A 2.5D MRI brain tumor segmentation method based on Mamba state space modeling
By improving the Selective Scan mechanism and dual-decoder structure of Mamba state space modeling, the problems of insufficient long-range dependency modeling and discontinuous boundary segmentation in MRI brain tumor segmentation were solved, achieving efficient and accurate 2.5D brain tumor segmentation to meet clinical needs.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- CHONGQING UNIV OF TECH
- Filing Date
- 2026-03-27
- Publication Date
- 2026-06-09
AI Technical Summary
Existing MRI brain tumor segmentation methods suffer from problems such as insufficient long-range dependency modeling, excessive computational overhead, and limited boundary segmentation accuracy. In particular, in 2.5D segmentation scenarios, the sequence traversal strategy is mismatched with the spatial structure, resulting in discontinuous tumor segmentation and easy missed detection of small lesions. Furthermore, a single decoder structure cannot take into account both global regional connectivity and boundary detail accuracy, and the simple feature fusion method leads to information coupling interference.
A 2.5D MRI brain tumor segmentation method based on Mamba state space modeling is adopted. By improving the Selective Scan mechanism, a cross-slice priority traversal module is designed, the feature fusion mechanism is optimized, and a dual decoder structure with global perception and edge enhancement is constructed. Combined with an adaptive dilated parallel convolution module, cross-slice semantic consistency and boundary accuracy are improved.
While reducing computational complexity, it improves the accuracy and continuity of brain tumor segmentation, enhances the ability to depict the details of boundaries, and meets the fine segmentation requirements of clinical needs.
Smart Images
Figure FT_1 
Figure FT_2 
Figure FT_3
Abstract
Description
Technical Field
[0001] This invention belongs to the field of medical image segmentation technology and proposes a 2.5D MRI brain tumor segmentation method based on Mamba state space modeling. This method is applicable to the automatic and accurate segmentation of clinical MRI brain tumor images and can provide technical support for brain tumor diagnosis, treatment planning and efficacy evaluation. Background Technology
[0002] With the rapid development of deep learning, automatic segmentation algorithms for MRI brain tumors have gradually become a research hotspot in the field of medical image segmentation. In particular, the application of 3D Convolutional Neural Networks (3D CNNs) has led to rapid development in this field. Although this method can segment brain tumors efficiently and accurately, 3D CNNs rely on local receptive fields for feature extraction, requiring multi-layer stacking to expand the receptive field. They lack the ability to model long-range spatial dependencies across slices, easily leading to problems such as discontinuous tumor segmentation and missed detection of small lesions. Simultaneously, 3D models suffer from high complexity and high computational resource requirements, resulting in excessively high deployment costs in clinical applications. While 2D CNNs can improve model inference efficiency, they cannot capture the spatial relationships between slices, easily leading to discontinuous segmentation.
[0003] Although subsequent researchers proposed lightweight segmentation networks to improve efficiency by reducing the number of parameters and computational complexity, this approach often leads to insufficient model feature representation and decreased segmentation accuracy, especially for complex brain tumor regions, where the model performs poorly. On the other hand, to improve the long-range spatial modeling capability of 3D CNNs across slices, researchers have introduced Transformers, which use their self-attention mechanism to establish long-range dependencies. However, the computational complexity increases quadratically with the sequence length, resulting in significant memory pressure and making it impossible to balance accuracy and efficiency.
[0004] The Mamba state-space model, which has emerged in recent years, offers a new technical approach for brain tumor segmentation by overcoming the limitations of long-range dependency modeling in CNNs and solving the problem of excessive computational overhead in Transformers, thanks to its ability to model long sequences with linear computational complexity. Therefore, balancing model complexity and segmentation accuracy is one of the key research focuses currently being explored.
[0005] However, existing Mamba-based segmentation methods mostly directly transfer general architectures without optimizing sequence traversal strategies for slice adjacency relationships in 2.5D segmentation scenarios. They cannot fully adapt to the spatial distribution characteristics of 2.5D neighborhood slice groups, resulting in problems such as insufficient spatial structure adaptability and lack of boundary characterization accuracy. They cannot fully leverage the dimensional advantages of 2.5D segmentation, nor can they fully unleash the long-range modeling potential of Mamba, and still cannot meet the clinical needs for refined segmentation of brain tumors. Summary of the Invention
[0006] This invention addresses the problems of insufficient long-range dependency modeling, high computational overhead, and limited boundary segmentation accuracy in existing MRI brain tumor segmentation methods. It proposes a 2.5D MRI brain tumor segmentation method based on Mamba state space modeling. Through innovative sequence modeling strategies, feature fusion mechanisms, and decoding structure design, the method reduces computational complexity while ensuring segmentation accuracy.
[0007] This invention aims to address the three core technical deficiencies of existing MRI brain tumor segmentation methods:
[0008] (1) Solve the problems of existing Mamba-based 2.5D segmentation methods, such as the mismatch between sequence traversal strategy and 2.5D slice spatial structure, insufficient cross-slice semantic consistency modeling ability, discontinuous tumor segmentation, and easy missed detection of small lesions.
[0009] (2) Solve the problem that the existing single decoder structure cannot take into account both the global connectivity of the tumor region and the accuracy of boundary details, resulting in blurred tumor boundaries, positioning deviations, and insufficient segmentation accuracy in complex scenarios.
[0010] (3) The method of fusing local details of convolution with long-range semantic features of Mamba is simple, which is prone to feature suppression and information coupling interference, and cannot adaptively balance local details and global semantic expression.
[0011] The specific solution of the present invention is as follows:
[0012] This invention proposes a 2.5D MRI brain tumor segmentation method based on Mamba state space modeling, comprising the following steps:
[0013] S1: Acquire multimodal MRI brain tumor data and preprocess it to construct a 2.5D neighborhood slice input tensor;
[0014] S2: Improve the Selective Scan mechanism based on the State Space Model (SSM) and design a cross-slice priority traversal module to adapt to the spatial structure characteristics of 2.5D brain tumor segmentation networks.
[0015] S3: Design a feature calibration and controlled fusion module to optimize the fusion effect of Mamba long-range features and convolutional local features in the coding layer.
[0016] S4: Construct a dual decoder structure for global perception and edge enhancement, consisting of a multi-branch assisted deep supervision decoder and a boundary refinement multi-scale decoder.
[0017] S5: Use the optimal model to predict the test set data.
[0018] Furthermore, the dataset used in step S1 is preferably the BraTS 2020 public dataset from the Medical Image Computing and Computer Assisted Intervention (MICCAI) series of public datasets. Preprocessing the brain tumor dataset effectively removes noise signals.
[0019] Furthermore, in step S2, the feature map is first extracted by 3×3 convolution, and then the designed cross-slice priority traversal module performs slice priority serialization on the encoded feature map, performs bidirectional Selective Scan operation to propagate semantic information, and after fusion, backfills to generate cross-slice semantic enhancement feature map.
[0020] Furthermore, in step S3, the feature calibration and controlled fusion module generates calibration weights to optimize convolutional features through three-way directional pooling, and then concatenates them with Mamba long-range features through channels, and adjusts the performance of the Mamba network in the coding layer by combining learnable coefficients.
[0021] Furthermore, in step S4, in the dual-decoder structure: multi-branch deep supervision is introduced into the multi-branch auxiliary deep supervision decoder to enhance the feature representation of different layers, and an adaptive atrous parallel convolution module (APD) is used to expand the receptive field. The boundary refinement multi-scale decoder combines the Sobel operator to enhance the accuracy of the boundary region, and multiple Bottleneck Blocks are stitched together to characterize features from coarse to fine.
[0022] Furthermore, in step S5, model prediction and evaluation involve using the best trained and optimized model to predict the validation set, and submitting the predicted segmentation results to the official website for online verification. When evaluating model performance, the Dice similarity coefficient and the 95% Hausdorff distance (HD95) are used as the main quantitative indicators. Attached Figure Description
[0023] Figure 1The cross-slice traversal process in this invention
[0024] Figure 2 This is the multi-branch assisted deep supervision decoder (the globally perceptive decoder in the dual-decoder structure) in this invention.
[0025] Figure 3 This invention relates to a boundary refinement multi-scale decoder (an edge enhancement decoder in a dual-decoder structure).
[0026] Figure 4 The adaptive dilated parallel convolution module in this invention Detailed Implementation
[0027] S1: Data Acquisition and Preprocessing
[0028] S11. Download the official dataset BraTS 2020 from the MICCAI website. This dataset contains 369 training cases and 125 validation cases. Each case contains four modalities: T1, T1ce, T2, and FLAIR, as well as three types of tumor region annotations: whole tumor (WT), tumor core (TC), and enhancing tumor (ET). The officially processed image size is 240×240×155.
[0029] S12. Remove non-brain tissue regions from the edges of the original 3D MRI volume data, retain the effective brain tissue regions, and uniformly scale each axial 2D slice to 160×160 resolution. Along the axis, take each 2D slice as the target slice, select 3 adjacent slices above and below it to form a neighborhood group of 7 slices, and use mirroring to fill the entire neighborhood in the edge region. The neighborhood slice groups of the four modalities are spliced along the channel dimension to construct the final 2.5D input tensor.
[0030] S2: Implementation of the slice-first traversal module
[0031] S21, such as Figure 1 As shown, for any spatial location (h, w), feature vectors of 7 slices are extracted in slice order to construct a slice-wise sequence.
[0032] S22. Perform a bidirectional Selective Scan on the sequence: the forward scan efficiently propagates cross-slice semantic information, and the reverse scan supplements the long-range semantic dependencies between slices, ensuring the integrity of cross-slice feature associations.
[0033] S23. Fuse the bidirectional scanning results to obtain cross-slice enhanced features, and backfill the enhanced features of all spatial locations to form a complete cross-slice semantic enhanced feature map.
[0034] S3: Feature calibration and controlled fusion module deployment
[0035] S31. The pre-calibration module extracts feature statistics through three-way global average pooling branches, aligns the dimensions of the three pooling results and concatenates them, then performs 1×1 convolution compression encoding to generate three types of calibration weights, which are applied to the height, width and slice dimension of the input features respectively to obtain the calibrated convolutional features.
[0036] S32. In the controlled fusion stage, the long-range features output from the Mamba branch are concatenated with the calibration feature channels, and a correction term G is generated by 1×1 convolution. The learnable coefficient α is obtained through Sigmoid, initially assigned a value of 0.5 and updated as the model is trained. It is used to adjust the injection intensity and finally output the fused features.
[0037] S4: Dual Decoder and Feature Fusion Unit Design
[0038] S41, such as Figure 2 As shown, the global awareness decoder is constructed, which includes four levels of upsampling operations. Each level doubles the resolution of the feature map through transposed convolution, embeds a boundary-aware attention module to calibrate semantic features, and fuses the skip connection features of the corresponding layers in the encoding stage to ensure the integrity of the global structure. Auxiliary loss branches are set in the four layers of the decoder to apply deep supervision constraints.
[0039] S42, such as Figure 3 As shown, the edge enhancement decoder is constructed using a four-level decoding structure matching the encoding module. Each level consists of two cascaded Bottleneck convolutional modules, formed by cascading 1×1, 3×3, and 1×1 convolutions. Residual connections are embedded between modules to prevent gradient vanishing. A boundary attention enhancement module (BAM) is embedded at the final output of the decoder. The Sobel operator uses a 3×3 convolution to extract gradient information in the X, Y, and Z directions to generate a gradient magnitude map. BAM weights are generated through a 3×3 convolution and Sigmoid activation to specifically enhance tumor edge features.
[0040] S43, such as Figure 4 As shown, in the adaptive dilated parallel convolution module, the adaptive dilation rate is calculated using the formula d = 16 / feature_size, where feature_size is the side length of the current feature map. The core module sets adaptive dilation rate parameters d1 and d2, which are negatively correlated with feature resolution, to dynamically adjust the receptive field range and balance global semantics with boundary detail features. By capturing features of different ranges through multi-scale dilated convolution, the features output from the dual decoders are concatenated and then integrated with channel information through a 1×1 convolution to output the final fused feature.
[0041] S5: Complete model training and optimization, use the selected optimal model to predict the test set data, and complete the performance evaluation of the segmentation results.
[0042] S51. The total loss function consists of a weighted sum of the main segmentation loss and the multi-branch deep supervision auxiliary loss. The main segmentation loss is a weighted combination of the Dice loss and the cross-entropy loss, with weight coefficients set to 0.6 and 0.4, respectively. The multi-branch deep supervision auxiliary loss uses isomorphic loss functions to the main segmentation loss and is applied to the auxiliary loss branches of the four upsampling levels of the global perception decoder. The auxiliary loss weights are set to 0.4, 0.3, 0.2, and 0.1 respectively, from deep to shallow decoder levels, to avoid gradient vanishing in deep networks and enhance multi-scale feature representation capabilities. The output features of each auxiliary loss branch are restored to the same resolution as the input image through bilinear upsampling and share the same set of ground truth labels as the main branch to calculate the loss value. Backpropagation is used to optimize network parameters.
[0043] The model training uses the Stochastic Gradient Descent (SGD) optimizer, with a training batch size of 8 and a total training epoch of 200. The initial learning rate is set to 1e-3, the weight decay coefficient is set to 1e-4, and the learning rate is dynamically adjusted using a cosine annealing learning rate scheduling strategy. The cosine annealing period is set to 50 epochs, and the minimum learning rate lower bound is set to 1e-6.
[0044] The formula for calculating the Dice loss used in training is as follows:
[0045]
[0046] In the formula, N is the total number of voxels. The model predicts the probability for the i-th voxel. ε is the truth label for the i-th voxel, and ε is a smoothing term to prevent the denominator from being 0, with a fixed value of 1e-5.
[0047] The formula for calculating the cross-entropy loss used in training is:
[0048]
[0049] The final formula for calculating the total loss function is:
[0050]
[0051] In the formula, The weight coefficients are the auxiliary loss coefficients of the k-th level decoder. This is the calculated value of the auxiliary loss for the k-th level decoder.
[0052] S52. During model training, after each round of training, offline validation is performed using the BraTS 2020 validation set. The average Dice similarity coefficient of the three types of tumor regions (WT, TC, and ET) is used as the core indicator, and the model with the highest average Dice coefficient is selected as the candidate optimal model. Finally, the candidate optimal model is used to complete the full test on the BraTS 2020 validation set, output the segmentation results of the three types of tumor regions (WT, TC, and ET), and upload the segmentation results to the official BraTS challenge validation platform for online validation to obtain official quantitative evaluation results. The final performance evaluation is completed by using the Dice similarity coefficient and the 95% Hausdorff distance (HD95).
[0053] S53, Dice similarity coefficient:
[0054] This measure assesses the overlap between predicted and ground truth segments at the voxel level, with values ranging from 0 to 1. Values closer to 1 indicate better overlap. It is suitable for class imbalance scenarios, and its formula is as follows:
[0055]
[0056] in, To predict the set of voxels, It is the set of truth voxels.
[0057] S54, 95% Hausdorf Distance:
[0058]
[0059] in, These represent the surface point sets for predicted segmentation and ground truth segmentation, respectively. Point Time The Euclidean distance; This represents the 95th percentile of all bidirectional shortest surface distances; the smaller the value, the better the boundary consistency.
Claims
1. A 2.5D MRI brain tumor segmentation method based on Mamba state space modeling, comprising the steps of preprocessing multimodal MRI brain tumor data to construct a 2.5D neighborhood slice input tensor, feature encoding, feature decoding, and segmentation result prediction, characterized in that, Includes the following steps: S2: Improve the Selective Scan mechanism based on the State Space Model (SSM) by performing slice-first serialization on the encoded feature map through the cross-slice priority traversal module, and then performing bidirectional SelectiveScan operation to generate cross-slice semantically enhanced feature maps. S3: Through the feature calibration and controlled fusion module, the convolutional local detail features and Mamba long-range semantic features are adaptively calibrated and fused to output the optimized fused features; S4: The fused feature input is used to complete the feature decoding of the global perception and edge enhancement dual decoder structure to obtain the brain tumor segmentation result; the dual decoder structure consists of a multi-branch assisted deep supervision decoder and a boundary refinement multi-scale decoder.
2. The 2.5D MRI brain tumor segmentation method based on Mamba state space modeling according to claim 1, characterized in that, In step S2, the processing flow of the cross-slice priority traversal module is as follows: for any spatial position (h, w) on the feature map, extract the feature vectors of the neighboring slice group according to the slice order to construct a slice-wise sequence; perform a bidirectional selective scan on the sequence, propagate cross-slice semantic information through forward scan and supplement long-range semantic dependencies between slices through reverse scan; The cross-slice enhanced features are obtained by fusing the bidirectional scanning results, and after backfilling, a complete cross-slice semantic enhanced feature map is formed.
3. The 2.5D MRI brain tumor segmentation method based on Mamba state space modeling according to claim 1, characterized in that, In step S3, the processing flow of the feature calibration and controlled fusion module is as follows: feature statistics of the input convolutional features are extracted through three global average pooling branches, and three types of calibration weights are generated after dimension alignment and convolutional encoding. These weights are applied to the height, width, and slice dimension of the input features to obtain the calibrated convolutional features. The long-range features output by the Mamba branch are concatenated with the calibrated convolutional feature channels, and a correction term is generated through convolution. The fused features are then output after the feature injection intensity is adjusted by learnable coefficients. The learnable coefficients are generated by Sigmoid activation, initially assigned a value of 0.5, and have a fixed value range of [0,1]. They are updated end-to-end during model training.
4. The 2.5D MRI brain tumor segmentation method based on Mamba state space modeling according to claim 1, characterized in that, The multi-branch assisted deep supervised decoder includes four levels of upsampling operations. Each level doubles the resolution of the feature map through transposed convolution, embeds a boundary-aware attention module to calibrate semantic features, and fuses skip connection features from the corresponding levels in the encoding stage. Each of the four levels of the decoder has an auxiliary loss branch to apply deep supervision constraints. The boundary refinement multi-scale decoder adopts a four-level decoding structure that matches the encoding module. Each level has two cascaded Bottleneck convolution modules with residual connections embedded between modules. The decoding output embeds a boundary attention enhancement module to strengthen tumor edge features.
5. The 2.5D MRI brain tumor segmentation method based on Mamba state space modeling according to claim 1, characterized in that, The features output by the dual decoders are finally fused by the adaptive dilated parallel convolution module. The adaptive dilated parallel convolution module is set with adaptive dilation rate parameters d1 and d2 that are negatively correlated with the feature resolution. It captures features of different ranges through multi-scale dilated convolution, and after concatenation, it integrates channel information to output the final fused feature. The basic calculation formula for the adaptive dilation rate is d=16 / feature_size, where feature_size is the side length of the current feature map.
6. The 2.5D MRI brain tumor segmentation method based on Mamba state space modeling according to claim 1, characterized in that, The total loss function used for model training consists of a weighted sum of the main segmentation loss and the multi-branch deep supervision auxiliary loss. The main segmentation loss is a weighted combination of the Dice loss and the cross-entropy loss, with weight coefficients of 0.6 and 0.4, respectively. The multi-branch deep supervision auxiliary loss and the main segmentation loss adopt isomorphic loss functions and are applied to the four levels of the multi-branch deep supervision decoder, respectively. The auxiliary loss weights are set to 0.4, 0.3, 0.2, and 0.1 as the decoder level decreases from deep to shallow.