Brain tumor medical image registration method based on improved Mama
By introducing an edge enhancement mechanism and an improved Mamba state space modeling module, the problem of insufficient accuracy in brain tumor image registration is solved, and efficient image registration effects are achieved. It is suitable for multi-phase MRI images and other high-precision non-rigid registration scenarios.
Patent Information
- Application Number
- CN202510772142.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-06-11
- Publication Date
- 2025-09-26
AI Technical Summary
Existing technologies have difficulty in achieving high-precision brain tumor image registration under conditions of complex anatomical changes and tumor growth, especially due to the limitations of CNN's local receptive field and insufficient global contextual information.
An edge enhancement mechanism, a lightweight U-Net feature extractor, and an improved Mamba state space modeling module are introduced. Combined with a lightweight U-Net encoder, a Mamba state space modeling module, and a deformation field prediction head, the image registration accuracy is improved through edge extraction, feature fusion, and global dependency modeling.
The accuracy and robustness of brain tumor image registration have been significantly improved, especially in the cases of non-rigid deformation and obvious deformation of the tumor area. It is superior to traditional methods and existing deep learning methods and has good generalization performance.
Smart Images
Figure CN120707606A_ABST
Abstract
Description
Technical field:
[0001] This invention belongs to the fields of computer vision and medical image processing, and more specifically, to a brain tumor medical image registration method based on an improved Mamba algorithm. By introducing an edge enhancement mechanism, a lightweight U-Net feature extractor, and an improved Mamba state-space modeling module, this method achieves high-precision deformation field prediction and spatial alignment between multimodal or time-series brain MRI images, improving the matching consistency of tumor regions and assisting clinical diagnosis. Background technology:
[0002] In modern medical image analysis, early detection and lesion tracking of brain tumors are crucial for treatment planning and efficacy evaluation. Medical image registration, a key step in image fusion, comparison, and analysis, aims to spatially align images from different time points or modalities. However, traditional registration methods, such as those based on affine transformations or mutual information, often struggle to achieve high-precision pixel-level alignment when faced with complex anatomical changes, non-rigid deformations, and tumor growth.
[0003] In recent years, deep learning-based image registration methods have made significant progress, especially the combination of convolutional neural networks (CNNs) and deformation networks (such as VoxelMorph). However, CNNs are limited by their local receptive field and have difficulty capturing long-range dependencies. Existing methods also rarely consider the impact of global context on deformation field generation.
[0004] Therefore, designing a brain tumor image registration method that can effectively model global features and generate high-quality deformation fields has become one of the key research directions. Summary of the invention:
[0005] To address these issues, this paper proposes a brain tumor medical image registration method based on an improved Mamba model. By incorporating an edge enhancement mechanism, a lightweight U-Net feature extractor, and an improved Mamba state-space modeling module, this method achieves efficient modeling and deformation field prediction of the tumor region and its surrounding structures in the input image.
[0006] 1. A brain tumor medical image registration method based on improved Mamba, characterized by the following specific steps:
[0007] S1: Model construction and initialization subsystem, which loads pre-trained or randomly initialized modular network structures, adopts a lightweight U-Net encoder with edge enhancement mechanism, Mamba state space modeling module and deformation field prediction head;
[0008] S2: Data management and distribution pipeline, which performs normalization, size adaptation, and multi-channel concatenation on input medical images to generate input tensors suitable for model training and inference;
[0009] S3: Model training and optimization framework, which uses a gradient descent optimizer to update model parameters during training and combines it with a learning rate scheduling strategy to improve model convergence efficiency.
[0010] S4: Model performance evaluation and visualization tool, using deformation consistency evaluation index and reconstruction error as the main evaluation criteria, providing deformation field visualization and image registration result comparison functions.
[0011] 2. The improved Mamba-based brain tumor medical image registration method according to claim 1, wherein the Mamba module is modified and integrated in a targeted manner in S1, specifically including the following contents:
[0012] S1-1: Image edge enhancement unit, which designs an edge extraction module based on the Sobel operator to enhance the local structural information of the input image to improve the registration accuracy;
[0013] S1-2: Feature fusion encoder, which adopts the encoder-decoder structure of lightweight U-Net architecture and combines the skip connection mechanism to fuse the feature representation of the original image and the edge map;
[0014] S1-3: Feature dimensionality mapping layer, which uses 1×1 convolution to upgrade the low-dimensional features extracted by CNN to the high-dimensional embedding space required by Mamba;
[0015] S1-4: Mamba sequence modeling component, which flattens the two-dimensional image into a one-dimensional sequence and inputs it into the Mamba module to capture the global dependencies in the image and restore it to a two-dimensional feature map for subsequent prediction.
[0016] 3. The improved Mamba-based brain tumor medical image registration method according to claim 1, wherein the data management and distribution pipeline in S2 specifically includes the following:
[0017] S2-1: Medical image preprocessing unit, which performs basic processing operations such as standardization and size cropping of input images to ensure consistency of data from different sources;
[0018] S2-2: Data stitching and channel fusion module, which stitches a fixed image with its corresponding edge map, and a moving image with its edge map, to form a dual-channel input, enhancing the model's perception capabilities;
[0019] S2-3: Image resizing adapter, which supports image scaling operations to ensure that images are of uniform size during training and can be restored to their original resolution during inference;
[0020] S2-4: Batch data generator, which extracts samples from the dataset based on the training batch size to ensure that each batch of data meets the principles of uniformity and randomness.
[0021] 4. The improved Mamba-based brain tumor medical image registration method according to claim 1, wherein the model training and optimization framework in S3 specifically includes the following:
[0022] S3-1: Forward propagation calculation unit, which performs the forward propagation process of the model on each batch of data and outputs the predicted deformation field tensor;
[0023] S3-2: Back propagation and gradient update unit, which calculates the gradient according to the deformation consistency loss function and updates the model parameters through an optimization algorithm (such as Adam);
[0024] S3-3: Dynamic learning rate decay strategy, dynamically adjusts the learning rate according to the number of training rounds and validation set performance to improve the model's generalization ability.
[0025] 5. The improved Mamba-based brain tumor medical image registration method according to claim 1, wherein the model performance evaluation and visualization tool in S4 specifically includes the following:
[0026] S4-1: Basic performance evaluation module, which calculates and reports basic registration quality indicators such as reconstruction error and pixel-level displacement consistency;
[0027] S4-2: Advanced deformation assessment method, which introduces more complex registration evaluation indicators such as structural similarity (SSIM) and mutual information (MI) to improve evaluation accuracy;
[0028] S4-3: Monitoring and visualization of the training process. Use charts to plot the loss curves and evaluation indicator trends during the training and validation phases to assist in observing the model convergence speed.
[0029] S4-4: Visual display of deformation field and image registration. The predicted deformation field is superimposed on the original image and compared with the actual registration result to provide an intuitive visual evaluation effect.
[0030] Technical effects:
[0031] This invention significantly improves the accuracy and robustness of brain tumor medical image registration by introducing an edge enhancement mechanism, a lightweight U-Net feature extractor, and an improved Mamba state-space modeling module. Particularly when dealing with non-rigid deformations and significant deformation of the tumor region, this method demonstrates superior registration performance compared to traditional methods and existing deep learning approaches. Through carefully designed data management and training processes, this invention achieves efficient model training and deployment capabilities, possesses excellent generalization performance, and has practical application value, making it suitable for automated image analysis systems in hospitals, research institutions, and medical imaging platforms.
[0032] The above description is only a specific embodiment of the present invention, but the present invention is not limited thereto. Any equivalent changes or substitutions based on the concept of the present invention should fall within the scope of protection of the claims of the present invention. Description of the drawings:
[0033] Figure 1 This is an overall flow chart of a brain tumor medical image registration method based on an improved Mamba provided by the present invention;
[0034] Figure 2 This is the improved Mamba network structure diagram provided by the present invention;
[0035] Figure 3 It is the MSDTask01 data set provided by the present invention;
[0036] Figure 4 The present invention provides a single MRI sequence of images A and B, a corresponding actual mask, and the registered sequence of images A and the mask.
[0037] Figure 5 This is the loss curve diagram of the improved Mamba training and verification provided by the present invention;
[0038] Figure 6 This is a Dice Score curve diagram of the improved Mamba training and verification provided by the present invention. Specific implementation method:
[0039] The present invention is further described below with reference to the accompanying drawings and examples. However, the present invention can be implemented in many different ways and should not be construed as limited to the illustrated embodiments; rather, these embodiments provide those skilled in the art with implementation methods that meet applicable legal requirements.
[0040] Example 1: Brain tumor image registration experiment based on improved Mamba
[0041] This example uses Task 1 (brain tumor task) from the publicly available MSD dataset. This dataset contains MRI images from multiple cases, including four sequences: FLAIR, T1W, T1QD, and T2W. The image resolution is 240x240. The data is divided into training, validation, and test sets in a 6:2:2 ratio.
[0042] During model training, the input image first uses the Sobel operator to extract edge features and then undergoes channel concatenation with the original image to form a dual-channel input. It is then fed into a lightweight U-Net encoder to extract multi-scale features. After dimensionality is increased via a 1×1 convolution, it is fed into the Mamba module for global modeling. Finally, the deformation field prediction head outputs a deformation displacement map, which is then combined with a spatial transformation network to generate the registered image.
[0043] The Adam optimizer is used during training, with an initial learning rate of 3e-4 and a learning rate decay strategy using the ReduceLROnPlateau strategy. The loss function uses the local normalized cross-correlation score, making it suitable for medical image registration tasks requiring precise alignment and can be embedded in deep learning models for end-to-end training.
[0044] During the validation and testing phases, the deformation field output by the model was used to align the moving image to the fixed image coordinate system. Registration accuracy was evaluated using the Dice Similarity Coefficient (DSC), Normalized Mutual Information (NMI), and Structural Similarity Index (SSIM). Furthermore, a visualization tool was used to overlay the predicted deformation field on the original image, visually demonstrating the changes before and after registration.
[0045] Experimental results demonstrate that our method achieves superior registration performance on Task 1 (brain tumor task) of the MSD dataset compared to traditional methods and mainstream deep learning methods. The method achieves particularly significant improvements in alignment accuracy in the tumor region, demonstrating its effectiveness and practicality. The specific results are shown in the table below.
[0046] Table 1 Comparative experiment
[0047]
[0048] In summary, this paper proposes a brain tumor medical image registration method based on an improved Mamba. By introducing an edge enhancement mechanism, a lightweight U-Net feature extractor, and an improved Mamba state-space modeling module, the method significantly improves the accuracy and stability of image registration. This method is not only applicable to multi-phase MRI image registration for brain tumor patients, but can also be extended to other medical image analysis scenarios requiring high-precision non-rigid registration, showing broad application prospects.
[0049] The above embodiments merely illustrate several implementations of the present invention, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention. It should be noted that a person skilled in the art would be able to make numerous variations and improvements without departing from the spirit of the present invention, and all such variations and improvements fall within the scope of protection of the present invention.
Claims
1. A brain tumor medical image registration method based on improved Mamba, characterized in that: The specific steps are as follows: S1: Model construction and initialization subsystem, which loads pre-trained or randomly initialized modular network structures, adopts a lightweight U-Net encoder with edge enhancement mechanism, Mamba state space modeling module and deformation field prediction head; S2: Data management and distribution pipeline, which performs normalization, size adaptation, and multi-channel concatenation on input medical images to generate input tensors suitable for model training and inference; S3: Model training and optimization framework, which uses a gradient descent optimizer to update model parameters during training and combines it with a learning rate scheduling strategy to improve model convergence efficiency. S4: Model performance evaluation and visualization tool, using deformation consistency evaluation index and reconstruction error as the main evaluation criteria, providing deformation field visualization and image registration result comparison functions.
2. The improved Mamba-based brain tumor medical image registration method according to claim 1, characterized in that: In S1, the Mamba module has been modified and integrated in a targeted manner, specifically including the following: S1-1: Image edge enhancement unit, which designs an edge extraction module based on the Sobel operator to enhance the local structural information of the input image to improve the registration accuracy; S1-2: Feature fusion encoder, which adopts the encoder-decoder structure of lightweight U-Net architecture and combines the skip connection mechanism to fuse the feature representation of the original image and the edge map; S1-3: Feature dimensionality mapping layer, which uses 1×1 convolution to upgrade the low-dimensional features extracted by CNN to the high-dimensional embedding space required by Mamba; S1-4: Mamba sequence modeling component, which flattens the two-dimensional image into a one-dimensional sequence and inputs it into the Mamba module to capture the global dependencies in the image and restore it to a two-dimensional feature map for subsequent prediction.
3. The improved Mamba-based brain tumor medical image registration method according to claim 1, characterized in that: The data management and distribution pipeline in S2 specifically includes the following: S2-1: Medical image preprocessing unit, which performs basic processing operations such as standardization and size cropping of input images to ensure consistency of data from different sources; S2-2: Data stitching and channel fusion module, which stitches a fixed image with its corresponding edge map, and a moving image with its edge map, to form a dual-channel input, enhancing the model's perception capabilities; S2-3: Image resizing adapter, which supports image scaling operations to ensure that images are of uniform size during training and can be restored to their original resolution during inference; S2-4: Batch data generator, which extracts samples from the dataset based on the training batch size to ensure that each batch of data meets the principles of uniformity and randomness.
4. The brain tumor medical image registration method based on improved Mamba according to claim 1, characterized in that: The model training and optimization framework in S3 specifically includes the following: S3-1: Forward propagation calculation unit, which performs the forward propagation process of the model on each batch of data and outputs the predicted deformation field tensor; S3-2: Back propagation and gradient update unit, which calculates the gradient according to the deformation consistency loss function and updates the model parameters through an optimization algorithm (such as Adam); S3-3: Dynamic learning rate decay strategy, dynamically adjusts the learning rate according to the number of training rounds and validation set performance to improve the model's generalization ability.
5. The brain tumor medical image registration method based on improved Mamba according to claim 1, characterized in that: The model performance evaluation and visualization tools in S4 specifically include the following: S4-1: Basic performance evaluation module, which calculates and reports basic registration quality indicators such as reconstruction error and pixel-level displacement consistency; S4-2: Advanced deformation assessment method, which introduces more complex registration evaluation indicators such as structural similarity (SSIM) and mutual information (MI) to improve evaluation accuracy; S4-3: Monitoring and visualization of the training process. Use charts to plot the loss curves and evaluation indicator trends during the training and validation phases to assist in observing the model convergence speed. S4-4: Visual display of deformation field and image registration. The predicted deformation field is superimposed on the original image and compared with the actual registration result to provide an intuitive visual evaluation effect.