Pathological cell segmentation method and device based on multi-scale Transformation double convolution

Abstract

The invention discloses a pathological cell segmentation method and device based on multi-scale Transformation double convolution, and relates to the technical field of image segmentation, and the method comprises the following steps: constructing a pathological cell segmentation model, and carrying out the training through a data set; utilizing the trained pathological cell segmentation model to realize pathological cell segmentation; the pathological cell segmentation model comprises a feature coding part, a middle connection part and a feature decoding part which are connected in sequence; the feature coding part sequentially performs multi-scale feature extraction on pathological cell images input into the pathological cell segmentation model by adopting dense double convolution operation; the intermediate connection layer receives the multi-scale features by using a plurality of jump connections, performs feature fusion by using LSTM coding, and outputs fused features; and the feature decoding part receives and decodes the fusion features, and outputs a pathological cell segmentation result. According to the method, the problem of low segmentation precision caused by factors such as fuzzy boundary, complex background and dense pathologic cell images can be solved, and the segmentation accuracy is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of image segmentation, and in particular to a method and device for pathological cell segmentation based on multi-scale transmission double convolution. Background Art

[0002] Pathological cell images provide an important basis for the diagnosis, prognosis and treatment of diseases. At present, pathology experts usually use stained pathological cell image diagnosis as the "gold standard" for cancer diagnosis. Among them, the morphological characteristics of pathological cell nuclei are one of the main bases for pathological diagnosis. In-depth understanding of the morphological changes of cell nuclei in pathological cell images is of great significance for diagnosis and identification. Therefore, the use of computers to automatically and accurately segment pathological cells not only helps medical experts to more accurately analyze the changes in cell nuclear morphology, but also provides a more reliable basis for the diagnosis, prognosis and treatment of cancer.

[0003] At present, medical cell segmentation methods can be mainly divided into two categories: traditional segmentation methods and deep learning-based segmentation methods. Traditional methods include threshold method, watershed method, region growing method, clustering method and morphological method. Although these methods are widely used in cell segmentation, they have some obvious limitations. First, traditional methods usually require cumbersome pre-processing and post-processing processes, which increases the complexity of the operation and the time cost of calculation. Secondly, the feature extraction process is more difficult, requiring manual design and selection of features, and its effect is often not as good as that of deep learning methods. In addition, these traditional methods often lack sufficient generalization ability when processing different types of cell images, and it is difficult to adapt to diverse and complex medical pathology image data.

[0004] The Chinese invention patent application "A U-Net Pathological Cell Segmentation Method and Device Based on Multi-scale Transformer Multi-attention" with publication number CN118379303A discloses a pathological cell segmentation method that introduces a dense attention module and a multi-scale Transformer attention module into a U-Net model. The dense attention module effectively reduces the loss of target information in the encoder downsampling process by improving the learning ability of target features, thereby achieving accurate distinction between background and target, while the multi-scale Transformer attention module further enhances the recognition ability of pathological cell boundaries by fully learning multi-scale features. However, this method also has some shortcomings: on the one hand, the superposition of the dense attention module and the multi-scale Transformer module increases the computational complexity of the model, limiting its efficiency in practical clinical applications; on the other hand, its segmentation accuracy still needs to be improved, and its generalization ability has not been fully verified. Summary of the invention

[0005] The purpose of the present invention is to solve the problems in the prior art.

[0006] The technical solution adopted by the present invention to solve the technical problem is: to provide a pathological cell segmentation method based on multi-scale transmission double convolution, comprising the following steps:

[0007] Preprocess the pathological cell images and construct the data set;

[0008] Build a pathological cell segmentation model and train it using the dataset;

[0009] Pathological cell segmentation is achieved using the trained pathological cell segmentation model;

[0010] Among them, the pathological cell segmentation model includes a feature encoding part, an intermediate connection part and a feature decoding part which are connected in sequence; the feature encoding part includes a multi-layer encoder, which sequentially extracts multi-scale features from the pathological cell image input into the pathological cell segmentation model, and dense double convolution operation is used to process the input features during the multi-scale feature extraction of each layer of encoder; the intermediate connection layer uses multiple jump connections to respectively receive the multi-scale features output by the multi-layer encoder, and uses LSTM encoding to fuse the features and output multiple fused features; the feature decoding part includes a multi-layer decoder, which respectively receives and decodes the multiple fused features and finally outputs the pathological cell segmentation result; the number of encoders, jump connections and decoders is consistent.

[0011] Preferably, each layer of encoder of the feature encoding part is composed of a convolution layer, a pooling layer and a dense double convolution module connected in sequence; the input of the first layer of encoder is the pathological cell image input to the pathological cell segmentation model, and the input of the encoders of other layers is the output of the encoder of the previous layer;

[0012] The output of the i-th layer encoder is expressed as:

[0013] ;

[0014] ;

[0015] ;

[0016] in, represents the input of the i-th layer encoder, represents the output of the i-th layer encoder; represents the convolutional layer operation, Represents the output of the convolutional layer; represents the maximum pooling layer operation, represents the output of the maximum pooling layer; Represents a dense dual convolution module operation.

[0017] Preferably, the dense dual convolution module includes multiple layers of nonlinear transformation layers, and the output of the kth nonlinear transformation layer is expressed as:

[0018] ;

[0019] in, represents the input of the kth nonlinear transformation layer, ; When k>0, represents the output of the kth nonlinear transformation layer; [] represents feature map concatenation; represents the k-th layer of nonlinear transformation function, including 1×1 convolution, batch normalization, activation function ReLU and grouped cross convolution operations; the output of the last nonlinear transformation layer is the final output of the dense double convolution module;

[0020] The group cross convolution operation comprises the following steps:

[0021] The input feature map x is received and normalized, and then mapped to the channel through a 1x1 convolution to obtain the feature map X, which is expressed as:

[0022] ;

[0023] in, Representation layer normalization operation, Represents a 1×1 convolution operation;

[0024] Split the feature map X into two parts and , expressed as:

[0025] ;

[0026] in, Indicates a split operation;

[0027] Will Split into two sub-parts and , respectively apply multi-scale large kernel convolution operations, and combine them with the corresponding weighted convolution operations to generate the weighted feature map b, expressed as:

[0028] ;

[0029] ;

[0030] ;

[0031] in, and Respectively and The result of applying a large kernel convolution operation; , and Represent 5×5, 7×7 and 9×9 convolution operations respectively; and Represents the weight coefficient, which weights the features accordingly; Represents a splicing operation;

[0032] The concatenated weighted feature map b is scaled by a learnable factor Scale it and perform residual connection with the input feature map x to obtain the final output feature map , expressed as:

[0033] .

[0034] Preferably, the intermediate connection part uses multiple jump connection layers to respectively receive the outputs of multiple encoder layers of the feature encoding part, and performs feature aggregation through the channel information transmission module to generate fused features and transmit them to the feature decoding part.

[0035] Preferably, the channel information transmission module implements the process of feature aggregation and feature fusion, including the following steps:

[0036] Receive image embedding and receive the feature map output by each layer encoder as input image , generate the corresponding image patch embedding;

[0037] LSTM encoding uses multiple LSTM modules to receive the image block embedding of each layer of encoder and update the state. Each LSTM module obtains the hidden state of the LSTM module at the current moment;

[0038] Feature fusion: Each LSTM module fuses the hidden states from the initial moment to the current moment to obtain the fused features.

[0039] Preferably, the input image embedding comprises the following steps:

[0040] For the input image Perform convolution operation to generate embedded feature E, expressed as:

[0041] ;

[0042] Add location embed , expressed as:

[0043] ;

[0044] in, represents the feature embedding of the i-th image patch of the input image, is the position embedding of the ith image patch, is a convolution operation.

[0045] Preferably, the LSTM encoding comprises the following steps:

[0046] Each LSTM module uses LSTM cell to calculate the current hidden state of the LSTM module. The specific process is as follows:

[0047] ;

[0048] ;

[0049] ;

[0050] ;

[0051] ;

[0052] ;

[0053] in, is the hidden state at the current moment, represents the hidden state at the previous moment, is the forget gate output, is the input gate output at the current moment, is the candidate cell state, is the cell state at the current moment, Indicates the cell state at the previous moment. is the output gate output; Both represent weight matrices, Both represent bias terms, is the Sigmoid activation function, is the hyperbolic tangent activation function.

[0054] Preferably, the feature fusion is expressed as:

[0055] ;

[0056] in, is the final output.

[0057] The present invention also provides a pathological cell segmentation device based on multi-scale transmission double convolution, comprising:

[0058] The data set construction module pre-processes the pathological cell images and constructs the data set;

[0059] Model training module, builds a pathological cell segmentation model and uses the data set for training;

[0060] Image segmentation module, which uses the trained pathological cell segmentation model to achieve pathological cell segmentation;

[0061] Among them, the pathological cell segmentation model includes a feature encoding part, an intermediate connection part and a feature decoding part which are connected in sequence; the feature encoding part includes a multi-layer encoder, which sequentially extracts multi-scale features from the pathological cell image input into the pathological cell segmentation model, and dense double convolution operation is used to process the input features during the multi-scale feature extraction of each layer of encoder; the intermediate connection layer uses multiple jump connections to respectively receive the multi-scale features output by the multi-layer encoder, and uses LSTM encoding to fuse the features and output multiple fused features; the feature decoding part includes a multi-layer decoder, which respectively receives and decodes the multiple fused features and finally outputs the pathological cell segmentation result; the number of encoders, jump connections and decoders is consistent.

[0062] The present invention has the following beneficial effects: by introducing a dense dual convolution module and a channel information transmission module, the pathological cell segmentation model significantly enhances the learning ability of the target feature information and effectively improves the segmentation accuracy. The dense dual convolution module reduces the loss of target information during encoder downsampling through multi-layer nonlinear transformation and feature fusion, thereby accurately distinguishing the background from the target; the channel information transmission module further enhances the capture of detailed information on the pathological cell boundaries through multi-scale feature aggregation. Therefore, the present invention can effectively solve the problem of low pathological cell segmentation accuracy caused by factors such as blurred boundaries, complex and dense image backgrounds in pathological cell images, enhance the feature information of pathological cells, improve the pathological cell segmentation accuracy, and achieve accurate segmentation of pathological cells.

[0063] The present invention is further described in detail below in conjunction with the accompanying drawings and embodiments, but the present invention is not limited to the embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0064] Figure 1 A method step diagram of an embodiment of the present invention;

[0065] Figure 2 is a detailed flow chart of an embodiment of the present invention;

[0066] Figure 3 is a model structure diagram of an embodiment of the present invention;

[0067] Figure 4 This is a structural diagram of a dense dual convolution module according to an embodiment of the present invention;

[0068] Figure 5This is a structural diagram of a channel information transmission module according to an embodiment of the present invention;

[0069] Figure 6 This is a comparison chart of the experimental results of the embodiment of the present invention and other segmentation algorithms;

[0070] Figure 7 FIG. 4 is a diagram showing the structure of a device according to an embodiment of the present invention. DETAILED DESCRIPTION

[0071] See also Figure 1 and Figure 2 The method step diagram and detailed flow chart of the embodiment of the present invention are shown, which include the following steps:

[0072] S101, preprocessing the pathological cell images and constructing a data set;

[0073] S102, construct a pathological cell segmentation model and use the dataset for training;

[0074] S103, implementing pathological cell segmentation using the trained pathological cell segmentation model;

[0075] Among them, the pathological cell segmentation model includes a feature encoding part, an intermediate connection part and a feature decoding part which are connected in sequence; the feature encoding part includes a multi-layer encoder, which sequentially extracts multi-scale features from the pathological cell image input into the pathological cell segmentation model, and dense double convolution operation is used to process the input features during the multi-scale feature extraction of each layer of encoder; the intermediate connection layer uses multiple jump connections to respectively receive the multi-scale features output by the multi-layer encoder, and uses LSTM encoding to fuse the features and output multiple fused features; the feature decoding part includes a multi-layer decoder, which respectively receives and decodes the multiple fused features and finally outputs the pathological cell segmentation result; the number of encoders, jump connections and decoders is consistent.

[0076] Specifically, the data set includes a test set and a training set. The training set is used to train the pathological cell segmentation model after data enhancement. The test set is input into the trained pathological cell segmentation model to obtain the segmentation prediction result. The pathological cell segmentation model adopts a network with a dense dual convolution module and a channel information transmission module added.

[0077] Specifically, since there is less pathological cell image data, data enhancement and image size normalization are used to increase the diversity of training samples, thereby improving the generalization ability and robustness of the model and ensuring that the model can fully learn the features in the data set. The data enhancement includes the following methods:

[0078] Contrast enhancement: adjust the grayscale value in the pathological cell image through grayscale mapping to make the grayscale distribution of the overall image more uniform;

[0079] Horizontal flip: perform 180° mirror flip on the pathological cell image;

[0080] Flip up and down, and flip the pathological cell image 180° vertically;

[0081] Image rotation: rotate the pathological cell image by 0 to 180 degrees, and fill the missing parts with zeros after rotation;

[0082] Image translation: Shift the pathological cell image up and down, left and right, and fill the missing parts with zeros after translation;

[0083] Add Gaussian noise to each pathological cell image in the training set. The probability density function of Gaussian noise is expressed as:

[0084] ;

[0085] in, is the standard deviation of z, is the mean.

[0086] The image size normalization specifically involves uniformly adjusting the original size of the original pathological image (eg, 1000×1000) to an input size of 224×224 to meet the training requirements of the model.

[0087] The U-net network can be divided into three parts: feature encoding, feature decoding, and the connection between the two. The feature encoding part is mainly composed of a convolution layer and a pooling layer connected in series. The convolution layer uses a 3×3 convolution kernel, and the pooling layer uses the maximum pooling method to participate in the calculation. The convolution layer and the pooling layer both use the linear correction function ReLu as the activation function. The feature decoding part completes pixel filling alternately through upsampling layers and jump connections. The embodiment of the present invention improves the encoder and jump connection parts of the U-Net network, introduces a dense double convolution mechanism in the encoder part, captures the target features, and introduces channel information Transmission in the jump connection stage to extract features at different scales, and then aggregates them. For details, see Figure 3 As shown, this embodiment adopts a 4-layer encoder and a 4-layer decoder.

[0088] Specifically, the feature encoding part uses several layers of encoders to extract features of different scales from the input image, and the output of the i-th layer encoder is expressed as:

[0089] ;

[0090] ;

[0091] ;

[0092] in, represents the input of the i-th layer encoder, Represents the output of the i-th layer encoder, when i=1, represents the input image; , , Respectively represent convolution operation, maximum pooling operation and dense double convolution operation. In this embodiment, .

[0093] Specifically, the dense double convolution operation is implemented by a dense double convolution module. Figure 4 As shown, the dense double convolution module includes multiple layers of nonlinear transformation layers, and the output of the kth nonlinear transformation layer is expressed as:

[0094] ;

[0095] in, represents the input of the k-th nonlinear transformation layer, [*] represents feature map concatenation; , when k>0, Represents the output of the kth nonlinear transformation layer; represents the kth layer of nonlinear transformation function, and the output of the last layer of nonlinear transformation layer is the final output of the dense double convolution module; in this embodiment, three layers of nonlinear transformation layers are used, namely .

[0096] Nonlinear transformation function It includes 1×1 convolution, batch normalization, activation function ReLU and group cross convolution operation. The group cross convolution operation includes the following steps: first, the input feature map x is normalized, and then it is mapped to the channel through a 1x1 convolution to obtain the feature map X; then, the feature map X is split into two parts a and x a , where a is further split into two sub-parts a1 and a2. For a1 and a2, multi-scale large kernel convolution operations (LKA5 and LKA7) are applied respectively to capture information of different scales through convolution and dilated convolution, and combined with the corresponding weighted convolution operations (X5 and X7) to generate a weighted feature map b. Then, the concatenated weighted feature map b is element-wise multiplied with the original feature map x, and mapped back to the original number of channels through 1x1 convolution. Finally, the output feature map is scaled by a learnable scaling factor scale and residually connected with the input feature map x to obtain the final output feature map, thereby effectively capturing multi-scale information in the image and enhancing the feature representation capability. The specific process is as follows:

[0097] ;

[0098] ;

[0099] ;

[0100] ;

[0101] ;

[0102] ;

[0103] in, , C is the number of channels, H and W are the height and width respectively, and out represents the final output result of the grouped cross convolution operation. Representation layer normalization operation, Indicates a split operation. Indicates an n×n convolution operation. and It means that features of different scales are captured through different convolution operations.

[0104] For details, see Figure 5 As shown in the figure, the intermediate connection part uses the channel information Transmission module to realize feature aggregation and feature fusion, which includes the following four steps: input image embedding, LSTM encoding and multi-scale fusion.

[0105] Specifically, the input image embedding includes the following steps: receiving the feature map output by each layer encoder (that is, the feature map ) as the input image and cut it into multiple image blocks, generate the embedded feature E through convolution operation, and add the position embedding Pos i The specific process is as follows:

[0106] ;

[0107] ;

[0108] Among them, W is the convolution kernel of the convolution operation, * represents the convolution operation, is the feature of the image block, is the position embedding of the image patch.

[0109] Specifically, the LSTM encoding includes the following steps: each LSTM cell is based on the current input And the hidden state h at the previous moment t−1 and cell state c t−1 Update the current hidden state. The specific process is as follows:

[0110] ;

[0111] ;

[0112] ;

[0113] ;

[0114] ;

[0115] ;

[0116] in: is the hidden state at the current moment. Is the cell state at the current moment. is the weight matrix, is the bias term. is the Sigmoid activation function, is the hyperbolic tangent activation function.

[0117] Specifically, the feature fusion is expressed as:

[0118] ;

[0119] in, is the final output.

[0120] An experimental comparative analysis was conducted on the embodiment of the present invention. The images in the validation set were segmented using the embodiment of the present invention and the prior art, respectively. The intersection over union (IoU) and Dice value were used as evaluation indicators. The calculation formula for the evaluation indicators is as follows:

[0121] ;

[0122] ;

[0123] Among them, true positives (TP) represent that the actual image is a cell nucleus, and the segmentation result predicts that it is also a cell nucleus. False positives (FP) represent that the actual image is background, and the segmentation result predicts that it is a cell nucleus. True negatives (TN) represent that the actual image is background, and the segmentation result predicts that it is also background. False negatives (FN) represent that the actual image is a cell nucleus, and the segmentation result predicts that it is background. IoU is the ratio of the intersection and union of the true value and the predicted value. When the value is 1, it means that the two are exactly the same; Dice represents the similarity between the true value and the predicted value. The larger the value, the higher the similarity between the two.

[0124] See the segmentation effect diagram for details. Figure 6 As shown, from left to right are the original image (Original Image), the ground truth (Ground Truth), the result graphs generated by the U-Net algorithm, the AttNet algorithm, the Swin-Unet algorithm, the UCtransnet algorithm, the U-Net++ algorithm, the U-Net+++ algorithm and the algorithm of the embodiment of the present invention. The first to third rows represent the segmentation results under the MoNuSeg dataset, the fourth to sixth rows represent the segmentation results under the GlaS dataset, and the seventh to eighth rows represent the segmentation results under the CoNSeP dataset.

[0125] As shown in Table 1, the segmentation results of seven segmentation algorithms, namely, the U-Net algorithm, the AttNet algorithm, the Swin-Unet algorithm, the UCtransnet algorithm, the U-Net++ algorithm, the U-Net+++ algorithm and the algorithm of the embodiment of the present invention, are shown in turn.

[0126] Table 1- Segmentation results of seven segmentation algorithms:

[0127]

[0128] As can be seen from Table 1, the example method of the present invention shows significant advantages in the pathological cell segmentation task. On the MoNuSeg dataset, the Dice value of the example of the present invention is 79.54%, and the IOU value is 66.34%, which is about 3%-4% higher than U-Net (Dice of U-Net is 76.45%, IOU is 62.86%), and compared with U-Net++ (Dice is 76.72%, IOU is 63.19%), it is 2.82% and 3.15% higher. Compared with AttNet (Dice is 75.67%, IOU is 62.47%), the improvement is 3%-4%. On the GlaS dataset, the Dice value of the example of the present invention is 92.12%, and the IOU value is 85.97%, which is 2.67% and 3.19% higher than U-Net (Dice is 89.45%, IOU is 82.78%), and 3.14% and 4.61% higher than Swin-Unet (Dice is 88.98%, IOU is 81.36%), far exceeding other methods, especially in terms of accuracy and reliability. In addition, on the CoNSeP dataset, the Dice value of the example of the present invention is 79.76%, and the IOU value is 66.47%, which is 1.5% and 2.01% higher than U-Net (Dice is 78.26%, IOU is 64.46%), and 1.61% and 2.14% higher than AttNet (Dice is 78.15%, IOU is 64.33%), showing its consistency and stability in different datasets. In general, the examples of the present invention have shown significant advantages over traditional and convolutional neural network models such as U-Net, AttNet, and Swin-Unet on all three datasets, especially on the GlaS dataset, with higher accuracy and stronger generalization ability, showing its superiority in cell segmentation tasks. The experimental results show that the method of the present invention has obvious advantages in pathological cell segmentation.

[0129] As shown in Table 2, the computational complexity and scale of seven segmentation algorithms, including the U-Net algorithm, the AttNet algorithm, the UCtransnet algorithm, the U-Net++ algorithm, the U-Net+++ algorithm, a U-Net pathological cell segmentation method and device based on multi-scale Transformer multi-attention, and the algorithm of the embodiment of the present invention, are shown in turn.

[0130] Table 2 Computational complexity and scale of seven segmentation algorithms:

[0131]

[0132] The example of the present invention shows significant advantages in terms of computational complexity and model scale. Although its parameter volume of 66.49M is relatively large, compared with other similar methods, such as the U-Net method based on multi-scale Transformer multi-attention, which has a parameter volume of 106.67M, the example of the present invention significantly reduces the consumption of computing resources through optimized design, reduces memory usage to 66.49M, significantly improves resource utilization, and is more adaptable to real-time needs in clinical environments. Therefore, the example of the present invention is not only competitive in segmentation accuracy, but can also effectively adapt to application scenarios with limited resources, and is particularly suitable for real-time pathological cell segmentation tasks that require higher computing resources, and has strong practicality and promotion value.

[0133] See also Figure 7 FIG. 1 is a diagram showing a structure of a device according to an embodiment of the present invention, comprising:

[0134] A data set construction module 701 pre-processes the pathological cell images to construct a data set;

[0135] A model training module 702 constructs a pathological cell segmentation model and performs training using a data set;

[0136] The image segmentation module 703 implements pathological cell segmentation using the trained pathological cell segmentation model.

[0137] It can be seen that the pathological cell segmentation method and device based on multi-scale transmission double convolution proposed in the present invention can improve the segmentation accuracy of the model and solve the problem of low pathological cell segmentation accuracy caused by factors such as blurred boundaries, complex and dense image backgrounds in pathological cell images.

[0138] The above are only preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A pathological cell segmentation method based on multi-scale transmission double convolution, characterized in that: The following steps are involved: Preprocess the pathological cell images and construct the data set; Build a pathological cell segmentation model and train it using the dataset; Pathological cell segmentation is achieved using the trained pathological cell segmentation model; Among them, the pathological cell segmentation model includes a feature encoding part, an intermediate connection part and a feature decoding part which are connected in sequence; the feature encoding part includes a multi-layer encoder, which sequentially extracts multi-scale features from the pathological cell image input into the pathological cell segmentation model, and dense double convolution operation is used to process the input features during the multi-scale feature extraction of each layer of encoder; the intermediate connection layer uses multiple jump connections to respectively receive the multi-scale features output by the multi-layer encoder, and uses LSTM encoding to fuse the features and output multiple fused features; the feature decoding part includes a multi-layer decoder, which respectively receives and decodes the multiple fused features and finally outputs the pathological cell segmentation result; the number of encoders, jump connections and decoders is consistent.

2. The pathological cell segmentation method based on multi-scale transmission double convolution according to claim 1, characterized in that: Each layer of encoders in the feature encoding part is composed of a convolutional layer, a pooling layer and a dense double convolution module connected in sequence; the input of the first layer of encoders is the pathological cell image input to the pathological cell segmentation model, and the input of the encoders in other layers is the output of the encoder in the previous layer; The output of the i-th layer encoder is expressed as: ； ； ； in, represents the input of the i-th layer encoder, represents the output of the i-th layer encoder; represents the convolutional layer operation, Represents the output of the convolutional layer; represents the maximum pooling layer operation, represents the output of the maximum pooling layer; Represents a dense dual convolution module operation.

3. The pathological cell segmentation method based on multi-scale transmission double convolution according to claim 2 is characterized in that: The dense dual convolution module includes multiple layers of nonlinear transformation layers, and the output of the kth nonlinear transformation layer is expressed as: ； in, represents the input of the kth nonlinear transformation layer, ; When k>0, represents the output of the kth nonlinear transformation layer; [] represents feature map concatenation; represents the k-th layer of nonlinear transformation function, including 1×1 convolution, batch normalization, activation function ReLU and grouped cross convolution operations; the output of the last nonlinear transformation layer is the final output of the dense double convolution module; The group cross convolution operation comprises the following steps: The input feature map x is received and normalized, and then mapped to the channel through a 1x1 convolution to obtain the feature map X, which is expressed as: ； in, Representation layer normalization operation, Represents a 1×1 convolution operation; Split the feature map X into two parts and , expressed as: ； in, Indicates a split operation; Will Split into two sub-parts and , respectively apply multi-scale large kernel convolution operations, and combine them with the corresponding weighted convolution operations to generate the weighted feature map b, expressed as: ； ； ； in, and Respectively and The result of applying a large kernel convolution operation; , and Represent 5×5, 7×7 and 9×9 convolution operations respectively; and Represents the weight coefficient, which weights the features accordingly; Represents a splicing operation; The concatenated weighted feature map b is scaled by a learnable factor Scale it and perform residual connection with the input feature map x to obtain the final output feature map , expressed as: 。 4. The pathological cell segmentation method based on multi-scale transmission double convolution according to claim 1, characterized in that: The intermediate connection part uses multiple jump connection layers to receive the outputs of multiple encoder layers of the feature encoding part respectively, and aggregates features through the channel information transmission module to generate fused features and pass them to the feature decoding part.

5. The pathological cell segmentation method based on multi-scale transmission double convolution according to claim 1, characterized in that: The channel information transmission module implements the process of feature aggregation and feature fusion, including the following steps: Receive image embedding and receive the feature map output by each layer encoder as input image , generate the corresponding image patch embedding; LSTM encoding uses multiple LSTM modules to receive the image block embedding of each layer of encoder and update the state. Each LSTM module obtains the hidden state of the LSTM module at the current moment; Feature fusion: Each LSTM module fuses the hidden states from the initial moment to the current moment to obtain the fused features.

6. The pathological cell segmentation method based on multi-scale transmission double convolution according to claim 5, characterized in that: The input image embedding comprises the following steps: For the input image Perform convolution operation to generate embedded feature E, expressed as: ； Add location embed , expressed as: ； in, represents the feature embedding of the i-th image patch of the input image, is the position embedding of the ith image patch, is a convolution operation.

7. The pathological cell segmentation method based on multi-scale transmission double convolution according to claim 5, characterized in that: The LSTM encoding includes the following steps: Each LSTM module uses LSTM cell to calculate the current hidden state of the LSTM module. The specific process is as follows: ； ； ； ； ； ； in, is the hidden state at the current moment, represents the hidden state at the previous moment, is the forget gate output, is the input gate output at the current moment, is the candidate cell state, is the cell state at the current moment, Indicates the cell state at the previous moment. is the output gate output; Both represent weight matrices, Both represent bias terms, is the Sigmoid activation function, is the hyperbolic tangent activation function.

8. The pathological cell segmentation method based on multi-scale transmission double convolution according to claim 5, characterized in that: The feature fusion is expressed as: ； in, is the final output.

9. A pathological cell segmentation device based on multi-scale transmission double convolution, characterized in that: include: The data set construction module pre-processes the pathological cell images and constructs the data set; Model training module, builds a pathological cell segmentation model and uses the data set for training; Image segmentation module, which uses the trained pathological cell segmentation model to achieve pathological cell segmentation; Among them, the pathological cell segmentation model includes a feature encoding part, an intermediate connection part and a feature decoding part which are connected in sequence; the feature encoding part includes a multi-layer encoder, which sequentially extracts multi-scale features from the pathological cell image input into the pathological cell segmentation model, and dense double convolution operation is used to process the input features during the multi-scale feature extraction of each layer of encoder; the intermediate connection layer uses multiple jump connections to respectively receive the multi-scale features output by the multi-layer encoder, and uses LSTM encoding to fuse the features and output multiple fused features; the feature decoding part includes a multi-layer decoder, which respectively receives and decodes the multiple fused features and finally outputs the pathological cell segmentation result; the number of encoders, jump connections and decoders is consistent.

Images