# Attention-enhanced segmentation network for automated cerebral microbleed detection and burden assessment

## TL;DR

This paper introduces a new deep learning model for detecting cerebral microbleeds in MRI scans, achieving high accuracy and reducing false positives.

## Contribution

The novel contribution is an attention-enhanced segmentation framework using RLK-UNet with CBAM modules to suppress false positives while maintaining sensitivity.

## Key findings

- The model achieved a precision of 0.891, recall of 0.887, and F1-score of 0.887 with a reduced false positive rate of 0.83 per subject.
- CBAM modules significantly improved precision while preserving sensitivity, as verified through ablation studies.
- Subject-level CMB counts correlated strongly with ground truth (Spearman ρ = 0.93), and severity classification aligned with ARIA-H intervals.

## Abstract

Cerebral microbleeds (CMBs) are small hemorrhagic lesions visible as hypointense foci on susceptibility-sensitive MRI and are established biomarkers of stroke risk and amyloid-related imaging abnormalities (ARIA-H) in patients receiving anti-amyloid therapy. However, automated detection remains challenging because true CMBs closely resemble veins, calcifications, and susceptibility artifacts. This visual ambiguity results in a persistent precision–recall trade-off, where models optimized for high sensitivity tend to generate excessive false positives, while precision-focused models risk missing clinically relevant lesions. To address this limitation, we propose an attention-enhanced segmentation framework designed to suppress confounding activations while preserving lesion sensitivity.

We developed RLK-UNet with Convolutional Block Attention Modules (CBAM), a single-stage encoder–decoder architecture that redefines skip connections as context-filtered pathways. The encoder incorporates large 13×13 residual local kernel (RLK) convolutions to capture broad contextual information for distinguishing spherical microbleeds from elongated vascular structures. CBAM modules are embedded in all skip connections to selectively enhance lesion-relevant features and suppress irrelevant background responses before feature fusion. The model was trained and evaluated on a multi-site dataset of 506 T2*-GRE and SWI scans, with lesion-level detection assessed using precision, recall, F1-score, and average false positives per scan. Subject-level burden estimation was further evaluated using ARIA-H severity intervals.

The proposed model achieved state-of-the-art lesion-level performance, with a precision of 0.891, recall of 0.887, F1-score of 0.887, and a markedly reduced false positive rate of 0.83 per subject. Five-fold cross-validation demonstrated stable performance with minimal variance across splits. In lesions ≤3 mm, the model maintained strong detection performance (F1-score 0.869) while effectively controlling false positives. Cross-modality evaluation between T2*-GRE and SWI confirmed robust generalization. Ablation studies verified that CBAM significantly improved precision while preserving sensitivity, and Grad-CAM visualizations demonstrated more spatially focused and clinically interpretable attention patterns. Subject-level CMB counts strongly correlated with ground truth (Spearman ρ = 0.93), and severity classification aligned with ARIA-H intervals.

RLK-UNet with CBAM provides a robust and interpretable solution for automated CMB detection by directly addressing false-positive propagation through attention-guided skip connections. The framework achieves balanced precision and sensitivity within a single-stage architecture and demonstrates reliable subject-level burden estimation aligned with clinically meaningful ARIA-H categories. These findings support its potential application in vascular risk stratification and treatment monitoring in patients undergoing anti-amyloid therapy.