Anatomical Brain Segmentation for False Positive Reduction and Localization in Automated Detection of Cerebral Microbleeds on Susceptibility-Weighted Imaging

• Introduction: Cerebral microbleeds (CMBs) have distinct clinical implications depending on their anatomical location. While deep learning-based detection achieves high sensitivity, false positives (FPs) remain a major limitation. We propose an anatomical location-aware brain segmentation model that enables both reduction of anatomically invalid FPs and localization of detected CMBs. This dualpurpose approach is designed to enhance both specificity and interpretability of automated CMB detection.
• Methods: This IRB-approved retrospective study utilized susceptibility-weighted imaging (SWI) data. The proposed segmentation model was implemented using 3D nnUNet and trained on 91 internal SWI scans with anatomical labels generated by FreeSurfer. The model segments the brain into lobar, deep, infratentorial, and CMBs-free regions. Validation was conducted on 23 internal and 92 external cases. CMB detection was performed using a 3D U-Net integrated with a Region Proposal Network (RPN), also trained on SWI. FP reduction was applied post-detection to discard candidates located in CMBs-free regions. A 3D U-Net served as the baseline segmentation model for comparison.
• Results: The proposed segmentation model achieved higher dice scores than the baseline across all brain regions: lobar (93.73% vs. 90.89%), deep (88.94% vs. 85.95%), and infratentorial (95.39% vs. 92.51%), with the overall mean dice improving from 89.78% to 92.69%. In internal validation, incorporating the baseline segmentation reduced the FP rate from 2.74 to 2.08 and improved precision from 53.67% to 58.97%, though sensitivity declined from 97.33% to 92.00%. With the proposed segmentation, the FP rate remained at 2.08, but precision further improved to 60.00%, and sensitivity recovered to 96.00%. Localization accuracy improved stepwise from 90.41% (baseline) to 95.89% (proposed). In external validation, the FP rate dropped from 1.76 to 1.13, precision increased from 63.18% to 72.49%, and sensitivity improved from 85.53% to 88.10%. Localization accuracy also increased from 91.73% to 94.96%. These results demonstrate the robustness and generalizability of the proposed anatomical brain segmentation framework across datasets.
• Discussion & Conclusion: Incorporating anatomical brain segmentation into CMB detection significantly enhances specificity by reducing anatomically invalid candidates and enables accurate localization of detected CMBs. This framework emulates expert radiological reasoning, thereby improving the clinical interpretability of CMB detection models and facilitating anatomy-informed analysis in both clinical and research settings.

Figure 1. The proposed framework integrates CMB detection with anatomical brain segmentation for improved false positive reduction and region-wise classification.