Biomarkers

Microbleed detection is significant in anti-amyloid therapy (AAT) monitoring to evaluate amyloid-related imaging abnormalities (ARIA) severity (mild: ≤4, moderate: 5-9, severe: ≥10), which directly affects treatment decisions. Although T2*/GRE is recommended for the primary diagnostic imaging sequence of ARIA-H monitoring, the susceptibility-weighted imaging (SWI) sequence is known to be more sensitive to microbleeds. However, manual assessments are time-consuming and prone to reader variability. Deep learning-based automated detection systems can improve the efficiency and reliability of ARIA-H evaluations.

565 SWI MRI scans (2mm slice thickness) from Asan Medical Center were analyzed, comprising 429 positive and 136 negative cases. The mean age of the cohort was 71.9 ± 10.2 years (202 males, 363 females). A neuroradiologist with 14 years' experience labeled microbleeds that were defined as hypointense lesions ranging from 2 to 10mm in diameter on SWI. The dataset was split into training, validation, and test subsets at a 3:1:1 ratio. An Attention U-Net architecture with deep supervision was employed to handle the small size and morphological similarity of cerebral microbleeds. Model validation was performed using the Dice coefficient and the lesion-level Matthews correlation coefficient (MCC).

A total of 114 test scans were evaluated (86 positive, containing 158 microbleeds, and 28 negative) using a 3 mm lesion center proximity threshold. The model detected 146 of 158 microbleeds, achieving an AUC of 0.872 (sensitivity=0.677, specificity=0.893). False-positive analysis revealed 103 occurrences in positive scans and 18 in negative scans. Patient-level metrics included 1.28 microbleeds per scan (95% CI: 1.02-1.54) and 1.06 false positives per scan (95% CI: 0.79-1.37), producing a minimal impact on the ARIA-H radiological severity classification standard (mild: ≤4 microbleeds, moderate: 5-9, severe: ≥10).

This study presents a robust automatic microbleed detection approach using SWI, facilitating ARIA-H assessment in diagnosis and severity categorization. Future work will involve multi-center external validation, supporting additional sequences, and incorporating additional ARIA-related factors for more comprehensive detection.