Attention-Enhanced Segmentation Network for Automated Cerebral Microbleed Detection and Burden Assessment

Cho Kwonhwi¹Jonghyun Jeon²Seonggyu Kim¹Young Seo KIM²Yu-Mi Kim²Mi Kyung Kim²Min-Ho Shin³Insung Chung⁴Sang Baek Koh⁵Hyeon Chang Kim⁶Chae Jung Park⁶Jong-Min Lee^1*

• ¹Hanyang University, Seoul, Republic of Korea
• ²Hanyang University College of Medicine, Seongdong-gu, Republic of Korea
• ³Chonnam National University Medical School, Gwangju, Republic of Korea
• ⁴Keimyung University School of Medicine, Dalseo-gu, Republic of Korea
• ⁵Yonsei University Wonju College of Medicine, Wonju-si, Republic of Korea
• ⁶Yonsei University College of Medicine, Seodaemun-gu, Republic of Korea

This is a provisional file, not the final typeset article Cerebral microbleeds (CMBs) are small hemorrhagic lesions visible as hypointense foci on susceptibility-sensitive MRI and serve as critical biomarkers for stroke risk and amyloid-related imaging abnormalities (ARIA-H) in anti-amyloid therapy. Accurate automated detection remains challenging because true CMBs closely resemble veins, calcifications, and other artifacts, creating a persistent trade-off: higher recall often leads to more false positives, while improved precision risks missing lesions. We propose RLK-UNet with Convolutional Block Attention Modules (CBAM), a single-stage segmentation framework that redefines skip connections as context-filtered pathways. Large-kernel residual local convolutions capture broad contextual cues, while CBAM selectively suppresses noise and amplifies lesion-specific features before fusion, directly addressing the root cause of false-positive propagation. Evaluated on a multi-site dataset of 506 T2*-GRE and SWI scans, our model achieved state-of-the-art performance with a precision of 0.891, recall of 0.887, and only 0.83 false positives per scan. Ablation studies confirmed CBAM's role in improving precision without compromising recall, and Grad-CAM visualizations demonstrated interpretable feature selection. Importantly, subject-level predictions aligned with ARIA-H severity categories, highlighting the framework's clinical relevance for treatment monitoring. These findings indicate that RLK-UNet with CBAM provides a robust, interpretable, and clinically meaningful solution for reliable CMB detection.