Background: Cerebral microbleeds (CMBs) are small, round or ovoid paramagnetic foci of hemosiderin that appear as signal loss on susceptibility-weighted MRI. They are markers of small-vessel disease, associated with future intracerebral hemorrhage and cognitive decline, and influence decisions about antithrombotic and anti-amyloid therapies. Susceptibility-weighted imaging (SWI) provides high sensitivity to CMB but also introduces challenges, such as blooming, venous or calcification mimics. Manual detection is labor-intensive, shows only moderate inter-reader agreement, and misses a proportion of small lesions, motivating robust automation. Hypothesis: We hypothesize that a machine learning model can achieve high performance in automatic CMB detection compared to traditional methods, with sensitivity exceeding 80%. Methods: We developed a fully automated SWI CMB-detection pipeline. Data comprised 952 cases of patients presenting with stroke or transient ischemic attack in Alberta (384 positive, 568 negative) with annotations from experienced observers. Preprocessing included skull stripping, N4 bias-field correction, and intensity normalization. Volumes were split into 12 patches fed to the model with an 80/20 train-test split. The network is a 3D attention U-Net with five encoder–decoder levels and a bottleneck, optimized using binary cross-entropy and focal loss. Performance was reported at two levels: case-level, which assessed how well the model classifies positive and negative cases, and cluster-level, which assessed how well the model detects each CMB. Results: A total of 76 positive and 78 negative unseen cases were evaluated for overall detection performance. The median number of CMBs in positive cases was 3 (IQR 2-8). For the case-level classifier, the model performed exceptionally well with 89.5% sensitivity and 76.9% specificity. In terms of cluster-wise performance, the model detected with high sensitivity, as shown in Table 1, for both large and small CMBs. However, there were still some false positives that reduced the model's precision, as seen in Fig. 2 and the sample images in Fig. 3. Conclusion: A 3D attention U-Net trained on standardized SWI demonstrates promising, fully automated CMB screening performance, which could potentially reduce reader burden. With the sensitivity of both classification and cluster-wise performance higher than 80%, this approach has the potential to support large-scale research cohorts and clinical decisions.
Charatpangoon et al. (Thu,) studied this question.
Synapse has enriched 5 closely related papers on similar clinical questions. Consider them for comparative context: