Deep Learning with Anatomical Guidance for Enhanced Cerebral Microbleeds Detection and Localization

Object Cerebral microbleeds (CMBs) have distinct clinical implications depending on their anatomical location. While deep learning–based detection has demonstrated high sensitivity, false positives (FPs) remain a major limitation. To address this, we extend the anatomical location–aware brain segmentation approach of J.-H. Kim et al.1) by improving segmentation performance, which in turn enhances FP reduction and localization accuracy within the detection framework. Our method relies solely on susceptibility-weighted imaging (SWI) data.
Methods This IRB-approved retrospective study utilized SWI data. The anatomical segmentation model was implemented with 3D nnU-Net and trained on 91 internal SWI scans with anatomical labels from FreeSurfer. The network partitions the brain into lobar, deep, infratentorial, and CMB-free regions, enabling location-aware analysis. For CMB detection, a 3D U-Net with a Region Proposal Network was employed to identify candidate lesions. FP reduction was performed by discarding candidates within CMB-free regions. A conventional 3D U-Net served as the baseline segmentation model. The overall workflow is illustrated in Figure 1.
Results The proposed segmentation achieved higher Dice scores than baseline across lobar (93.73% vs. 90.89%), deep (88.94% vs. 85.95%), and infratentorial (95.39% vs. 92.51%) regions, raising the mean from 89.78% to 92.69% (Table 1, Figure 2). In internal validation, baseline segmentation reduced the FP rate from 2.74 to 2.08 and increased precision from 53.67% to 58.97%, though sensitivity declined. With our enhanced segmentation, the FP rate remained 2.08, precision further rose to 60.00%, and sensitivity recovered to 96.00%, while localization accuracy improved from 90.41% to 95.89%. In external validation, our method reduced the FP rate from 1.76 to 1.13, increased precision from 63.18% to 72.49%, and improved sensitivity from 85.53% to 88.10%, achieving localization accuracy of 94.96% (Table 2).
Conclusions Incorporating anatomical segmentation into CMB detection significantly reduces anatomically implausible FPs, enhances localization accuracy, and thereby improves both specificity and clinical interpretability.