Multi-modal MRI (Magnetic Resonance Imaging) brain tumor image segmentation method based on channel rearrangement lightweight network

Abstract

The invention discloses a multi-modal MRI (Magnetic Resonance Imaging) brain tumor image segmentation method based on a channel rearrangement lightweight network, relates to the field of medical image analysis, and provides a series of innovative solution strategies for a medical image segmentation task, especially for the problem of efficient utilization of computing resources in a brain tumor image segmentation task. By introducing a channel rearrangement and grouping strategy and a quaternion grouping convolution method and combining with cavity convolution with gradually increased expansion rate, the feature expression ability and receptive field are effectively improved, the light weight of the model is kept, and information in and between feature layers is effectively utilized. While the computing resource and memory requirements are reduced, the information between the feature layers and in the feature layers can be efficiently utilized.

Description

Technical Field

[0001] The present invention relates to a multimodal MRI brain tumor image segmentation method based on a channel rearrangement lightweight network, and mainly relates to the field of medical image analysis. Background Art

[0002] In the field of medical image analysis, especially in the task of brain tumor image segmentation, deep learning models have made significant progress. However, this also comes with the need for efficient use of computing resources. In the face of this challenge, there are currently three main strategies:

[0003] First, develop efficient feature map compression technology to reduce redundant information, reduce computational complexity, and improve model performance. For example, channel pruning technology reduces the number of parameters and computational complexity by removing unnecessary channels, and sparse constraints optimize the sparsity of the model.

[0004] Secondly, a channel-level attention mechanism is introduced for feature map compression. This mechanism can dynamically adjust the importance of channels and help the model focus on more critical information. The latest work integrates channel pruning and knowledge distillation, as well as the comprehensive application of channel pruning, quantization, and tensor decomposition.

[0005] Finally, reduce redundancy in neural networks. Group convolution and depth-separable convolution (including depth-wise convolution and point-wise convolution) have been widely used in mobile and edge computing networks as effective strategies. These techniques help create lighter models by reducing the redundancy of filters.

[0006] However, increasing the width of the network may lead to an increase in memory access. For example, the network proposed in Chapter 3 uses depthwise separable convolution, which leads to long training time and high memory usage. On the other hand, although partial convolution methods effectively utilize the redundancy in feature maps and reduce the number of floating-point operations and memory access, they perform poorly in terms of feature utilization efficiency.

[0007] In addition, in the classic design framework of segmentation models, the pursuit of high accuracy is often accompanied by the construction of a deeper network structure. This deep network extracts features with a large receptive field, enabling the model to capture information more comprehensively, thereby more effectively understanding and processing segmentation tasks. In order to expand the receptive field without significantly increasing the depth of the network, a common practice is to stack dilated convolutions, but this method mainly focuses on the independent pixels within each layer, and does not sufficiently consider the information interaction between layers. Summary of the invention

[0008] In view of the above deficiencies in the prior art, the present invention proposes a multimodal MRI brain tumor image segmentation method based on a channel rearrangement lightweight network.

[0009] To achieve the above object, the technical solution of the present invention is: comprising the following steps:

[0010] S1 first performs a channel rearrangement strategy on the input feature map;

[0011] S2 then extracts features from the selected group, first grouping the feature maps of the selected group again, so that the number of channels in each group meets the requirements of quaternion convolution;

[0012] S3 uses dilated convolution with different expansion rates to convolve the regrouped feature maps by multiplying quaternion matrices.

[0013] S4 finally concatenates each group and performs 1×1 convolution to complete the feature extraction operation.

[0014] Preferably, step S1: grouping is achieved by introducing a grouping ratio parameter r, which divides the total number of channels C into several groups, each group containing C / r channels.

[0015] Preferably, in step S2, the unique structure of quaternion is applied to the scene of MRI three-dimensional medical image segmentation, and quaternion can be expressed as:

[0016] R=t+yi+uj+pk

[0017] Where t, y, u, and p are real numbers, and i, j, and k are the unit bases of quaternions.

[0018] The base has the following properties:

[0019] i 2 =j 2 =k 2 =ijk=-1

[0020] ki=j, jk=i, ij=k, ji=-k, ik=-j, kj=-i

[0021] Suppose there are two quaternions:

[0022] R1=t1+y1i+u1j+p1k

[0023] R2=t2+y2i+u2j+p2k

[0024] Then their addition and scalar multiplication can be defined as:

[0025] R1+R2=t1+t2+(y1+y2)i+(u1+u2)j+(p1+p2)k

[0026] αR1=αt1+αy1i+αu1j+αp1k

[0027] αR2=αt2+αy2i+αu2j+αp2k

[0028] The conjugate of R * It can be expressed as

[0029] R * =t-yi-uj-pk

[0030] The Hamilton product is used instead of the standard real-valued dot product, which represents the multiplication of two quaternions and is defined as:

[0031]

[0032] In brain tumor segmentation, the four modalities of an MRI image can be divided into four parts, each of which is represented by a quaternion component.

[0033] The T1-weighted, enhanced T1-weighted, T2-weighted, and fluid-attenuated inversion recovery of the input image correspond to t, y, u, and p in Formula 3-1;

[0034] Brain tumor image segmentation requires output composed of real numbers; it converts quaternion features into ordinary real number features, that is, It is guaranteed that each element is a real number.

[0035] Preferably, in step S2, the GeLU activation function can also be used as the activation function in the quaternion group convolution, which is defined as:

[0036] y=GeLU(x0)+GeLU(x1)i+GeLU(x2)j+GeLU(x3)k

[0037] x0 represents the real component of the input, while x1, x2, and x3 represent the imaginary components of the input associated with the imaginary units i, j, and k, respectively;

[0038] The matrix multiplication between the feature layer and the weight matrix generates the corresponding output feature layer according to the Hamilton product of the quaternion;

[0039] The quaternion-valued 3D convolution is defined as:

[0040] if: E=E0+E1i+E2j+E3k,

[0041] x=x0+x1i+x2j+x3k,

[0042] b=b0+b1i+b2j+b3k,

[0043]

[0044] then:

[0045] Where E and x represent the weight matrix of the quaternion value and the input of the quaternion value (feature layer). Indicates that the matrix multiplication between the feature layer and the weight matrix is ​​based on the Hamilton product of quaternions, where the E0x0, E1x1... operations are performed by 3D convolution;

[0046] Let x in is the input image, and Represents each part of the image, x out It is obtained by performing 3D convolution on the parts between E and x; through the Hamilton product process, the quaternion weights interact with multiple quaternion input parts to explore the hidden relationships among the elements.

[0047] Preferably, in step S2, assuming that the number of input channels is C1, the grouping ratio parameter is C1 / r, the number of channels to be feature extracted is C2 / r, the number of channels after feature extraction is, the height and width of the input and output are H and w respectively, and the bias is 0; the parameter (F) can be defined as:

[0048] F=C2 / r×(C1 / r×H×W)

[0049] In the grouped convolution setting, the input layer and convolution kernels are divided into groups, and then convolution operations are performed within their respective groups. The parameters after grouped convolution are defined as follows:

[0050] F=C2 / r×(C1 / r / h×H×W)

[0051] In the brain tumor segmentation task, the number of input and output channels will change after the feature map is upsampled and downsampled; the number of weight matrices is adjusted to ensure that even if the number of channels changes, the number of channels in each group remains a multiple of four.

[0052] Preferably, step S3, quaternion grouping dilated convolution module:

[0053] Use multiple mixed dilated convolutions with different dilation rates to expand the receptive field. The dilated convolution increases the receptive field by setting the dilation rate.

[0054] The technical principles and beneficial effects of the present invention are as follows:

[0055] By introducing the channel rearrangement and grouping strategy, the implicit expansion of the model training data volume is achieved, thereby improving the efficiency of feature usage. This strategy can identify and extract key information in the brain tumor segmentation task with higher accuracy. A quaternion grouped convolution method is proposed, which reduces the model parameters and effectively utilizes the features between layers by grouping the convolution operations and applying quaternion convolution to each group. In order to effectively improve the feature expression and expand the receptive field, the quaternion grouped convolution is combined with the dilated convolution with gradually increasing dilation rate. This strategy not only enhances the model's ability to capture information within and between feature layers, but also effectively utilizes affinity pixels while ensuring the lightweight of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only five of the drawings of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0057] Figure 1 Provide channel rearrangement and grouping strategies;

[0058] Figure 2 It is the feature extraction process;

[0059] Figure 3 It is a quaternion value 3D convolution;

[0060] Figure 4 Visualize pixel contributions in dilated convolutions at different dilation rates;

[0061] Figure 5 Dilated convolution module for quaternion grouping. DETAILED DESCRIPTION

[0062] The technical solutions in the present invention will be described clearly and completely below in conjunction with the accompanying drawings. Obviously, the described embodiments are only preferred embodiments of the present invention, not all embodiments. Based on the embodiments in 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.

[0063] Example

[0064] The steps include:

[0065] S1 first performs a channel rearrangement strategy on the input feature map;

[0066] S2 then extracts features from the selected group, first grouping the feature maps of the selected group again, so that the number of channels in each group meets the requirements of quaternion convolution;

[0067] S3 uses dilated convolution with different expansion rates to convolve the regrouped feature maps by multiplying quaternion matrices.

[0068] S4 finally concatenates each group and performs 1×1 convolution to complete the feature extraction operation.

[0069] The details are as follows:

[0070] 1. Refined design of segmentation model

[0071] 1.1 Channel rearrangement and grouping strategy

[0072] In order to optimize the use of data and minimize redundancy, this embodiment provides a strategy that combines channel reordering and grouping. Through this strategy, not only can the capacity of data processing be increased, but also the required number of parameters can be reduced by processing a portion of the channels in groups. In this way, the model can focus on some channels in the input feature map, thereby improving the specificity and efficiency of processing. In addition, channel reordering also enhances the diversity of training data, further improving the performance of the model.

[0073] Figure 1 Four proposed channel reordering and grouping methods are shown. There is a high similarity between the feature maps of different channels, so the input channels are grouped to select some channels. Grouping is achieved by introducing a grouping ratio parameter r, which divides the total number of channels C into several groups, each group contains C / r channels.

[0074] Figure 1 (a) shows the first channel rearrangement and grouping strategy, which first increases the diversity of data by randomly adjusting the channel order of the input feature map, then divides these channels into several groups, and finally selects one group from these groups for feature extraction as output. This operation of randomizing the channel order introduces a wider range of data changes for model training, which helps improve the model's adaptability to different data characteristics.

[0075] Figure 1 (b) shows the second channel rearrangement and grouping strategy. First, the module splits the input feature map into equal-sized blocks along the channel dimension, and then performs a flip operation on the channels within each block, that is, the first channel becomes the last, and so on. This step breaks the original order between channels within each block, introducing a simple but effective form of data perturbation. Then, these flipped blocks are further rearranged and the order of the blocks is adjusted, which increases the complexity of the processing and the diversity of the data.

[0076] Figure 1(c) shows the third channel rearrangement and grouping strategy, which cyclically shifts the channels of the input feature map instead of randomly shuffling the order of the channels as in the first strategy. Then, one group is selected from these groups for output as the input of the feature extraction module. The cyclic shift strategy maintains the relative order between channels but changes their absolute positions. In this way, the model can learn features from different channel configurations and exploit the spatial relationship between channels.

[0077] Figure 1 (d) shows the fourth channel rearrangement and grouping strategy, which first completes the cyclic shift and grouping of the overall feature map according to the third strategy, and then performs a cyclic shift operation within each group. In this way, the channels of each group are also reordered according to a certain shift amount. This method combines the advantages of group processing and channel shifting, improving the flexibility and efficiency of the model in processing input data.

[0078] After completing the channel rearrangement and grouping operations, the feature extraction module performs extraction operations on part of the channels of the rearranged feature map, while keeping the other channels unchanged. The process is as follows: Figure 2 shown.

[0079] The feature extraction module of the first branch of this embodiment uses a quaternion grouped convolution module with a gradually increasing dilation rate, and the feature extraction modules of the remaining branches still use the 3D convolution combination used in Chapter 2. In this way, the module can use convolution to extract and learn new features while retaining some original features. This method of combining channel flipping, rearrangement and partial convolution operations is effective in improving the performance of brain tumor segmentation tasks.

[0080] 2Quaternion Grouped Convolution

[0081] Quaternions are an extension of the concept of complex numbers, consisting of a real part and three imaginary parts. They show special applicability when dealing with rotation problems in three-dimensional space. Inspired by this unique structure, this embodiment applies the unique structure of quaternions to the scene of MRI three-dimensional medical image segmentation. By combining the advantages of quaternions and grouped convolution, the lightweight model is further achieved. Quaternions can be expressed as:

[0082] R=t+yi+uj+pk

[0083] Where t, y, u, and p are real numbers, and i, j, and k are the unit bases of quaternions.

[0084] The base has the following properties:

[0085] i 2 =j 2 =k 2=ijk=-1

[0086] ki=j, jk=i, ij=k, ji=-k, ik=-j, kj=-i

[0087] Suppose there are two quaternions:

[0088] R1=t1+y1i+u1j+p1k

[0089] R2=t2+y2i+u2j+p2k

[0090] Then their addition and scalar multiplication can be defined as:

[0091] R1+R2=t1+t2+(y1+y2)i+(u1+u2)j+(p1+p2)k

[0092] αR1=αt1+αy1i+αu1j+αp1k

[0093] αR2=αt2+αy2i+αu2j+αp2k

[0094] The conjugate of R * It can be expressed as

[0095] R * =t-yi-uj-pk

[0096] The Hamilton product is used instead of the standard real-valued dot product, which represents the multiplication of two quaternions and is defined as:

[0097]

[0098] In brain tumor segmentation, the four modalities of an MRI image can be divided into four parts, each of which is represented by a quaternion component.

[0099] The T1-weighted (T1), enhanced T1-weighted (T1ce), T2-weighted (T2), and fluid-attenuated inversion recovery (FLAIR) images of the input images correspond to t, y, u, and p in Formula 3-1;

[0100] However, in actual calculations, quaternions are converted to real numbers because most segmentation tasks, such as brain tumor image segmentation, require outputs consisting of real numbers; the quaternion features are converted into ordinary real number features, that is, It is guaranteed that each element is a real number.

[0101] The GeLU activation function can also be used as the activation function in quaternion group convolution, defined as:

[0102] y=GeLU(x0)+GeLU(x1)i+GeLU(x2)j+GeLU(x3)k

[0103] x0 represents the real component of the input, while x1, x2, and x3 represent the imaginary components of the input associated with the imaginary units i, j, and k, respectively;

[0104] Hamilton product intuitively promotes the potential interaction between two quaternions R1 and R2. This inspires us to apply Hamilton product in 3D convolution operation. That is, the matrix multiplication between feature layer and weight matrix generates the corresponding output feature layer according to Hamilton product of quaternions.

[0105] The quaternion-valued 3D convolution is defined as:

[0106] if: E=E0+E1i+E2j+E3k,

[0107] x=x0+x1i+x2j+x3k,

[0108] b=b0+b1i+b2j+b3k,

[0109]

[0110] then:

[0111] Where E and x represent the weight matrix of the quaternion value and the input of the quaternion value (feature layer). Indicates that the matrix multiplication between the feature layer and the weight matrix is ​​based on the Hamilton product of quaternions, where the E0x0, E1x1... operations are performed by 3D convolution;

[0112] Let x in is the input image, and Represents each part of the image, x out It is obtained by performing 3D convolution on each part between E and x; through the process of Hamilton product, the quaternion weights interact with multiple quaternion input parts to explore the hidden relationships in the elements. Therefore, stacking 3D convolutions with different expansion rates and performing correlation learning in the first branch in the manner of Hamilton product becomes the core strategy to enhance the ability of inter-layer feature interaction. At the same time, assuming that the input feature map contains four channels and the number of channels of each part is one, then through the quaternion numerical 3D convolution, only four weight matrices (3D convolution with one channel per matrix) are needed to obtain the output feature map with four channels. Compared with traditional convolution, which requires 16 weight matrices to achieve the same output effect, quaternion convolution can reduce the number of parameters by 75%.

[0113] Combination Figure 3As shown in the figure, the quaternion value 3D convolution reduces 3 / 4 of the parameters. In addition, the quaternion value 3D convolution can further reduce the parameters through channel rearrangement and grouping strategy and grouped convolution: let the number of input channels be C1, the grouping ratio parameter be r, then the number of channels to be feature extracted is C1 / r, the number of channels after feature extraction is C2 / r, the height and width of the input and output are H and w respectively, and the bias is 0; the parameter (F) can be defined as:

[0114] F=C2 / r×(C1 / r×H×W)

[0115] In the grouped convolution setting, the input layer and convolution kernels are divided into groups, and then convolution operations are performed within their respective groups. The parameters after grouped convolution are defined as follows:

[0116] F=C2 / r×(C1 / r / h×H×W)

[0117] The number of parameters of group convolution is 1 / h of regular convolution. In order to minimize the number of parameters, the minimum number of channels allowed in quaternion value 3D convolution is 4, so when the number of groups is C / 4, it has the minimum number of parameters.

[0118] Quaternion numerical 3D convolution is used for each group of feature maps, and then all output layers of each group are connected in series to obtain the final output feature map. This method can be further lightweight. The parameters of quaternion group convolution are defined as follows:

[0119]

[0120] In the brain tumor segmentation task, the number of input and output channels will change after the feature map is upsampled and downsampled; the number of weight matrices is adjusted to ensure that even if the number of channels changes, the number of channels in each group remains a multiple of four.

[0121] 3 Quaternion grouping hole convolution module:

[0122] Using large convolution kernels to expand the receptive field will increase the model training time, which is especially obvious in the case of 3D convolution. Therefore, in this section, we use dilated convolution to expand the receptive field. Dilated convolution increases the receptive field by setting the dilation rate, but this method introduces a grid effect. This effect means that in the process of continuous convolution, if the dilation rate remains consistent, the adjacent pixels used in each layer will be processed independently, thus affecting the effect. To solve this problem, recent work has proposed using multiple mixed dilated convolutions with different dilation rates to optimize the effect of continuous convolution. Figure 4As shown in the figure, we can clearly see the contribution of each pixel by using continuous dilation convolutions with different dilation rates. The colored grid in the figure shows the pixels involved when the convolution kernel slides through the feature layer to output a pixel after three layers of convolution operations. The darker the color, the more frequently the pixel is used.

[0123] Figure 3 The complete process of the quaternion grouped atrous convolution module is shown. The module first groups the input features, and then sequentially applies 3D quaternion convolutions with dilation rates of 1, 2, and 3 to each group. After processing all the groups, the module concatenates the outputs and fuses the information of all channels through 1×1 convolution, thereby achieving efficient feature integration.

[0124] After all the grouping is completed, the module concatenates the outputs and fuses the information of all channels through 1×1 convolution to achieve efficient integration of features.

[0125] 4 Experiments

[0126] 4.1 Experimental design and results analysis

[0127] In this embodiment, we use the Dice coefficient to evaluate the model performance. In view of the reduction of model parameters, we use a small learning rate to train the network to ensure the stability and efficiency of the training.

[0128] The effectiveness of the method in this chapter is verified by comparing different mainstream models. The evaluation indicator used is the Dice coefficient, which targets three sub-regions: enhanced tumor (ET), tumor core (TC) and whole tumor (WT). The experimental results are shown in Table 1. Although the number of parameters of the method in this chapter is only 4.32M, which is much less than other comparative methods such as SegResNet (19.92M), Attention UNet (6.44M), V-Net (45.51M) and UNETR (92.58M), in terms of the performance of the Dice coefficient, the scores of the method in this chapter in ET, TC and WT are 0.878, 0.916 and 0.927 respectively, and the average Dice coefficient is 0.907, which surpasses most comparative models. This result shows that our method can effectively capture the details of the tumor area while keeping the model lightweight, and thus achieves excellent performance in the tumor segmentation task.

[0129] Table 1 Comparison of segmentation results of different models

[0130]

[0131] 4.2 Ablation Experiment

[0132] Table 2 shows the impact of different channel rearrangement and grouping strategies on model performance. The four different schemes proposed in this embodiment are compared. Scheme 1 adopts a random shuffle rearrangement strategy, which did not achieve a good improvement. Scheme 2 adopts an exchange channel rearrangement strategy, which significantly improves the model performance, and the average Dice coefficient reaches 0.907. Schemes 3 and 4 explored two strategies for cyclic shifting of channels, and the average Dice coefficients were 0.896 and 0.905, respectively. The experimental results show that compared with the traditional channel rearrangement method, the optimized channel rearrangement and grouping strategy can more effectively improve the model's segmentation accuracy of the tumor area.

[0133] Table 2 Comparison of results of different channel rearrangement and grouping strategies

[0134]

[0135] Table 3 shows the use of channel rearrangement and quaternion technology, as well as the impact of different dilation rate configurations on model performance. Through comparative experiments, it is found that the simultaneous use of channel rearrangement and quaternion technology, and adjustment of the dilation rate configuration, can significantly improve the prediction accuracy of the model in the tumor core area, the overall tumor area, and the enhanced tumor area. The experimental results show that the use of this comprehensive strategy is beneficial to improving the overall performance of the model.

[0136] Table 3 Comparison of results of different composition strategies

[0137]

[0138] The experimental results in Table 4 show that as the value increases, the number of channels in each group decreases, which can effectively reduce redundant calculations and implicitly increase the amount of training data. However, a larger value does not necessarily lead to better results. When it is too large, it leads to fewer parameters. This insufficient number of parameters may make it difficult to capture complex data information.

[0139] Table 4 Comparison of results of different grouping ratio parameters

[0140]

[0141] In Table 5, the effect of the number of channels per group in the quaternion grouped atrous convolution module on the model performance is discussed. By fixing other parameters, different numbers of channels per group have a significant impact on the model performance. In particular, when the number of channels per group is set to 4, the model achieves higher performance in various indicators. As the number of channels increases to 8, 16, and 32, the model performance fluctuates slightly in some indicators, but the overall performance does not improve significantly.

[0142] Table 5 Comparison of the results of the number of channels in each group

[0143]

[0144] The technical principles and beneficial effects of the present invention are as follows:

[0145] By introducing the channel rearrangement and grouping strategy, the implicit expansion of the model training data volume is achieved, thereby improving the efficiency of feature usage. This strategy can identify and extract key information in the brain tumor segmentation task with higher accuracy. A quaternion grouped convolution method is proposed, which reduces the model parameters and effectively utilizes the features between layers by grouping the convolution operations and applying quaternion convolution to each group. In order to effectively improve the feature expression and expand the receptive field, the quaternion grouped convolution is combined with the dilated convolution with gradually increasing dilation rate. This strategy not only enhances the model's ability to capture information within and between feature layers, but also effectively utilizes affinity pixels while ensuring the lightweight of the model.

[0146] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

Claims

1. A multimodal MRI brain tumor image segmentation method based on a channel rearrangement lightweight network, characterized in that: The steps include: S1 first performs a channel rearrangement strategy on the input feature map; S2 then extracts features from the selected group, first grouping the feature maps of the selected group again, so that the number of channels in each group meets the requirements of quaternion convolution; S3 uses dilated convolution with different expansion rates to convolve the regrouped feature maps by multiplying quaternion matrices. S4 finally concatenates each group and performs 1×1 convolution to complete the feature extraction operation.

2. The multimodal MRI brain tumor image segmentation method based on channel rearrangement lightweight network according to claim 1 is characterized by: Step S1: Grouping is achieved by introducing a grouping ratio parameter r, which divides the total number of channels C into several groups, each group containing C / r channels.

3. The multimodal MRI brain tumor image segmentation method based on channel rearrangement lightweight network according to claim 1 is characterized by: Step S2, the unique structure of quaternion is applied to the scene of MRI three-dimensional medical image segmentation. Quaternion can be expressed as: R=t+yi+uj+pk Where t, y, u, and p are real numbers, and i, j, and k are the unit basis of the quaternion. The unit basis in the quaternion has the following properties: i 2 =j 2 =k 2 =ijk=-1 ki=j, jk=i, ij=k, ji=-k, ik=-j, kj=-i Suppose there are two quaternions: R1=t1+y1i+u1j+p1k R2=t2+y2i+u2j+p2k Then their addition and scalar multiplication can be defined as: R1+R2=t1+t2+(y1+y2)i+(u1+u2)j+(p1+p2)k αR1=αt1+αy1i+αu1j+αp1k αR2=αt2+αy2i+αu2j+αp2k The conjugate of R * It can be expressed as R * =t-yiuj-pk The Hamilton product is used instead of the standard real-valued dot product, which represents the multiplication of two quaternions and is defined as: In brain tumor segmentation, the four modalities of an MRI image can be divided into four parts, each of which is represented by a quaternion component. The T1-weighted, enhanced T1-weighted, T2-weighted, and fluid-attenuated inversion recovery of the input image correspond to t, y, u, and p in Formula 3-1; Brain tumor image segmentation requires output composed of real numbers; it converts quaternion features into ordinary real number features, that is, It is guaranteed that each element is a real number.

4. The multimodal MRI brain tumor image segmentation method based on channel rearrangement lightweight network according to claim 1, characterized in that: Step S2, the GeLU activation function can also be used as the activation function in the quaternion group convolution, defined as: y=GeLU(x0)+GeLU(x1)i+GeLU(x2)j+GeLU(x3)k x0 represents the real component of the input, while x1, x2, and x3 represent the imaginary components of the input associated with the imaginary units i, j, and k, respectively; The matrix multiplication between the feature layer and the weight matrix generates the corresponding output feature layer according to the Hamilton product of the quaternion; The quaternion-valued 3D convolution is defined as: if: E=E0+E1i+E2j+E3k, x=x0+x1i+x2j+x3k, b=b0+b1i+b2j+b3k, Where E and x represent the weight matrix of the quaternion value and the input of the quaternion value (feature layer). Indicates that the matrix multiplication between the feature layer and the weight matrix is ​​based on the Hamilton product of quaternions, where the E0x0, E1x1... operations are performed by 3D convolution; Let x in is the input image, and Represents each part of the image, x out It is obtained by performing 3D convolution on the parts between E and x; through the Hamilton product process, the quaternion weights interact with multiple quaternion input parts to explore the hidden relationships among the elements.

5. The multimodal MRI brain tumor image segmentation method based on channel rearrangement lightweight network according to claim 1, characterized in that: Step S2, assuming that the number of input channels is C1, the grouping ratio parameter is C1 / r, the number of channels to be feature extracted is C2 / r, the number of channels after feature extraction is, the height and width of the input and output are H and w respectively, and the bias is 0; the parameter (F) can be defined as: F=C2 / r×(C1 / r×H×W) In the grouped convolution setting, the input layer and convolution kernels are divided into groups, and then convolution operations are performed within their respective groups. The parameters after grouped convolution are defined as follows: F=C2 / r×(C1 / r / h×H×W) In the brain tumor segmentation task, the number of input and output channels will change after the feature map is upsampled and downsampled; The number of weight matrices is adjusted to ensure that even if the number of channels changes, the number of channels in each group remains a multiple of four.

6. The multimodal MRI brain tumor image segmentation method based on channel rearrangement lightweight network according to claim 1, characterized in that: Step S3, quaternion grouping dilated convolution module: Use multiple mixed dilated convolutions with different dilation rates to expand the receptive field. The dilated convolution increases the receptive field by setting the dilation rate.

Images