Medical image segmentation model and medical image segmentation method

Abstract

The invention provides a medical image segmentation model and a medical image segmentation method, a coding layer and / or a decoding layer in a U-net are / is improved, so that the coding layer and / or the decoding layer of the medical image segmentation model adopt a multi-scale feature processing module, and the multi-scale feature processing module comprises a plurality of parallel convolution layers with different scales. When the coding layer comprises a multi-scale feature processing module, an output result of the multi-scale feature processing module is obtained by splicing outputs of a plurality of parallel convolution layers with different scales corresponding to the coding layer; and when the decoding layer comprises the multi-scale feature processing module, the output result of the multi-scale feature processing module is obtained by splicing the outputs of the plurality of parallel convolution layers with different scales corresponding to the decoding layer, so that the medical image segmentation model provided by the invention can accurately position the target area and improve the segmentation precision during image segmentation.

Description

Technical Field

[0001] This application relates to the field of medical image segmentation technology, and in particular to a medical image segmentation model and a medical image segmentation method. Background Technology

[0002] Medical image segmentation technology is a technique in the field of computer vision that is widely used in medical image analysis, intelligent diagnosis and treatment. Its purpose is to separate lesions, tumors or organs from the background by classifying medical images at the pixel level.

[0003] Among them, the U-Net model is a convolutional neural network specifically designed for biomedical image segmentation; however, the U-Net model has the following problems: (1) The U-Net model gradually increases the receptive field to extract semantic context information and global information through continuous downsampling operations. This will compress the dimension of the feature map and lose image detail information, resulting in the loss of some features when the decoder restores the image resolution, which makes the image segmentation inaccurate.

[0004] (2) The U-Net model uses skip connections to achieve direct fusion between features of the same scale. However, this forced fusion strategy ignores the potential semantic differences between features at different levels, resulting in feature maps lacking detailed features and reducing segmentation accuracy.

[0005] (3) The feature maps extracted by the convolutional layer lack spatial and positional information, making it difficult for the network to extract detailed features in complex environments, resulting in unclear segmentation of the foreground boundary region and incomplete segmentation results.

[0006] Therefore, there is an urgent need for a medical image segmentation model and a medical image segmentation method. Summary of the Invention

[0007] This application provides a medical image segmentation model and a medical image segmentation method to solve the technical problems mentioned in the background art.

[0008] In a first aspect, this application provides a medical image segmentation model, including: an encoder and a decoder, wherein the number of encoding layers contained in the encoder is the same as the number of decoding layers contained in the decoder, and the decoding layers correspond one-to-one with the encoding layers, the encoding layers are connected to other encoding layers in adjacent levels and the corresponding decoding layers, and the decoding layers are also connected to other decoding layers in adjacent levels. The coding layer is used to extract features from the input data corresponding to the coding layer, obtain the corresponding output result, and then send the output result to the next coding layer and the corresponding decoding layer. The input data corresponding to the coding layer is the input target medical image or the output result of the previous coding layer. The decoding layer is used to perform feature fusion on the input data corresponding to the decoding layer, and after obtaining the corresponding output result, it sends the output result corresponding to the decoding layer to the next decoding layer or obtains the image segmentation result of the target medical image based on the output result corresponding to the decoding layer. The input data corresponding to the decoding layer includes: the output result sent by the previous decoding layer and the corresponding output result sent by the decoding layer. The encoding layer and / or the decoding layer employ a multi-scale feature processing module, which includes multiple convolutional layers of different scales connected in parallel. When the coding layer includes the multi-scale feature processing module, the input data of the multi-scale feature processing module and the input data of the multiple parallel convolutional layers of different scales are consistent with the input data of the coding layer. The output result of the multi-scale feature processing module is consistent with the output result of the coding layer. The output result of the multi-scale feature processing module is obtained by concatenating the outputs of the multiple parallel convolutional layers of different scales corresponding to the coding layer. When the decoding layer includes a multi-scale feature processing module, the input data of the multi-scale feature processing module and the input data of the multiple parallel convolutional layers of different scales are consistent with the input data of the decoding layer. The output result of the multi-scale feature processing module is consistent with the output result of the decoding layer, and the output result of the multi-scale feature processing module is obtained by splicing the outputs of the multiple parallel convolutional layers of different scales corresponding to the decoding layer.

[0009] Secondly, this application provides a medical image segmentation method, comprising: Acquire target medical images; The target medical image is segmented according to the trained medical image segmentation model to obtain the image segmentation result, wherein the medical image segmentation model is any one of the medical image segmentation models described in the first aspect.

[0010] Thirdly, this application provides a medical image segmentation apparatus, comprising: The acquisition module is used to acquire the target medical image; The image segmentation module is used to segment the target medical image according to a trained medical image segmentation model to obtain the image segmentation result, wherein the medical image segmentation model is the medical image segmentation model described in the first aspect.

[0011] Fourthly, this application provides an electronic device, including: a processor and a memory; The memory stores the instructions that the computer executes; The processor executes computer execution instructions stored in memory, causing the processor to perform the method described in the second aspect.

[0012] Fifthly, embodiments of this application provide a readable storage medium including a program or instructions that, when run on a computer, execute the method described in the second aspect above.

[0013] In a sixth aspect, embodiments of this application provide a computer program product, including a computer program that, when executed by a processor, implements the method described in the second aspect.

[0014] The medical image segmentation model and method provided in this application improve the encoding and / or decoding layers in U-net by adopting a multi-scale feature processing module in the encoding and / or decoding layers. This multi-scale feature processing module includes multiple convolutional layers of different scales connected in parallel. When the encoding layer includes the multi-scale feature processing module, its output is obtained by concatenating the outputs of the corresponding convolutional layers of different scales. Similarly, when the decoding layer includes the multi-scale feature processing module, its output is obtained by concatenating the outputs of the corresponding convolutional layers of different scales. This allows the medical image segmentation model provided in this application to accurately locate target regions and improve segmentation accuracy during image segmentation. Attached Figure Description

[0015] Figure 1 This is a schematic diagram of the structure of the encoding layer and / or decoding layer of a medical image segmentation model proposed in an embodiment of this application; Figure 2 This is a transformation diagram of a medical image segmentation model based on structural reparameterization technology proposed in an embodiment of this application; Figure 3 This is a schematic diagram of the overall structure of a medical image segmentation model proposed in an embodiment of this application; Figure 4 This is a schematic diagram of the structure of the DEB in a medical image segmentation model proposed in an embodiment of this application; Figure 5 A flowchart of a medical image segmentation method provided in an embodiment of this application; Figure 6 This is a schematic diagram illustrating the cropping and comparison of sample images from a dataset provided in one embodiment of this application. Figure 7 A comparison chart of network segmentation performance metrics provided for an embodiment of this application; Figure 8 Visual comparison results of different networks provided in an embodiment of this application; Figure 9 Provided for an embodiment of this application Figure 1Comparison of ablation experimental performance indicators of the medical image segmentation model corresponding to (c); Figure 10 Provided for an embodiment of this application Figure 1 Visual comparison of the effectiveness of the medical image segmentation model corresponding to (c); Figure 11 A comparison chart of the ablation performance indicators of the DEB module provided in an embodiment of this application; Figure 12 A visual comparison diagram of DEB effectiveness provided in an embodiment of this application; Figure 13 This is a schematic diagram of the structure of a medical image segmentation device provided in an embodiment of this application; Figure 14 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Detailed Implementation

[0016] The U-Net model has achieved remarkable results in multiple medical segmentation tasks. However, the U-Net encoder only constructs two cascaded 3×3 convolutional layers to extract image features. This extraction method is limited in function and insufficient to adapt to complex medical image data.

[0017] Therefore, this application proposes a medical image segmentation model and a medical image segmentation method. In the encoding and / or decoding layers of the U-Net model, richer contextual information is mined from feature maps at different levels through multi-scale feature extraction, and multi-level features are fused, thereby significantly enhancing the network's ability to capture targets of different sizes and detailed information. This enables the network to better identify the boundaries of target structures when processing complex medical images, especially in areas with blurred edges and complex details.

[0018] The solution of this application will be explained and illustrated below through examples: Figure 1 This is a schematic diagram of the encoding and / or decoding layers of a medical image segmentation model proposed in an embodiment of this application. Figure 1 As shown, the medical image segmentation model includes an encoder and a decoder. The number of coding layers in the encoder is the same as the number of decoding layers in the decoder, and the decoding layers correspond one-to-one with the coding layers. The coding layers are connected to other coding layers in adjacent levels and to the corresponding decoding layers. The decoding layers are also connected to other decoding layers in adjacent levels. The coding layer is used to extract features from the input data corresponding to the coding layer, and after obtaining the corresponding output results, it sends the output results to the next coding layer and the corresponding decoding layer. The input data corresponding to the coding layer is the target medical image or the output results of the previous coding layer. The decoding layer is used to perform feature fusion on the input data corresponding to the decoding layer, and after obtaining the corresponding output result, it sends the output result corresponding to the decoding layer to the next decoding layer or obtains the image segmentation result of the target medical image based on the output result corresponding to the decoding layer. The input data corresponding to the decoding layer includes: the output result sent by the previous decoding layer and the output result sent by the corresponding decoding layer. The encoding layer and / or decoding layer employ a multi-scale feature processing module, which includes multiple convolutional layers of different scales connected in parallel. When the encoding layer includes a multi-scale feature processing module, the input data of the multi-scale feature processing module and the input data of the multiple convolutional layers of different scales connected in parallel are consistent with the input data of the encoding layer. The output result of the multi-scale feature processing module is consistent with the output result of the encoding layer. Moreover, the output result of the multi-scale feature processing module is obtained by concatenating the outputs of the multiple convolutional layers of different scales connected in parallel corresponding to the encoding layer. When the decoding layer includes a multi-scale feature processing module, the input data of the multi-scale feature processing module and the input data of the multiple parallel convolutional layers of different scales are consistent with the input data of the decoding layer. The output result of the multi-scale feature processing module is consistent with the output result of the decoding layer, and the output result of the multi-scale feature processing module is obtained by splicing the outputs of the multiple parallel convolutional layers of different scales corresponding to the decoding layer.

[0019] In this embodiment, the medical image segmentation model is obtained by improving the structure of the existing U-Net model. The U-Net model includes an encoder, a bottleneck layer, a decoder, skip connection layers, and an output layer. This embodiment improves the encoder and / or decoder in the U-Net model. Therefore, the structure of the medical image segmentation model provided in this embodiment is similar to that of the U-Net model. That is, the number of encoding layers in the encoder is the same as the number of decoding layers in the decoder, and the corresponding levels of the decoding and encoding layers correspond one-to-one through skip connection layers. The encoding layers are connected to other encoding layers in adjacent levels and to their corresponding decoding layers, and the decoding layers are also connected to other decoding layers in adjacent levels.

[0020] For the U-Net model and the medical image segmentation model provided in this embodiment, each encoding layer is used to extract features from the received input data, obtain the corresponding output results, and then send them to the next encoding layer and to the corresponding decoding layer through skip connection layers. The input data is the target medical image or the output result of the previous encoding layer.

[0021] Correspondingly, each decoding layer is used to perform feature fusion on the received input data, obtain the corresponding output result, and then send it to the next decoding layer or to the input layer, so that the input layer can obtain the image segmentation result of the target medical image based on the corresponding output result. The input data corresponding to the decoding layer includes: the output result sent by the previous decoding layer and the output result sent by the corresponding decoding layer.

[0022] The existing U-Net model uses two cascaded 3×3 convolutional layers in its encoder to extract image features. The structure of the medical image segmentation model provided in this embodiment is obtained by improving the structure of the encoding layer and / or decoding layer based on the structure of the U-Net model.

[0023] Specifically, the encoding layer and / or decoding layer employ a multi-scale feature processing module, which includes multiple convolutional layers of different scales connected in parallel. When the encoding layer uses multiple convolutional layers of different scales connected in parallel, the input data of the multi-scale feature processing module and the input data of the multiple convolutional layers of different scales connected in parallel are consistent with the input data of the encoding layer; the output of the multi-scale feature processing module is consistent with the output of the encoding layer, and the output of the multi-scale feature processing module is obtained by concatenating the outputs of the multiple convolutional layers of different scales connected in parallel corresponding to the encoding layer.

[0024] When the decoding layer uses multiple convolutional layers of different scales in parallel, the input data of the multi-scale feature processing module and the input data of the multiple convolutional layers of different scales in parallel are consistent with the input data of the decoding layer. The output result of the multi-scale feature processing module is consistent with the output result of the decoding layer. Moreover, the output result of the multi-scale feature processing module is obtained by splicing the outputs of the multiple convolutional layers of different scales in parallel corresponding to the decoding layer.

[0025] For other parts of the medical image segmentation model provided in this embodiment, existing U-Net models or other improved U-Net models, such as U-Net++ and Attention U-Net, can be referenced. This application does not limit this.

[0026] When performing image segmentation on a target medical image, the neural network "perceives" the size of the input medical image region through its receptive field. The size of the input medical image region perceived by the receptive field is related to the size of the semantic feature region of the image that can be extracted by the convolution operation.

[0027] In the medical image segmentation model provided in this embodiment, when the coding layer uses multiple convolutional layers of different scales in parallel, each scale of the convolutional layer extracts features from the input data. Since the scales of the convolutional layers are different, each convolutional layer extracts different features from the input data. Then, the features extracted by each scale of the convolutional layer are concatenated, so that the concatenated features integrate local detail features and global features, thereby improving the accuracy of the final image segmentation.

[0028] Correspondingly, when the decoding layer employs multiple convolutional layers of different scales in parallel, different features are identified through these convolutional layers of varying scales, avoiding the omission of features and effectively preserving more detailed information while restoring the spatial resolution of the image. This multi-scale feature fusion design enables the decoder network to recover the complex boundaries of the target region more completely, especially when dealing with regions with blurred boundaries or complex structures, exhibiting stronger robustness and higher segmentation accuracy.

[0029] Optionally, the multiple convolutional layers of different scales in parallel can be of different numbers or different scales. For example, Example 1 is three convolutional layers of different scales in parallel, or Example 2 is three convolutional layers in parallel, at least one of which has a different scale than in Example 1.

[0030] Specifically, such as Figure 1 As shown in (a), the multiple convolutional layers of different scales in parallel are a first convolutional layer of 3×3, a second convolutional layer of 5×5, and a third convolutional layer of 7×7 in parallel.

[0031] Optional, in Figure 1 Based on the embodiment shown in (a), the second convolutional layer is replaced by a first convolutional layer and a fourth convolutional layer of the first scale arranged in series; The third convolutional layer is replaced by a first convolutional layer, a fourth convolutional layer, and a fifth convolutional layer of the first scale arranged in series. The input data of the first convolutional layer is the same as the input data of the multi-scale feature processing module. The input data of the fourth convolutional layer is the input data of the first convolutional layer. The input data of the fifth convolutional layer is the input data of the fourth convolutional layer. The output results of the first convolutional layer, the fourth convolutional layer and the fifth convolutional layer are concatenated to form the output results of multiple convolutional layers of different scales in parallel. The multi-scale feature processing module also includes: a sixth convolutional layer at the fourth scale, where the fourth scale is smaller than the first scale; The input data of the sixth convolutional layer is the same as that of the first convolutional layer. The output of the multi-scale feature processing module is obtained by fusing the output of the sixth convolutional layer with the outputs of multiple convolutional layers of different scales in parallel.

[0032] In this embodiment, the receptive field of two cascaded 3×3 convolutions is equivalent to one 5×5 convolution operation, and three cascaded 3×3 convolutions are equivalent to one 7×7 convolution operation. Therefore, in Figure 1 Based on the embodiment shown in (a), Figure 1 In (a), the second convolutional layer is replaced with a first convolutional layer and a fourth convolutional layer at a first scale arranged in series, and the third convolutional layer is replaced with a first convolutional layer, a fourth convolutional layer, and a fifth convolutional layer at a first scale arranged in series, which is reflected in Figure 1 In (b), three 3×3 convolutional layers are connected in series. The first convolutional layer is a parallel branch, the first and fourth convolutional layers are a parallel second branch, and the first, fourth, and fifth convolutional layers are a parallel third branch. The input data of each branch is the same.

[0033] Specifically, such as Figure 1 As shown in (b), the output of the first convolutional layer is used as the input data of the fourth convolutional layer. The output of the first convolutional layer is also output to the channel-dimensional concatenation layer, and the output of the fourth convolutional layer is also output to the channel-dimensional concatenation layer, so that the branch formed by the fourth convolutional layer and the first convolutional layer is connected in parallel with the first convolutional layer; the output of the fourth convolutional layer is used as the input data of the fifth convolutional layer, and the output of the fifth convolutional layer is also output to the channel-dimensional concatenation layer, so that the branch formed by the first convolutional layer, the fourth convolutional layer and the fifth convolutional layer is connected in parallel with the branch formed by the first convolutional layer, and the branch formed by the fourth convolutional layer and the first convolutional layer.

[0034] Among them, such as Figure 1 As shown in (b), the multi-scale feature processing module also includes a sixth convolutional layer at the fourth scale, where the fourth scale is smaller than the first scale. Specifically, the fourth scale is 1×1, therefore, the sixth convolutional layer is a 1×1 convolutional layer. This 1×1 convolutional layer extracts features from the input data of the first convolutional layer, capturing more localized detailed features. The output is then concatenated with the outputs of the first, fourth, and fifth convolutional layers in the channel domain using a concatenation layer. The resulting concatenation is then performed element-wise to obtain the output of the multi-scale feature processing module, i.e., the output corresponding to that level.

[0035] In this embodiment, the second convolutional layer is replaced by a series of first and fourth convolutional layers, and the third convolutional layer is replaced by a series of first, fourth, and fifth convolutional layers. Since the smaller the scale of the convolutional layers, the fewer parameters are required. Therefore, through the above operations, the implementation... Figure 1 Based on the convolution operation shown in (a), the number of parameters is reduced, thus reducing the amount of computation and improving the model's computation speed when using this medical image segmentation model to segment medical images.

[0036] Optionally, to extract more comprehensive image detail features, in Figure 1 Based on (b), such as Figure 1 As shown in (c), the multi-scale feature processing module further includes: a seventh convolutional layer of the fourth scale and a first identity mapping branch connected in parallel with the first convolutional layer; an eighth convolutional layer of the fourth scale and a second identity mapping branch connected in parallel with the fourth convolutional layer; and a ninth convolutional layer of the fourth scale and a third identity mapping branch connected in parallel with the fifth convolutional layer. The input data of the first convolutional layer is the same as the input data of the multi-scale feature processing module. The input data of the fourth convolutional layer is the output result obtained by fusing the output results of the first convolutional layer, the seventh convolutional layer, and the first identity mapping branch. The input data of the fifth convolutional layer is the output result obtained by fusing the output results of the fourth convolutional layer, the eighth convolutional layer, and the second identity mapping branch. The first output result is obtained by fusing the output results of the first convolutional layer, the seventh convolutional layer, and the first identity mapping branch. The output result obtained by fusing the output results of the fourth convolutional layer, the eighth convolutional layer, and the second identity mapping branch is the second output result; The third output result is obtained by fusing the output results of the fifth convolutional layer, the ninth convolutional layer, and the third identity mapping branch. The output of the multi-scale feature processing module is obtained by fusing the output of the first, second and third outputs with the output of the sixth convolutional layer.

[0037] Specifically, such as Figure 1 As shown in (c), for Figure 1 In (b), each 3×3 convolutional layer is configured with a parallel 1×1 convolutional layer and an identity mapping branch. The output results obtained by the 3×3 convolutional layer from the input data, the output results obtained by the 1×1 convolutional layer from the input data, and the input data directly transmitted through the identity mapping branch (which can also be called the output results of the identity mapping branch) are added and fused together, and used as the input data for the next 3×3 convolutional layer, 1×1 convolutional layer, and identity mapping branch.

[0038] Furthermore, for ease of description, the output obtained by fusing the outputs of the first convolutional layer, the seventh convolutional layer, and the first identity mapping branch is denoted as the first output result; the output obtained by fusing the outputs of the fourth convolutional layer, the eighth convolutional layer, and the second identity mapping branch is denoted as the second output result; and the output obtained by fusing the outputs of the fifth convolutional layer, the ninth convolutional layer, and the third identity mapping branch is denoted as the third output result. The first, second, and third output results are concatenated along the channel dimension to serve as the outputs of multiple convolutional layers of different scales in parallel. Then, the outputs of the multiple convolutional layers of different scales in parallel are added to and fused with the output of the sixth convolutional layer to obtain the output of the multi-scale feature processing module.

[0039] In this embodiment, Figure 1 Based on (b), the input features are directly passed through the identity mapping branch, which helps to preserve the original information. The 1×1 convolutional layer is used to capture local detail features, while the 3×3 convolutional layer is responsible for extracting broader contextual information. By adding and fusing the outputs of the identity mapping branch, the 1×1 convolutional layer and the 3×3 convolutional layer, not only is the ability to extract multi-scale features preserved, but the expressive power of features is also enhanced through residual connections, which can generate richer and more accurate feature representations.

[0040] Among them, such as Figure 1 The model structure shown in (c), while capable of generating richer and more accurate feature representations, is still inferior to... Figure 1 The model structure shown in (b) increases the number of parameters, thus increasing computational complexity and memory requirements. Therefore, as Figure 2 As shown, structural reparameterization can be used to separate the model structure during the training phase from that during the inference phase, transforming the multi-branch parallel structure into a branchless structure. Therefore, for Figure 1 The model structure shown in (c) can be referred to as the multi-scale feature processing module as the heavy parameter multi-scale residual module.

[0041] Therefore, for such Figure 1 The model structure shown in (c) extracts rich feature representations through a multi-branch structure during the model training phase, and through... Figure 1 When the model shown in (c) performs image segmentation on the target medical image, it transforms the parameters of the multi-branch parallel structure into the parameters of a branchless structure through parameter transformation, thereby achieving equivalent reconstruction of the structure. This design not only retains the advantages of the multi-branch structure in the training phase, but also significantly reduces the computational overhead in the inference phase.

[0042] Specifically, for a given input imagex ∈ R B×C×H×W ,one k × k convolution W ∈ R D×C×k×k Convolution result y It can be represented as: y=x·W+b Formula 1 in, b ∈ R D It's a bias, and after passing through the Batch Normalization (BN) layer, the output... y BN It can be represented as: Formula 2 in, μ and σ 2 These are the batch normalized mean and variance. γ and β These are learnable parameters. Combining Equation 1 and Equation 2 yields the reparameterized output. y reparam : Formula 3 Among them, the reparameterized convolution kernel bias .

[0043] Formula 3 is Figure 1 The reparameterized output corresponding to any convolutional layer in (c).

[0044] Based on the above formula, the unit branch can be viewed as a 1×1 convolution with a kernel parameter of 1, and the 1×1 convolution can be viewed as a degenerate 3×3 convolution transformation. For the output of a three-branch parallel convolution structure... Y reparam It can be represented as: Formula 4 in, W 1. W 2. W 3 corresponds to the convolution kernels of the identity mapping branch, the 1×1 convolution branch, and the 3×3 convolution branch, respectively. γ 1. γ 2. γ 3 and , , These are the batch normalization parameters for each branch. B reparamIt is the total bias term after reparameterization.

[0045] During the training phase, use Figure 1 The three-branch parallel convolutional structure shown in (c) is used for training to fully utilize the feature extraction capabilities of the multi-branch structure. During the inference phase, the multi-branch structure is equivalently transformed into... Figure 1 The single-branch structure shown in (b) demonstrates this structural transformation. This transformation not only fully leverages the powerful feature extraction capabilities of multi-branch structures during training but also enables efficient multi-scale feature extraction with lower computational cost during inference.

[0046] Optional, such as Figure 3 As shown, the medical image segmentation model also includes a detail enhancement block (DEB) set between the one-to-one coding layer and the decoding layer. The detail enhancement module is used to increase the attention to the detailed features of the target region in the output of the corresponding encoding layer through the set attention mechanism, obtain the corresponding output, and send the corresponding output to the corresponding decoding layer.

[0047] In this embodiment, the existing U-Net model uses skip connection layers to achieve direct fusion of features at the same scale between the one-to-one corresponding encoding and decoding layers. However, this fusion ignores the potential semantic differences between features at different levels, resulting in feature maps lacking detailed features and reducing segmentation accuracy. Therefore, this embodiment sets up a DEB between the one-to-one corresponding encoding and decoding layers. Through the set attention mechanism, it increases the attention to the detailed features of the target region in the output of the corresponding encoding layer, obtains the corresponding output, and sends the corresponding output to the corresponding decoding layer.

[0048] Optionally, the detail enhancement module includes: a splitting unit and a fusion unit. The fusion unit includes: a global feature extraction subunit, a dimensional attention channel, a spatial attention channel, and a global feature map fusion subunit. The input of the splitting unit is connected to the encoding layer, and its output is connected to the input of the global feature extraction subunit. The output of the global feature extraction subunit is connected to the inputs of the dimensional attention channel and the spatial attention channel. The outputs of the dimensional attention channel and the spatial attention channel are connected to the input of the global feature map fusion subunit. The output of the global feature map fusion subunit is connected to the decoding layer. The splitting unit is used to extract features from the output of the coding layer through the tenth convolutional layer at the first scale, the eleventh convolutional layer at the second scale, and the twelfth convolutional layer at the third scale, respectively, to obtain feature maps at three different scales. The global feature extraction subunit is used to extract global information from feature maps at three different scales to obtain global feature maps. The dimensional attention channel is used to perform feature transformation on the global feature map through global average pooling and the thirteenth convolutional layer of the first scale to obtain the channel feature map, and to obtain the channel attention map based on the channel feature map and the feature maps of three different scales. The spatial attention channel is used to perform feature transformation on the global feature map through global average pooling, global max pooling, and the fourteenth convolutional layer at the third scale to obtain the spatial feature map, and to obtain the spatial attention map based on the spatial feature map and the feature maps at three different scales. The global feature map fusion subunit is used to obtain the corresponding output results based on the channel attention map and the spatial attention map.

[0049] Specifically, such as Figure 4 As shown in (a), the DEB mainly consists of two parts: splitting units and fusion units, where, as Figure 4 As shown in (b), the splitting unit uses convolutional layers with different receptive fields to extract multi-scale features of the image, extracting different scale features of the image through three different sizes of convolution operations. input i , i =1, 2, 3}: Formula 5 in, input This indicates the output received by the DEB from the encoding layer.

[0050] The splitting unit feeds the input data at three different scales to the global feature extraction subunit. The global feature extraction subunit then extracts global information from the input data at these three scales to obtain the global feature map F. Formula 6 The global feature extraction subunit feeds the global feature map F to the dimensional attention channel and the spatial attention channel. The dimensional attention channel and the spatial attention channel are two parallel branches, which are used to calculate the attention weights of image features in the channel dimension and the spatial dimension, respectively.

[0051] In the dimensional attention channel, the global feature map F is subjected to global average pooling (GAP), and then further transformed by a 1×1 thirteenth convolutional layer to obtain the channel feature map. F C The number of thirteenth convolutional layers can be two, meaning the global feature map F is processed by global average pooling and then passed through two 1×1 thirteenth convolutional layers: Formula 7 In the spatial attention channel, the global feature map F undergoes a global average pooling (GAP) and a global max pooling (GMP) operation, followed by a 7×7 convolutional layer for feature transformation to generate the spatial attention map. F S : Formula 8 generated F C and F S These are two parallel attention maps, representing the attention information in the channel dimension and the spatial dimension, respectively. To apply these weights to the multi-scale feature map, the dimensional attention channel... F C Evenly divided into three parts, the spatial attention channel also includes the channel dimension. F S Divide evenly into three parts: Formula 9 Formula 10 In the dimensional attention channel, the split { C i , i =1, 2, 3} is multiplied by the split multi-scale feature map, i.e., the input data at 3 scales, to obtain the channel attention map of the channel dimension at each scale; in the spatial attention channel, the split { S i , i =1, 2, 3} are multiplied with the split multi-scale feature map, i.e., the input data at 3 scales, to obtain the spatial attention map of the channel dimension at each scale.

[0052] Then, the global feature map fusion subunit fuses the channel attention map and the spatial attention map to obtain the corresponding output: Formula 11 in, This represents a channel attention map. Represents a spatial attention map.

[0053] In this embodiment, by introducing the DEB module, the model not only breaks through the limitations of traditional single-scale feature extraction but also enhances its ability to capture global information through the comprehensive representation of multi-scale information. This multi-scale feature mapping strategy improves the model's ability to express details, especially in complex medical images, making the identification of blurred edges more accurate. Furthermore, by parallelly fusing channel attention and spatial attention, the DEB module introduces attention to both spatial and channel dimensions on the feature map, allowing the model to dynamically adjust the degree of attention given to different regions and channel features. This significantly improves the model's performance, enabling it to better address the challenges of medical image segmentation, particularly demonstrating stronger robustness and accuracy when dealing with blurred boundaries.

[0054] Figure 5 A flowchart of a medical image segmentation method provided in an embodiment of this application is shown below. Figure 5 As shown, the subject executing this method can be, for example, an electronic device such as a laptop, tablet, or server, and the method includes: S501. Acquire the target medical image.

[0055] In this step, the target medical image can be a CT image, or a magnetic resonance imaging (MRI), magnetic resonance angiography (MRA), or other medical images.

[0056] S502. Perform image segmentation on the target medical image based on the trained medical image segmentation model to obtain the image segmentation result.

[0057] The medical image segmentation model is the medical image segmentation model described in any of the above embodiments.

[0058] In this step, the target regions to be segmented are different for medical images of different parts of the body. Therefore, when it is necessary to segment the target regions corresponding to specific parts of the body, it is necessary to use medical image samples of the corresponding body parts to train the medical image segmentation model described in any of the above embodiments.

[0059] For example, if the target medical image is a CT image of a throat cancer tumor, then samples of throat cancer tumor CT images need to be used to train the medical image segmentation model described in any of the above embodiments.

[0060] by Figure 3 The structure of the medical image segmentation model shown illustrates the image segmentation process for CT images of pharyngeal cancer tumors. The structure of the encoding and decoding layers in the medical image segmentation model used in this step is as follows: Figure 1 As shown in (c), a DEB is set between the encoding layer and the decoding layer, and the image segmentation process is as follows: A center-cropping strategy was used to crop CT images of pharyngeal cancer tumors to obtain images with a pixel size of 256×256. The image data corresponding to this image was used as the input data for the first coding layer. The first convolutional layer, the seventh convolutional layer, and the first identity mapping branch of this coding layer used Equation 4 to extract features from the input data to obtain the first output result. The fourth convolutional layer, the eighth convolutional layer, and the second identity mapping branch used Equation 4 to extract features from the first output result to obtain the second output result. The fifth convolutional layer, the ninth convolutional layer, and the third identity mapping branch used Equation 4 to extract features from the second output result to obtain the third output result.

[0061] Among them, such as Figure 3 As shown, the first, second, and third output results each have 21 channels. Channel-wise fusion is performed on these three output results, resulting in a feature map with 63 channels. Then, this feature map with 63 channels is element-wise added to the output of the sixth convolutional layer to obtain the output of this coding layer with 63 channels. This output is then used as the input data for the next coding layer and transmitted to the corresponding DEB module.

[0062] The working process of the next coding layer is the same as that of the previous coding layer, and will not be repeated here.

[0063] After receiving the input data, the DEB module processes it according to Formulas 5-11. This enhances the DEB module's output data, making it more accurate in capturing global information about image features in CT images of pharyngeal cancer tumors and improving the recognition of blurred edges in the target area. The number of channels in the DEB module's output data is consistent with the number of channels in its received input data.

[0064] for Figure 3 The fifth encoding layer in the process obtains the output result and then sends it directly to the first decoding layer.

[0065] Since the structure of the decoding layer is the same as that of the encoding layer, the processing of input data by the decoding layer can be referenced from that of the encoding layer. The first decoding layer sends the output to the second decoding layer. The decoding layer upsamples the image features, so the output of the first decoding layer has 504 channels.

[0066] The second decoding layer concatenates the input data received from the first decoding layer with the input data received from the DEB module in the channel dimension. Therefore, the number of channels for the input data of the second decoding layer is 1008.

[0067] The fifth decoding layer processes the input data to obtain the image segmentation result.

[0068] In this embodiment, since the medical image segmentation model described in any of the above embodiments is used, the medical image segmentation method can more accurately identify the target region and improve the accuracy of image segmentation when segmenting the target medical image.

[0069] The following section uses a CT image of a pharyngeal cancer tumor as an example to verify the medical image segmentation model and method proposed in this application.

[0070] To ensure the rigor and reproducibility of the experimental results, this section will detail the experimental setup, including the dataset used, the selected optimizer, the configuration of experimental parameters, and the model evaluation criteria. These experiments will not only comprehensively verify the advantages of the medical image segmentation model (hereinafter referred to as MEDE-Net) presented in this application, but also further reveal its design principles and innovations, thus promoting the research and development of medical image segmentation technology.

[0071] 1. Setting up the training dataset In training the MEDE-Net model, this application uses a private dataset of manually labeled CT images of pharyngeal cancer tumors, with the original data being 512×512×512 three-dimensional CT images. To facilitate subsequent experimental processing and analysis, the original 3D data was sliced, extracting 5551 2D images, each 512×512 pixels. Experimental analysis shows that pharyngeal tumors are relatively small compared to other regions and occupy a very limited area in the entire CT image, resulting in a very low proportion of the foreground tumor to be segmented. Specifically, in the entire dataset, the smallest foreground tumor area occupies only 0.097% of the entire image, the largest is 11.68%, and the average is 2.57%. This phenomenon reflects a significant class imbalance problem in the dataset, which often poses a significant challenge in medical image segmentation, potentially leading to insufficient learning of the tumor region during model training and affecting segmentation performance.

[0072] To solve this problem, a center-cropping strategy is adopted, cropping the original image to a size of 256×256. For example... Figure 6 As shown, although the image resolution is reduced in this way, the cropped image can better focus on displaying the tumor region, thereby increasing the proportion of the foreground tumor. After cropping, the average proportion increases to 10.32%. The preprocessing results demonstrate that this method effectively alleviates the imbalance problem in the dataset, enhances the importance of the tumor region in the image data, and enables the model to better focus on key areas, thus improving the accuracy of segmentation.

[0073] In terms of dataset partitioning, 551 images were randomly selected from 5551 images in a 9:1 ratio as the validation set, while the remaining 5000 images were used to construct the training set. This partitioning method ensures that the distribution of the training and validation sets is representative, while also effectively improving the model's generalization ability. This data preprocessing and partitioning strategy maximizes the model's training performance and minimizes the negative impact of data imbalance on model performance.

[0074] 2. Experimental Setup The MEDE-Net model was constructed and configured using the PyTorch framework and trained on an NVIDIA GeForce RTX 3090 GPU. The optimizer was Adam, the batch size was 16, the learning rate was 1e-4, the input image size was 256×256, and the Dice loss function was selected.

[0075] To comprehensively evaluate the performance of MEDE-Net in laryngeal cancer tumor CT image segmentation, four performance metrics were used to evaluate the model's segmentation effect: mean absolute error (MAE), Dice coefficient, intersection-over-union ratio (IoU), and weighted f-measure. These evaluation metrics have different focuses and can comprehensively reflect the segmentation accuracy and effectiveness of the model from multiple aspects.

[0076] 3. Comparison and Analysis of Experimental Results To evaluate the effectiveness of the proposed MEDE-Net network model in CT image segmentation for laryngeal cancer, this section designs and implements relevant experiments. First, under the same dataset and training environment, the network model is compared with five related medical segmentation networks. The segmentation performance of MEDE-Net is evaluated from both quantitative and qualitative visual perspectives. To analyze the contribution of key MEDE-Net modules to segmentation performance, ablation experiments are then designed and implemented to explore... Figure 1 The actual effects of the multi-scale feature processing module and detail enhancement module (DEB) corresponding to (c) are evaluated to assess their role in improving the segmentation accuracy of the model.

[0077] 3.1 Model Comparison Results and Analysis To verify the superior segmentation performance of MEDE-Net, five state-of-the-art CNN-based network models were selected for comparison: U-Net, UNet++, Attention U-Net, and M... 2SNet and Repunet. To ensure the fairness and rigor of the experiment, these networks were retrained on the same laryngeal cancer tumor CT image dataset based on their source code. The segmentation performance differences between MEDE-Net and these models were compared from both quantitative and qualitative visual evaluation perspectives.

[0078] 3.1.1 Quantitative Evaluation According to Table 1 and Figure 7 The experimental results show that the proposed MEDE-Net model significantly outperforms other comparative network models in segmentation performance on the pharyngeal cancer tumor CT image dataset. Specifically, MEDE-Net achieves Dice coefficients of 0.854 and IoU of 0.777, demonstrating its powerful performance in pharyngeal cancer tumor image segmentation.

[0079] Table 1. Quantitative comparison of CT image datasets of the pharyngeal region

[0080] Compared to the baseline model U-Net, MEDE-Net achieves significant improvements across four common segmentation performance metrics, with an IoU improvement of 10.7% and a mean absolute error (MAE) reduction of 33.7%. This significant performance improvement is attributed to MEDE-Net's innovative introduction of the RepMultiRes Block and DEB module. Figure 1 The multi-scale feature processing module (c) can simultaneously capture local details and global contextual information of an image through a multi-scale feature extraction mechanism, thereby enhancing the model's ability to represent complex structures. DEB, on the other hand, combines multi-scale strategies and attention mechanisms to further improve the model's ability to focus on tumor regions. These improvements enable MEDE-Net to outperform U-Net in multi-scale feature extraction, attention allocation, and feature fusion, resulting in a significant overall performance improvement.

[0081] Compared to the second-place Attention U-Net, the MEDE-Net model improved the IoU metric by 2.7 percentage points and reduced the MAE by 4.8 percentage points. This improvement indicates that while Attention U-Net effectively enhances model performance in medical image segmentation tasks by introducing an attention mechanism, it performs poorly in throat tumor segmentation tasks, primarily due to limitations in its network structure. Attention U-Net only assigns weights to different regions of the feature map in the spatial dimension, ignoring feature variations and their correlations in the channel dimension, resulting in insufficient feature representation capabilities. Furthermore, Attention U-Net's use of single-scale convolutional operations makes it difficult to simultaneously capture the details and global structure of throat tumors, especially when dealing with small tumors or regions with blurred boundaries, easily leading to the loss of important features. These factors collectively cause Attention U-Net to underperform MEDE-Net when handling complex and varied throat tumor images.

[0082] Compared to UNet++, MEDE-Net improves the Dice coefficient and IoU by 1.5 and 3 percentage points, respectively. UNet++ introduces more skip connections and denser nested structures on top of U-Net, enabling the model to fully utilize multi-level feature information while capturing both global contextual information and local details of the image. However, UNet++'s feature fusion method primarily focuses on reusing features between layers, failing to adequately consider the interrelationships between features at the same layer. This can lead to a lack of feature representation and reduced segmentation accuracy when dealing with complex scenes.

[0083] With Repunet and M 2 Compared to SNet, MEDE-Net also demonstrated a significant performance advantage. In terms of Dice coefficients, MEDE-Net improved by 3.5 and 1.5 percentage points respectively, and in terms of IoU, by 5 and 1.8 percentage points respectively. Repunet only reduced the model structure through structural parameter refactoring, but its encoder and decoder remained consistent with U-Net. The model did not specifically optimize for feature extraction and boundary localization, resulting in its performance shortcomings in complex images. 2 SNet enhances the robustness of the model through multi-scale feature fusion, but its segmentation performance is still limited because it fails to fully consider the multi-dimensional interaction of local and global information in the image during feature extraction.

[0084] In summary, MEDE-Net introduces... Figure 1The multi-scale feature processing module (c) and DEB module innovatively improve the model's performance in laryngeal cancer tumor CT image segmentation. The parallel extraction of multi-scale features and the combination of a hybrid attention mechanism enable MEDE-Net to more effectively extract key features of the pharynx and accurately locate boundaries, thus demonstrating superior segmentation performance when processing complex medical images. These innovations allow MEDE-Net to outperform other mainstream medical image segmentation networks in segmentation accuracy on pharyngeal CT datasets, demonstrating its superiority and broad application potential in medical image segmentation tasks.

[0085] 3.1.2 Qualitative Evaluation In medical image segmentation tasks, CT images of the pharynx are particularly challenging due to their unique anatomical structure and image characteristics. Figure 8 This presentation showcases the prediction results of different models for pharyngeal cancer tumor CT image segmentation, including U-Net, UNet++, Attention U-Net, and M... 2 SNet, Repunet, and MEDE-Net (proposed in this paper) are compared and analyzed. This comparison clearly demonstrates the advantages and disadvantages of each model in throat region segmentation.

[0086] First, CT images of pharyngeal tumors have several notable characteristics: small size, very low foreground (i.e., tumor area) proportion, and minimal pixel difference between the tumor and background. These factors all increase the difficulty of segmentation. Under such circumstances, accurately identifying the pharyngeal region from the image while avoiding interference from background noise becomes a key challenge in the segmentation task.

[0087] from Figure 8 As shown in the segmentation results (a, c, e), U-Net, Repunet, M²SNet, and Attention U-Net exhibit significant segmentation errors when segmenting pharyngeal cancer CT image data. The segmentation results in the figures show a large deviation between the foreground region predicted by the model and the actually labeled tumor region, especially in the identification of irregular tumor boundaries, where significant errors occur. Figure 8 As shown in (a), in complex structural regions, the model's ability to capture detailed information is insufficient, resulting in unsmooth segmentation boundaries that do not match the actual tumor structure. However, for small tumor regions, such as... Figure 8 As shown in (b), these models often fail to accurately capture their boundary information, resulting in fragmented or blurred segmentation results.

[0088] These issues may stem from inherent limitations in the structural design of these models, hindering their ability to effectively segment complex throat tumor images with small regions and low contrast. While U-Net can preserve spatial information to some extent, its single-scale feature extraction method limits its tumor feature extraction capabilities. UNet++, despite its strong feature extraction capabilities, often fails to accurately extract detailed information when dealing with fine-grained structures, especially small-volume regions with minimal foreground-background contrast. RepuNet, while enhancing feature transfer through residual connections, still lacks sufficient fusion of multi-scale features, resulting in poor performance when handling complex structures. AttentionU-Net, while improving performance to some extent through its attention mechanism, primarily focuses on the spatial dimension, neglecting feature associations along the channel dimension, thus limiting its performance in low-contrast images.

[0089] In contrast, the proposed MEDE-Net demonstrates significant advantages in this task. MEDE-Net can more accurately locate the pharyngeal region and excels particularly in boundary recognition. By introducing a multi-scale feature fusion module and an attention mechanism, MEDE-Net can simultaneously capture both global structural and local detail information of the image, thus significantly improving the model's ability to capture fine-grained structures. These improvements enable MEDE-Net to more accurately segment the target region when processing complex and varied pharyngeal tumor images, and exhibit higher accuracy and robustness in boundary recognition.

[0090] 3.2 Ablation Experiment MDEM-Net is an improvement and optimization of the general medical segmentation network U-Net, designed to enhance its segmentation accuracy in laryngeal tumor tasks. To validate the two modules proposed in MDEM-Net— Figure 1 The effectiveness of the multi-scale feature processing module and detail enhancement module (DEB) corresponding to (c) is analyzed step by step using U-Net as the backbone network to analyze the contribution of each component of MDEM-Net to the segmentation task.

[0091] 3.2.1 Figure 1 The effectiveness of the multi-scale feature processing module corresponding to (c) This paper uses U-Net as the backbone network and designs four control networks by replacing new modules in the network to verify the effectiveness of RepMutiRes Block segmentation. The control networks are as follows: the original U-Net, the one using... Figure 1 (c) The corresponding multi-scale feature processing module replaces the encoder's U-Net (U-Net + Replaced Encoder), using Figure 1 (c) The corresponding multi-scale feature processing module replaces the U-Net decoder (U-Net + Replaced Decoder), using Figure 1 The multi-scale feature processing module corresponding to (c) replaces the encoding and decoding U-Net (equivalent to MEDE-Net without DEB, i.e., MEDE-Net - DEB). The mean absolute error (MAE), Dice coefficients, intersection-over-union ratio (IoU), and weighted f-measure are used. The evaluation is based on four indicators.

[0092] Table 2 Figure 1 Validation of the multi-scale feature processing module corresponding to (c)

[0093] The experimental results are shown in Table 2 and Figure 9 As shown in the results. It can be seen from the results that, with... Figure 1 The multi-scale feature processing module corresponding to (c) gradually replaces the encoder and decoder in U-Net, and the segmentation performance of the model is significantly improved. Figure 1 After the multi-scale feature processing module corresponding to (c) replaced the U-Net encoder, the IoU increased from 0.702 to 0.735, an increase of 3.3%, indicating that the replaced encoder's ability in multi-scale feature extraction was enhanced. And when... Figure 1 When the multi-scale feature processing module corresponding to (c) is applied to the decoder, the improvement in IoU is more significant, reaching 0.741. This is because, compared to the encoder extracting only multi-scale features at the same level, the introduction of... Figure 1 The decoder of the multi-scale feature processing module corresponding to (c) can not only extract multi-level multi-scale features, but also perform more effective multi-scale feature integration, thereby better restoring image detail features, especially when processing tumor boundaries and small regions.

[0094] when Figure 1 When the multi-scale feature processing module corresponding to (c) simultaneously replaces both the encoder and decoder, the IoU increases to 0.767, an increase of 6.5 percentage points. This result indicates that... Figure 1 The multi-scale feature processing module corresponding to (c) significantly enhances the model's multi-scale feature extraction and fusion capabilities through synergistic effects in the encoder and decoder. The encoder extracts feature information at different levels through multi-scale convolution, while the decoder further optimizes detail recovery through multi-scale feature fusion, resulting in stronger robustness and higher segmentation accuracy when processing complex medical images. This improvement not only verifies... Figure 1The effectiveness of the multi-scale feature processing module corresponding to (c) in the overall network also provides a more powerful solution for medical image segmentation tasks.

[0095] Qualitative analysis and visual effects can further verify this. Figure 1 The effectiveness of the multi-scale feature processing module corresponding to (c). Figure 10 Qualitative results showed that the replaced network was able to identify the outline of laryngeal tumors more clearly, especially in areas with indistinct boundaries or complex structures. Figure 1 The introduction of the multi-scale feature processing module corresponding to (c) significantly improves the clarity of the segmentation results. With Figure 1 The multi-scale feature processing module corresponding to (c) gradually replaces the traditional modules in U-Net, making the boundaries of the tumor region more complete and the localization more accurate, especially in areas with blurred boundaries. Figure 1 The multi-scale feature processing module corresponding to (c) helps the model better capture detailed information through its multi-scale feature extraction and fusion capabilities, thereby improving the accuracy and clarity of segmentation.

[0096] The quantitative comparison results in Table 2 are consistent with Figure 10 The qualitative visual comparison results are consistent, further illustrating... Figure 1 The introduction of the multi-scale feature processing module corresponding to (c) effectively improves the network's segmentation performance in laryngeal cancer tumor image segmentation. Through multi-scale feature extraction, Figure 1 The multi-scale feature processing module corresponding to (c) not only enhances the model's ability to capture details, but also improves the clarity and accuracy of the segmentation results.

[0097] 3.2.2 Validity of DEB This paper uses U-Net as the backbone network and designs four control networks by replacing new modules in the network to verify the segmentation effectiveness of the DEB module. The control networks are as follows: the original U-Net, U-Net with DEB added to the skip links (U-Net + DEB), MEDE-Net and MDEM-Net without DEB. By comparing the different networks, the effect of DEB on improving segmentation performance is analyzed in depth.

[0098] Table 3 Validation of DEB

[0099] The experimental results are shown in Table 3 and Figure 11As shown, U-Net with DEB exhibits a comprehensive improvement in segmentation accuracy compared to the original U-Net. The IoU index significantly increases from 0.702 in the original U-Net to 0.742, an increase of 5.7%; the Dice coefficient increases from 0.801 to 0.831, an improvement of 3.7%. The MAE error jumps from 0.790 to 0.835, while decreasing from 0.406 to 0.323, a reduction of 20.4%. This result demonstrates that the DEB module significantly improves the segmentation accuracy of U-Net by enhancing the model's focus on tumor region features, especially when processing complex images, where it can better capture detailed information and reduce errors.

[0100] For MEDE-Net without DEB integration, the model already demonstrated high segmentation performance for tumor images, with IoU and Dice coefficients reaching 0.767 and 0.844, respectively. Adding DEB improved the model's Dice coefficient and IoU by approximately 1%, indicating that even in high-performing models, DEB can still provide additional performance gains.

[0101] As can be seen from the qualitative experimental data in Table 3, the DEB module provides an effective feature enhancement mechanism. Whether in the basic network model or in complex networks with good segmentation performance, it can further optimize the model performance and improve the segmentation accuracy.

[0102] Visual comparison results of network segmentation performance are as follows: Figure 12 As shown in the figure, it is clear that the network with the added DEB module has a significant improvement in segmentation performance. Compared with the original U-Net, the U-Net model with integrated DEB shows stronger ability to handle complex and small regions, especially in the segmentation of small tumor boundaries, which is clearer and more accurate, effectively avoiding the problem of segmentation blur.

[0103] from Figure 12 As can be seen, for small tumors or regions with unclear boundaries, the introduction of DEB effectively improves the robustness and accuracy of the model, enabling it to accurately identify and delineate the outer contour of the tumor. Without DEB, the segmentation results often exhibit jagged boundaries or unnatural curves, affecting the reliability and visual appeal of the segmentation results. However, with the introduction of the DEB module, the boundaries of the segmentation results are smoother and more continuous, reducing boundary irregularities. This improved smoothness not only enhances the visual appeal of the segmentation results but also further improves their accuracy, making them more suitable for practical applications.

[0104] Experimental results show that the DEB module significantly improves the performance of tumor segmentation tasks. (See Table 3 and...) Figure 10As shown, this module has achieved significant improvements in several key indicators, mainly in the following three aspects: First, the localization accuracy of small tumors is significantly improved; second, the clarity of tumor edge contours is significantly improved; and finally, the overall segmentation results exhibit better smoothness and continuity. These improvements fully demonstrate the important application value of the DEB module in the field of medical image segmentation, especially in high-precision tumor detection and segmentation tasks. The introduction of this module not only improves segmentation results but also provides more reliable auxiliary evidence for clinicians' diagnostic decisions.

[0105] Therefore, based on the above experimental comparisons, it can be seen that the medical image segmentation model and method proposed in this application, by introducing a multi-scale feature processing module and a detail enhancement module (DEB), achieve the following effects: First, to address the shortcomings of the traditional U-Net model in feature extraction, a multi-scale feature processing module is proposed. This module extracts richer contextual information from feature maps at different levels through multi-scale feature extraction and fuses features from multiple levels, thereby significantly enhancing the network's ability to capture targets of different sizes and detailed information. This multi-scale feature extraction strategy enables the network to better identify the boundaries of tumor structures when processing complex pharyngeal tumor CT images, especially in areas with blurred edges and complex details.

[0106] Secondly, to address the issue of blurred edges in CT images of pharyngeal cancer, this application proposes a DEB module. The DEB module progressively corrects features in both channel and spatial dimensions, adaptively repairing and enhancing blurred edge information to help the network more accurately capture the boundary features of the target region. Through this progressive enhancement process, MEDE-Net can effectively learn missing boundary information, thus enabling more refined and accurate segmentation of the specific shape of the pharynx during image segmentation.

[0107] In the experimental section, a dataset of pharyngeal cancer CT images was used to evaluate the performance of MEDE-Net using multiple segmentation networks. Experimental results show that MEDE-Net outperforms several state-of-the-art segmentation methods in terms of segmentation accuracy and boundary detail restoration. Through the combined effect of multi-scale feature extraction and detail enhancement modules, MEDE-Net can better handle the complex features in pharyngeal CT images, thus achieving more accurate segmentation results. Ablation experiments further demonstrate the synergistic effect of the multi-scale feature processing module and the DEB module; the Dice coefficient of the complete model is improved by 7.5% compared to U-Net, showcasing the significant contribution of both modules to improving network performance.

[0108] In summary, the proposed MEDE-Net network structure can achieve multi-scale feature extraction and detail enhancement, making it suitable for complex scenarios in medical image segmentation.

[0109] Figure 13 This is a schematic diagram of the structure of a medical image segmentation device provided in one embodiment of this application. Figure 13 As shown, the medical image segmentation device includes an acquisition module 1301 and an image segmentation module 1302.

[0110] The acquisition module 1301 is used to acquire the target medical image; The image segmentation module 1302 is used to segment the target medical image according to a trained medical image segmentation model to obtain an image segmentation result, wherein the medical image segmentation model is the medical image segmentation model described in any of the above embodiments.

[0111] The medical image segmentation device provided in this application embodiment can be referred to the above method embodiment for its specific implementation process. The implementation principle and technical effect are similar, and will not be repeated here.

[0112] Figure 14 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. The electronic device may be a server, such as... Figure 14 As shown, the electronic device includes a processor 1401 and a memory 1402.

[0113] The memory 1402 stores computer-executed instructions.

[0114] The processor 1401 executes the computer execution instructions stored in the memory 1402, causing the processor 1401 to perform the method described in any of the above embodiments.

[0115] The electronic device provided in this application embodiment can be referred to the above method embodiment for its specific implementation process. The implementation principle and technical effect are similar, and will not be repeated here.

[0116] In the above Figure 14In the illustrated embodiments, it should be understood that the processor can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), etc. The general-purpose processor can be a microprocessor or any conventional processor. The steps of the method disclosed in this invention can be directly implemented by a hardware processor, or implemented by a combination of hardware and software modules within the processor.

[0117] This application also provides a computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, implement the method shown in the above-described method embodiments.

[0118] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of this application.

Claims

1. A medical image segmentation model, characterized by, The application relates to a coding and decoding system, comprising: an encoder and a decoder, the number of coding layers contained in the encoder is consistent with the number of decoding layers contained in the decoder, and the decoding layers correspond to the coding layers one by one, the coding layers are connected with other coding layers of adjacent levels and the decoding layers corresponding to the coding layers one by one, and the decoding layers are further connected with other decoding layers of adjacent levels; the coding layers are used for feature extraction on input data corresponding to the coding layers, and the coding layers send corresponding output results to the next coding layer and the corresponding decoding layer after obtaining the output results, the input data corresponding to the coding layers is a target medical image input or an output result corresponding to the previous coding layer; the decoding layers are used for feature fusion on input data corresponding to the decoding layers, and the decoding layers send output results corresponding to the decoding layers to the next decoding layer or obtain an image segmentation result of the target medical image according to the output results corresponding to the decoding layers after obtaining the output results, the input data corresponding to the decoding layers comprises output results sent by the previous decoding layer and output results sent by the corresponding decoding layer; wherein the coding layers and / or the decoding layers adopt a multi-scale feature processing module, the multi-scale feature processing module comprises a plurality of convolution layers of different scales connected in parallel; when the coding layers comprise the multi-scale feature processing module, the input data of the multi-scale feature processing module and the input data of the plurality of convolution layers of different scales connected in parallel are consistent with the input data of the coding layers, the output result of the multi-scale feature processing module is consistent with the output result of the coding layers, and the output result of the multi-scale feature processing module is obtained by splicing the outputs of the plurality of convolution layers of different scales corresponding to the coding layers connected in parallel; when the decoding layers comprise the multi-scale feature processing module, the input data of the multi-scale feature processing module and the input data of the plurality of convolution layers of different scales connected in parallel are consistent with the input data of the decoding layers, the output result of the multi-scale feature processing module is consistent with the output result of the decoding layers, and the output result of the multi-scale feature processing module is obtained by splicing the outputs of the plurality of convolution layers of different scales corresponding to the decoding layers connected in parallel.

2. The medical image segmentation model of claim 1, wherein, the plurality of convolution layers of different scales connected in parallel are a first convolution layer of a first scale, a second convolution layer of a second scale and a third convolution layer of a third scale, wherein the first scale is smaller than the second scale, and the second scale is smaller than the third scale.

3. The medical image segmentation model of claim 2, wherein, the second convolution layer is replaced by the first convolution layer and a fourth convolution layer of a first scale arranged in series; the third convolution layer is replaced by the first convolution layer, the fourth convolution layer and a fifth convolution layer of a first scale arranged in series; The input data of the first convolutional layer is consistent with the input data of the multi-scale feature processing module, the input data of the fourth convolutional layer is the input data of the first convolutional layer, and the input data of the fifth convolutional layer is the input data of the fourth convolutional layer. The multi-scale feature processing module further comprises a sixth convolutional layer of a fourth scale, wherein the fourth scale is smaller than the first scale. The input data of the sixth convolutional layer is consistent with the input data of the first convolutional layer, and the output result of the multi-scale feature processing module is obtained by fusing the output result of the sixth convolutional layer and the output result of the plurality of convolutional layers of different scales in parallel.

4. The medical image segmentation model of claim 3, wherein, The multi-scale feature processing module further comprises a seventh convolutional layer of a fourth scale and a first identity mapping branch in parallel with the first convolutional layer, an eighth convolutional layer of a fourth scale and a second identity mapping branch in parallel with the fourth convolutional layer, and a ninth convolutional layer of a fourth scale and a third identity mapping branch in parallel with the fifth convolutional layer. The input data of the first convolutional layer is consistent with the input data of the multi-scale feature processing module, the input data of the fourth convolutional layer is the output result of the first convolutional layer, the output result of the seventh convolutional layer, and the output result of the first identity mapping branch, and the input data of the fifth convolutional layer is the output result of the fourth convolutional layer, the output result of the eighth convolutional layer, and the output result of the second identity mapping branch. The output result of the first convolutional layer, the output result of the seventh convolutional layer, and the output result of the first identity mapping branch are fused to obtain a first output result. The output result of the fourth convolutional layer, the output result of the eighth convolutional layer, and the output result of the second identity mapping branch are fused to obtain a second output result. The output result of the fifth convolutional layer, the output result of the ninth convolutional layer, and the output result of the third identity mapping branch are fused to obtain a third output result. The output result of the multi-scale feature processing module is obtained by fusing the output result of the first output result, the second output result, and the third output result after splicing, and the output result of the sixth convolutional layer.

5. The medical image segmentation model of any one of claims 1-4, wherein, Further comprising: A detail enhancement module arranged between the one-to-one corresponding encoding layer and the decoding layer; The detail enhancement module is configured to increase the attention degree of the target region in the output result of the corresponding encoding layer by the attention mechanism, obtain a corresponding output result, and send the corresponding output result to the corresponding decoding layer.

6. The medical image segmentation model of claim 5, wherein, The detail enhancement module comprises a splitting unit and a fusion unit, the fusion unit comprises a global feature extraction unit, a dimension attention channel, a spatial attention channel and a global feature map fusion unit, the input end of the splitting unit is connected with the encoding layer, the output end is connected with the input end of the global feature extraction unit, the output end of the global feature extraction unit is connected with the input end of the dimension attention channel and the input end of the spatial attention channel, the output end of the dimension attention channel and the output end of the spatial attention channel are connected with the input end of the global feature map fusion unit, and the output end of the global feature map fusion unit is connected with the decoding layer. The splitting unit is used for performing feature extraction on the output result of the encoding layer through a tenth convolutional layer of a first scale, an eleventh convolutional layer of a second scale and a twelfth convolutional layer of a third scale respectively, and obtaining feature maps of three different scales. The global feature extraction unit is used for extracting global information of the feature maps of the three different scales respectively, and obtaining global feature maps. The dimension attention channel is used for performing feature transformation on the global feature maps through global average pooling and a thirteenth convolutional layer of the first scale, obtaining a channel feature map, and obtaining the channel attention map according to the channel feature map and the feature maps of the three different scales. The spatial attention channel is used for performing feature transformation on the global feature maps through global average pooling, global maximum pooling and a fourteenth convolutional layer of the third scale, obtaining a spatial feature map, and obtaining the spatial attention map according to the spatial feature map and the feature maps of the three different scales. The global feature map fusion unit is used for obtaining the corresponding output result according to the channel attention map and the spatial attention map.

7. The medical image segmentation model of claim 6, wherein, The dimension attention channel obtains the channel attention map according to the channel feature map and the feature maps of the three different scales, comprising: The channel feature map is averagely split in the channel dimension to obtain 3 channel feature sub-maps with the same number of channels; The channel attention map is obtained according to the 3 channel feature sub-maps and the feature maps of the three different scales.

8. The medical image segmentation model of claim 3 or 4, wherein, The first scale is 3*3, the second scale is 5*5, the third scale is 7*7, and the fourth scale is 1*1.

9. A medical image segmentation method characterized by, Comprising: Obtaining a target medical image; Performing image segmentation on the target medical image according to the trained medical image segmentation model to obtain an image segmentation result, wherein the medical image segmentation model is the medical image segmentation model in any one of claims 1-8.

10. An electronic device, comprising: Comprising: A processor and a memory; The memory stores computer execution instructions; The processor executes the computer execution instructions stored in the memory, so that the processor executes the method in claim 9.

Images