3D stroke lesion segmentation method based on multi-scale convolution and bidirectional mamba network

Abstract

The application discloses a 3D cerebral apoplexy lesion segmentation method based on a multi-scale convolution and a bidirectional Mamba network, relates to the field of computer vision and medical image processing, and solves the problems of limited feature extraction receptive field, high global modeling calculation complexity and semantic conflict during feature fusion in existing cerebral apoplexy lesion segmentation. The method comprises the following steps: acquiring brain MRI image sequence data to be evaluated and performing preprocessing; constructing a multi-scale feature encoder for outputting four layers of encoding feature maps with different resolutions; constructing a bidirectional global-local state space decoder, performing bidirectional adaptive gated feature fusion on the fused branch information, and conveying the information to a convolution layer for channel adjustment; mapping the feature map output by the decoder into a segmentation probability map through the convolution layer; and adopting a hybrid loss function to train the network, combining a Dice loss and a weighted cross-entropy loss to minimize the loss function through a back propagation algorithm, updating network parameters, and stopping until the model converges.

Description

Technical Field

[0001] This invention relates to the fields of vehicle computer vision and medical image processing technology, specifically to a 3D stroke lesion segmentation method based on multi-scale convolution and bidirectional Mamba networks. Background Technology

[0002] Stroke is a leading cause of death and long-term disability worldwide. In clinical practice, automated lesion segmentation based on 3D magnetic resonance imaging (MRI) is crucial for quantifying lesion volume, developing treatment plans, and assessing prognosis. However, stroke lesions exhibit extremely high morphological heterogeneity: their volume varies widely, ranging from tiny ischemic points to large hemorrhages; and lesion boundaries are often indistinct and their shapes highly irregular. This high degree of complexity presents significant challenges to the practical application of existing automated segmentation technologies, primarily due to the following technical bottlenecks: The limitations of traditional convolutional neural networks (CNNs) in multi-scale lesion capture: CNN methods, represented by U-Net and its variants (such as 3DU-Net, U-Net++), are currently the mainstream approach. However, these methods primarily rely on fixed-size convolutional kernels to extract features and suffer from the following problems: Limited receptive field: Fixed local receptive fields are difficult to effectively capture large-scale global contextual information, resulting in "holes" or incomplete segmentation of diffuse or large-volume lesions.

[0003] Insensitive to small lesions: When dealing with small lesions, the lack of a multi-scale feature refinement mechanism leads to the loss of features after multiple downsampling steps, resulting in a high false negative rate. Although some studies have attempted to use multi-path CNNs (such as Dense Multi-path CNN), they often focus on 2D features and fail to fully utilize the spatial continuity of 3D volumetric data.

[0004] Computational and Optimization Challenges of Transformer Architecture in 3D Medical Imaging: To address the limited receptive field of CNNs, the Transformer was introduced to capture long-range dependencies. However, in stroke segmentation tasks, the Transformer faces two main challenges: High computational complexity: The Transformer's self-attention mechanism has quadratic computational complexity. Processing high-resolution 3D voxel data consumes enormous computational resources, making it difficult to meet the needs of rapid clinical processing.

[0005] Poor boundary delineation performance: Transformers typically require massive amounts of data for pre-training. However, stroke MRI datasets are relatively limited, and lesion boundaries are often ambiguous. In such cases, Transformers are prone to overfitting, resulting in lower segmentation accuracy for lesion edges compared to CNNs, often exhibiting jagged edges or localization errors.

[0006] The existing state-space model (Mamba) suffers from structural corruption; in recent years, the Mamba architecture based on the state-space model (SSM) has been criticized for its linear complexity. It has gained attention for its global modeling capabilities. Current medical image segmentation methods (such as U-Mamba and SegMamba) typically directly "flatten" 3D image patches into 1D sequences and input them into the Mamba model.

[0007] Spatial neighborhood loss: This forced serialization process disrupts the local neighborhood relationships of voxels in 3D space. For tasks like lesion segmentation, which heavily rely on local texture and edge details, direct serialization leads to the loss of crucial spatial structural information, thus affecting the segmentation's precision.

[0008] The "semantic gap" problem in encoder-decoder feature fusion: Existing hybrid architectures (such as TransUNet, Swin-UNETR) usually use simple skip connections, which directly concatenate the features of the encoder and the features of the decoder.

[0009] Feature conflict: Encoder features are semantically strong but have low resolution, while decoder features have high resolution but weak semantics and contain noise. A significant "semantic gap" exists between the two. Direct concatenation leads to feature distribution mismatch, introduces redundant information and even noise, making the model prone to misjudgment or artifacts when dealing with ambiguous boundaries, and lacks the ability to adaptively select effective features.

[0010] In summary, existing technologies struggle to simultaneously achieve robust capture of multi-scale lesions (especially very small lesions), efficient modeling of 3D global context, and accurate preservation of local spatial details within a single framework. Therefore, a novel network architecture is urgently needed that integrates the advantages of multi-scale convolution with linear global modeling capabilities, while resolving feature fusion conflicts. Summary of the Invention

[0011] This invention addresses the problems of limited receptive field in feature extraction, high computational complexity in global modeling, and semantic conflicts during feature fusion in existing stroke lesion segmentation methods. To this end, it proposes a 3D stroke lesion segmentation method based on multi-scale convolution and a bidirectional Mamba network. The invention achieves these technical problems through the following technical solution: Option 1: This invention proposes a 3D stroke lesion segmentation method based on multi-scale convolution and bidirectional Mamba networks. The method includes the following steps: Step 1: Obtain the brain MRI image sequence data to be evaluated, and store the brain MRI image sequence data in 3D voxel format, denoted as the input tensor. ,in, For the number of channels, These are depth, height, and width, respectively. Step 2: Preprocess the brain MRI image sequence data to be evaluated obtained in Step 1; Step 3: Construct a multi-scale feature encoder based on the preprocessed brain MRI image sequence data in Step 2, which is used to output four layers of encoded feature maps at different resolutions. Step 4: Construct a bidirectional global-local state space decoder based on the four layers of encoded feature maps with different resolutions output in Step 3, including adaptive spatial convolutional local enhancement preprocessing and bidirectional sequence scanning and fusion; Step 5: Perform bidirectional adaptive gating feature fusion on the information from each branch fused in Step 4, and then input the fused features to... The convolutional layer performs channel adjustment and serves as the input to the next level decoder; Step 6: Transfer the feature map output from the decoder in Step 5 through... Convolutional layers are mapped to segmentation probability maps The network is trained using a hybrid loss function, which combines Dice loss and weighted cross-entropy loss and minimizes the loss function through backpropagation to update the network parameters until the model converges.

[0012] Furthermore, a preferred embodiment is provided in which the preprocessing of the acquired brain MRI image sequence data to be evaluated in step 2 includes steps of resampling, intensity normalization, and cropping.

[0013] Furthermore, a preferred embodiment is provided, wherein the method for constructing the multi-scale feature encoder in step 3 is as follows: Step 3.1, the input tensor mentioned in Step 1 The convolution adjustment splits the flow into two parallel depthwise separable convolutional branches. The two branches extract spatial context at different scales, and an element-wise addition method is used for preliminary fusion to obtain features containing multi-scale information. ; Step 3.2: Merge the features Send in one The standard convolutional layer is used to locally refine and smooth features by utilizing its small receptive field, generating features. ; Step 3.3: Construct residual connections, connecting the original input... Compared with the refined features Perform element-wise addition to output the final encoded features. :

[0014] Step 3.4: Repeat the above steps to generate four layers of encoded feature maps with different resolutions.

[0015] Furthermore, a preferred embodiment is provided, wherein the method for constructing the bidirectional global-local state space decoder in step 4 is as follows: Step 4.1, the adaptive spatial convolution local enhancement preprocessing, is used to process the input tensor from step 1. Through parallel The deep convolutional branch generates channel weights through the SE attention module and multiplies them element-wise with the features of the other branch to achieve channel-level feature recalibration. Step 4.2, the bidirectional sequence scanning and fusion, flattens the enhanced 3D feature volume into a 1D sequence. And establish two independent state-space model branches, SSM, including a forward branch and a backward branch; Step 4.3: Transfer the forward features and backward features The sequence is concatenated along the channel dimension, and information is fused through linear layer projection. The sequence is then restored to a 3D voxel format, and residual connections are introduced again. These residual connections are then added to the input features to obtain the decoded output.

[0016] in, These are the input features used for residual summation.

[0017] Furthermore, a preferred embodiment is provided, wherein the method for bidirectional adaptive gating feature fusion in step 5, which involves fusing information from each branch in step 4, is as follows: Step 5.1: In the encoder measurement, the spatial attention module is used to calculate the spatial saliency map, which is used to highlight the location features of the lesion, and obtain... ; Step 5.2, on the decoder side: Using the channel attention module, calculate the channel importance weights, and select the feature channels most sensitive to lesion segmentation, thus obtaining... ; Step 5.3: Enhance the two-way features and The layers are concatenated by compressing the channels to 2 using a shared gated convolutional layer, followed by a Softmax activation function to generate two complementary scalar weight maps. and ; Step 5.4: Adaptively weighted summation of features using the generated weights to achieve soft feature selection.

[0018] The fusion feature was fed into The convolutional layer performs channel adjustment, which serves as the input to the next level of decoder.

[0019] Furthermore, a preferred implementation is provided, in step 6, the method of minimizing the loss function by combining Dice loss and weighted cross-entropy loss through backpropagation algorithm, updating the network parameters, until the model converges is as follows: .

[0020] Option 2: A computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the method described in Option 1.

[0021] Option 3: A computer device, including a memory and a processor, wherein the memory stores a computer program, and when the processor runs the computer program stored in the memory, the processor executes the method described in Option 1.

[0022] The advantages of this invention are: This invention presents a 3D stroke lesion segmentation method based on multi-scale convolution and bidirectional Mamba networks. The method employs a hierarchical strategy of "large kernel coarse capture + small kernel fine correction" through an LSC encoder to dynamically adapt to significant scale changes in lesions from small to large, thus addressing the problem of limited receptive fields in single convolutional kernels. Next, a BSC-Mamba decoder is used, introducing a state-space model with linear computational complexity. A bidirectional scanning mechanism captures omnidirectional global dependencies, and local convolution enhancement compensates for spatial information loss caused by serialization. Finally, the BAGF module performs asymmetric adaptive feature selection and gating fusion at skip connections, effectively bridging the semantic gap between the encoder (high semantics, low resolution) and the decoder (low semantics, high resolution), outputting accurate quantized segmentation results.

[0023] The present invention discloses a 3D stroke lesion segmentation method based on multi-scale convolution and bidirectional Mamba networks, which designs a dual-branch parallel large kernel (…). ) is used to capture global semantics, immediately following the little core ( This allows for local detail refinement. This hierarchical design enables the model to explicitly establish a "coarse-grained to fine-grained" feature extraction mechanism, significantly improving the ability to capture features of lesions of different sizes (especially small lesions).

[0024] This invention describes a 3D stroke lesion segmentation method based on multi-scale convolution and a bidirectional Mamba network. It develops an adaptive spatial convolution (ASC) preprocessing module to enhance local texture and edge information before serialization. Simultaneously, it introduces a bidirectional scanning mechanism, enabling feature updates to simultaneously perceive forward and backward contextual information. This design maintains... While achieving linear computational complexity, it realizes dual modeling of 3D lesions, both globally and locally.

[0025] This invention presents a 3D stroke lesion segmentation method based on multi-scale convolution and bidirectional Mamba networks. It abandons the traditional simple concatenation approach and proposes an asymmetric attention mechanism: spatial attention is used at the encoder to suppress background noise, and channel attention is used at the decoder to filter lesion features. Weighted fusion is then performed using learnable gating coefficients. This mechanism effectively solves the feature conflict problem and enhances the model's segmentation determinism at ambiguous boundaries. It significantly improves the accuracy and robustness of stroke lesion segmentation, especially demonstrating excellent performance in the highly challenging detection of small lesions.

[0026] This invention is also applied to a 3D stroke lesion segmentation method based on hybrid multiscale convolution and state space model (Mamba) in MRI (magnetic resonance imaging) image analysis. Attached Figure Description

[0027] Figure 1 This is a flowchart of the network architecture of a 3D stroke lesion segmentation method based on multi-scale convolution and bidirectional Mamba network as described in Implementation Method 1.

[0028] Figure 2 This is a schematic diagram of the structure of the multi-scale convolutional coding module (LSC Block) described in Implementation Method 1.

[0029] Figure 3 This is a schematic diagram of the bidirectional global-local state space decoding module (BSC-Mamba) described in Implementation Method 1.

[0030] Figure 4 This is a schematic diagram of the structure of the bidirectional adaptive gated feature fusion (BAGF) module described in Implementation Method 1.

[0031] Figure 5 This is a qualitative comparison of the model of the present invention described in Implementation Method 1 with existing mainstream models on the ATLAS v2.0 dataset.

[0032] Figure 6 This is a comparison chart showing the trend of the model segmentation performance (Dice coefficient) as described in Implementation Method 1 with the change of lesion volume.

[0033] Figure 7 This is a Bland-Altman analysis diagram of the volume prediction bias of the model described in Implementation Method 1.

[0034] Among them, (a) represents small lesions, (b) represents medium-sized lesions, and (c) represents large lesions.

[0035] Figure 8 This is a comparison chart showing the segmentation performance of the model of the present invention described in Implementation Method 1 on the ISLES 2022 dataset for small lesions. Detailed Implementation

[0036] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them.

[0037] Implementation Method 1: This implementation method proposes a 3D stroke lesion segmentation method based on multi-scale convolution and bidirectional Mamba networks. Through a unique "large-small" convolution strategy, a bidirectional state-space model, and an adaptive gating fusion mechanism, it achieves high-precision segmentation of multi-scale, irregular lesions. The method specifically includes the following steps: Step 1: Data Acquisition: Acquire the brain MRI image sequence data to be evaluated. Data modalities may include T1-weighted imaging (T1w), diffusion-weighted imaging (DWI), or apparent diffusion coefficient (ADC) maps, etc. These image data are stored in 3D voxel format and denoted as the input tensor. ,in, For the number of channels, These are depth, height, and width, respectively.

[0038] Step 2, Data Preprocessing: To eliminate differences caused by different equipment and scanning parameters, the acquired raw MRI data undergoes standardized preprocessing: Resampling: Resamples the voxel spacing of all 3D volume data to a fixed resolution (e.g., ...). The image is processed using cubic spline interpolation, and the labels are processed using nearest neighbor interpolation.

[0039] Intensity Normalization: Z-score normalization is performed on the image for each modality, and the mean and standard deviation are calculated based only on the foreground voxels (non-zero regions) to highlight the contrast of the lesion area.

[0040] Trim: Crops the image to a non-zero bounding box while retaining certain edges (e.g., 5 voxels), removing background redundancy and reducing computational cost.

[0041] Step 3: Construct a multi-scale feature encoder (LSC Encoder), which is an improved hierarchical feature extraction method. Addressing the issue that existing CNNs with fixed convolutional kernels cannot simultaneously capture the global contours of large lesions and the fine boundaries of small lesions, this step improves upon traditional convolutional units by proposing a parallel hybrid strategy of "large kernel for coarse capture + small kernel for fine correction," which includes the following steps: First, parallel large-kernel feature extraction and fusion: input features First, go through a Convolution adjustment channels.

[0042] Secondly, the flow splits into two parallel depth-wise separable convolution branches, with kernel sizes of [sizes to be filled in]. and These two branches extract spatial context at different scales and perform initial fusion using element-wise summation to obtain features containing multi-scale information. :

[0043] Serial little kernel refinement: refining the fused features Send in one The standard convolutional layer. This layer utilizes its small receptive field to locally refine and smooth the texture of features, generating features. .

[0044] Residual Fusion Output: To maintain gradient propagation and fuse the original information, a residual connection is constructed. The original input is then processed... Compared with the refined features (through After projection and batch normalization, element-wise summation is performed to output the final encoded features. :

[0045] Repeat the above steps to generate four layers of encoded feature maps at different resolutions.

[0046] Step 4: Construct a bidirectional global-local state-space decoder (BSC-Mamba Decoder), which is an improved global modeling and space preservation method. Addressing the issue of Mamba models destroying 3D spatial structure during serialization and the limitation of unidirectional scanning failing to perceive the full-image context, this step designs a local enhancement preprocessing and bidirectional scanning mechanism, specifically including the following steps: Adaptive Spatial Convolution (ASC) Local Augmentation: Before serialization, the input features are first processed in parallel... The deep convolutional branch generates channel weights through the Squeeze-and-Excitation (SE) attention module, which are then element-wise multiplied with the features from the other branch to achieve channel-level feature recalibration. Subsequently, a large kernel convolution further expands the receptive field, ensuring that local spatial information is encoded into the feature channels.

[0047] Bidirectional sequence scanning and fusion: flattening the enhanced 3D feature volume into a 1D sequence To achieve omnidirectional modeling, two independent SSM (State-Space Model) branches are established: Forward branch: Scan from the beginning of the sequence to the end to capture forward dependencies. .

[0048] Backward branch: Reverse the sequence and scan to capture backward dependencies. .

[0049] Feature reconstruction and residual fusion: combining forward features and backward features Concatenation is performed along the channel dimension, and information is fused through linear layer projection. The sequence is then reshaped into a 3D voxel format, residual connections are introduced again, and the results are added to the input features to obtain the decoded output.

[0050] Step 5: Bidirectional Adaptive Gated Feature Fusion (BAGF), also known as improved skip connections, addresses the "semantic gap" problem caused by direct concatenation in traditional U-Net (encoder features have strong semantics but low resolution, while decoder features have weak semantics but high resolution). This step proposes an asymmetric attention-guided gated fusion mechanism to replace simple concatenation. Specifically, it includes the following steps: Asymmetric attention enhancement: encoder side ( ): Through the Spatial Attention module, a spatial saliency map is calculated to suppress background noise and highlight the location features of lesions, thus obtaining... .

[0051] Decoder side ( The channel attention module calculates channel importance weights and selects the feature channels most sensitive to lesion segmentation. .

[0052] Gating coefficient calculation: The enhanced two-way features and The layers are concatenated by compressing the channels to 2 using a shared gated convolutional layer (Gate Conv), followed by a Softmax activation function to generate two complementary scalar weight maps. (Corresponding encoder) and (Corresponding decoder).

[0053] Weighted fusion: This method uses generated weights to adaptively weight and sum the features, achieving soft feature selection.

[0054] The fusion feature was then fed into The convolutional layer performs channel adjustment, which serves as the input to the next level of decoder.

[0055] Step Six: Loss Calculation and Model Optimization. The feature map output by the decoder is then processed... Convolutional layers are mapped to segmentation probability maps A hybrid loss function is used for network training, combining Dice loss (focusing on overlap) and weighted cross-entropy loss (focusing on classification accuracy) to address the foreground-background class imbalance problem.

[0056] The loss function is minimized using the backpropagation algorithm, and the network parameters are updated until the model converges.

[0057] See Figure 1 As shown, this network employs an improved U-shaped encoder-decoder architecture, comprising the following key components: The left encoder path consists of four stacked stages (Stages 1-4), each containing the multi-scale large-small convolutional module (LSC Block) proposed in this invention. After initial processing by the STEM layer, the spatial resolution of the feature maps decreases layer by layer as the layers deepen. The number of channels increases layer by layer ( The encoder and decoder are connected via a bidirectional global-local state space module (BSC-Mamba) designed in this invention to extract high-level semantic features. The right-side decoder path consists of four stages, each constructed using a trilinear interpolation method to recover the spatial resolution of the feature maps layer by layer, and utilizes the Mamba architecture for global context modeling. The intermediate connection path does not directly perform simple feature concatenation between the encoder and decoder; instead, they are connected via a bidirectional adaptive gated feature fusion (BAGF) module. This module receives shallow features from the encoder and deep features from the decoder, and outputs the fused features to the corresponding level of the decoder. The final decoded features are then processed... The 3D convolutional layer mapping outputs a lesion segmentation prediction map.

[0058] Figure 2 The image illustrates the internal structure of the LSC module, which uses a "large kernel for coarse capture + small kernel for fine correction." The input features first undergo... The convolution adjusts the channels, then splits into two parallel depthwise separable large-kernel convolution branches (with kernel sizes of [missing information] and [missing information]). and ), used to capture multi-scale global context; after the two branches' outputs are fused, they are serially passed through a The convolutional layers are used for local texture refinement; finally, the result is added to the original input via residual connections. This structure effectively solves the problem of limited receptive field of a single convolutional kernel.

[0059] Figure 3 The input features are first locally enhanced using Adaptive Spatial Convolution (ASC), then flattened into a sequence, and subjected to Mamba processing for both forward and backward scans. Finally, the sequences are concatenated and restored to 3D features. The flowchart below refines the ASC structure: utilizing... Convolution, combined with SE attention mechanism, enhances local spatial information, and further expands the local receptive field through gated convolutional units and large kernel convolution.

[0060] Figure 4 The left path processes encoder features (E) and applies spatial attention to suppress background noise; the right path processes decoder features (D) and applies channel attention to filter lesion-related channels. The enhanced features from both sides (E' and D') are concatenated and input into the gating network, where they are processed by Softmax to generate weight coefficients. and Finally, a weighted summation (integration) is performed on E' and D'. This mechanism bridges the semantic gap between encoder and decoder features.

[0061] Figure 5This paper aims to visually demonstrate the morphological segmentation advantages of the proposed LBMNet model in handling chronic stroke lesions. The figure shows a typical test sample (Case) from the ATLAS v2.0 dataset. This sample is characterized by lesions located deep within the brain parenchyma, exhibiting highly irregular shapes, blurred edges, and varying degrees of gray-scale unevenness, classifying them as "complex morphological lesions" that are clinically difficult to segment. The first two columns show the original T1-weighted MRI slices overlaid with the ground truth labels (red areas). It can be seen that the lesions are not simply circular but possess a complex topological structure with bifurcations and elongated protrusions. The second column shows the segmentation results (white mask) of this invention (Ours / Model5). Visual comparison reveals a very high degree of overlap between the segmentation results of this invention and the ground truth labels. Particularly noteworthy is the edge processing of the lesions: the mask generated by this invention has smooth and compact edges, accurately depicting the small protruding structures above the lesions, without any breaks or obvious "island" noise. This is mainly due to the "large kernel coarse capture + small kernel fine correction" strategy of the LSC encoder in this invention, which enables the model to locate the main lesion while preserving subtle edge features. Simultaneously, the BAGF module effectively suppresses background noise during feature fusion, ensuring the purity of the segmentation. Limitations of existing mainstream models (subsequent columns): Transformer-type models (such as SwinUNet, TransUNet, UNETR): Observing the segmentation results of these models, the phenomena of "undersegmentation" and "overly smooth edges" can be clearly observed. For example, the mask contours generated by SwinUNet and UNETR are relatively rigid, losing the jagged details of the lesion edges, and even missing the small branches at the end of the lesion. This is because the Transformer architecture typically uses a patch-sequential processing method. Although it excels at capturing global context, it is often less refined than convolutional networks in reconstructing local details, resulting in blurred edges. Traditional CNN-like models (such as U-Net++, 3DUnet): While these models retain some shape information in their segmentation results, they are prone to producing "holes" inside lesions or false "false positive" pixels on the outside. For example, U-Net++ results appear relatively fragile at the main connections of lesions, with a risk of breakage. This is due to the limited local receptive field of pure convolutional networks, which cannot effectively establish long-distance voxel dependencies like Mamba or Transformer, resulting in insufficient understanding of the overall connectivity of lesions. Other hybrid / state-of-the-art models (such as UMamba, FRPNet, STU-Net): Although models like UMamba outperform pure CNNs or pure Transformers, in a horizontal comparison with this invention, they still show slight inferiority in the accuracy of microstructure reconstruction.For example, UMamba's partition boundaries are slightly widened and not compact enough in some local areas. Figure 5 It is intuitively demonstrated that, thanks to the organic combination of the global modeling capability of bidirectional Mamba and the local refinement capability of multi-scale convolution, LBMNet can achieve "precise morphology, clear boundaries, and good connectivity" segmentation results when dealing with complex and irregular chronic stroke lesions, which is significantly better than various existing state-of-the-art algorithms.

[0062] Figure 6 Six scatter regression subplots were used to quantify the performance differences between our invention (LBMNet) and six mainstream algorithms (UMamba, FRPNet, UNETR, TransUNet, SwinUNet, U-Net++) under different lesion volumes. The horizontal axis represents the lesion volume after logarithmic processing, and the vertical axis represents the Dice similarity coefficient (0.0-1.0). Each point represents a test sample, and the curve represents the performance trend line based on scatter fitting. All subplots show a common pattern: as the lesion volume increases (horizontal axis from left to right), the Dice score increases. This aligns with the general rules of medical image segmentation: large lesions are relatively easy to segment due to their significant features and large number of pixels; while small lesions are extremely difficult to segment due to their weak features and susceptibility to noise.

[0063] Figure 6 This demonstrates the significant advantage of small lesion regions and reveals the performance differences of the models in the "small volume region" (left side of the horizontal axis). Observing the left half of all subplots, it can be clearly seen that the intercept and initial slope of the present invention are significantly higher than all the comparison models.

[0064] Robustness across all scales: Not only does it lead in small lesions, but the curve of this invention also remains at the highest level in the large lesion range on the right side of the horizontal axis, and the scatter distribution is more compact (smaller variance). This means that LBMNet can not only "see" small lesions, but also leverages the global modeling advantages of bidirectional Mamba when dealing with large lesions, avoiding the voids inside the lesions caused by insufficient receptive field, and achieving performance suppression across the entire size range. Figure 6 From a statistical perspective, it is strongly demonstrated that the hybrid architecture proposed in this invention is not only optimized for a specific type of sample, but also achieves robust segmentation of lesions of all sizes, from tiny to huge, by solving the balance problem between local and global features. In particular, it has established a significant technical barrier in the detection of small lesions that are most easily missed in clinical practice.

[0065] In clinical practice, it is not enough to simply pursue shape overlap (Dice). Doctors pay more attention to the accuracy of lesion volume calculation, because this is an important indicator for determining the thrombolytic treatment window, assessing the degree of brain injury, and prognosis and rehabilitation. Figure 7 The consistency between the predicted volume and the actual labeled volume of the model in this invention was evaluated using the Bland-Altman analysis method, a standard method in medical statistics. The graph is divided into three sub-graphs, corresponding to (a) small lesions, (b) medium-sized lesions, and (c) large lesions, respectively. The horizontal axis represents the average volume of the two measurement methods, and the vertical axis represents the volume difference percentage. Overall bias and confidence interval analysis: From the overall data, the mean bias of the labeled data in the graph is -5.22%. The negative value indicates that the model tends to slightly underestimate the lesion volume, but a bias of around 5% is very good in the field of medical image segmentation (the difference between human annotators can often reach over 10%). The red dashed line in the graph represents the 95% limit of agreement; the vast majority of sample points (circles) fall within this interval, demonstrating the statistical reliability of the model's predictions and indicating no systematic serious errors. Figure 7 (a) Small Lesions: This is the most challenging group. The longitudinal distribution of the sample points is relatively scattered (large variation), with some points showing volume differences approaching ±40%. This is technically reasonable, as for extremely small lesions (e.g., only a few dozen voxels), misjudgments of just a few pixels at the edge can lead to huge percentage errors. However, even so, the vast majority of points converge within an acceptable range, and no large-scale outliers appear, indicating that the model is stable in detecting small lesions, with the main error originating from the partial volume effect. Figure 7 (b) Medium-sized lesions vs. large lesions: As the volume increases, the sample points rapidly converge towards the center line (Bias line). Especially in the large lesion group, the data points are tightly clustered near the 0 error line with minimal fluctuation. This indicates that for lesions of a certain size, the bidirectional Mamba decoder of this invention can reconstruct the lesion volume with extremely high accuracy. Its global context modeling capability effectively ensures the integrity of the lesion's internal structure, avoiding volume underestimation due to voids. Clinical significance: This distribution trend from "high dispersion" to "high convergence" verifies the safety of LBMNet in clinical applications. For small lesions, although the percentage error is slightly larger, the absolute volume error is extremely small and will not mislead clinical grading; for large lesions, the extremely high volume accuracy can provide accurate data support for cerebral edema assessment and space-occupying effect analysis. Compared to models that only provide Dice scores, Bland-Altman analysis further confirms the potential of this invention as a clinical auxiliary quantitative tool.

[0066] like Figure 5 The model's performance in the chronic phase (ATLAS v2.0) is demonstrated. Figure 8 By introducing acute / subacute phase multimodal data (ISLES 2022), an extremely rigorous "stress test" was conducted to verify the model's limits in cross-domain generalization and small target detection. The imaging modality shown in the small lesion and multimodal difference map is DWI (diffusion-weighted imaging), the gold standard for diagnosing acute ischemic stroke. Compared with T1-weighted images, DWI images typically have lower resolution and are accompanied by characteristic speckle noise. The target in the image is a punctate small lesion, possibly smaller than 1 inch in volume. Extremely difficult to spot in a noisy background Demonstration of generalization ability: ISLES 2022 and the training data ATLAS v2.0 differ significantly in scanning parameters, pathological stages (acute vs. chronic), and imaging modalities (DWI vs. T1). Despite such a large domain shift, LBMNet can still accurately capture the most difficult-to-segment small lesions, fully demonstrating its strong generalization ability and robustness. This means that this patented technology is not only applicable to specific datasets but also has practical clinical application value in different hospitals and equipment environments.

[0067] Example 1: To fully verify the effectiveness and advancement of the technical solution of this invention, we built an experimental platform in a unified hardware and software environment. This platform is based on the Linux operating system, Python 3.10 programming language, and PyTorch deep learning framework, and is equipped with an NVIDIA GeForce RTX 3090 graphics card (24GB VRAM) and CUDA 11.8 parallel computing library for model training and inference. In the data preprocessing stage, to eliminate the spatial heterogeneity of multi-center medical images, we uniformly resampled all input 3D MRI images to... The median voxel spacing was determined, and cubic spline interpolation was applied to the images, while nearest neighbor interpolation was applied to the labels. Subsequently, Z-score intensity normalization based on the foreground non-zero region was performed on the MRI images for each modality. Finally, the 3D volume data was cropped to non-zero bounding boxes, retaining 5 voxel edges around the edges to reduce computational overhead.2 A stochastic gradient descent (SGD) optimizer was used during training, with a momentum parameter set to 0.99 and a weight decay coefficient of [value missing]. The initial learning rate was set to 0.01 with a multinomial decay strategy, the batch size was 2, the total number of training rounds was 300, and end-to-end supervised training was performed using a weighted combination of Dice loss and cross-entropy loss as the objective function.

[0068] Example 2: To verify the core performance of this invention in lesion segmentation of chronic stroke, we selected the large-scale public dataset ATLAS v2.0 containing 1271 T1-weighted MRI images and divided the dataset in an 8:1:1 ratio. Experimental results show that the proposed LBMNet model achieved an excellent Dice similarity coefficient of 67.57% on the test set. This result is significantly better than the traditional 3D U-Net model (50.06%) and U-Net++ model (45.85%), demonstrating the significant leap in feature extraction capability of this invention. Compared with the current state-of-the-art Transformer architecture model, this invention improves performance by 11.34% over UNETR (56.23%) and by 17.56% over TransUNet (50.01%), effectively overcoming the problem of overfitting of Transformer on limited medical data. Even compared with the latest hybrid models in the same category, this invention is 4.26% and 7.41% higher than UMamba (63.31%) and FRPNet (60.16%), respectively. Furthermore, the recall rate of this invention is as high as 69.38%, the F2 score reaches 72.39%, and the HD95 distance is reduced to 22.32 pixels. These indicators are the best values ​​among all comparison models, which fully demonstrates that the segmentation mask generated by this invention has tight boundaries, good internal connectivity, and very few missed diagnoses, and has extremely high clinical application value.

[0069] Table 1. Comparison of different models on the ATLAS v2.0 dataset

[0070] Example 3: To further verify the model's generalization ability across different pathological stages and imaging modalities, we extended the experiment to the ISLES 2022 acute / subacute ischemic stroke dataset. This dataset contains 250 cases of DWI and ADC multimodal images, with blurred lesion boundaries and severe penumbra effects. Even without modifying the network structure and only retraining, this invention still achieved state-of-the-art (SOTA) performance, with a Dice coefficient as high as 82.03%, significantly outperforming STU-Net (76.92%), FRPNet (77.16%), and UMamba (74.31%). Simultaneously, the HD95 distance of this invention was only 21.15 pixels, far lower than U-Net's 36.35 pixels and SwinUNet's 35.52 pixels. This result strongly demonstrates that this invention is not only applicable to chronic phase T1w single-modal images but also exhibits strong robustness to complex acute phase multimodal data, effectively addressing domain shift challenges in clinical scenarios.

[0071] Table 2. Comparison of different models on the ISLES 2022 dataset

[0072] Example 4: Addressing the challenge of detecting small lesions that are extremely difficult to segment clinically, this invention performs stratified evaluation on the ATLAS v2.0 test set. For lesions smaller than 10... For small lesions (accounting for 61.8% of the test set), the present invention achieved a Dice score of 58.47% and a recall rate as high as 66.52%, while Transformer models such as SwinUNet and TransUNet, compared at the same time, often suffered from serious missed detections on such samples due to feature loss; for samples larger than 50 For large lesions, the Dice score of this invention reached 74.18%, indicating that the model did not sacrifice its ability to model the integrity of large lesions while focusing on small targets. Further Bland-Altman consistency analysis showed that the average deviation between the predicted volume and the actual labeled volume was only -5.22%, and the vast majority of sample points fell within the 95% confidence interval. As the lesion volume increased, the deviation distribution converged rapidly, verifying the high reliability of this invention in providing accurate clinical quantitative assessment indicators.

[0073] Table 3. Performance evaluation of the ATLAS v2.0 test set under different lesion sizes.

[0074] Example 5: Ablation Experiment: To fully demonstrate the necessity and inventiveness of the various innovative modules and their internal design in the proposed LBMNet model, we designed and executed four rigorous controlled-variable experiments on the ATLAS v2.0 dataset. The experiments aimed to quantitatively analyze the specific contributions of techniques such as multi-scale convolutional coding (LSC), bidirectional state-space decoding (BSC-Mamba), and adaptive gated fusion (BAGF) to the final segmentation performance.

[0075] Component Ablation: To verify the synergistic effect of the three core components of LBMNet, the standard 3D U-Net was used as the baseline model (Baseline / Model1). Five variant models were constructed by adding components one by one, and the changes in their Dice coefficients and IoU scores were recorded. Baseline Model Performance: Due to the limitation of a single-scale convolutional kernel, the basic 3D U-Net model had a Dice of only 50.06% and an IoU of 41.39%. Qualitative results showed that it had a significant "under-segmentation" phenomenon in segmenting irregular lesions and could not capture complex boundary details. Contribution of LSC Encoder: After introducing the LSC multi-scale encoder of this invention into the baseline model (Model2), the Dice coefficient jumped to 60.09% (an improvement of 10.03 percentage points), and the IoU improved to 47.01%. This significant performance leap proves that the LSC module, through the combination of large kernel convolution and small kernel refinement, successfully solves the problem of insufficient receptive field in the feature extraction stage of traditional CNNs, and can effectively cover the complete region of the lesion. Contribution of the BSC-Mamba decoder: Replacing only the decoder with the BSC-Mamba module in the baseline model (Model 3) improves the Dice coefficient to 63.11% and IoU to 51.38%. Qualitative analysis shows that Model 3 can successfully detect discrete lesions located in the contralateral hemisphere, thanks to the excellent long-range dependency modeling capability of the Mamba architecture, which compensates for the deficiency of pure convolutional networks in being unable to perceive global context. Combination of LSC and BSC-Mamba: Combining the two (Model 4) further improves the Dice to 65.03%. At this point, the model has the ability to extract features at multiple scales and model global context, resulting in sharper segmentation contours. Decisive role of the BAGF fusion module: Finally, introducing the BAGF module to replace the traditional skip connection concatenation operation (Model 5, i.e., the complete model of this invention) finally achieves a Dice of 67.57% and an IoU of 54.97%. Compared to Model 4, BAGF brings an additional improvement of 2.54%. This result strongly demonstrates that simple feature concatenation can lead to semantic conflicts between the encoder and decoder, while the adaptive gating fusion mechanism of this invention effectively eliminates such conflicts and further refines the segmentation accuracy of fuzzy boundaries.

[0076] Table 4. Performance Comparison of Different Components

[0077] Those skilled in the art will understand that the above description is merely a preferred embodiment of the present invention, and the features described in the various embodiments and / or technical solutions of this disclosure can be combined or combined in various ways, even if such combinations or combinations are not explicitly described in this disclosure. This is not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

[0078] Although preferred embodiments of the present invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended technical solutions are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the present invention. Clearly, those skilled in the art can make various modifications and variations to the present invention without departing from its spirit and scope. Thus, if these modifications and variations of the present invention fall within the scope of the present invention and its equivalents, the present invention also intends to include these modifications and variations.

Claims

1. A 3D stroke lesion segmentation method based on multi-scale convolution and bidirectional Mamba networks, characterized in that, The method includes the following steps: Step 1: Obtain the brain MRI image sequence data to be evaluated, and store the brain MRI image sequence data in 3D voxel format, denoted as the input tensor. ,in, For the number of channels, These are depth, height, and width, respectively. Step 2: Preprocess the brain MRI image sequence data to be evaluated obtained in Step 1; Step 3: Construct a multi-scale feature encoder based on the preprocessed brain MRI image sequence data in Step 2, which is used to output four layers of encoded feature maps at different resolutions. Step 4: Construct a bidirectional global-local state space decoder based on the four layers of encoded feature maps with different resolutions output in Step 3, including adaptive spatial convolutional local enhancement preprocessing and bidirectional sequence scanning and fusion; Step 5: Perform bidirectional adaptive gating feature fusion on the information from each branch fused in Step 4, and then input the fused features to... The convolutional layer performs channel adjustment and serves as the input to the next level decoder; Step 6: Transfer the feature map output from the decoder in Step 5 through... Convolutional layers are mapped to segmentation probability maps The network is trained using a hybrid loss function, which combines Dice loss and weighted cross-entropy loss and minimizes the loss function through backpropagation to update the network parameters until the model converges. The method for constructing the multi-scale feature encoder in step 3 is as follows: Step 3.1, the input tensor mentioned in Step 1 The convolution adjustment splits the flow into two parallel depthwise separable convolutional branches. The two branches extract spatial context at different scales, and an element-wise addition method is used for preliminary fusion to obtain features containing multi-scale information. ; Step 3.2: Merge the features Send in one Standard convolutional layers are used to locally refine and detail features using a small receptive field, generating features. ; Step 3.3: Construct residual connections, connecting the original input... Compared with the refined features Perform element-wise addition to output the final encoded features. : Step 3.4: Repeat the above steps to generate four layers of encoded feature maps with different resolutions; The method for constructing the bidirectional global-local state space decoder in step 4 is as follows: Step 4.1, the adaptive spatial convolution local enhancement preprocessing, is used to process the input tensor from step 1. Through parallel The deep convolutional branch generates channel weights through the SE attention module and multiplies them element-wise with the features of the other branch to achieve channel-level feature recalibration. Step 4.2, the bidirectional sequence scanning and fusion, flattens the enhanced 3D feature volume into a 1D sequence. And establish two independent state-space model branches, SSM, including a forward branch and a backward branch; Step 4.3: Transfer the forward features and backward features The sequence is concatenated along the channel dimension, and information is fused through linear layer projection. The sequence is then restored to a 3D voxel format, and residual connections are introduced again. These residual connections are then added to the input features to obtain the decoded output. in, These are the input features used for residual summation; Step 5 involves using the bidirectional adaptive gating feature fusion method from step 4, where information from each branch is fused. Step 5.1: In the encoder measurement, the spatial attention module is used to calculate the spatial saliency map, which is used to highlight the location features of the lesion, and obtain... ; Step 5.2, on the decoder side: Using the channel attention module, calculate the channel importance weights, and select the feature channels most sensitive to lesion segmentation, thus obtaining... ; Step 5.3: Enhance the two-way features and The layers are concatenated by compressing the channels to 2 using a shared gated convolutional layer, followed by a Softmax activation function to generate two complementary scalar weight maps. and ; Step 5.4: Adaptively weighted summation of features using the generated weights to achieve soft feature selection. The fusion feature was fed into The convolutional layer performs channel adjustment, which serves as the input to the next level of decoder.

2. The 3D stroke lesion segmentation method based on multi-scale convolution and bidirectional Mamba network according to claim 1, characterized in that, Step 2 involves preprocessing the acquired brain MRI image sequence data to be evaluated, including resampling, intensity normalization, and cropping.

3. The 3D stroke lesion segmentation method based on multi-scale convolution and bidirectional Mamba network according to claim 1, characterized in that, Step 6, which combines Dice loss and weighted cross-entropy loss to minimize the loss function using backpropagation and updates the network parameters until the model converges, is as follows: 。 4. A computer storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, it implements the method described in any one of claims 1-3.

5. A computer device, characterized in that, include: A memory, a processor, and a computer program stored in the memory and executable on the processor, the processor executing the program to implement the method of any one of claims 1-3.

Images