Abstract Number: Esoc2026a809 Automatic Cerebral Microbleed Detection and Localization on Swi Using 3d Attention U-Net With Mars Region-Based Evaluation

Abstract Background and aims Cerebral microbleeds (CMBs) on susceptibility-weighted imaging (SWI) are important markers of cerebral small-vessel disease associated with future intracerebral hemorrhage and cognitive decline and inform decisions regarding antithrombotic and anti-amyloid therapies. However, CMB assessment remains challenging in routine practice due to variability in reader expertise. To address this gap, we developed an automated CMB detection pipeline for SWI. Methods Pooled SWI sequences from an Albertan population-based stroke-TIA cohort (ACTIVATION, n=952) and the Alteplase compared to Tenecteplase (AcT) stroke trial (n=134) with expert CMB annotations were split 70/15/15 into training, validation, and test sets. Preprocessing generated three image channels (SWI, vessel-removed SWI, and fast radial symmetry transform FRST maps), which were input to a 3D attention U-Net with four encoder-decoder. Test-set performance was evaluated at the case-level (CMB present), cluster-level (lesion detection), and by Magnetic Resonance Anatomical Region Segmentation (MARS) territories. Results In the test set (n=163), 41% were CMB-positive, with a median of 3 CMBs per positive case (IQR1-7). Case-level sensitivity/specificity were 80.6%/81.3%. Cluster-level sensitivity/precision were 80.1%/75.1%, with 0.56 false-positive clusters per case. Territory-level performance (MAE-bias and sensitivity-specificity) is summarized in Figure1, showing a location-based trade-off between sensitivity and specificity/MAE, with both sensitivity/MAE highest for lobar. An example case is shown in Figure2. Conclusions The results demonstrated stable automated CMB screening on SWI, achieving 80% sensitivity at both case and cluster levels. MARS region-based evaluation identified higher sensitivity and MAE in lobar. With further refinement and validation, this approach may support large-scale cohort studies and clinical workflows. Conflict of interest All authors: nothing to disclose Figure 1 - belongs to Results