Medical Image Segmentation Method, System and Device Based on Convolution Reparameterization

Abstract

The present invention discloses a medical image segmentation method, system and device based on convolution reparameterization. The method includes the following steps: S1, collecting multi-modal medical images; S2, preprocessing the multi-modal medical images to construct a training data set; S3, constructing a medical image segmentation model based on convolution reparameterization; S4, training the medical image segmentation model using the training data set; S5, inputting the medical image to be segmented into the trained medical image segmentation model to obtain a segmentation result. The medical image segmentation method and system based on convolution reparameterization provided by the present invention can enhance the skeleton parameters of the convolution kernels in the segmentation network model, thereby improving the expression ability of the segmentation network. The present invention can help improve the accuracy of glioma segmentation and has potential application value in glioma morphological analysis, high-grade and low-grade identification, and prognosis evaluation.

Description

Technical Field

[0001] The present invention relates to the field of medical image processing, and particularly relates to a medical image segmentation method, system and device based on convolutional reparameterization. Background Art

[0002] Glioma is the most common primary brain tumor in adults. It is usually treated by microsurgical resection. Magnetic resonance imaging (MRI) can be used to accurately locate the position of glioma and determine the impact of the tumor on the structure and function of the brain. The goal of glioma segmentation is to segment the tumor area into enhancing tumor (ET), tumor core (TC), and whole tumor (WT). Preoperatively, accurate automatic tumor segmentation based on magnetic resonance imaging is crucial for the clinical diagnosis, treatment, and supervision of glioma patients. However, manual segmentation is time-consuming, laborious, and unstable. With the development of deep learning technology and the increase of clinical glioma image data, the U-net-based automatic brain tumor segmentation model has achieved great success. Researchers have dedicated to improving the network based on U-net, designing architectures such as cascading [1] and multi-branch [2], improving various convolutional processing modules such as residual blocks [3], dense connection modules, and attention mechanisms, and improving the feature abstraction ability and judgment ability of the network model by modifying the feature fusion method and enhancing the network flow ability of the gradient flow.

[0003] In recent years, convolutional reparameterization technology has received attention. Among them, the asymmetric convolutional reparameterization technology can effectively reduce the computational complexity by factorizing the standard symmetric convolution into an asymmetric convolution, thereby accelerating the training and inference process [4], especially for larger convolutional kernels. However, in deep network models, its equivalence cannot be fully established. Some researchers have studied this defect [5, 6, 7]. However, due to the fact that three-dimensional convolutional kernels (such as a convolutional kernel with a size of 3*3*3) do not have the characteristics of factorizable convolution, the factorized form of asymmetric convolution faces challenges in three-dimensional convolution applications. For this problem, researchers have proposed a new form of applying asymmetric convolution [8], which performs parallel and corresponding normalization operations on the input 3×3, 1×3, and 3×1 convolutions through multi-path convolution and feature fusion. The sum of these three operations is equivalent to a 3×3 convolution operation, and the kernel parameters can be calculated through mathematical formulas. This method can enhance the skeleton parameters of the convolutional kernel in the segmentation network model, thereby improving the expression ability of the segmentation network. Researchers have verified this method on the publicly available dataset of infant brain tissue segmentation and achieved good segmentation results [9]. However, there are still problems such as limited receptive field caused by a single convolutional kernel scale and the inability to adaptively and directly transfer feature information in the downsampling and upsampling of U-net in the application of asymmetric convolution in medical image segmentation that need to be further solved.

[0004] Most existing glioma segmentation methods use classical deep learning networks for tumor segmentation, such as: U-net [Method, system and device for multi-modal imaging brain glioma target area segmentation, Chinese Patent CN202210550215.5], cascaded neural network [A brain glioma segmentation method based on cascaded neural network structure, Chinese Patent CN202111404516.9], convolutional neural network [Glioma segmentation method and system, Chinese Patent CN202110843725.7], etc. The convolution modules used in the above methods are all general convolutions. In the case of a large convolution kernel, the computational complexity is high, and the expression ability of the convolution network is limited first, resulting in difficulty in improving the segmentation accuracy.

[0005] References:

[0006] [1] S. Hussain, S. M. Anwar and M. Majid, “Brain tumor segmentation using cascaded deep convolutional neural network,” 2017 39th Annual International Conference of the IEEE Engineering in Medicine and Biology Society (EMBC), 2017, pp. 1998 - 2001.

[0007] [2] Sedlar, S. (2018). “Brain Tumor Segmentation Using a Multi-path CNN Based Method.” In: Crimi, A., Bakas, S., Kuijf, H., Menze, B., Reyes, M. (eds) Brain lesion: Glioma, Multiple Sclerosis, Stroke and Traumatic Brain Injuries. BrainLes 2017. Lecture Notes in Computer Science, vol 10670. Springer, Cham.

[0008] [3]Gupta, V., Bibhu, V. “Deep residual network based brain tumor segmentation and detection with MRI using improved invasive bat algorithm.” Multimed Tools Appl (2022).

[0009] [4]Min Wang, Baoyuan Liu, Hassan Foroosh, “Factorized Convolutional Neural Networks,” Computer Vision and Pattern Recognition (cs.CV), 2017.

[0010] [5]Emily L Denton, Wojciech Zaremba, et al., “Exploiting linear structure within convolutional networks for efficient evaluation.” In Advances in neural information processing systems, pages 1269–1277, 2014.

[0011] [6]Max Jaderberg, Andrea Vedaldi, and Andrew Zisserman. “Speeding up convolutional neural networks with low rank expansions.” arXiv preprint arXiv:1405.3866, 2014.

[0012] [7]Jonghoon Jin, Aysegul Dundar, and Eugenio Culurciello. “Flattened convolutional neural networks for feedforward acceleration.” arXiv preprint arXiv:1412.5474, 2014.

[0013] [8]Xiaohan Ding, Yuchen Guo et al., “ACNet: Strengthening the Kernel Skeletons for Powerful CNN via Asymmetric Convolution Blocks,” arXiv:1908.03930 [cs.CV], 2019.

[0014] [9]Zilong Zeng, Tengda Zhao et al., “3D-MASNet: 3D Mixed-scale Asymmetric Convolutional Segmentation Network for 6-month-old Infant Brain MR Images,” bioRxiv preprint, 2021. Summary of the Invention

[0015] The technical problem to be solved by the present invention is to provide a medical image segmentation method, system and device based on convolutional reparameterization for the deficiencies in the above-mentioned prior art.

[0016] To solve the above technical problems, the technical solution adopted by the present invention is: A medical image segmentation method based on convolutional reparameterization, comprising the following steps:

[0017] S1. Collect multi-modal medical images;

[0018] S2. Preprocess the multi-modal medical images to construct a training data set;

[0019] S3. Construct a medical image segmentation model based on convolutional reparameterization, the medical image segmentation model includes an input module, a downsampling part, an intermediate connection unit, an upsampling part and a convolutional output module connected in sequence;

[0020] The downsampling part includes 3 first optimized residual modules and 3 multi-scale asymmetric convolution modules connected alternately, and the input module is connected to the first first optimized residual module, and the last multi-scale asymmetric convolution module is connected to the intermediate connection unit, and the intermediate connection unit includes two second optimized residual modules;

[0021] The upsampling part includes 3 upsampling modules and 3 third optimized residual modules connected alternately, and the first upsampling module is connected to the intermediate connection unit, and the last third optimized residual module is connected to the convolutional output module;

[0022] Among them, the output of each first optimized residual module in the downsampling part is cross-connected with the output of an upsampling module in the upsampling part through a self-attention module;

[0023] S4. Use the training dataset to train the medical image segmentation model;

[0024] S5. Input the medical image to be segmented into the trained medical image segmentation model to obtain the segmentation result.

[0025] Preferably, the step S2 specifically includes: performing co-registration, resampling, bias field correction, image cropping, normalization, and data augmentation on the multi-modal medical image in sequence to obtain the training dataset.

[0026] Preferably, the input module is an initial asymmetric convolution module composed of 16 convolution kernels of size 3;

[0027] The first optimized residual module, the second optimized residual module, and the third optimized residual module have the same structure and all include convolution kernels of size 3.

[0028] Preferably, the multi-scale asymmetric convolution module includes a first convolution unit and a second convolution unit. The first convolution unit includes a 5*5*5 convolution module in parallel, and three 5*5*1, 5*1*5, 1*5*5 asymmetric convolution modules obtained by convolution reparameterization decomposition of the 5*5*5 convolution module; the second convolution unit includes a 3*3*3 convolution module in parallel, and three 3*3*1, 3*1*3, 1*3*3 asymmetric convolution modules obtained by convolution reparameterization decomposition of the 3*3*3 convolution module;

[0029] Among them, the output of the 5*5*5 convolution module is merged with the outputs of the 5*5*1, 5*1*5, 1*5*5 asymmetric convolution modules as the output of the first convolution unit, and the output of the 3*3*3 convolution module is merged with the outputs of the 3*3*1, 3*1*3, 1*3*3 asymmetric convolution modules as the output of the second convolution unit. Finally, the output of the first convolution unit and the output of the second convolution unit are merged as the output of the multi-scale asymmetric convolution module.

[0030] Preferably, the self-attention module adds the output of a first optimized residual module in the downsampling part to the output of an upsampling module at the same level in the upsampling part, and then inputs it into a third optimized residual module connected to the output end of the upsampling module for convolution;

[0031] The self-attention module is defined as follows:

[0032] Y = sigmoid(conv(conv(X)));

[0033] Among them, X is the final feature map output by the downsampling part, conv represents the convolution operation, the sigmoid function is used as the activation function of the neural network to map the variable to between 0 and 1, and the final output of the self-attention module is the attention feature map Y.

[0034] Preferably, the loss function used for training the medical image segmentation model in step S4 is Dice-Loss, and the Adam optimizer is used. The data formula of the loss function is defined as follows:

[0035] Dice_Loss(A,B) = 1 - 2|A∩B| / (|A| + |B|)

[0036] Among them, A is the segmentation gold standard, B is the segmentation result obtained by the medical image segmentation model, and the learning rate is changed by the Cosine-Annealing strategy to train the network.

[0037] Preferably, the multi-modal medical image is a glioma magnetic resonance image, including images of four modalities: T1, FLAIR, T2, and T1ce;

[0038] Step S2 specifically includes:

[0039] S2-1. Co-register the glioma magnetic resonance images of four modalities, namely T1, FLAIR, T2, and T1ce, into a unified template space;

[0040] S2-2. Perform resampling to make the voxel size 1*1*1;

[0041] S2-3. Perform bias field correction processing;

[0042] S2-4. Perform image cropping to remove the background image;

[0043] S2-5. Perform normalization processing on the images of each modality;

[0044] S2-6. Use random flipping and intensity transformation to enhance the data to obtain a training data set.

[0045] The present invention also provides a medical image segmentation system based on convolution reparameterization, which is characterized in that it uses the method described above for medical image segmentation. The system includes:

[0046] An image acquisition module, which is used to acquire multi-modal medical images;

[0047] An image preprocessing module, which preprocesses the multi-modal medical images by using the method of step S2;

[0048] and a trained medical image segmentation model, which is used to segment the input medical image.

[0049] The present invention also provides a storage medium, on which a computer program is stored, and characterized in that the program is used to implement the method as described above when executed.

[0050] The present invention also provides a computer device, including a memory, a processor, and a computer program stored on the memory and executable on the processor, and characterized in that the processor implements the method as described above when executing the computer program.

[0051] The beneficial effects of the present invention are:

[0052] The medical image segmentation method and system based on convolution reparameterization provided by the present invention can enhance the absolute value of the equivalent convolution kernel skeleton parameters, which are proven to be the most important in the network model, so as to improve the expression ability of the convolution network; through the convolution reparameterization technology, the present invention improves the expression ability and robustness of the segmentation network, and at the same time introduces the scale technology to design a relatively small network on the premise of maintaining good segmentation performance, so that the computing cost is significantly reduced; the present invention can help improve the accuracy of glioma segmentation and has potential application value in glioma morphological analysis, high-grade and low-grade identification, and prognosis evaluation. Description of the Drawings

[0053] Figure 1 It is a schematic structural diagram of the medical image segmentation model based on convolution reparameterization of the present invention;

[0054] Figure 2 It is a schematic structural diagram of the multi-scale asymmetric convolution module of the present invention;

[0055] Figure 3 It is the visualization result of the application of the medical image segmentation method of the present invention in glioma segmentation. Detailed Embodiments

[0056] The following further describes the present invention in detail with reference to embodiments, so that those skilled in the art can implement it according to the description in the specification.

[0057] It should be understood that the terms such as "having", "comprising", and "including" used herein do not exclude the presence or addition of one or more other elements or their combinations.

[0058] Embodiment 1

[0059] This embodiment provides a medical image segmentation method based on convolution reparameterization, including the following steps:

[0060] S1. Collect multi-modal medical images;

[0061] S2. Preprocess the multi-modal medical images to construct a training dataset:

[0062] Perform co-registration, resampling, bias field correction, image cropping, normalization, and data augmentation on the multi-modal medical images in sequence to obtain the training dataset;

[0063] S3. Construct a medical image segmentation model based on convolutional reparameterization;

[0064] Refer to Figure 1 , the medical image segmentation model includes an input module, a downsampling part, an intermediate connection unit, an upsampling part, and a convolutional output module connected in sequence;

[0065] The input module is an initial asymmetric convolutional module composed of 16 convolutional kernels of size 3;

[0066] The downsampling part includes 3 first optimized residual modules and 3 multi-scale asymmetric convolutional modules connected alternately, and the input module is connected to the first first optimized residual module, and the last multi-scale asymmetric convolutional module is connected to the intermediate connection unit. The intermediate connection unit includes two second optimized residual modules;

[0067] The upsampling part includes 3 upsampling modules and 3 third optimized residual modules connected alternately, and the first upsampling module is connected to the intermediate connection unit, and the last third optimized residual module is connected to the convolutional output module;

[0068] Among them, the output of each first optimized residual module in the downsampling part is cross-connected with the output of an upsampling module in the upsampling part through a self-attention module;

[0069] Among them, the first optimized residual module, the second optimized residual module, and the third optimized residual module have the same structure, all of which include convolutional kernels of size 3.

[0070] Among them, the multi-scale asymmetric convolutional module includes a first convolutional unit and a second convolutional unit. The first convolutional unit includes a 5*5*5 general convolutional module in parallel, and 3 5*5*1, 5*1*5, 1*5*5 asymmetric convolutional modules obtained by convolutional reparameterization decomposition of the 5*5*5 convolutional module; the second convolutional unit includes a 3*3*3 convolutional module in parallel, and 3 3*3*1, 3*1*3, 1*3*3 asymmetric convolutional modules obtained by convolutional reparameterization decomposition of the 3*3*3 convolutional module;

[0071] Among them, the outputs of the 5*5*5 convolution module and the 5*5*1, 5*1*5, 1*5*5 asymmetric convolution modules are combined as the output of the first convolution unit, and the outputs of the 3*3*3 convolution module and the 3*3*1, 3*1*3, 1*3*3 asymmetric convolution modules are combined as the output of the second convolution unit. Finally, the outputs of the first convolution unit and the second convolution unit are combined as the output of the multi-scale asymmetric convolution module.

[0072] Among them, the self-attention module adds the output of a first optimized residual module in the downsampling part to the output of an upsampling module at the same level in the upsampling part, and then inputs it into a third optimized residual module connected to the output end of the upsampling module for convolution;

[0073] The self-attention module is defined as follows:

[0074] Y = sigmoid(conv(conv(X)));

[0075] Among them, X is the final feature map output by the downsampling part, conv represents the convolution operation, the sigmoid function is used as the activation function of the neural network to map the variable to between 0 and 1, and the final output of the self-attention module is the attention feature map Y. The advantage of the self-attention module can extract more tumor region information by using the global gray information of the image itself as a guide.

[0076] S4. Train the medical image segmentation model using the training data set;

[0077] In step S4, the loss function used for training the medical image segmentation model is Dice-Loss, and the Adam optimizer is used. The data formula of the loss function is defined as follows:

[0078] Dice_Loss(A, B) = 1 - 2|A ∩ B| / (|A| + |B|)

[0079] Among them, A is the segmentation gold standard, B is the segmentation result obtained by the medical image segmentation model, and the learning rate is changed through the Cosine-Annealing strategy for network training.

[0080] S5. Input the medical image to be segmented into the trained medical image segmentation model to obtain the segmentation result.

[0081] Among them, the medical image to be segmented can be a multi-modal image or a single-modal image. And in the preferred embodiment, the medical image to be segmented is preprocessed first and then input into the trained medical image segmentation model. The steps of preprocessing mainly include co-registration (this step is omitted for single-modal images), resampling, bias field correction, image cropping, normalization, etc.

[0082] Example 2

[0083] In this example, the multimodal medical image is a glioma magnetic resonance image, including four modalities: T1, FLAIR, T2, and T1ce. Taking this as an example, the medical image segmentation method based on convolutional reparameterization of the present invention will be described in detail as follows.

[0084] A medical image segmentation method based on convolutional reparameterization includes the following steps:

[0085] S1. Collect multimodal glioma magnetic resonance images;

[0086] S2. Preprocess the multimodal medical images to construct a training dataset:

[0087] S2-1. Co-register the four-modal glioma magnetic resonance images of T1, FLAIR, T2, and T1ce into a unified template space;

[0088] S2-2. Perform resampling to make the voxel size 1*1*1;

[0089] S2-3. Perform bias field correction to make the gray uniformity of the image better;

[0090] S2-4. Perform image cropping to remove the background image; considering that there is no brain scan data at the edge of the magnetic resonance image, we crop the peripheral background part of the image to minimize its impact on the overall tumor segmentation, and use the method of random cropping to obtain a three-dimensional image of the specified size;

[0091] S2-5. Normalize the images of each modality;

[0092] S2-6. Augment the data using random flipping and intensity transformation to obtain a training dataset.

[0093] S3. Construct a medical image segmentation model based on convolutional reparameterization. In this model:

[0094] (1) First, input the multimodal medical image data of 4*128*128*96 into the initial asymmetric convolution module for fusion. The structure of this module is composed of 16 convolution kernels of size 3, and the output is an initial feature map of 16*128*128*96;

[0095] (2) Then, enter the downsampling part. First, input the above-obtained initial feature map of 16*128*128*96 into the first first-optimized residual module. In this module, further convolutional encoding is performed using a convolution kernel of size 3, and the output is a feature map of 32*128*128*96, and then it is input into the first multi-scale asymmetric convolution module connected to it;

[0096] The multi-scale asymmetric convolution module reparameterizes a general convolution into multiple asymmetric convolutions at different scales. The specific process is as follows Figure 2 , first use a 5*5*5 general convolution module, which is decomposed into three asymmetric convolution modules of 5*5*1, 5*1*5, and 1*5*5 through convolution reparameterization. Four modules are connected in parallel as the first convolution unit, and the outputs of the four modules are combined as the output of the first convolution unit; then use a 3*3*3 general convolution module, which is decomposed into three asymmetric convolution modules of 3*3*1, 3*1*3, and 1*3*3 through convolution reparameterization. Four modules are connected in parallel as the second convolution unit, and the outputs of the four modules are combined as the output of the second convolution unit; finally, the output of the first convolution unit and the output of the second convolution unit are combined as the output of the multi-scale asymmetric convolution module; the output of the first multi-scale asymmetric convolution module obtains a feature map of 64*64*64*48; finally, through three groups of alternately connected first optimized residual modules and multi-scale asymmetric convolution modules, a feature map of 256*16*16*12 is obtained, which is used as the output of the downsampling part. After passing through two second optimized residual modules, it is input into the upsampling part;

[0097] (3) In the upsampling part, the output of a first optimized residual module in the downsampling part is added to the output of an upsampling module at the same level in the upsampling part through a self-attention module, and then input into a third optimized residual module connected to the output end of the upsampling module for convolution; through three groups of alternately connected upsampling modules and third optimized residual modules, repeated operations are performed to obtain a feature map of size 32*128*128*96, and finally output through a convolution output module.

[0098] The self-attention module is defined as follows:

[0099] Y = sigmoid(conv(conv(X)));

[0100] Among them, X is the final feature map output by the downsampling part, conv represents the convolution operation, the sigmoid function is used as the activation function of the neural network to map the variable to between 0 and 1, and the final output of the self-attention module is the attention feature map Y. The advantage of the self-attention module can extract more tumor region information by using the global gray information of the image itself as a guide.

[0101] S4. Train the medical image segmentation model using the training dataset;

[0102] In step S4, the loss function used for training the medical image segmentation model is Dice-Loss, and the Adam optimizer is used. The data formula of the loss function is defined as follows:

[0103] Dice_Loss(A,B) = 1 - 2|A∩B| / (|A| + |B|)

[0104] Where A is the gold standard for segmentation, B is the segmentation result obtained by the medical image segmentation model, and the learning rate is changed by the Cosine-Annealing strategy to train the network.

[0105] S5. Input the medical image to be segmented into the trained medical image segmentation model to obtain the segmentation result.

[0106] In this embodiment, the five-fold cross-validation method is used to verify the segmentation accuracy of the model. The medical image segmentation metrics dice coefficient (DICE) and 95% Hausdorff distance (HD95) are used to evaluate the accuracy of the segmentation method proposed in the present invention, and the overall training time (total training cost, TTC) is used to evaluate the efficiency of the segmentation method proposed in the present invention; the comparison results of the performance metrics of the method of the present invention and other methods for glioma segmentation (proposed is the method proposed in this patent) are as Figure 1 and Table 1 below:

[0107] Table 1

[0108]

[0109] It can be seen that the method of the present invention is superior to the conventional scheme.

[0110] Example 3

[0111] This embodiment provides a medical image segmentation system based on convolutional reparameterization, which uses the method of Example 1 for medical image segmentation. The system includes:

[0112] An image acquisition module for acquiring multi-modal medical images;

[0113] An image preprocessing module for preprocessing the multi-modal medical images using the method of step S2;

[0114] And a trained medical image segmentation model for segmenting the input medical image.

[0115] This embodiment also provides a storage medium, on which a computer program is stored, characterized in that the program is used to implement the method of Example 1 when executed.

[0116] This embodiment also provides a computer device, including a memory, a processor, and a computer program stored on the memory and executable on the processor, characterized in that the processor implements the method of Example 1 when executing the computer program.

[0117] Although the embodiments of the present invention have been disclosed as above, they are not limited to the applications listed in the specification and embodiments. It can be fully applied to various fields suitable for the present invention. For those skilled in the art, additional modifications can be easily achieved. Therefore, without departing from the general concept defined by the claims and the equivalent scope, the present invention is not limited to specific details.

Claims

1. A medical image segmentation method based on convolutional reparameterization, characterized in that It includes the following steps: S1. Collect multi-modal medical images; S2. Preprocess the multi-modal medical images to construct a training data set; S3. Construct a medical image segmentation model based on convolutional reparameterization. The medical image segmentation model includes an input module, a downsampling part, an intermediate connection unit, an upsampling part, and a convolutional output module connected in sequence; The downsampling part includes 3 first optimized residual modules and 3 multi-scale asymmetric convolutional modules connected alternately. The input module is connected to the first first optimized residual module, and the last multi-scale asymmetric convolutional module is connected to the intermediate connection unit. The intermediate connection unit includes 2 second optimized residual modules; The upsampling part includes 3 upsampling modules and 3 third optimized residual modules connected alternately. The first upsampling module is connected to the intermediate connection unit, and the last third optimized residual module is connected to the convolutional output module; Among them, the output of each first optimized residual module in the downsampling part is cross-connected with the output of an upsampling module in the upsampling part through a self-attention module; S4. Train the medical image segmentation model using the training data set; S5. Input the medical image to be segmented into the trained medical image segmentation model to obtain a segmentation result; The step S2 specifically includes: performing co-registration, resampling, bias field correction, image cropping, normalization, and data augmentation on the multi-modal medical images in sequence to obtain a training data set; The input module is an initial asymmetric convolutional module composed of 16 convolutional kernels of size 3; The structures of the first optimized residual module, the second optimized residual module, and the third optimized residual module are the same, and all include convolutional kernels of size 3; The multi-scale asymmetric convolutional module includes a first convolutional unit and a second convolutional unit. The first convolutional unit includes a 5*5*5 convolutional module in parallel, and 3 5*5*1, 5*1*5, 1*5*5 asymmetric convolutional modules obtained by convolutional reparameterization decomposition of the 5*5*5 convolutional module; the second convolutional unit includes a 3*3*3 convolutional module in parallel, and 3 3*3*1, 3*1*3, 1*3*3 asymmetric convolutional modules obtained by convolutional reparameterization decomposition of the 3*3*3 convolutional module; Among them, the output of the 5*5*5 convolutional module is combined with the outputs of the 5*5*1, 5*1*5, and 1*5*5 asymmetric convolutional modules as the output of the first convolutional unit. The output of the 3*3*3 convolutional module is combined with the outputs of the 3*3*1, 3*1*3, and 1*3*3 asymmetric convolutional modules as the output of the second convolutional unit. Finally, the output of the first convolutional unit and the output of the second convolutional unit are combined as the output of the multi-scale asymmetric convolutional module.

2. The medical image segmentation method based on convolution reparameterization according to claim 1, wherein The self-attention module adds the output of a first optimized residual module in the downsampling part to the output of an upsampling module at the same level in the upsampling part, and then inputs it into a third optimized residual module connected to the output end of the upsampling module for convolution; The self-attention module is defined as follows: Y = sigmoid(conv(conv(X))); Where X is the final feature map output by the downsampling part, conv represents the convolution operation, and the sigmoid function is used as the activation function of the neural network, mapping the variable to between 0 and 1. The final output of the self-attention module is the attention feature map Y.

3. The medical image segmentation method based on convolutional reparameterization according to claim 2, characterized in that In step S4, the loss function used for training the medical image segmentation model is Dice-Loss, and the Adam optimizer is used. The data formula of the loss function is defined as follows: Dice_Loss(A,B) = 1 - 2|A∩B| / (|A| + |B|) Where A is the segmentation gold standard and B is the segmentation result obtained by the medical image segmentation model. The learning rate is changed through the Cosine-Annealing strategy for network training.

4. The medical image segmentation method based on convolution reparameterization according to claim 3, characterized in that, The multi-modal medical image is a glioma magnetic resonance image, including images of four modalities: T1, FLAIR, T2, and T1ce; Step S2 specifically includes: S2-1. Co-register the glioma magnetic resonance images of four modalities, T1, FLAIR, T2, and T1ce, into a unified template space; S2-2. Perform resampling to make the voxel size 1*1*1; S2-3. Perform bias field correction processing; S2-4. Perform image cropping to remove the background image; S2-5. Normalize the images of each modality; S2-6. Augment the data by using random flipping and intensity transformation to obtain the training data set.

5. A medical image segmentation system based on convolutional reparameterization, characterized in that, It performs medical image segmentation by using the method described in any one of claims 1-4. The system includes: An image acquisition module, which is used to acquire multi-modal medical images; An image preprocessing module, which preprocesses the multi-modal medical images by using the method of step S2; And a trained medical image segmentation model, which is used to segment the input medical images.

6. A storage medium having a computer program stored thereon, characterized in that, When the program is executed, it is used to implement the method described in any one of claims 1-4.

7. A computer device, comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the method described in any one of claims 1-4.

Images