMUSE promotes wider utilization of advanced brain segmentation techniques.