Adenoid segmentation method using multi-scale feature fusion with deep supervision strategy

Abstract

The application discloses a multi-scale feature fusion gland segmentation method adopting a deep supervision strategy. The network proposed in the application adopts an encoder-decoder network structure, residual connection is added to adjacent convolutional parts in feature extraction and analysis, and a coordinate parallel attention mechanism and a dense hollow convolution block are used on a splicing path from an encoding layer to a decoding layer, so that noise in a pathological image is inhibited, the network is guided to focus on a salient region of the image, and multi-scale image features can be acquired. Meanwhile, a deep supervision strategy is adopted, the output of each decoding layer is converted into a prediction map, a combined loss function based on pixel classification, region and edge segmentation is adopted, loss is calculated with a label in advance, a shallow layer of the network is supervised and constrained, finally, prediction maps of different scales are fused, missing values are complemented, and the network achieves better segmentation effect, and the accuracy of gland segmentation is effectively improved.

Description

Technical Field

[0001] This invention relates to the fields of deep learning and image segmentation, specifically to a gland segmentation method employing a deep supervised strategy with multi-scale feature fusion. Background Technology

[0002] Deep learning-based medical image segmentation, as an emerging biomedical image processing technology, has made significant contributions to sustainable healthcare and has now become an important research direction in the field of computer vision. With the rapid development of deep learning, medical image processing based on deep neural networks has become a research hotspot.

[0003] Glands are important histological structures distributed throughout many tissues and organs. Adenocarcinoma, a malignant tumor originating from the salivary gland epithelium, varies in structure and is the most common form of cancer. Clinical morphological analysis of histopathological samples is an effective method for cancer diagnosis. Professional pathologists frequently use gland shape to assess the malignancy of adenocarcinoma. Accurate segmentation of glands is a crucial step in obtaining reliable morphological statistics. Therefore, precise segmentation of gland shape in histopathology is of great significance and urgently needed in clinical treatment. However, gland segmentation currently requires manual annotation by experienced pathologists, which is time-consuming, labor-intensive, costly, and subject to human subjectivity. Therefore, automated gland segmentation methods are of great importance.

[0004] Early automated segmentation methods for glands included threshold-based and morphological segmentation methods. However, these methods often suffered from low segmentation accuracy when dealing with the complex content of pathological slide images of glands.

[0005] In recent years, due to the advantages of deep learning in automatic feature learning and function simulation construction, it has rapidly become the main method for analyzing medical images, especially in the field of digital pathology.

[0006] However, the current deep learning-based glandular cell segmentation method still faces the following challenges: (1) Artifacts inevitably occur during the preparation of pathological slide samples of glandular cells, reducing image quality. Uneven staining of pathological slides leads to blurred boundaries, making it difficult to distinguish them from the background, which can easily result in inaccurate segmentation of glandular cell edges and missegmentation. (3) Glandular cells of different histologies vary in shape and size, and their density and scale vary greatly, which can easily lead to missing segmentation of glandular cells. Summary of the Invention

[0007] The first objective of this invention is to propose a gland segmentation method with multi-scale feature fusion using a deep supervision strategy, which solves the problems of large missing values ​​and missegmentation in gland cell segmentation, and ultimately improves the accuracy of gland cell segmentation based on deep learning.

[0008] A multi-scale feature fusion gland segmentation method employing a deep supervision strategy includes the following steps:

[0009] Step 1: Obtain digital pathological images of the glands;

[0010] Step 2: Preprocess the digital pathological images and labels of the glands as training data;

[0011] Step 3: Building and training the neural network;

[0012] 3-1. Construct a multi-scale feature fusion gland segmentation network using a deep supervision strategy.

[0013] The multi-scale feature fusion gland segmentation network employing a deep supervision strategy includes an encoder-decoder network structure;

[0014] The encoder comprises five cascaded residual convolutional blocks ResConvBlock L1-L5; the decoder comprises four cascaded residual convolutional blocks ResConvBlock L6-L9.

[0015] The residual convolutional blocks ResConvBlock L1-L9 have the same structure. They all perform residual processing on the convolutional operations in encoding and decoding, and connect the input parts of each layer to the output parts of each layer in a skip connection.

[0016] in:

[0017] The residual convolutional blocks ResConvBlock L1-L3 are connected to the residual convolutional blocks ResConvBlock L7-L9 through the coordinate parallel attention module CPA, respectively. The coordinate parallel attention is used to process the feature map output by the current coding layer to obtain a new feature map with attention weights.

[0018] The residual convolutional block ResConvBlock L4 is connected to the residual convolutional block ResConvBlock L6 using a densely dilated convolutional block DAC to perform multi-scale feature extraction on the feature map output by the current coding layer.

[0019] The residual convolutional block ResConvBlock L5 performs residual processing on the feature map output by the current coding layer and connects it to the residual convolutional block ResConvBlock L6;

[0020] The four outputs of the residual convolutional blocks ResConvBlock L6-L9 are converted into four segmentation prediction maps Sup1, Sup2, Sup3, and Sup4. The four segmentation prediction maps Sup1, Sup2, Sup3, and Sup4 are fused and their missing values ​​are complemented to obtain segmentation prediction map Sup5. The combined loss function is used to supervise and constrain the segmentation prediction maps Sup1, Sup2, Sup3, Sup4, and Sup5 to achieve backpropagation of model parameters and complete the update of model parameters.

[0021] 3-2. Model Training

[0022] The image data after preprocessing in step 2 is input into the multi-scale feature fusion gland segmentation network built in step 3-1 using a deep supervision strategy for training; the loss value of each segmentation prediction map and label is calculated by combining the loss function; the network iteratively updates the parameters through backpropagation of the loss value to continuously reduce the loss value until the preset number of iterations is reached;

[0023] Step 4: Use the multi-scale feature fusion gland segmentation network with deep supervision strategy built and trained in Step 3 to achieve gland segmentation.

[0024] As a preferred approach, the inputs to the residual convolutional blocks ResConvBlock L2-L5 are all feature maps downsampled by the previous residual convolutional block ResConvBlock through 2*2 max pooling.

[0025] Preferably, the residual convolutional blocks ResConvBlock L1-L9 have the same structure, each including the following in sequence:

[0026] 1) Kernel size = 3x3, stride = 1, using a convolution kernel with same padding;

[0027] 2) BatchNorm2d normalization layer;

[0028] 3) ReLU activation function;

[0029] 4) Kernel size = 3x3, stride = 1, using a convolution kernel with same padding;

[0030] 5) BatchNorm2d normalization layer;

[0031] 6) ReLU activation function;

[0032] 7) The original input is skipped to the output of two convolutions, which is achieved by adjusting the dimensions through a 1x1 convolution.

[0033] Preferably, the coordinate parallel attention module (CPA) performs the following calculation:

[0034] The input feature map is fed to the soft pooling branch and the average pooling branch, respectively. Then, in these two branches, pooling and channel compression operations are performed on the feature map along the width W and height H, respectively, to obtain a new feature vector F. softpool and F avgpool ;

[0035] The feature vector F softpool and F avgpool The feature vector F is obtained by adding and fusing the features. total ;

[0036] Along the spatial dimension, the feature vector F total Decomposed into two feature vectors F related to coordinate positions h and F w ;

[0037] Adjust the number of channels in the feature vectors related to the two coordinate positions mentioned above, activate them, and finally combine the input feature map (Input) with the weights in the height direction (Weight). h Weight in the width direction w Multiplying these together yields the feature map Output, which is weighted by attention.

[0038] Preferably, the dense dilated convolutional block includes four branches, each using dilated convolution to expand the model's receptive field. Finally, the feature map obtained after feature extraction through these four branches is added to the original input feature map; wherein:

[0039] The first branch includes a convolutional block with a kernel size of 3x3 and an inflation rate of 1. The receptive field of the first branch is 3.

[0040] The second branch consists of two convolutional blocks cascaded in sequence. The kernel size of the first convolutional block is 3x3 and its dilation rate is set to 3. The kernel size of the second convolutional block is 1x1 and its dilation rate is set to 1. The receptive field of the second branch is 7.

[0041] The third branch consists of three convolutional blocks cascaded in sequence. The kernel size of the first convolutional block is 3x3 and its dilation rate is set to 1. The kernel size of the second convolutional block is 3x3 and its dilation rate is set to 3. The kernel size of the third convolutional block is 1x1 and its dilation rate is set to 1. The receptive field of the third branch is 9.

[0042] The fourth branch consists of four convolutional blocks cascaded in sequence. The first convolutional block has a kernel size of 3x3 and an inflation rate of 1. The second convolutional block has a kernel size of 3x3 and an inflation rate of 3. The third convolutional block has a kernel size of 3x3 and an inflation rate of 5. The fourth convolutional block has a kernel size of 1x1 and an inflation rate of 1. The receptive field size of the fourth branch is 19.

[0043] Preferably, the combined loss function Pred is used to calculate the loss value of each segmentation prediction map and label, as shown in equation (1):

[0044]

[0045] in Sup represents the segmentation prediction map. k Binary cross-entropy loss with labels; Sup represents the segmentation prediction map. k Loss of the active outline of the label;

[0046]

[0047] Where N represents the total number of pixels in the image. Sup represents the segmentation prediction map. k The real label corresponding to the nth pixel, Sup represents the segmentation prediction map. k The prediction result corresponding to the nth pixel;

[0048]

[0049] Segmentation is a pixel-by-pixel classification task, therefore the loss is calculated pixel by pixel:

[0050]

[0051] Where Length is used to constrain the edge contour, and Ω represents the image domain, that is, the set of pixels (i,j). This indicates the gradient, where x and y represent the horizontal and vertical directions, and ∈ represents the minima to prevent the square root from being zero. Sup represents the segmentation prediction map. k The actual label corresponding to the pixel (i,j) in the y-direction. Sup represents the segmentation prediction graph. k The prediction result corresponding to pixel (i,j) in the x direction;

[0052]

[0053] Where Region represents the region similarity loss between P and Y, used to improve the segmentation of missing target regions; Sup represents the segmentation prediction map. k The actual label corresponding to pixel (i,j). Sup represents the segmentation prediction map. k The prediction result corresponding to pixel (i,j);

[0054] Total network loss total The sum of the losses calculated for each segmentation prediction map and the label:

[0055]

[0056] The second objective of this invention is to provide a gland segmentation device, characterized by including a trained multi-scale feature fusion gland segmentation network employing a deep supervision strategy.

[0057] A third objective of this invention is to provide a computer-readable storage medium having a computer program stored thereon, which, when executed in a computer, causes the computer to perform the method described thereon.

[0058] A fourth objective of this invention is to provide a computing device, including a memory and a processor, wherein the memory stores executable code, and the processor executes the executable code to implement the method described herein.

[0059] The beneficial effects of this invention are as follows:

[0060] 1) This invention addresses noise and artifacts in gland images by employing a coordinate parallel attention mechanism to suppress noise and highlight salient areas, thereby guiding the network to better focus on key pixel regions.

[0061] 2) This invention addresses the issues of narrow gaps between adjacent glands, different cell sizes and densities in gland images, by employing densely dilated convolutional blocks to extract features at multiple scales, which can effectively detect targets with large morphological differences.

[0062] 3) This invention addresses the problem of missing values ​​in cell segmentation caused by differences in the shape, size, and density of different glands. It fuses the prediction outputs at different scales to complement each other's missing values. Furthermore, it employs a deep supervision strategy to supervise and constrain the prediction outputs at different scales using a combined loss function, which to some extent solves the problem of missing values ​​in some regions during gland segmentation. Attached Figure Description

[0063] Figure 1 This is a structural architecture diagram of the gland segmentation network of the present invention.

[0064] Figure 2 This is a structural diagram of the ResConvBlock residual convolution block in this invention.

[0065] Figure 3 This is a structural diagram of the coordinate parallel attention mechanism (CPA) in this invention.

[0066] Figure 4 This is a model structure diagram of dense dilated convolution in this invention.

[0067] Figure 5 This is a diagram showing the effect of gland segmentation on the Glas dataset in this invention.

[0068] Figure 6 This is a diagram showing the effect of gland segmentation on the Crag dataset in this invention. Detailed Implementation

[0069] The present invention will be further analyzed below with reference to specific embodiments.

[0070] A gland segmentation network with multi-scale fusion features employing a deep supervision strategy includes the following:

[0071] Step 1: Conduct experiments on two public datasets, namely the gland datasets Glas and Crag.

[0072] Step 2: Read the digital pathological images of glands from the Glas training set, normalize them by dividing by 255, and perform data augmentation operations such as random rotation to obtain the preprocessed image x. origin Its label;

[0073] Step 3, as follows Figure 1 A multi-scale feature fusion gland segmentation network employing a deep supervision strategy was constructed.

[0074] 3-1 The preprocessed gland image X origin The input is fed into the network. First, the feature map is obtained through feature extraction processing using the L1 layer residual convolutional block ResConvBlock 1. Then, a downsampling operation is performed using 2x2 max pooling to obtain the input feature map of the L2 layer. This is used as the input to the next encoding layer, L2. Layers L2, L3, L4, and L5 all perform further feature extraction operations based on the 2*2 max pooling downsampling of the feature map output from the previous layer. These feature extraction operations are all performed by residual convolutional blocks (ResConvBlock). This completes the feature extraction operations along the encoding path.

[0075] like Figure 2The residual convolutional blocks ResConvBlocks in the encoding path are set sequentially as follows:

[0076] 1) ResConvBlock 1 contains the following in sequence:

[0077] A 3x3 convolutional block with a stride of 1 and 32 filters;

[0078] A 3x3 convolutional block with a stride of 1 and 32 filters;

[0079] A 1x1 convolutional block with a stride of 1 and 32 filters;

[0080] 2) ResConvBlock 2 contains the following in sequence:

[0081] 3x3 convolutional blocks with a stride of 1 and 64 filters;

[0082] 3x3 convolutional blocks with a stride of 1 and 64 filters;

[0083] A 1x1 convolutional block with a stride of 1 and 64 filters;

[0084] 3) ResConvBlock 3 contains the following in sequence:

[0085] A 3x3 convolutional block with a stride of 1 and 128 filters;

[0086] A 3x3 convolutional block with a stride of 1 and 128 filters;

[0087] A 1x1 convolutional block with a stride of 1 and 128 filters;

[0088] 4) ResConvBlock 4 contains the following in sequence:

[0089] A 3x3 convolutional block with a stride of 1 and 256 filters;

[0090] A 3x3 convolutional block with a stride of 1 and 256 filters;

[0091] A 1x1 convolutional block with a stride of 1 and 256 filters;

[0092] 5) ResConvBlock 5 contains the following in sequence:

[0093] A 3x3 convolutional block with a stride of 1 and 256 filters;

[0094] A 3x3 convolutional block with a stride of 1 and 256 filters;

[0095] A 1x1 convolutional block with a stride of 1 and 256 filters.

[0096] Each of the above convolutional blocks consists of a convolutional layer, a BatchNorm layer, and a ReLU activation function.

[0097] 3-2 On the feature concatenation path between the encoding and decoding layers, taking the L1 layer as an example, the output of L1... Feature maps are obtained through a coordinate-parallel attention mechanism (CPA) that suppresses noise in the image, assigns greater weight to important pixel regions, and guides the network to focus on key regions. Similarly, the outputs of L2 and L3 layers and Obtained through coordinate parallel attention mechanism CPA and L4 layer output Feature maps are obtained after multi-scale feature extraction using a densely dilated convolutional block DAC with a large receptive field. The densely dilated convolutional block DAC has four branches, each with a different receptive field of 3, 7, 9, and 19, respectively. These operations prepare the features for concatenation between the output of the coding layer and the input of the decoding layer.

[0098] like Figure 3 The coordinate parallel attention mechanism (CPA) includes, in sequence:

[0099] 1) A one-dimensional soft pooling layer that performs pooling along the width direction of the feature map;

[0100] 2) A one-dimensional soft pooling layer that performs pooling along the feature map height direction;

[0101] 3) A one-dimensional average pooling layer that performs pooling along the width direction of the feature map;

[0102] 4) A one-dimensional average pooling layer that performs pooling along the feature map height direction;

[0103] 5) A 1*1 convolutional block with a stride of 1, which consists of a convolutional layer, a BatchNorm layer, and a ReLU activation function in sequence;

[0104] 6) A 1*1 convolutional block with a stride of 1, which consists of a convolutional layer, a BatchNorm layer, and a Sigmoid activation function.

[0105] like Figure 4 The densely dilated convolutional block DAC contains, in sequence:

[0106] 1) First branch: 3*3 convolutional block with stride of 1, filters = 256, and dilation rate of 1.

[0107] 2) Second branch: 3*3 convolutional block with stride of 1, filters = 256, and dilation rate of 1;

[0108] A 3x3 convolutional block with a stride of 1, 256 filters, and a dilation rate of 3.

[0109] 3) The third branch: a 3*3 convolutional block with a stride of 1, filters = 256, and an inflation rate of 1;

[0110] A 3x3 convolutional block with a stride of 1, 256 filters, and a dilation rate of 3;

[0111] A 1x1 convolutional block with a stride of 1, 256 filters, and a dilation rate of 1.

[0112] 4) Fourth branch: 3*3 convolutional block with stride of 1, filters = 256, and dilation rate of 1;

[0113] A 3x3 convolutional block with a stride of 1, 256 filters, and a dilation rate of 3;

[0114] A 3x3 convolutional block with a stride of 1, 256 filters, and a dilation rate of 5;

[0115] A 1x1 convolutional block with a stride of 1, 256 filters, and a dilation rate of 1.

[0116] Each of the above convolutional blocks consists of a convolutional layer and a ReLU activation function.

[0117] 3-3 Feature Map Upsampled by a factor of 2 using bilinear interpolation and then... The feature map is obtained by splicing.

[0118] 3-4 Decoding stage.

[0119] As input to the L6 layer, it is obtained through feature parsing of ResConvBlock 6. Feature map After bilinear interpolation and upsampling by a factor of 2, the number of channels is adjusted to 128 using a 3x3 convolution. The input to layer L7 is obtained by splicing the data.

[0120] Input of L7 layer After feature analysis of ResConvBlock 7, the following was obtained: Feature map After bilinear interpolation and upsampling by a factor of 2, the number of channels is adjusted to 128 using a 3x3 convolution. The input to the L8 layer is obtained by concatenation.

[0121] Input of L8 layer After feature analysis of ResConvBlock 8, the following was obtained: After bilinear interpolation and upsampling by a factor of 2, the number of channels is adjusted to 128 using a 3x3 convolution. The input to layer L9 is obtained by concatenation.

[0122] L9 layer input After feature analysis of ResConvBlock 9, the following was obtained:

[0123] The settings for ResConvBlock 6 are the same as those for ResConvBlock 4; the settings for ResConvBlock 7 are the same as those for ResConvBlock 3; the settings for ResConvBlock 8 are the same as those for ResConvBlock 2; and the settings for ResConvBlock 9 are the same as those for ResConvBlock 1.

[0124] 3-5 Prediction Phase.

[0125] L6 floor The output is converted to the predicted output by a 1x1 convolution. And Upsampled to the same size as the label That is, the segmentation prediction graph Sup1.

[0126] L7 floor The output is converted to the predicted output by a 1x1 convolution. And Upsampled to the same size as the label That is, the segmentation prediction graph Sup2.

[0127] L8 floor The output is converted to the predicted output by a 1x1 convolution. And Upsampled to the same size as the label That is, the segmentation prediction graph Sup3.

[0128] L9 floor The output is converted to the predicted output by a 1x1 convolution. That is, the segmentation prediction graph Sup4 because of the prediction output The size is already the same as the label size, so there is no need to upsample.

[0129] Ultimately The fusion and splicing operation is performed to obtain x. result_fusion That is, the segmentation prediction graph Sup5.

[0130] 3-6 employs a deep supervision strategy to calculate the loss for each predicted output, such as... Figure 1 As shown:

[0131] Combination loss function The loss value used to calculate the segmentation prediction map and the label is shown in Equation (1):

[0132]

[0133] in Sup represents the segmentation prediction map. k Binary cross-entropy loss with labels; Sup represents the segmentation prediction map. k Loss of the active outline of the label;

[0134]

[0135] Where N represents the total number of pixels in the image. Sup represents the segmentation prediction map. k The real label corresponding to the nth pixel, Sup represents the segmentation prediction map. k The prediction result corresponding to the nth pixel;

[0136]

[0137] Segmentation is a pixel-by-pixel classification task, therefore the loss is calculated pixel by pixel:

[0138]

[0139] Where Length is used to constrain the edge contour, and Ω represents the image domain, that is, the set of pixels (i,j). This indicates the gradient, where x and y represent the horizontal and vertical directions, and ∈ represents the minima to prevent the square root from being zero. Sup represents the segmentation prediction map. k The actual label corresponding to the pixel (i,j) in the y-direction. Sup represents the segmentation prediction graph. k The prediction result corresponding to pixel (i,j) in the x direction;

[0140]

[0141] Where Region represents the region similarity loss between P and Y, used to improve the segmentation of missing target regions; Sup represents the segmentation prediction map. k The actual label corresponding to pixel (i,j). Sup represents the segmentation prediction map. k The prediction result corresponding to pixel (i,j);

[0142] Total network loss total The sum of the losses calculated for each segmentation prediction map and the label:

[0143]

[0144] 3-7 The total loss calculated above total Perform backpropagation to update the parameters in the network, thus completing one training iteration. Repeat steps 3-1 to 3-7 for iterative updates. As the iterations proceed, the total loss function... total It will be continuously optimized until the preset number of iterations is reached, and then the trained weight model will be saved.

[0145] 3-8 Load the trained weight model into the network, and input the gland image x to be segmented into the network. input The network processes the data to obtain the final gland segmentation result.

[0146] To verify the effectiveness of this invention, the method was compared with various existing segmentation methods, such as U-Net, U-Net++, KiU-Net, MultiResUNet, and UT-Net. Experiments were conducted using the Glas and Crag datasets for training and testing. On the Glas dataset, the training set consisted of 85 gland images, used to train the network, while the test set consisted of 20 gland images, used to test the network. On the Crag dataset, the training set consisted of 138 gland images, used to train the network, while the test set consisted of 40 gland images, used to test the network. The final evaluation metric was obtained from the testing on the test set.

[0147] To verify the effectiveness of the segmentation method proposed in this invention, F1 score, Intersection over Union (IOU), Hausdorff 95, and Accuracy were selected as evaluation indicators to measure the segmentation effect from multiple aspects. The calculation formulas for the four types of indicators are shown in Equations 7 and 8.

[0148]

[0149]

[0150] Where TP represents correctly predicted positive examples, TN represents correctly predicted negative examples, FP represents incorrectly predicted positive examples, FN represents incorrectly predicted negative examples, SR represents the output of the segmentation network, and GT represents the input label.

[0151] Hausdorff(SR,GT)=max(h(SR,GT),h(GT,SR)) (9)

[0152]

[0153]

[0154] Where ||·|| is the distance normal form between point set X and point set Y. Hausdorff95 is the value taken at the 95th percentile during the calculation of Hausdorff, in order to eliminate the influence of outlier bias on the calculation results.

[0155]

[0156] The experimental results are shown in Table 1. The network model of this invention achieves better segmentation results for glandular cells compared to existing segmentation methods. On the Glas dataset, compared to the classic U-Net, the F1 score and IOU are improved by 3.15% and 5.09%, respectively, the Hausdorff95 score is reduced by 17.2109, and the accuracy is improved by 3.49%. On the Crag dataset, compared to U-Net, the F1 score and IOU are improved by 3.61% and 6.01%, respectively, the Hausdorff95 score is reduced by 32.3601, and the accuracy is improved by 2.44%. Furthermore, the proposed method achieves higher evaluation metrics on both datasets than other methods. The segmentation results on the Glas dataset are shown in the figure below. Figure 5 As shown in the figure, the segmentation effect of the Crag dataset is as follows: Figure 6 As shown.

[0157] Table 1 Comparison of segmentation results of different models on the Glas and Crag datasets.

[0158]

Claims

1. A multi-scale feature fusion gland segmentation method using a deep supervision strategy, characterized by The method comprises the following steps: Step one, obtaining a digital pathology image of the gland; Step two, pre-processing the digital pathology image of the gland and the label as training data; Step three, neural network building and training; 3-1, building a multi-scale feature fusion gland segmentation network adopting a deep supervision strategy The multi-scale feature fusion gland segmentation network adopting the deep supervision strategy comprises an encoder-decoder network structure. The encoder comprises five residual convolution blocks ResConvBlock L1-L5 connected in series; and the decoder comprises four residual convolution blocks ResConvBlock L6-L9 connected in series. The residual convolution blocks ResConvBlock L1-L9 are of the same structure, and each is a residual processing of the convolution operation in the encoding and decoding, and a jump connection of the input part of each layer to the output part of each layer. Among them: The residual convolution blocks ResConvBlock L1-L3 are connected to the residual convolution blocks ResConvBlock L9-L7 through coordinate parallel attention modules CPA respectively, and use coordinate parallel attention to process the feature map output by the current encoding layer to obtain a new feature map with attention weight; The residual convolution block ResConvBlock L4 is connected to the residual convolution block ResConvBlock L6 using a dense atrous convolution block DAC, and performs multi-scale feature extraction on the feature map output by the current encoding layer; The residual convolution block ResConvBlock L5 performs residual processing on the feature map output by the current encoding layer and is connected to the residual convolution block ResConvBlock L6; The four outputs of the residual convolution block ResConvBlock L6-L9 are converted into four segmentation prediction maps 、 、 、 The four segmentation prediction maps are fused and complementary missing values are obtained 、 、 、 The segmentation prediction map is obtained by fusing and complementary missing values ; and the segmentation prediction map is supervised and constrained by using a combined loss function 、 、 、 、 The model parameters are updated by back propagation. 3-2, model training The image data pre-processed in step two is input into the multi-scale feature fusion gland segmentation network adopting the deep supervision strategy built in step 3-1 for training; the loss value of each segmentation prediction map and the label is calculated through a combined loss function; the network iteratively updates the parameters to continuously reduce the loss value through back propagation until a preset number of iterations is reached; Step four, using the multi-scale feature fusion gland segmentation network adopting the deep supervision strategy built and trained in step three to realize gland segmentation.

2. The method of claim 1, wherein The input of the residual convolution blocks ResConvBlock L2-L5 is the feature map of the 2x2 maximum pooling downsampling operation of the previous residual convolution block ResConvBlock.

3. The method of claim 1, wherein The residual convolution blocks ResConvBlock L1-L9 are of the same structure and each comprises in turn: 1) a convolution kernel with kernel size=3x3, stride=1, and same padding; 2) a BatchNorm2d normalization layer; 3) a Relu activation function; 4) a convolution kernel with kernel size=3x3, stride=1, and same padding; 5) a BatchNorm2d normalization layer; 6) a Relu activation function; 7) a jump connection of the original input to the output of the two times of convolution, and a dimension adjustment through a 1x1 convolution.

4. The method of claim 1, wherein The specific calculation process of the coordinate parallel attention module CPA is as follows: The input feature map is transmitted to a soft pooling branch and an average value pooling branch respectively, and then the feature map is subjected to pooling aggregation and channel compression operation along the width W and height H directions in the two branches respectively to obtain a new feature vector and ; The feature vector and are added to obtain a feature vector ; along the spatial dimension decomposed into two coordinate position dependent feature vectors and ; The channel number of the feature vector related to the above two coordinate positions is adjusted, and is activated, and finally the input feature map Input is multiplied by the weight in the height direction , the weight in the width direction to obtain the feature map Output with attention weight.

5. The method of claim 1 wherein The dense hollow convolution block includes 4 branches, and a hollow convolution is used in each branch to expand the receptive field of the model, and finally the feature maps obtained after feature extraction through the 4 branches are added to the original input feature maps; wherein: The first branch includes one convolution block, the convolution kernel size of the convolution block is 3x3, the dilation rate is set to 1, and the receptive field of the first branch is 3; The second branch includes two convolution blocks connected in sequence, the convolution kernel size of the first convolution block is 3x3, the dilation rate is set to 3, the convolution kernel size of the second convolution block is 1x1, the dilation rate is set to 1, and the receptive field of the second branch is 7; The third branch includes three convolution blocks connected in sequence, the convolution kernel size of the first convolution block is 3x3, the dilation rate is set to 1, the convolution kernel size of the second convolution block is 3x3, the dilation rate is set to 3, the convolution kernel size of the third convolution block is 1x1, the dilation rate is set to 1, and the receptive field of the third branch is 9; The fourth branch includes four convolution blocks connected in sequence, the convolution kernel size of the first convolution block is 3x3, the dilation rate is set to 1, the convolution kernel size of the second convolution block is 3x3, the dilation rate is set to 3, the convolution kernel size of the third convolution block is 3x3, the dilation rate is set to 5, the convolution kernel size of the fourth convolution block is 1x1, the dilation rate is set to 1, and the receptive field of the fourth branch is 19.

6. The method of claim 1, wherein The combined loss function For calculating the loss value of each segmentation prediction map and label, see formula (1): （1） wherein denotes the segmentation prediction map binary cross-entropy loss with the label; denotes the segmentation prediction map active contour loss with the label; （2） wherein denotes the total number of image pixels, denotes a segmentation prediction map the true label corresponding to the nth pixel of the segmentation prediction map, denotes a segmentation prediction map the predicted result corresponding to the nth pixel of the segmentation prediction map; （3） Segmentation is a pixel-by-pixel classification task, so the loss is calculated pixel by pixel: （4） Length is used to constrain the edge contour. This represents the image domain, that is, the set of pixels (i,j). This indicates calculating the gradient. , Indicates horizontal and vertical directions. It is a minimal term that prevents the square root from being zero. Represents segmentation prediction map The actual label corresponding to the pixel (i,j) in the y-direction. This represents the segmentation prediction graph. The prediction result corresponding to the pixel (i,j) in the x direction; （5） wherein Region represents a region similarity loss between P and Y, used to improve the missing of the segmented target region; represents a pixel point (i, j) of a segmentation prediction map corresponding to a real label, represents a prediction result of a pixel point (i, j) of a segmentation prediction map . Total loss of the network The sum of the losses computed for each split prediction map and label: （6）。 7. An apparatus for gland segmentation based on a method according to any one of claims 1 to 6, characterized in that The computer readable storage medium has stored thereon a computer program, when the computer program is executed in a computer, the computer program causes the computer to execute the method of any one of claims 1-6.

9. A computing device comprising a memory and a processor, the memory having stored therein executable code, the processor executing the executable code to implement the method of any one of claims 1-6. ​

Images