Cerebral microbleeds (CMBs) are key markers of small-vessel disease and inform eligibility for anti-amyloid therapies, but manual detection can be time-consuming and challenging, particularly for small lesions. We developed a 3D Attention U-Net using three inputs (original SWI, vessel-removed SWI, and FRST map) to automatically detect and count CMBs.
The model was trained and tested on pooled, expert-annotated data from two cohorts, then externally validated on an independent dataset. It demonstrated strong performance across both internal and external validation. The model enables generalizable, scalable CMB detection on SWI, supporting neuroimaging research and therapy-related decision-making.