3D medical image registration method based on optimization iteration and spatial feature Mama

Abstract

The invention discloses a 3D medical image registration method based on optimization iteration and spatial feature Mamba, which belongs to the technical field of medical image processing, and utilizes a double-flow pyramid architecture and an optimization iteration technology to convert a registration task into a multi-scale gradual optimization process. In the process, a spatial feature extraction module SMB and an optimization iteration mechanism are combined, the spatial relationship among multi-scale features is effectively captured, and accurate alignment of complex deformation is ensured through incremental updating of a deformation field. Meanwhile, the content consistency principle ensures the anatomical structure integrity between the source image and the registration result. Through the method, the registration result with higher precision and smooth deformation field can be generated, the registration quality is obviously superior to that of various existing advanced algorithms, and excellent applicability is shown in clinical tasks. Meanwhile, the method shows excellent performance in specific tasks of medical image analysis, such as anatomical structure segmentation, and provides powerful technical support for improving medical image processing efficiency and diagnosis accuracy.

Description

Technical Field

[0001] The present invention belongs to the technical field of medical image processing, and in particular relates to a 3D medical image registration method based on optimization iteration and spatial feature Mamba. Background Art

[0002] In the field of medical image registration algorithm research, many scholars have proposed a variety of registration methods. Early methods were mainly based on classical optimization algorithms, which achieved alignment by solving nonlinear transformation functions between images. These methods include free deformation model (FFD), large deformation differential homeomorphism metric mapping (LDDMM), and optical flow-based registration technology. Although these methods are theoretically highly applicable, they require pixel-by-pixel optimization and have high computational complexity, making it difficult to meet the real-time processing requirements of high-resolution medical images.

[0003] With the development of deep learning, data-driven registration methods have been widely studied. Among them, registration models based on convolutional neural networks (such as VoxelMorph) have significantly improved registration efficiency and accuracy through end-to-end training. However, such methods often find it difficult to balance global consistency and local accuracy when dealing with large-scale complex deformations. In addition, the introduction of Transformer provides a new idea for medical image registration. Methods such as TransMorph use the self-attention mechanism to capture long-range dependencies and achieve good registration results. However, the quadratic complexity of the Transformer model brings significant computational costs on high-resolution medical images.

[0004] As an efficient sequence modeling method based on the state space model (SSM), the Mamba framework has gradually become an important tool in generation tasks due to its linear complexity and long sequence processing capabilities. Some studies have attempted to apply the Mamba framework to medical image registration. For example, MambaMorph has shown excellent performance in multimodal image registration. However, directly using Mamba for registration tasks may lead to the loss of some spatial information because it uses a one-dimensional sequence flattening operation, while the three-dimensional spatial features in medical images are crucial for accurate registration. The iterative module is gradually refined at different scales, and the spatial feature extraction module SMB is introduced to enhance the modeling capability of three-dimensional spatial features. Summary of the invention

[0005] In order to overcome the problems in the background technology, the present invention provides a 3D medical image registration method based on optimization iteration and spatial feature Mamba, which combines a dual-stream pyramid architecture and an optimization iteration module to gradually refine the deformation field at different scales, and introduces a spatial feature extraction module SMB to enhance the modeling capability of three-dimensional spatial features.

[0006] To achieve the above object, the present invention is implemented through the following technical solutions:

[0007] S1, using a dual-stream weight sharing feature encoder to transform the motion image I m and a fixed image I f As input, a 5-layer convolutional block is used to extract hierarchical features and obtain feature maps M of different scales. i and F i , where i represents the i-th scale;

[0008] S2, according to S1 m and I f The smallest feature map F 5 and M 5 The spatial features are then sent to the spatial feature extraction module to retain the spatial features and combined with F 5 and M 5 Fusion obtains the feature image F;

[0009] S3, according to the processing of the feature image F in S2, the feature image F incorporating the spatial features is processed using the Mamba module, and the initial deformation field is formed through the registration head

[0010] S4. According to The optimization iteration module performs multiple iterations to form a temporary deformation field φ 1 ;

[0011] S5, M 4 Through the temporary deformation field φ 1 After deformation and F 4 The final deformation field φ is used to extract the spatial features of the moving image I using the feature extraction module. m The registered image can be obtained by deformation

[0012] S6. Apply the registration results to advanced medical image analysis tasks (such as anatomical structure segmentation, lesion detection, or functional region analysis) for verification to evaluate the practical application effect and reliability of the registration method.

[0013] The beneficial effects of the present invention are as follows:

[0014] In view of the problems that traditional medical image registration methods are inefficient when dealing with large-scale complex deformations and insufficient modeling of spatial relationships of anatomical structures, the present invention proposes a 3D medical image registration method based on optimization iteration and spatial feature Mamba, called PSMamba. By combining the dual-stream pyramid architecture with the optimization iteration module, PSMamba extracts global and local information from multi-scale features, and effectively models the spatial relationships of anatomical structures by introducing the spatial feature extraction module SMB. In order to resolve the contradiction between accuracy and smoothness in deformation field prediction, the present invention designs a deformation field update strategy based on incremental optimization, which gradually refines the registration results at different scales, significantly improving the registration efficiency and accuracy.

[0015] The present invention also introduces a content consistency loss function to ensure the structural integrity between the registration result and the source image, so as to better maintain the spatial consistency of anatomical features. The present invention verifies the effectiveness of its method on medical image datasets such as LPBA40 and Mindboggle, and has been successfully applied to clinical tasks such as anatomical structure segmentation and lesion detection, demonstrating strong generalization ability and practicality. Through comparative experiments with the 15 most advanced registration methods currently available, the results show that PSMamba exhibits excellent registration efficiency and robustness in both qualitative and quantitative evaluation, especially in large-scale deformation and complex structure processing, which is significantly superior to existing technologies. This innovative achievement provides a more efficient and accurate solution for medical image analysis, which helps to improve the accuracy of clinical diagnosis and the scientific nature of treatment plans. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 It is a flowchart of the 3D medical image registration method provided by the present invention;

[0017] Figure 2 It is a schematic diagram of the spatial feature Mamba module provided by the present invention;

[0018] Figure 3 It is a schematic diagram of the optimization iteration module provided by the present invention;

[0019] Figure 4 This is a comparison chart of the registration effect of the present invention on the LPBA dataset with a variety of the most advanced medical image registration methods;

[0020] Figure 5 It is a result diagram of each registration pair of different medical image registration methods on the Mindboggle dataset provided by the present invention. DETAILED DESCRIPTION

[0021] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0022] like Figure 1-Figure 5 As shown, a 3D medical image registration method based on optimization iteration and spatial feature Mamba includes the following steps, wherein the symbols used are described as shown in Table 1:

[0023] Table 1

[0024]

[0025] S1, using a dual-stream weight sharing feature encoder to transform the motion image I m and a fixed image I f As input, a 5-layer convolutional block is used to extract hierarchical features and obtain feature maps M of different scales. i and F i , where i represents the i-th scale;

[0026] S2, according to S1 m and I f The smallest feature map F 5 and M 5 The spatial features are then sent to the spatial feature extraction module to retain the spatial features and combined with F 5 and M 5 Fusion obtains the feature image F;

[0027] S3, according to the processing of the feature image F in S2, the feature image F incorporating the spatial features is processed using the Mamba module, and the initial deformation field is formed through the registration head

[0028] S4. According to The optimization iteration module performs multiple iterations to form a temporary deformation field φ 1 ;

[0029] S5, M 4 Through the temporary deformation field φ 1 After deformation and F 4 The final deformation field φ is used to extract the spatial features of the moving image I using the feature extraction module. m The registered image can be obtained by deformation

[0030] S6. Apply the registration results to advanced medical image analysis tasks (such as anatomical structure segmentation, lesion detection, or functional region analysis) for verification to evaluate the practical application effect and reliability of the registration method.

[0031] Furthermore, the principle of the spatial feature extraction module in S2 is to filter the local area of ​​the image by using spatial convolution, which can extract local spatial information in the image, such as edges, textures, local shapes, etc. This step can be expressed as:

[0032]

[0033] where C denotes the convolutional block, μ(*) and σ(*) are the mean and variance of each channel, γ and β are learnable scaling and translation parameters, ε is a small constant to prevent division by zero, and k is the size of the convolution kernel.

[0034] Then, the information flow in the feature map is dynamically adjusted through the gating mechanism, so that the network can selectively retain or suppress spatial features, and finally reuse the features with residual links. This process is expressed as:

[0035] (F u ,F v )=S(C sfe (F 5 +M 5 )

[0036]

[0037] Where S(*) is the partition function, F u and F v The input feature is divided into two independent parts along the channel dimension, and ° represents element-by-element multiplication.

[0038] Furthermore, the principle of the Mamba module in step S3 to process the feature image is: based on the mathematical framework of the state space model (SSM), it captures the long-term dependency and global feature relationship in the input sequence, and combines with the subsequent registration head module to provide more accurate global information modeling for the medical image registration task; wherein, the SSM generates hidden states through the input sequence and outputs a feature sequence, and the specific mathematical expression is:

[0039] h'(t)=Ah(t)+Bx(t)

[0040] y(t)=Ch(t)+Dx(t)

[0041] Where x(t)∈R is the one-dimensional sequence of the flattened feature image, h((t)) is the intermediate representation of the hidden state used to describe the system, and A∈RN×N is the transfer matrix, B, C∈R N×1 is the projection parameter, D∈R is the skip connection, y(t) is the output sequence, and represents the feature map after modeling.

[0042] Since SSM is a mathematical model of a continuous-time system, and the data in deep learning is discrete, SSM needs to be discretized. Here, the zero-order hold method (ZOH) is used to convert the continuous-time equation into discrete-time form:

[0043]

[0044] in represents the discretization of the state transfer matrix, represents the discretization of the input matrix, Indicates that the projection matrix remains unchanged.

[0045] Finally, the feature map after Mamba modeling is input into the registration head RH(*) to obtain the initial deformation field Right now:

[0046] Furthermore, the principle of the optimization iteration module in step S4 is: using the characteristics of low-resolution feature maps that pay more attention to the global structure, multiple iterations are performed on the minimum-scale feature maps to quickly align the global deformation, and each iteration only updates the incremental deformation field φ' k , superimposed on the previous deformation field φ k This incremental update method can gradually refine the deformation field while avoiding excessive deformation deviation. The process can be expressed as:

[0047] φ 1 =SMB(F m ,F f )

[0048]

[0049] φ k =φ k-1 +φ′ k ,k=2,3,…,K

[0050] Where SMB(*) is the integrated representation of steps S2 and S3, and K is the number of iterations.

[0051] Furthermore, the principle of pyramid cascading in step S5 is: using a lower resolution feature map (such as F 5 ,M 5 ) to make a rough deformation field prediction, and then gradually build on higher resolution feature maps (such as F 4 ,M4 , until F 1 ,M 1 ) to further refine the deformation field. The deformation field prediction of each layer is superimposed on the previous deformation field, thus gradually generating a more accurate deformation field.

[0052] Furthermore, after step S6 is completed, the proposed registration method is integrated into the medical image acquisition and processing system, directly embedded in the existing medical imaging equipment (such as MRI or CT scanner) or post-processing module, to ensure that the registration method can operate efficiently in various real medical environments to meet actual clinical needs. Among them, steps S2-S3 are obtained Figure 2 , step S4 gives Figure 3 By calculating the fusion results of different advanced methods and applying them to different objective indicators, we can obtain Figure 4 and Figure 5 .

[0053] Specifically Figure 2 The structure and working principle of the spatial feature Mamba module proposed in the present invention are shown. The module uses the linear state space model (SSM) of Mamba for feature extraction and combines spatial feature modeling to enhance the accuracy of medical image registration. Figure 3 The optimization iteration module proposed in the present invention is used for the optimization calculation process of medical image registration. The module adopts a step-by-step optimization strategy and continuously adjusts the deformation field through multiple iterations to make the registration result more accurate.

[0054] Figure 4 The registration effect of the method of the present invention on the LPBA dataset is demonstrated and compared with multiple state-of-the-art medical image registration methods. From left to right, each column represents a registration method, and the method of the present invention is located in the last column. From top to bottom, each row corresponds to: distorted image, segmentation label of distorted image, color heat map of distorted image, deformation field and deformation grid. It can be clearly seen from the figure that the present invention has better accuracy.

[0055] Figure 5 The performance of the method of the present invention and other registration methods on the Mindboggle dataset is shown. The curves in the figure reflect the changes in DSC of different methods on different image pairs. It can be seen from the figure that the present invention has better stability and generalization ability than other methods.

[0056] The preferred embodiments of the present invention disclosed above are only used to help illustrate the present invention. The preferred embodiments do not describe all the details in detail, nor do they limit the invention to the specific implementation methods described. Obviously, many modifications and changes can be made according to the content of this specification. This specification selects and specifically describes these embodiments in order to better explain the principles and practical applications of the present invention, so that those skilled in the art can understand and use the present invention well. The present invention is limited only by the claims and their full scope and equivalents.

Claims

1. A 3D medical image registration method based on optimization iteration and spatial feature Mamba, characterized in that: The following steps are involved: S1, using a dual-stream weight sharing feature encoder to transform the motion image I m and a fixed image I f As input, a 5-layer convolutional block is used to extract hierarchical features and obtain feature maps M of different scales. i and F i , where i represents the i-th scale; S2, according to S1 m and I f The smallest-scale feature maps F5 and M5 are sent to the spatial feature extraction module to retain the spatial features, and are fused with F5 and M5 to obtain the feature image F; S3, according to the processing of the feature image F in S2, the feature image F incorporating the spatial features is processed using the Mamba module, and the initial deformation field is formed through the registration head S4. According to The processing is performed in the optimization iteration module for multiple iterations to form a temporary deformation field φ1; S5, after M4 is deformed by the temporary deformation field φ1, it is input into the feature extraction module together with F4 to obtain spatial features, and this process is repeated at different scales by means of pyramid cascade to obtain the final deformation field φ, and the final deformation field φ is used to extract the motion image I m The registered image can be obtained by deformation S6. Apply the registered images to advanced medical image analysis tasks for verification to evaluate the practical application effect and reliability of the registration method. The advanced medical image analysis tasks include: anatomical structure segmentation, lesion detection and functional area analysis.

2. A 3D medical image registration method based on optimization iteration and spatial feature Mamba according to claim 1, characterized in that: The extraction process of the spatial feature extraction module in S2 is as follows: First, we filter the local area of ​​the image using spatial convolution. Spatial convolution can extract local spatial information in the image, including edges, textures, and local shapes. The extraction process is expressed as: Where C represents the convolution block, μ(*) and σ(*) are the mean and variance of each channel, γ and β are learnable scaling and translation parameters, ε is a small constant to prevent division by zero, and k is the size of the convolution kernel; Then, the information flow in the feature map is dynamically adjusted through the gating mechanism, so that the network can selectively retain or suppress spatial features, and finally reuse the features with residual links. This process is expressed as: (F u ,F v )=S(C sfe (F5+M5)) Where S(*) is the partition function, F u and F v The input feature is divided into two independent parts along the channel dimension, and ° represents element-by-element multiplication.

3. The 3D medical image registration method based on optimization iteration and spatial feature Mamba according to claim 1, characterized in that: The Mamba module in S3 processes the feature image, specifically including: Firstly, based on the mathematical framework of the state space model SSM, the long-term dependency and global feature relationship in the input sequence are captured, and combined with the subsequent registration head module, a more accurate global information modeling is provided for the medical image registration task; wherein, SSM generates hidden states through the input sequence and outputs a feature sequence, and the specific mathematical expression is: h'(t)=Ah(t)+Bx(t) y(t)=Ch(t)+Dx(t) Where x(t)∈R is the one-dimensional sequence of the flattened feature image, h(t) is the intermediate representation of the hidden state used to describe the system, and A∈R N×N is the transfer matrix, B, C∈R N×1 is the projection parameter, D∈R is the skip connection, y(t) is the output sequence, and represents the feature map after modeling; Secondly, the zero-order hold method ZOH is used to discretize the SSM to convert the continuous time equation into discrete time form, and its expression is as follows: in represents the discretization of the state transfer matrix, represents the discretization of the input matrix, Indicates that the projection matrix remains unchanged; Finally, the feature map after Mamba modeling is input into the registration head RH(*) to obtain the initial deformation field Right now:

4. The 3D medical image registration method based on optimization iteration and spatial feature Mamba according to claim 1, characterized in that: The optimization iteration in S4 specifically includes: Taking advantage of the fact that low-resolution feature maps pay more attention to the global structure, multiple iterations are performed on the smallest-scale feature maps F5 and M5 to quickly align the global deformation, and only the incremental deformation field φ' is updated in each iteration. k , superimposed on the previous deformation field φ k The process can be expressed as: φ1=SMB(F m ,F f ) f k =φ k-1 +φ′ k ,k=2,3,…,K Where SMB(*) is the integrated representation of steps S2 and S3, and K is the number of iterations.

5. The 3D medical image registration method based on optimization iteration and spatial feature Mamba according to claim 1, characterized in that: The pyramid cascading in S5 specifically includes: A rough deformation field prediction is made using a lower-resolution feature map, and then the deformation field is further refined on a higher-resolution feature map. The deformation field prediction of each layer is superimposed on the previous deformation field, thereby gradually generating a more accurate deformation field.

6. A medical image acquisition and processing system, characterized in that: The system integrates a 3D medical image registration method based on optimization iteration and spatial feature Mamba as described in any one of claims 1 to 5.

7. A medical imaging device, characterized in that: The medical image acquisition and processing system described in claim 6 is embedded in the device.

Images