Alzheimer's Imaging Consortium.

Area of Science:

• Neuroimaging
• Artificial Intelligence
• Radiology

Background:

• Microbleed detection is crucial for monitoring anti-amyloid therapy (AAT) and assessing amyloid-related imaging abnormalities (ARIA) severity.
• Susceptibility-weighted imaging (SWI) is more sensitive to microbleeds than T2*/GRE, but manual assessment is time-consuming and variable.
• Automated deep learning systems offer improved efficiency and reliability for ARIA-H evaluations.

Purpose of the Study:

• To develop and validate a deep learning model for automated microbleed detection using SWI MRI.
• To enhance the efficiency and reliability of ARIA-H assessment in clinical practice.
• To support accurate severity classification of ARIA for treatment decisions in AAT.

Main Methods:

• Analysis of 565 SWI MRI scans from Asan Medical Center.
• Development of an Attention U-Net architecture with deep supervision for microbleed detection.
• Model validation using Dice coefficient and lesion-level Matthews correlation coefficient (MCC).

Main Results:

• The model achieved an AUC of 0.872 with a sensitivity of 0.677 and specificity of 0.893.
• Detected 146 out of 158 microbleeds in test scans with minimal impact on ARIA-H severity classification.
• Reported 1.28 microbleeds per scan and 1.06 false positives per scan.

Conclusions:

• A robust automated microbleed detection approach using SWI was presented, aiding ARIA-H assessment and severity categorization.
• The developed system facilitates diagnosis and severity classification for AAT monitoring.
• Future work includes multi-center validation and integration of additional ARIA-related factors.