Medical image segmentation method based on fusion of multilayer perceptron, terminal device and medium

Abstract

The application relates to a medical image segmentation method based on a fusion multilayer perceptron, a terminal device and a medium. The method constructs a medical image segmentation model, adopts a U-shaped encoder-decoder architecture, combines the advantages of convolution and MLP, and fuses the two into each layer. A multi-axis and multi-window MLP module is designed to capture feature map information from multiple windows and multiple scales. The input does not need to crop the picture, can receive pictures of any size, and always has a global receptive field. A Conv-MCG module is developed to cross-fuse the two outputs of the convolution module, realize the interaction of multiple paths and multiple information, and achieve good segmentation results with small calculation time and space consumption.

Description

Technical Field

[0001] This invention relates to the field of image processing, and more particularly to a medical image segmentation method, terminal device, and medium that integrates multilayer perceptrons. Background Technology

[0002] In recent years, convolutional neural networks (CNNs) have made significant progress in computer vision applications, such as image classification, object detection, semantic segmentation, object tracking, and image enhancement. Thanks to the development of CNNs, computer vision technology has been widely applied in the field of medical image processing. Image semantic segmentation is an important component of medical image processing and can play a crucial role in computer-aided diagnosis and image-guided clinical surgery.

[0003] The most typical existing deep learning-based medical image segmentation network is U-Net, which consists of a symmetric encoder-decoder architecture and skip connections. In the encoder, a series of convolutional layers and consecutive downsampling layers are used to increase the receptive field and extract more semantic features. Then, the decoder upsamples the extracted semantic features to obtain pixel-level semantic predictions and fuses shallow features from the encoder with deep features from the decoder through skip connections to mitigate the loss of semantic information caused by depthwise convolution and downsampling. Thanks to its excellent structural design, U-Net has achieved great success in medical image processing. Many variant networks have been developed based on this network for various types of medical image segmentation. The excellent performance of these U-Net-based methods in medical image segmentation has demonstrated the strong feature-learning ability of CNNs. However, the inherent locality of the receptive field and weight sharing in convolutional operations make it difficult for U-Net-based methods to learn explicit global information and perform long-range semantic information interaction, which to some extent fails to meet the stringent requirements for segmentation accuracy in the field of medical image segmentation.

[0004] Applying Transformers to computer vision can alleviate long-range dependencies to some extent. Simultaneously, Transformers possess powerful global relation modeling capabilities, achieving good results in medical image analysis tasks. Existing research has proposed the Vision Transformer (ViT) to perform image recognition tasks, taking 2D image patches with location embeddings as input, achieving excellent performance. Research based on ViT has also promoted the study of Multi-Layer Perceptron (MLP) structures, such as MLP-Mixer, gMLP, and aMLP. In particular, MLP-Mixer, a completely MLP-based network, delivers performance comparable to Transformers with less computation. Although existing Transformer and MLP-based methods have proven promising in image analysis tasks, including medical image segmentation, several significant challenges remain:

[0005] (1) It only accepts fixed image sizes and requires the image to be rigidly divided into fixed sizes, which may not be able to capture fine-grained spatial details of the image.

[0006] (2) When applied to larger images, it will inevitably cause boundary artifacts.

[0007] (3) Because the input is treated as a one-dimensional sequence and only global information is modeled at all stages, detailed localization information is lacking. When performing semantic analysis, the ability to locate the position of interest may be lacking.

[0008] These issues highlight the shortcomings of Transformer and MLP networks compared to CNN networks in extracting low-level features and visual structures. Summary of the Invention

[0009] To address the aforementioned issues, this invention proposes a medical image segmentation method, terminal device, and medium that integrates a multilayer perceptron.

[0010] The specific plan is as follows:

[0011] A medical image segmentation method incorporating multilayer perceptron includes the following steps:

[0012] S1: Collect medical image data to construct a training set;

[0013] S2: Construct a medical image segmentation model and train the model using a training set;

[0014] The model's network backbone adopts a U-shaped network, including N encoders, N decoders and one output module; each encoder, decoder and output module includes three convolutional blocks, one MLP cross-gating module and one splicing module;

[0015] In the encoder, data is input from the first convolutional block. The output of the first convolutional block is downsampled and then input to the second convolutional block. The output of the second convolutional block is upsampled and then input to the third convolutional block. The outputs of the first and third convolutional blocks are input together to the MLP cross-gating module. The output of the MLP cross-gating module is connected to the decoder via a skip connection, and after being downsampled, it is concatenated with the output of the second convolutional block in the stitching module. The output of the stitching module is the output of the encoder.

[0016] In the decoder, the output data of the previous layer is input from the first convolutional block. The output of the first convolutional block is upsampled and then input together with the skip connection data of the previous layer encoder into the second convolutional block. The output of the second convolutional block is downsampled and then input together with the skip connection data of the corresponding layer encoder into the third convolutional block. The outputs of the first and third convolutional blocks are input together into the MLP cross-gating module. The output of the MLP cross-gating module is upsampled and then concatenated with the output of the second convolutional block in the splicing module. The output of the splicing module is the output of the decoder.

[0017] In the output module, the decoder's output data is input from the first convolutional block. The output of the first convolutional block is directly input to the second convolutional block. The outputs of the first and second convolutional blocks are input together to the MLP cross-gating module. The output of the MLP cross-gating module is upsampled and then concatenated with the data from the first layer encoder in the concatenation module. The output of the concatenation module is used as the input to the third convolutional block, and the output of the third convolutional block is the model output result.

[0018] S3: Segment the medical image to be segmented using the trained model.

[0019] Furthermore, the MLP cross-gating module includes two input paths. For each path, the input data is sequentially processed through layer normalization, a fully connected layer, and a GELU activation function before being input into a multi-axis multi-window MLP structure. The output of the multi-axis multi-window MLP structure is multiplied by the input of the multi-axis multi-window MLP structure. The multiplication result is sequentially processed through a fully connected layer and activated by Dropout. The Dropout activation result is then added to the input data of that path to obtain the output result of that path. The output results of the two paths are then added together to obtain the output result of the MLP cross-gating module.

[0020] Furthermore, the multi-axis multi-window MLP structure divides the input feature map into four heads according to channels. In each head, MLP is used to fuse information. Two heads are sent to the local feature branch, and the other two heads are sent to the global feature branch. In the local feature branch, the two heads are sent to two local feature processing axes with different window sizes. In the global feature branch, the two heads are sent to two global feature processing axes with different window sizes.

[0021] Furthermore, the convolutional block consists of a combination of two 3×3 convolutional layers, a BN layer, and a ReLU activation function layer.

[0022] Furthermore, step S1 also includes augmenting the acquired medical image data by random rotation or by adjusting the hue, brightness, or cropping, and constructing a training set based on the augmented data.

[0023] A medical image segmentation terminal device integrating a multilayer perceptron includes a processor, a memory, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the steps of the method described in the embodiments of the present invention.

[0024] A computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the method described above in the embodiments of the present invention.

[0025] The present invention employs the above technical solution, which achieves the best segmentation performance while maintaining the learnable parameters of the model and keeping the computational cost relatively low. Attached Figure Description

[0026] Figure 1 The diagram shown is a flowchart of a method according to an embodiment of the present invention.

[0027] Figure 2 The diagram shown is a schematic of the network structure of the model in this embodiment.

[0028] Figure 3 The diagram shown is a structural schematic of each module inside the model in this embodiment.

[0029] Figure 4 The diagram shown is a structural schematic of the MCG module in this embodiment.

[0030] Figure 5 The diagram shown is a schematic of the multi-axis multi-window MLP structure in this embodiment. Detailed Implementation

[0031] To further illustrate the various embodiments, the present invention provides accompanying drawings. These drawings are part of the disclosure of the present invention, primarily used to illustrate the embodiments, and can be used in conjunction with the relevant descriptions in the specification to explain the operating principles of the embodiments. With reference to these drawings, those skilled in the art should be able to understand other possible implementations and the advantages of the present invention.

[0032] The present invention will now be further described in conjunction with the accompanying drawings and specific embodiments.

[0033] Example 1:

[0034] This invention provides a medical image segmentation method incorporating multilayer perceptron, such as... Figure 1 As shown, the method includes the following steps:

[0035] S1: Collect medical image data to construct a training set.

[0036] The medical image data acquired in this embodiment comes from the following three existing datasets:

[0037] Breast Ultrasound Images (BUSI) is a medical image dataset of breast cancer scans. The BUSI dataset contains breast ultrasound images of women aged 25 to 75 years, collected in 2018. The patient population was 600 women. The dataset consists of 780 images, all of which were cropped to different sizes to remove unused and unimportant borders. The images are in PNG format and categorized into three types: normal, benign, and malignant. Each image has its own Ground Truth (GT). In this example, images of both benign and malignant tumors were selected, and a total of 647 images were uniformly resized to 256×256 RGB images.

[0038] The International Skin Imaging Collaboration (ISIC 2018) is the world's largest dataset for a skin image analysis challenge. This dataset contains 2594 images across three distinct categories: 20.0% melanoma, 72.0% nevi, and 8.0% seborrheic keratosis. The dataset comprises images at various resolutions; in this embodiment, all images have been resized to 512×512 RGB images.

[0039] CVC-ClinicDB is a colonoscopy image dataset consisting of endoscopic images. These images were extracted from video sequences of colonoscopy examinations. The ground truth images consist of masks corresponding to the regions covered by polyps in the images. Since the segmentation task only considers images with polyps, a total of 612 images from 31 colonoscopy sequences were obtained. In this embodiment, all images were resized to 256×256 RGB images.

[0040] To increase the diversity of training samples and enhance the generalization ability of the trained model, this embodiment also augments the acquired raw medical image data, performing random rotations and random adjustments to the images, including hue, brightness, and cropping.

[0041] S2: Construct a medical image segmentation model and train the model using a training set.

[0042] For medical image segmentation, numerous studies have shown that the U-Net-like network architecture is a highly competitive foundation. Therefore, in this embodiment, the main framework of the medical image segmentation model (hereinafter referred to as MsMCG for ease of description) still adopts a U-shaped architecture, as follows: Figure 2 As shown, the U-shaped architecture has significant advantages in extracting low-level features, but its local receptive field and weak long-range dependencies still hinder the accuracy of segmentation results. A good network, in addition to possessing the advantages of U-Net, should also incorporate an MLP to increase the network's global receptive field. Based on this idea, this embodiment designs a Conv-MLP encoder module and a Conv-MLP decoder module to build the encoder and decoder of the MsMCG network. The MLP Cross Gating (MCG) in the Conv-MLP encoder and decoder modules is used to fuse information from two convolutional layers, obtaining the global receptive field from the fine local features extracted by the convolutional information through the MLP. The output of the MGP in the Conv-MLP encoder module is connected to the decoder via skip connections; each MGP sub-module is connected twice to different Conv-MLP decoder modules. The proposed architecture adopts an end-to-end training strategy, relying on the hybrid model design of each block to adaptively obtain local and global receptive fields at each layer. Conv is used for local receptive fields, and MLP is used for long-range interactions to fully utilize every piece of information.

[0043] The core building blocks of MsMCG are the Conv-MLP encoder and the Conv-MLP decoder, which form the backbone of the U-shaped network. For example... Figure 2As shown, MsMCG includes N encoders, N decoders, and one output module. In this embodiment, N is preferably set to 4. Each module uses a convolutional block design, consisting of two Conv3×3 (3×3 convolutional layers) + BN + ReLU.

[0044] In Conv-MLP encoders (such as...) Figure 3 As shown in (a), the U-Net network encoder design concept is still adopted. However, considering the information loss during downsampling, a downsampling followed by upsampling operation is used for convolutional blocks. This allows for the acquisition of information from the next layer while minimizing information loss. Specifically, after data is input from the first convolutional block, its output is downsampled and input to the second convolutional block. The output of the second convolutional block is upsampled and input to the third convolutional block. Furthermore, the outputs of two convolutional blocks (the first and third) in the same layer enter the MLP Cross Gating (MCG) module, which performs cross-fusion of the convolutional block information. The output of the MLP Cross Gating module is used for skip connections with the decoder. Simultaneously, the feature map size of the downsampled output of the MLP Cross Gating module is halved, and it is concatenated with the output of the downsampled convolutional block (the second convolutional block) from the previous convolutional operation (Concat module) for use in the subsequent module.

[0045] The Conv-MLP decoder also employs a U-Net network decoder design, progressively restoring image resolution and fusing it with decoder information. This embodiment uses a convolutional block upsampling followed by downsampling to obtain information from the previous layer at a higher resolution. This design makes the current layer's information more forward-looking and able to perceive more detailed information. Specifically, the output data of the previous layer (either the output data of the previous layer's decoder or the output data of the last layer's encoder) is input to the first convolutional block. The output of the first convolutional block is upsampled and then input to the second convolutional block along with the skip connection data from the previous layer's encoder. The output of the second convolutional block is downsampled and then input to the third convolutional block along with the skip connection data from the current layer's encoder. The outputs of the first and third convolutional blocks are input to the MLP cross-gating module. The output of the MLP cross-gating module is upsampled and then concatenated with the output of the second convolutional block in the stitching module. The output of the stitching module is the decoder's output. The Conv-MLP decoder structure is as follows: Figure 3 As shown in (b) above. Similar to the encoder design, the convolutional blocks in each layer of the decoder are fed into the MCG module twice simultaneously for cross-fusion.

[0046] To better integrate the information from convolution and MLP, this embodiment incorporates an MLP cross-gating (MCG) module, such as... Figure 4 As shown. This module effectively preserves the ability of convolutional operations to extract local and fine features, while also expanding the feature map to the global receptive field. The MCG module receives input from two paths, where each path is the output of one convolutional block. Let the inputs of the two paths be U1 and V1, where U1, V1 ∈ R. H×W×C The input has been processed as follows:

[0047] U2=σ(Ds(LN(U1)))V2=σ(Ds(LN(V1))) (1)

[0048] Where LN (LayerNorm) represents layer normalization, Ds (Dense) represents a fully connected layer, and σ represents the GELU activation function. After this series of operations, U2 and V2 are obtained, and then they are respectively fed into a multi-axis, multi-window MLP structure. First, the processed data is split into heads:

[0049] (L1,L2,G1,G2)=S(σ(Ds(LN(X)))) (2)

[0050] Here, S represents splitting the input feature map into four heads for operation. These four heads can be used for operations on different axes, as shown in Equation (3).

[0051] M(X)=Do(Ds([Local1(L1),Local2(L2),Global1(G1),Global2(G2)])) (3)

[0052] Here, square brackets [] indicate concatenation operations, and M(X) represents the output of the multi-axis, multi-window MLP module, which are M u and M v It is used to merge two paths.

[0053] U3=U2⊙M v (4)

[0054] V3=V2⊙M u (5)

[0055] Here, ⊙ represents the product operation of corresponding elements, and the output of each path is obtained by connecting it with the residual of the input.

[0056] U4=U1+(ρ(Ds(U3))) (6)

[0057] V4=V1+(ρ(Ds(V3))) (7)

[0058] Where ρ represents Dropout activation. To obtain two inputs and one output that satisfy the design, this embodiment performs an addition operation on the two paths, therefore:

[0059] M = U⁴ + V⁴ (8)

[0060] In 2021, researchers in this field proposed the idea of ​​sparse self-attention operation, which performs different forms of sparse self-attention on two axes of a segmented image to capture local and global information in parallel. However, this module requires a fixed-size test image as input, which easily leads to boundary blurring and artifacts, and is not friendly to training models with large images. Based on this, this embodiment proposes a multi-axis multi-window MLP framework, which establishes a four-axis image information fusion framework, such as... Figure 5 As shown.

[0061] The multi-axis multi-window MLP structure divides the input feature map into four heads according to channels. In each head, MLP is used to fuse information. Two heads are sent to the local feature branch, and the other two heads are sent to the global feature branch.

[0062] In the local feature branch, the two heads are fed into two feature processing axes, with the input tensor size of each axis being (H, W, C / 4). Simultaneously, one axis is divided into H / b × W / b non-overlapping windows of size [b × b]. This changes the dimension of the input tensor, resulting in a tensor of (H / b × W / b, b × b, C / 4). The other axis is divided into H / 2b × W / 2b non-overlapping windows of size [2b × 2b], again changing the dimension of the input tensor to a tensor of (H / 2b × W / 2b, b × b, C / 4).

[0063] In the global feature branch, one axis is divided into b×b non-overlapping windows of size H / b×W / b according to a grid size of [b×b], with each window being H / b×W / b. After reshaping the dimension, a tensor of (b×b, H / b×W / b, C / 4) is obtained. Similarly, the other axis is divided into 2b×2b windows of size H / 2b×W / 2b according to a grid size of [2b×2b], and after reshaping the dimension, a tensor of (2b×2b, H / 2b×W / 2b, C / 4) is obtained.

[0064] exist Figure 5In this diagram, we set b=2, with the upper half representing the local branch and the lower half representing the global branch. As shown in the diagram, after the input feature map passes through LayerNorm, Dense, and GELU layers, it is divided into four axes based on channels. Half of these axes enters the local branch head, and the other half enters the global branch head. Different sized windows are used for partitioning in both the local and global branches, and MLP is used for varying degrees of information fusion. The local branch corresponds to local feature fusion, and the global branch corresponds to global feature fusion. Different colored squares in the diagram represent the fusion status of information within that window. Finally, the four processed axes are concatenated and output through Dense and Dropout layers. This module's design is completely end-to-end from input to output, requiring no specific size and eliminating the need for image cropping, thus avoiding the negative effects of cropping. Furthermore, the model's complexity is linear, with low computational cost.

[0065] In the output (Conv-MLP) module (such as...) Figure 3 As shown in (c), the decoder's output data is input from the first convolutional block. The output of the first convolutional block is directly input to the second convolutional block. The outputs of the first and second convolutional blocks are then input together into the MLP cross-gating module. The output of the MLP cross-gating module, after upsampling, is concatenated with the data from the first layer encoder in the concatenation module. The output of the concatenation module is input to the third convolutional block, and the output of the third convolutional block is the model output.

[0066] S3: Segment the medical image to be segmented using the trained model.

[0067] To integrate the advantages of convolutional neural networks and Multiprocessor Logic (MLP), this invention proposes a novel U-shaped deep convolutional neural network architecture for medical image segmentation. The core idea is to fuse the respective strengths of convolutional and MLP, integrating them into each layer of the encoder and decoder. Simultaneously, it can capture feature map information from multiple windows and dimensions, and the network always possesses a global receptive field. The proposed MsMCG architecture efficiently builds the network with relatively few parameters and computational cost. Experimental results demonstrate that the proposed architecture outperforms state-of-the-art baselines on various benchmarks for segmenting biomedical images from different types. Furthermore, our network achieves excellent results for image segmentation in complex backgrounds.

[0068] The embodiments of the present invention have the following technical effects:

[0069] (1) A general framework for medical image segmentation, called MsMCG, is proposed. It uses a U-shaped encoder-decoder architecture, combining the advantages of convolution and MLP, and integrating both into each layer.

[0070] (2) A multi-axis and multi-window MLP module was designed to capture local and global information from multiple dimensions and distribute it equally to the obtained branches according to the channels. The information is mixed on their respective axes according to different mechanisms. Its input does not require image cropping, can receive images of any size, and always has a global receptive field.

[0071] (3) A Conv-MCG module was developed to perform feature cross-fusion of the two outputs of the convolution module, which is also a kind of multi-path, multi-information interaction.

[0072] (4) The designed network achieves some improvements in segmentation accuracy while also being lightweight. The computational cost of the MLP module is linearly related to the size of the input feature map. By inputting the feature maps output from two convolutional blocks into the MCG module, local and global receptive fields are obtained at each layer, enabling deeper information fusion. Compared to most existing networks, the model has superior performance in terms of parameter count and running speed. That is, it achieves good segmentation results with relatively low computational time and space overhead.

[0073] Example 2:

[0074] The present invention also provides a medical image segmentation terminal device integrating a multilayer perceptron, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the steps in the method embodiment described above in Embodiment 1 of the present invention.

[0075] Furthermore, as an executable solution, the medical image segmentation terminal device integrating a multilayer perceptron can be a computing device such as a desktop computer, laptop, handheld computer, or cloud server. The medical image segmentation terminal device integrating a multilayer perceptron may include, but is not limited to, a processor and memory. Those skilled in the art will understand that the above-described composition of the medical image segmentation terminal device integrating a multilayer perceptron is merely an example and does not constitute a limitation on the medical image segmentation terminal device integrating a multilayer perceptron. It may include more or fewer components than described above, or combine certain components, or different components. For example, the medical image segmentation terminal device integrating a multilayer perceptron may also include input / output devices, network access devices, buses, etc., and this embodiment of the invention does not limit this.

[0076] Furthermore, as an executable solution, the processor can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices. The general-purpose processor can be a microprocessor or any conventional processor. This processor is the control center of the medical image segmentation terminal device with fused multilayer perceptrons, connecting all parts of the device via various interfaces and lines.

[0077] The memory can be used to store the computer programs and / or modules. The processor implements various functions of the medical image segmentation terminal device fused with a multilayer sensor by running or executing the computer programs and / or modules stored in the memory and calling data stored in the memory. The memory may mainly include a program storage area and a data storage area. The program storage area may store the operating system and at least one application program required for a function; the data storage area may store data created based on the use of the mobile phone, etc. In addition, the memory may include high-speed random access memory and may also include non-volatile memory, such as hard disk, RAM, plug-in hard disk, smart media card (SMC), secure digital (SD) card, flash card, at least one disk storage device, flash memory device, or other volatile solid-state storage device.

[0078] The present invention also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps of the method described in the embodiments of the present invention.

[0079] If the modules / units integrated in the medical image segmentation terminal device incorporating a multilayer perceptron are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, all or part of the processes in the methods of the above embodiments of the present invention can also be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable medium can include: any entity or device capable of carrying the computer program code, a recording medium, a USB flash drive, a portable hard drive, a magnetic disk, an optical disk, a computer memory, a read-only memory (ROM), a random access memory (RAM), and a software distribution medium, etc.

[0080] Although the invention has been specifically shown and described in conjunction with preferred embodiments, those skilled in the art should understand that various changes in form and detail may be made to the invention without departing from the spirit and scope of the invention as defined in the appended claims, all of which shall be within the scope of protection of the invention.

Claims

1. A medical image segmentation method integrating multilayer perceptron, characterized in that, Includes the following steps: S1: Collect medical image data to construct a training set; S2: Construct a medical image segmentation model and train the model using a training set; The model's network backbone adopts a U-shaped network, including N encoders, N decoders and one output module; each encoder, decoder and output module includes three convolutional blocks, one MLP cross-gating module and one splicing module; In the encoder, data is input from the first convolutional block. The output of the first convolutional block is downsampled and then input to the second convolutional block. The output of the second convolutional block is upsampled and then input to the third convolutional block. The outputs of the first and third convolutional blocks are input together to the MLP cross-gating module. The output of the MLP cross-gating module is connected to the decoder via a skip connection, and after being downsampled, it is concatenated with the output of the second convolutional block in the stitching module. The output of the stitching module is the output of the encoder. In the decoder, the output data of the previous layer is input from the first convolutional block. The output of the first convolutional block is upsampled and then input together with the skip connection data of the previous layer encoder into the second convolutional block. The output of the second convolutional block is downsampled and then input together with the skip connection data of the corresponding layer encoder into the third convolutional block. The outputs of the first and third convolutional blocks are input together into the MLP cross-gating module. The output of the MLP cross-gating module is upsampled and then concatenated with the output of the second convolutional block in the splicing module. The output of the splicing module is the output of the decoder. In the output module, the decoder's output data is input from the first convolutional block. The output of the first convolutional block is directly input to the second convolutional block. The outputs of the first and second convolutional blocks are input together to the MLP cross-gating module. The output of the MLP cross-gating module is upsampled and then concatenated with the data from the first layer encoder in the concatenation module. The output of the concatenation module is used as the input to the third convolutional block, and the output of the third convolutional block is the model output result. The MLP cross-gating module includes two input paths. For each path, the input data is sequentially processed through layer normalization, a fully connected layer, and a GELU activation function before being input into a multi-axis multi-window MLP structure. The output of the multi-axis multi-window MLP structure is multiplied by its input. The multiplication result is sequentially processed through a fully connected layer and activated by Dropout. The Dropout activation result is then added to the input data of that path to obtain the output result for that path. The output results of the two paths are then added together to obtain the output result of the MLP cross-gating module. The multi-axis multi-window MLP structure divides the input feature map into four heads according to channels. Information fusion is performed using MLP in each head. Two heads are fed into the local feature branch, and the other two heads are fed into the global feature branch. In the local feature branch, the two heads are fed into two local feature processing axes with different window sizes. In the global feature branch, the two heads are fed into two global feature processing axes with different window sizes. S3: Segment the medical image to be segmented using the trained model.

2. The medical image segmentation method fused with multilayer perceptron according to claim 1, characterized in that: The convolutional block consists of a combination of two 3×3 convolutional layers, a BN layer, and a ReLU activation function layer.

3. The medical image segmentation method fused with multilayer perceptron according to claim 1, characterized in that: Step S1 also includes augmenting the acquired medical image data by random rotation or by adjusting the hue, brightness, or cropping, and constructing a training set based on the augmented data.

4. A medical image segmentation terminal device integrating a multilayer perceptron, characterized in that: It includes a processor, a memory, and a computer program stored in the memory and running on the processor, wherein the processor executes the computer program to implement the steps of the method as described in any one of claims 1 to 3.

5. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, it implements the steps of the method as described in any one of claims 1 to 3.

Images