Deep Learning-Based Automated Detection of Cerebral Microbleeds on Susceptibility-Weighted MRI: Validation Study
Abstract
Background
Cerebral microbleeds (CMBs) are imaging biomarkers of small vessel disease and carry prognostic significance in stroke and dementia. Manual detection is time-consuming and subject to inter-rater variability. We developed and validated a deep learning model for automated CMB detection on SWI-MRI.
Methods
A convolutional neural network (3D U-Net architecture) was trained on 1,240 SWI-MRI scans annotated by two experienced neuroradiologists from three tertiary neurology centres. The model was validated on an independent test set of 310 scans. Performance metrics: sensitivity, specificity, false positive rate per scan, and Dice coefficient. Inter-rater agreement assessed using Cohen's kappa.
Results
The deep learning model achieved sensitivity of 91.3%, specificity of 88.7%, and false positive rate of 1.8 per scan. Dice coefficient was 0.84. Inter-rater kappa for manual annotation was 0.79 (substantial agreement). Model performance was superior in supratentorial regions (sensitivity 93.1%) compared to infratentorial (84.6%).
Conclusion
The proposed deep learning model demonstrates high diagnostic accuracy for automated CMB detection, approaching expert neuroradiologist performance. Integration into clinical radiology workflows could significantly reduce reporting time and improve consistency.
Conflict of Interest
No conflicts of interest. No commercial funding.