Convolutional neural networks are considered the most appropriate models for medical image segmentation. U-Net successfully segments medical images using an encoder-decoder design and skip-connection architecture. R2U-Net (11) enhances feature representation by utilizing recurrent convolution and residual modules; Attention U-Net (12). Unlapping irrelevant features of skip connections in an attention gating mode. The 3D U-Net proposed by Çiçek et al. directly works for volumetric data by fully leveraging the spatial continuity of 3D scans. However, convolutional neural networks cannot capture the long-term dependency involved in segmenting morphologically complex lesions. Due to the large size variations of the lesions, small lesions do not segment well, and large lesions only partially perform segmentation, which is considered a major problem in segmentation failure.

To adapt to the scale variations of visual targets, inception networks (13), Res2Net (14), and HRNet (15) employ parallel or multi-branch convolutions for feature extraction from 2D classification and detection tasks. While these methods demonstrate excellent performance, they exhibit limitations in stroke lesion segmentation, failing to capture fine-grained lesion features due to a lack of voxel consistency. In our LSC module complete preservation of semantic information through a dual mechanism of global modeling (large-kernel convolutions) and local refinement (small-kernel convolutions).

There are also some works reported in studies suggesting CNN architectures for stroke lesion segmentation. For instance, Dolz et al. (16) presented a dense multi-path CNN that passes features of different scales in parallel. It is able to capture rich 2D features but cannot make good use of the highly 3D spatial context. Another method, ML-Net, is strong in terms of distance-based metrics (HD95) but not stable in the Dice score, especially for small lesions. Unlike FRPNet (17), which employs a feature refinement pyramid to enhance the segmentation accuracy, these methods are based on fixed-scale kernels and thus are limited to dealing with the high heterogeneity of stroke lesion sizes. These methods still lack the adaptability to cover the wide range of lesion sizes. It is still a challenging task to accurately segment both small lesions and large diffuse lesions (18).

Before sequence flattening, the input feature map x is processed by the ASC module. This module is designed to enhance local spatial information that may be lost during the serialization process. Initially, a depth-separable 3 × 3 × 3 convolution is applied to capture direct neighborhood relationships. Then, the SE attention mechanism adapts the channel feature information based on the local spatial context, effectively “preprocessing” the features before passing them to the global Mamba stage. This approach ensures the encoding and preservation of fine details, such as the edges of minute lesions.

Subsequently, a gated convolutional unit and a large kernel (7 × 7 × 7) depthwise convolution further expand the local receptive field, thereby creating a robust feature representation with local perceptual capabilities y.

After local enhancement by ASC, the 3D feature map y is flattened into a one-dimensional sequence y_seq. Standard unidirectional Mamba scans are limited in that they can only capture context from preceding elements. This approach is inadequate for spatial data, where the spatial context is isotropic. The features of a lesion at a given voxel depend on its global context, which must be considered to capture the full spatial relationship. Therefore, we adopted a bidirectional scanning mechanism to process the sequence in both forward and backward directions:

Forward scanning captures dependencies from the “beginning” to the “end” of the flattened sequence, while backward scanning tracks dependencies from the “end” to the “beginning.” By concatenating y_fwd and y_bwd, we ensure that the feature representation at each position is informed by the complete global context from both directions along the scan axis. This bidirectional approach is essential for accurately modeling complex structures, such as stroke lesions.

Finally, the concatenated features are projected back to the original channel dimension, reshaped into 3D volumetric data, and subsequently linked through residual connections to preserve the integrity of the information.

In order to assess LBMNet's performance, we trained the model on the ATLAS v2.0 dataset and compared LBMNet's results with a conventional CNN-based network [U-Net (6), 3DUnet (29), U-Net++ (30)], a Transformer-based network UNETR, TransUNet, SwinUNet, and recently proposed hybrid models [STU-Net (31), FRPNet, UMamba]. As shown in Table 1, LBMNet receives a Dice score of 67.57% as compared to the baseline U-Net (Dice = 48.34%), which improves by 17.51%. In comparison with the previous state-of-the-art UMamba (Dice = 63.31%), LBMNet improves by 4.26%.

This impressive performance boost is due to our proposed architectural design. First, the performance gap with respect to CNN models shows the effectiveness of our LSC encoder. By dynamically modeling features from various receptive fields, the LSC module effectively alleviates the limitation of great variation of lesion-size existing in fixed-kernels CNN models. Second, compared with the Transformer-based models (note they are strong at modeling global context but are likely to ignore fine-grained information), LBMNet shows the effectiveness of our method. Specifically, the cooperation between BSC-Mamba decoder and BAGF skip-connection module is strong. BSC-Mamba decoder is able to model effective global dependencies and local spatial contexts efficiently, while BAGF module can filter and aggregate most informative features from both encoder and decoder adaptively. Thus, the information loss in pure Transformer model is alleviated. The global-connected segmentation result of BSC-Mamba decoder can help to refine the local segmentation results to accurately segment the lesion boundary.

In addition, as can be seen from the four metrics, LBMNet shows the best performance. As shown in Figure 6, the IoU of 52.03% is the largest, which means the segmentation result is closer to the ground truth. What is more interesting is that the F2-score of LBMNet achieves 72.39% and the Recall rate of LBMNet is 69.38%. In a medical application, high recall indicates that the model is unlikely to lead to false negative, which means the model is unlikely to miss the actual lesion area and further misdiagnose the disease. Because the lesion area to segment is very complex, having a globally connected segmentation result helps to minimize the false negatives in local segmentation.

In addition, in Figure 6, we give the qualitative results and the results are visually convincing for the quantitative results above. Specifically, it is obvious that compared with pure CNN-based model, U-Net++, due to the limitation of receptive field, the global continuity of large lesions cannot be well captured, which leads to the undersegmentation. While, TransUNet, although it is better at modeling global information of lesions, tends to smooth the global connection or even omit the detailed boundary information, which leads to unsatisfactory segmentation. Globally connected segmentation result of U-Net++ and local detailed segmentation result of TransUNet are used to refine the final segmentation result of LBMNet.

To evaluate the model's adaptability to lesion size, we specifically categorized the ATLAS v2.0 test set into three groups based on lesion volume: small lesions (< 10 cm³), medium lesions (10–50 cm³), and large lesions (>50 cm³). Due to the limited sample size of the ISLES 2022 dataset, we did not conduct lesion size stratification experiments on it. The detailed results for ATLAS v2.0 are provided in Table 2.

In the segmentation of small lesions, we obtain satisfactory model performance, and the Dice coefficient and recall are 58.47 and 66.52%, respectively. The above results demonstrate that LBMNet can still discover and retain most of the lesion areas for minute targets. The hierarchical strategy of LSC module “coarse-grained capture and then fine-grained refinement” is very important to avoid losing the local details, which are usually ignored by ordinary convolutional layers. As expected, the evaluation metrics increase gradually with the increase of lesion volume. For the medium-sized lesions, the Dice score achieves 70.06%. And for the large lesions, the Dice score can also reach an excellent 74.18%. The above consistent and robust performance for all the lesion sizes further demonstrates the strength of our model on handling such large scale variation for stroke lesions.

The effectiveness of LBMNet is also verified in the comparison shown in Figure 7. Specifically, for the case of small lesions (< 10 cm³). In addition to Umamba, our LBMNet also achieves state-of-the-art performance. It is of great importance that missing small lesions are highly risky, and the effectiveness of our model in maintaining details is remarkable due to its multi-scale feature extraction. For medium-sized lesions, our model still has an advantage over the models, while some models also get similar performance for this range. For large lesions, our model has its own advantage, which shows the ability of our model to integrate global context with local details and keep the segmentation part intact. To rigorously validate the model's robustness on small lesions (< 10 cm³), we conducted a comparative analysis against state-of-the-art methods (Table 3). Consequently, we utilized Recall (Sensitivity) and Hausdorff Distance (HD95) as more clinically relevant metrics to experimentally validate the performance on the small lesion subset. As shown in Table 3, while baseline models like SwinUNet struggle with a low Recall of 47.55%, LBMNet achieves a superior Recall of 66.52%. This indicates a significantly lower rate of missed diagnoses, which is critical for early stroke screening. Furthermore, our HD95 is reduced to 24.12 px (compared to < 29 px for UMamba), demonstrating that LBMNet preserves the topological structure of small lesions more accurately than competitors, even when the Dice metric fluctuates.

To further assess the accuracy of volume segmentation, we conducted a Bland–Altman analysis to evaluate segmentation errors across lesions of various sizes, as illustrated in Figure 8. The Bland–Altman analysis revealed a mean bias of –5.22%, indicating a slight systematic underestimation of lesion volume. The 95% confidence interval ranged from –68.11 to 57.68%, with larger relative deviations mainly observed in small lesions. This bias was most pronounced in small lesions (Figure 8a), where boundary ambiguity likely contributed to larger relative volume discrepancies. However, even within this challenging subgroup, the majority of data points remained within the confidence interval. As the lesion size increased to medium (Figure 8b) and large (Figure 8c), the variability in segmentation error notably decreased, with data points clustering tightly around the mean bias line. This progressively improved consistency underscores the model's reliability for clinically relevant lesions. Overall, the Bland–Altman analysis confirms that LBMNet provides robust and accurate volume assessments across the full range of lesion sizes.

To further analyze the effectiveness of each proposed component, we conducted a series of ablation experiments on the ATLAS v2.0 dataset. These experiments were performed in independent runs and were primarily used to analyze the relative performance contributions of each component. We use 3D U-Net as a baseline model and successively add our approach, namely multi-scale LSC encoder, BSC-Mamba decoder, and adaptive BAGF skip-connection module, to evaluate the improvement in both quantitative and qualitative results shown in Table 6 and Figure 10, respectively.

Our baseline model, the standard 3D U-Net, obtains Dice = 50.06% and IoU = 41.39%. As shown in Figure 10, the model can roughly locate the lesions, but it clearly undersegments them and fails to capture the complicated boundaries of the lesions. These results demonstrate the limitations of traditional convolutional networks in modeling the variety of sizes and irregular shapes of stroke lesions.

To alleviate the limitations of the baseline in multi-scale feature extraction, we replaced the conventional encoder with our proposed LSC encoder (Model 2). The single modification leads to a remarkable improvement in the Dice score (+10.03 percentage points, 60.09%). As shown in Figure 10, the qualitative results also support this improvement, where the model outlines a more complete region of the primary lesion area. These results demonstrate the effectiveness of the LSC module in learning features from various receptive fields, which are crucial for modeling both core and peripheral areas of lesions. However, the model is still unable to detect the lesions in the opposite hemisphere.

To validate the effectiveness of global context modeling, we swapped out the baseline decoder for BSC-Mamba (Model 3). With this setup, we obtained Dice score of 63.11%. From Figure 10, we can see that Model 3 was able to model bilateral lesions, which demonstrated the power of BSC-Mamba to model long-range dependencies and utilize distant spatial information. Meanwhile, the model lost its fine-grained local information and thus led to slight under-segmentation in boundary regions.

After independently validating the effectiveness of the encoder and decoder, we concatenate the encoder and decoder to build Model 4. This new model takes advantage of LSC's ability to extract multi-scale local features and BSC-Mamba's ability to integrate global–local information. The Dice coefficient achieves 65.03%. Visualization results show that the lesion region is more completely captured by Model 4, and the contour is sharper and clearer. However, directly fusing features may cause over-segmentation at blurred boundaries, which means that conflicting or redundant features are probably forwarded via skip connections.

To address the suboptimal feature fusion issue, we further developed the BAGF module to obtain the complete LBMNet (Model 5). Specifically, BAGF module adaptively chooses and fuses salient features from encoder and decoder pathways while suppressing noise and redundant information. With the further improvement of feature fusion, the final model achieves the best segmentation performance as shown in (Model 5: Dice = 67.57%, IoU = 54.97%). As shown in Figure 10, the segmentation masks generated by Model 5 are most accurate. That is, the boundaries of segmentation masks are clear and there is neither over-segmentation nor under-segmentation. The above results also demonstrate that the BAGF module plays a key role in integrating multi-scale and global representations to obtain robust and accurate segmentation results.

To analyze the influence of our particular design choices in the LSC module, we perform an ablation study in Table 7. The baseline model in Table 7 is different from Model 1 in Figure 9. Specifically, the baseline is the complete LBMNet model (Model 5), where we replace its LSC module with a typical 3D residual block with only 3³ convolutions. With this controlled setup, the only part of the network that varies from the baseline is the LSC module, meaning that if the LSC design choices lead to any improvement, this excludes the possibility that other components of the decoder or skip connections might be causing it.

Replacing the baseline 3³ convolutional kernel with a single large kernel (Models A and B) resulted in substantial performance improvements, with gains of +3.94% for 5³ and +6.72% for 7³. These findings confirm the importance of large receptive fields in effectively capturing stroke lesions. However, increasing the kernel size from 7³ to 9³ (Model C) provided only a marginal gain of +0.19% in Dice score, accompanied by higher computational costs. More importantly, a single large kernel struggles to balance global context with local details. In contrast, the dual-branch design, which combines 5³ and 7³ convolutional kernels (Model D), significantly outperformed any single-kernel configuration, achieving a Dice score of 66.17%. This highlights the superiority of explicitly modeling features across multiple scales, rather than relying on a single large convolutional kernel. Our final LSC module design, which incorporates a 3³ refinement path after the dual-branch output (Ours), achieves the highest Dice score of 67.57%. The +1.4% improvement over Model D strongly supports our core “coarse-to-fine” hypothesis: capturing broad context with large kernels, followed by refinement with small kernels, is the optimal strategy.

To rigorously validate the design of our BSC-Mamba module, we conducted two sets of targeted ablation experiments.

First, our goal is to analyze the independent and synergistic advantages of two parts in BSC-Mamba: the ASC module in local enhancement and the bidirectional state-space model in global modeling. We designed four variants on top of a standard convolutional decoder baseline. The baseline (Baseline) uses LBMNet with a basic 3D convolutional decoder (Table 8) shows the necessity of these two parts. With only the ASC module (1.14 Dice gain), the network can still greatly improve local feature representations and thus proves the advantage of its design to preserve local information. With only Bi-directional Mamba (2.22 Dice gain), the network demonstrates the necessity of global context modeling. But the performance of Bi-directional Mamba is limited due to the information loss from sequence flattening. The whole BSC-Mamba that applies ASC in local preprocessing and Bi-SSM in global modeling obtains the best performance.

Next, we evaluated the necessity of the proposed bidirectional scanning strategy by comparing it with alternative scanning methods, as illustrated in the Table 9. Transitioning from unidirectional to bidirectional scanning (our approach) produced a substantial performance improvement, with the Dice score increasing by 1.9%. This result highlights the critical role of reverse contextual information in accurately segmenting non-causal objects. Further extending the approach to quad-directional scanning yielded negligible additional gains in Dice score while introducing higher computational overhead. These findings strongly support the effectiveness of the bidirectional scanning strategy, which captures nearly all essential global information without the unnecessary computational burden of multi-axis scanning.

Test the performance of our proposed BAGF module and its components, we conducted an ablation study as follows: we started with a key baseline model in which we replaced the BAGF module with a concatenation operation following the design of the standard U-Net. Then, we successively added each of our proposed components to this baseline: Enc-Res: residual depthwise separable convolutional blocks added to the encoder path to improve feature representation; Enc-SA: spatial attention operation added to the encoder path; and Dec-CA: channel attention operation added to the decoder path. As illustrated in Figure 11, introducing the Enc-Res component alone improves the Dice score by 1.18%, demonstrating that strengthening local feature representations in the encoder path prior to feature fusion is advantageous. Incorporating the Enc-SA component results in an even greater gain, enhancing the Dice score by 2.02%, which confirms the effectiveness of selectively emphasizing spatially relevant regions within the encoder's coarse feature maps. Applying the Dec-CA component independently also produces a notable improvement, raising the IoU by 1.22%. This outcome highlights its capacity to refine decoder features by emphasizing lesion-relevant channels while suppressing noise-induced oversegmentation. The complete BAGF model achieves the best overall performance, with Dice score of 67.57% and an IoU of 54.97%. These results exceed the additive contributions of individual components, suggesting strong synergistic interactions among them. Collectively, these findings demonstrate that optimizing feature information within the skip-connection framework is essential for achieving accurate stroke lesion segmentation.