Attention-enhanced segmentation network for automated cerebral microbleed detection and burden assessment
Kwon Hwi Cho1, Jonghyun Jeon2, Seonggyu Kim3
1Department of Artificial Intelligence, Hanyang University, Seoul, Republic of Korea.
Frontiers in Neuroscience
|March 20, 2026
Summary
This study introduces RLK-UNet with attention mechanisms to accurately detect cerebral microbleeds (CMBs) on MRI scans. The model improves precision and recall, aiding stroke risk assessment and monitoring anti-amyloid therapy.
Area of Science:
- Neuroimaging
- Artificial Intelligence
- Medical Diagnostics
Background:
- Cerebral microbleeds (CMBs) are crucial biomarkers for stroke risk and amyloid-related imaging abnormalities (ARIA-H), particularly in patients undergoing anti-amyloid therapy.
- Automated CMB detection is challenging due to visual similarities with other brain structures and artifacts, leading to a precision-recall trade-off in existing models.
Purpose of the Study:
- To develop an attention-enhanced deep learning framework for robust and accurate automated detection of cerebral microbleeds (CMBs).
- To address the limitations of current methods by improving both sensitivity and precision in CMB identification, thereby reducing false positives.
Main Methods:
- Development of RLK-UNet, an encoder-decoder architecture incorporating Convolutional Block Attention Modules (CBAM) and residual local kernel (RLK) convolutions.
- Training and evaluation on a multi-site dataset of 506 T2*-GRE and SWI scans, assessing lesion-level and subject-level performance metrics.
- Utilizing attention mechanisms within skip connections to filter irrelevant information and enhance focus on relevant lesion features.
Main Results:
- The RLK-UNet model achieved state-of-the-art performance with a precision of 0.891, recall of 0.887, and F1-score of 0.887, alongside a significantly reduced false positive rate of 0.83 per subject.
- The model demonstrated robust generalization across different MRI modalities (T2*-GRE and SWI) and maintained strong performance for small lesions (≤3 mm).
- Ablation studies confirmed the efficacy of CBAM in improving precision and sensitivity, with Grad-CAM visualizations showing interpretable attention patterns.
Conclusions:
- RLK-UNet with CBAM offers a reliable and interpretable solution for automated CMB detection, effectively balancing precision and sensitivity.
- The framework's ability to provide accurate subject-level burden estimation supports its clinical utility in vascular risk stratification and treatment monitoring for patients on anti-amyloid therapy.
