Automated Detection of Candidate Subjects With Cerebral Microbleeds Using Machine Learning

Area of Science:

• Neuroimaging
• Medical Image Analysis
• Machine Learning

Background:

• Cerebral microbleeds (CMBs) are small lesions detected on MRI scans.
• Manual identification of CMBs in large datasets is time-consuming and challenging.
• Automated CMB detection requires accurate training and testing datasets.

Purpose of the Study:

• To develop a computationally light, machine learning-based pipeline for preselecting subjects with potential CMBs.
• To improve the efficiency of creating datasets for training and validating automated CMB detection algorithms.

Main Methods:

• A machine learning pipeline was developed to detect CMB candidate subjects.
• The pipeline was evaluated on three diverse datasets, including UK Biobank and clinical data, using SWI or GRE images.
• Performance was compared against alternative preselection methods.

Main Results:

• The developed pipeline achieved subject-level detection accuracy exceeding 80% across all evaluated datasets.
• The method demonstrated good generalizability, maintaining consistent accuracy (>80%) across different imaging modalities and scanners.
• The approach proved effective in preselecting subjects for manual CMB labeling.

Conclusions:

• A novel, efficient machine learning pipeline can accurately preselect subjects with cerebral microbleeds.
• This method significantly aids in the creation of robust datasets for training and testing automated CMB detection tools.
• The pipeline's generalizability across datasets and modalities makes it a valuable tool for large-scale neuroimaging studies.