All input magnetic resonance imaging data underwent a standardized preprocessing pipeline to ensure consistency and improve model performance. First, N4ITK bias field correction was applied to reduce low-frequency intensity inhomogeneity, thereby enhancing tissue contrast and supporting more stable feature learning (Tustison et al., 2010). Subsequently, non-brain tissues such as the skull and surrounding fat were removed using the Brain Extraction Tool (BET) from FSL (Jenkinson et al., 2012). To optimize memory usage and computational efficiency, each image was tightly cropped around the brain region, and all slices were resampled to a fixed resolution of 512 × 512 pixels.

During training, various data augmentation techniques were applied to improve generalization and mitigate overfitting. Specifically, random geometric transformations including ±15° rotations, horizontal flipping, and vertical flipping were used. To further focus the training on relevant anatomical areas, slices were cropped around non-zero mask regions, ensuring that lesion-containing areas were included in the input. Additionally, bezier curve-based nonlinear intensity distortion was introduced to simulate contrast variability across subjects and scanners, thus improving the model’s robustness to imaging heterogeneity.

All images were standardized using z-score normalization to have zero mean and unit variance. The data were processed on a 2D slice-wise basis. For model input, each sample was constructed as a three-channel input by stacking the center slice with its immediate preceding and succeeding slices.

This input configuration is anatomically motivated, as cerebral microbleeds are small, focal lesions that typically extend across only a limited number of contiguous axial slices. Incorporating adjacent slices allows the model to capture local inter-slice continuity relevant to microbleed detection while remaining within a 2D-based framework. In contrast, using all slices along the z-dimension as input channels would introduce substantial variability in input dimensionality due to differences in slice thickness and acquisition protocols, potentially impairing generalization in multi-center datasets. The specific input configuration and how it integrates with the network are described in the following section.

As illustrated in Figure 1, the proposed model adopts a modified encoder–decoder architecture based on U-Net, designed specifically for the detection of cerebral microbleeds (CMBs). Although CMBs are three-dimensional anatomical entities, they typically appear as small, focal lesions spanning only a limited number of slices along the axial direction. Based on this observation, the network is built upon a 2D-based framework that effectively captures in-plane spatial features while maintaining robustness across heterogeneous imaging protocols.

The network integrates Residual Local Kernel (RLK) blocks and Convolutional Block Attention Modules (CBAM) to enhance discriminative feature learning and suppress irrelevant activations. The model is optimized for the segmentation of small and sparse CMBs, which are often challenging to distinguish due to their similarity to surrounding anatomical structures.

To effectively supervise the segmentation of cerebral microbleeds (CMBs), which are typically small in size and sparsely distributed, we adopt a composite loss function that combines Dice Loss and Focal Loss. This formulation is designed to balance the need for high sensitivity with the suppression of false positives, addressing the class imbalance inherent in lesion segmentation tasks.

The Dice loss ℒ Dice directly optimizes the overlap between the predicted and ground truth masks, and is particularly effective for handling small lesion regions:

where p i and g i denote the predicted probability and ground truth label for voxel i , respectively, and ϵ is a small constant for numerical stability.

To further enhance the model’s focus on hard-to-classify voxels, the Focal Loss ℒ Focal is employed, defined as:

where p t is the predicted probability corresponding to the true class, α t is a balancing parameter, and γ is the focusing parameter. In our experiments, we set γ = 2.0 to emphasize difficult samples and reduce the influence of well-classified background voxels.

Importantly, although subjects without cerebral microbleeds (zero-CMB cases) were not explicitly included during training, the use of Focal loss effectively mitigates the dominance of background voxels by down-weighting easy negative samples. This design encourages the network to focus on hard-to-classify lesion candidates ans duppress spurious false positives, thereby alleviating class imbalance even in the absence of explicit negative-only subjects during training.

The final training loss ℒ total is the weighted sum of Dice and Focal losses, computed across the main output and three auxiliary outputs from the decoder:

where λ k denotes the weight assigned to the output at the k-th decoder level. In our implementation, the weights were set to λ = {0.4,0.3,0.2,0.1} from the highest-resolution (final) output to the deepest auxiliary output, respectively. This weighting scheme places greater emphasis on the final prediction while allowing auxiliary outputs to contribute to training stability through deep supervision.

This study utilized a total of 506 brain MRI scans collected from three sources: the MICCAI2021 Challenge dataset, a multi-center cohort referred to as Cavas, and clinical data from Yongin Severance Hospital. The combined dataset includes both T2*-weighted Gradient Echo (T2*-GRE) and Susceptibility Weighted Imaging (SWI) sequences, reflecting real-world variability in acquisition protocols and scanner manufacturers.For evaluation, a test set comprising 72 subjects was constructed, while the remaining scans were used for training. Notably, the test set included subjects without cerebral microbleeds, ensuring a comprehensive assessment of the model’s ability to distinguish between positive and negative cases.

To evaluate the lesion-level performance of the proposed segmentation model, four standard metrics were employed: precision, recall, F1-score, and the average number of false positives per scan ( F P avg ). These metrics were selected to comprehensively assess both sensitivity and specificity, as well as the practical viability of the model in clinical applications.

All evaluation metrics were computed on a lesion-wise basis, where a detected lesion was considered a true positive if the Euclidean distance between its centroid and the nearest ground truth lesion was within 4 voxels. This threshold was selected to account for small lesion sizes and spatial localization uncertainty during annotation.

To implement and evaluate the proposed model, all experiments were conducted in a controlled computational environment summarized in Table 1.

A warm-up strategy was applied at the beginning of training, where the learning rate was gradually increased from 1e-7 to 1e-4 over the first 10 epochs. After the warm-up phase, a cosine annealing schedule with warm restarts was used to adjust the learning rate dynamically throughout the remaining training epochs.

The model was trained using 2D slice-based input with a batch size of 2 for 1,000 epochs. All experiments were performed on a single NVIDIA RTX A5000 Ada GPU. At inference time, the proposed model required an average of 0.22 s per subject, measured on the test set. This runtime includes forward inference and post-processing for lesion detection.

Early stopping was not applied during training. Instead, model selection was performed based on validation performance within each fold of the K-fold cross-validation framework. Specifically, for each fold, the checkpoint achieving the highest validation F1-score was retained and used for final evaluation. Training and validation curves were monitored to ensure stable optimization, and the best-performing checkpoints were not consistently obtained at the final epoch, indicating no systematic overfitting during prolonged training.

To assess the effectiveness of the proposed method, we conducted a comparative evaluation against existing cerebral microbleed (CMB) detection approaches, including both 3D CNN-based and segmentation-based models. The compared models were trained using the dataset employed in this study, with training conducted under conditions that reflected the environments originally proposed for each model. The evaluation metrics include precision, sensitivity (recall), F1-score, and the average number of false positives per subject ( F P avg ). A detailed summary of these results is provided in Table 2.

The proposed method demonstrates superior performance across all evaluation criteria. In terms of precision, our model achieves a value of 0.891, which is substantially higher than that of Dou et al. (0.521), Fan et al. (0.714), and even more recent methods such as Wei et al. (0.822) and Tsuchida et al. (0.767). This indicates that the proposed model is highly effective at reducing false positive predictions while maintaining accurate identification of true CMBs. Sensitivity also improves noticeably, reaching 0.887, which suggests the model is capable of detecting a large proportion of true lesions. Compared to other methods, which show recall values in the range of 0.730 to 0.863, this represents a meaningful gain in lesion retrieval ability.

The F1-score, which reflects the balance between precision and recall, is also the highest among all compared methods, with our model achieving a score of 0.887. This confirms that the model does not sacrifice sensitivity for precision or vice versa, but rather achieves a strong trade-off between both. Furthermore, the average number of false positives per subject ( F P avg ) is significantly reduced to 0.83, while previous methods report considerably higher values, such as 5.22 for Dou et al., 2.35 for Fan et al., and 1.71 for Tsuchida et al. This dramatic reduction in F P avg highlights the model’s robustness and its ability to suppress incorrect activations.

Overall, the consistent improvements in all metrics underscore the practical advantages of integrating CBAM into the RLK-UNet architecture. The attention mechanism contributes to better localization and discrimination of lesion-relevant features while reducing interference from non-lesion signals. These results strongly support the effectiveness of the proposed method for reliable and accurate detection of CMBs in complex neuroimaging data.

To evaluate the generalizability and robustness of the proposed RLK-UNet with CBAM, we conducted a 5-fold cross-validation on the dataset. Table 3 presents the detailed lesion-level detection performance for each fold in terms of precision, sensitivity (recall), F1-score, and the average number of false positives per subject ( F P avg ). Across all folds, the model consistently demonstrated high detection accuracy with an average precision of 0.891 ± 0.015, sensitivity of 0.887 ± 0.011, and F1-score of 0.887 ± 0.013, while maintaining a low false positive rate of F P avg = 0.83 ± 0.016.

Notably, the variation across folds was minimal, indicating stable performance regardless of data split. Fold 4 achieved the best results across all metrics, with the highest precision (0.910) and F1-score (0.903), as well as the lowest F P avg (0.68). These findings support the robustness of the proposed model architecture and the effectiveness of incorporating CBAM into the skip connections for reducing false positives while maintaining sensitivity. The low standard deviation across metrics further suggests that the model can reliably detect cerebral microbleeds across different subsets of the data.

To assess the performance of our model on small cerebral microbleeds (CMBs), we conducted a comparative analysis using lesions with a maximum diameter of 3 mm. Table 4 summarizes the quantitative results across different methods. Our model achieved the highest performance across all key metrics, with a precision of 0.854, sensitivity of 0.883, and F1-score of 0.869. It also reported the lowest average number of false positives per subject (FP_avg = 0.91), indicating effective suppression of incorrect detections.

Compared to the method by Wei et al., our model showed an F1-score improvement of 0.043 and a substantial reduction in FP_avg from 1.66 to 0.91, demonstrating its ability to detect subtle and low-contrast lesions with higher precision. These results suggest that the attention-based skip connection integrated into our model effectively enhances the network’s focus on small lesion features.

Overall, the proposed model provides accurate and reliable detection performance even for tiny lesions, supporting its potential for clinical application in scenarios where high sensitivity and precision are essential.

To further evaluate the performance of the proposed model, qualitative visualization results are presented in Figures 2, 3. These figures provide representative examples across a diverse range of imaging appearances and lesion characteristics.

Figure 2 illustrates qualitative detection results for multiple subjects, where each column corresponds to an axial slice from a different case. The top row shows the original T2*-GRE or SWI images without annotations, serving as a reference for image contrast and quality. The bottom row overlays the model’s prediction results, with detections color-coded as follows: true positives (TP, blue), false positives (FP, yellow), and false negatives (FN, red).

As observed in Figure 2, the proposed model successfully detects the majority of cerebral microbleeds with high localization accuracy. Most true positive detections are well aligned with the manually annotated ground truth, indicating precise lesion localization. While minor variations in lesion shape or size may be present, the predicted lesions generally exhibit strong visual correspondence with the ground truth annotations. A small number of false positives are observed, typically arising from low-intensity regions or imaging artifacts that visually resemble microbleeds. False negatives are relatively rare and, when present, are primarily associated with extremely small or faint lesions.

While Figure 2 demonstrates the overall detection behavior of the proposed method, Figure 3 provides additional qualitative examples focusing specifically on very small cerebral microbleeds (<3 mm), which are known to be particularly challenging to detect even for experienced human readers. In these cases, lesions occupy only a few voxels and often exhibit low contrast relative to surrounding tissue.

As shown in Figure 3, the proposed model is able to reliably localize the presence of these very small microbleeds, even when slight discrepancies in lesion extent or boundary definition are observed. Although the predicted segmentation may not perfectly match the precise voxel-level annotations, the model consistently identifies the correct anatomical locations of the lesions. This behavior suggests that the proposed approach prioritizes robust lesion detection over exact boundary delineation in extremely small lesions, which is appropriate for automated CMB screening and burden assessment tasks.

Overall, the qualitative results presented in Figures 2, 3 are consistent with the quantitative findings reported earlier, demonstrating that the attention-enhanced architecture effectively captures subtle CMB patterns while suppressing spurious activations, even in challenging small-lesion scenarios.

To evaluate the robustness and generalizability of the proposed method across different imaging modalities, we conducted a cross-modality evaluation between T2*-weighted gradient echo (T2*-GRE) and susceptibility-weighted imaging (SWI), as summarized in Table 5. In this experiment, the model was trained exclusively on one modality and tested on the other, thereby assessing its ability to generalize beyond the training domain.

When trained on T2*-GRE and tested on SWI, the proposed model achieved a precision of 0.758, a sensitivity of 0.771, and an F1-score of 0.765, with an average false positive rate ( F P avg ) of 2.18 per subject. Conversely, training on SWI and testing on T2*-GRE resulted in improved performance, yielding a precision of 0.813, a sensitivity of 0.827, an F1-score of 0.820, and a reduced F P avg of 1.83.

The relatively stronger performance observed when training on SWI can be attributed, at least in part, to the higher number of slices typically available in SWI acquisitions, which provides a larger and more diverse set of training samples. This increased data availability may allow the model to learn more robust representations of susceptibility-related features, which subsequently transfer more effectively to T2*-GRE images despite differences in contrast characteristics and artifact profiles.

Importantly, although a moderate performance gap is observed between the two training configurations, the proposed method maintains stable detection performance in both cross-modality settings. This indicates that the model does not rely on modality-specific cues alone, but instead captures modality-invariant characteristics of cerebral microbleeds. Furthermore, when compared with existing CMB detection approaches reported in Table 2, the proposed method demonstrates competitive or superior performance even under cross-modality evaluation, highlighting its strong generalization capability.

Overall, these results suggest that the proposed attention-enhanced architecture exhibits robust cross-modality generalization between T2*-GRE and SWI, supporting its potential applicability in heterogeneous clinical environments where imaging protocols and acquisition settings may vary.

To evaluate the clinical relevance and consistency of our model’s performance, we analyzed its ability to estimate the number of cerebral microbleeds (CMBs) per subject using four severity intervals defined by the ARIA-H radiographic classification: 0, 1–4, 5–9, and ≥10 (Hampel et al., 2023). This categorization reflects clinically meaningful thresholds, as outlined in anti-amyloid immunotherapy trials where the CMB burden is used to guide treatment eligibility and safety monitoring. Table 6 presents the subject-level confusion matrix based on this classification scheme.

The results show that the model’s predictions are predominantly aligned with the ground truth. Among the 50 subjects with 1–4 true CMBs, 46 (92.0%) were correctly categorized. In the 5–9 group, the majority of cases (6 out of 9) were accurately predicted, while minor misclassifications occurred only in adjacent classes (e.g., 1–4 or ≥10), which are clinically tolerable given the gradational nature of severity. Importantly, no subject without CMBs was mistakenly predicted to have a moderate (5–9) or severe (≥10) lesion burden. Similarly, all subjects with ≥10 CMBs were either correctly classified or placed in the adjacent group (5–9), with no cases underestimated into the 1–4 or 0 category.

Such prediction patterns highlight the model’s ability to avoid extreme overestimation or underestimation of CMB burden—an essential quality for reliable clinical deployment. This behavior suggests that the model is not only detecting individual lesions accurately but also maintaining count-level consistency that reflects the true severity of each subject.

We attribute this strength to the CBAM-based skip connection design integrated into our model. While conventional U-Net skip connections may indiscriminately transfer both relevant and irrelevant features—including vessels or background structures—our model leverages spatial and channel attention to selectively enhance lesion-specific representations. This mechanism appears to suppress false positives and improve true lesion detection across a wide spectrum of lesion counts.

Overall, the results demonstrate that our model maintains a clinically aligned, component-level understanding of lesion burden. This capability is particularly critical in the context of ARIA-H–related safety evaluations, where accurate and conservative quantification of microhemorrhages directly impacts trial eligibility and patient risk stratification.

To further assess the reliability of the proposed method at the subject level, we performed a correlation analysis between the predicted and ground-truth cerebral microbleed (CMB) counts on the test set (n = 72). For each subject, the total number of detected CMBs was compared with the manually annotated CMB count.

As shown in Figure 4, the predicted CMB counts exhibit a strong positive correlation with the ground-truth counts. The Spearman correlation coefficient was ρ  = 0.93, indicating a high degree of monotonic agreement between predicted and true CMB burdens. To quantify the statistical reliability of this correlation, a 95% confidence interval (CI) was computed using Fishter z-transformation, yielding a CI of [0.89–0.96] (p < 0.001).

In addition to the correlation coefficient, the regression line demonstrates a slope close to unity and a near-zero intercept, suggesting that the proposed model does not systematically overestimate or underestimate the CMB burden across subjects. While minor dispersion is observed in low-burden cases, the overall trend remains consistent across the entire burden range.

These results indicate that the proposed method provides stable and reliable subject-level CMB burden estimation, supporting the robustness of the reported performance beyond single point estimates and reinforcing its potential applicability for clinical CMB severity assessment.

To assess the impact of the Convolutional Block Attention Module (CBAM) in our RLK-UNet framework, we conducted a series of ablation experiments focused on both quantitative performance and qualitative attention visualization (Selvaraju et al., 2017).