Using transfer learning for automated microbleed segmentation.

Area of Science:

• Neuroimaging
• Medical image analysis
• Neurology

Background:

• Cerebral microbleeds are small hemorrhages indicating cerebrovascular pathology and dementia risk.
• Current identification relies on manual segmentation of MRI, which is time-consuming and variable.
• Existing automated methods struggle with false positives, limiting their clinical utility.

Purpose of the Study:

• To develop an automated, precise microbleed segmentation tool.
• To utilize standardizable MRI contrasts for improved harmonization across scanners.
• To overcome limitations of manual segmentation and existing automated techniques.

Main Methods:

• A ResNet50 network was trained using transfer learning on T1-weighted, T2-weighted, and T2* MRIs.
• Morphological operators and rules were applied to reduce false positives.
• The system was trained and validated using manual microbleed segmentations from 78 participants.

Main Results:

• The automated method achieved high patch-level performance: 99.57% sensitivity, 99.16% specificity, and 99.93% accuracy.
• Per-lesion analysis showed high sensitivity (89.1-100%) across different brain regions (cortical GM, deep GM, WM).
• The system demonstrated excellent performance in deep gray matter (100% sensitivity, precision, and Dice index).

Conclusions:

• The developed automated method is highly sensitive for microbleed detection.
• This tool offers a more precise and efficient alternative to manual segmentation.
• The approach has the potential to improve the diagnosis and monitoring of cerebrovascular diseases.