A pathological slice image segmentation detection method with multi-level lesion detection optimization

Abstract

The present invention discloses a pathological slice image segmentation detection method for multi-level lesion detection optimization. The method constructs a multi-level feature extraction structure, uses a lightweight network to perform preliminary image feature extraction, performs multi-scale feature extraction based on a hollow space pyramid pooling module, and superimposes and fuses the multi-scale features with the shallow feature map in the feature preliminary extraction network, thereby achieving sufficient lesion feature extraction, improving the algorithm's recognition accuracy for lesions, and reducing the number of network layers, thereby avoiding feature information loss caused by an overly deep network structure. Based on a multi-layer attention mechanism, lesion features are weighted, thereby improving the algorithm's ability to detect lesions. A trained multi-component segmentation algorithm module based on a Sobel operator is added in parallel after the preliminary feature extraction network to assist in optimizing the loss function. Through multi-level lesion feature extraction, enhancement, and optimization training, the lesion detection effect of the detection algorithm can be effectively improved.

Description

Technical Field

[0001] The present invention belongs to the field of medical image processing and relates to a pathological slice image segmentation detection method, and specifically relates to a pathological slice image lesion detection and segmentation detection method based on a multi-level feature extraction structure, a multi-layer attention mechanism module and an optimized loss function. Background Art

[0002] In the field of clinical medical lesion detection, the commonly used detection methods in clinical practice mainly include B-ultrasound examination, X-ray examination and biopsy pathological diagnosis, but the most effective lesion detection method is still biopsy pathological diagnosis. Biopsy pathological diagnosis is to remove some epithelial cells or tissues from the inspection site, fix them, stain them and make them into pathological sections, and then the pathologist will perform lesion detection and diagnosis. Compared with other detection methods, the results of biopsy pathological diagnosis are more accurate and more reliable, which can be used as the basis for doctors to formulate treatment plans. However, in this diagnostic process, the diagnostic process is cumbersome, time-consuming and labor-intensive, and the diagnostic results basically rely on the diagnostic experience and subjective judgment of the pathologist, and the diagnostic results are easily affected by external factors. In addition, the number of slice samples required for biopsy diagnosis is large, which leads to a very large diagnostic workload for pathologists, which is not conducive to the timely development of treatment. Now with the development of artificial intelligence technology, methods such as deep learning have also been widely used in medical research, especially in the field of medical image processing, and medical image lesion detection algorithms based on deep neural networks have received more and more attention. The feature distribution of medical pathology images is often very complex, and the images contain a lot of feature information, which has a great impact on the image lesion detection and segmentation tasks. The current medical pathology image lesion detection algorithms based on deep learning networks all extract image features through a convolutional layer structure, and then upsample the feature map to restore it to the original image size. Among them, the more convolutional layers there are, the greater the algorithm's ability to extract feature information from the image. However, too many network convolutional layers will lead to irreversible loss of image features, which will greatly reduce the network's lesion recognition accuracy. Therefore, it is difficult to obtain accurate lesion detection and segmentation results by using only a single convolutional neural network module as the feature extraction structure of the lesion detection algorithm.

[0003] With the research on feature extraction network structure, many lightweight neural networks have been launched year by year. Among them, the MobileNetV3 network uses a lightweight convolution module composed of deep convolution, which not only reduces the number of model parameters and computation, but also ensures that the model accuracy does not decrease. In addition, with the introduction of the DeepLab series of neural networks, the dilated space pyramid pooling module used by it as a multi-scale feature extraction structure has achieved good feature extraction effects on medical images. The dilated space pyramid pooling module uses parallel dilated convolution modules as its unique multi-scale feature extraction module. By setting different dilated convolution expansion rates, a larger receptive field can be obtained without increasing the number of network convolution layers, avoiding the problem of irreversible loss of feature information, making the network structure simpler and more efficient. In image edge feature detection, the Sobel algorithm can achieve better edge feature extraction effects. Based on the Sobel operator, it can calculate the grayscale approximation of the image brightness function.

[0004] The ECA (Efficient Channel Attention Module) channel attention mechanism module can calculate a corresponding weight value for each channel in the feature map through global average pooling or maximum pooling operations, and then learn the enhanced weight values ​​corresponding to each channel through adaptive calculation of the one-dimensional convolution of the convolution kernel and the Sigmoid function. The SAM (Spatial Attention Module) spatial attention mechanism module can obtain a feature map that has been optimized and enhanced with spatial information through global average pooling, global maximum pooling, two-dimensional convolution and Sigmoid function. By adding a multi-layer attention mechanism structure consisting of ECA modules and SAM modules to the network, the network can comprehensively and effectively enhance the features related to the lesion, so as to improve the network's ability to extract information about the features of the lesion area, and further enhance the image segmentation network's ability to detect lesions. Summary of the invention

[0005] In order to overcome the problem that the feature distribution of pathological slice images is complex, the feature information is redundant, and it is difficult to accurately segment the lesion area, the present invention provides a pathological slice image segmentation detection method with multi-level lesion detection optimization. The method constructs a multi-level feature extraction structure, first uses a lightweight network to perform preliminary extraction of image features, and then performs multi-scale feature extraction based on a hollow space pyramid pooling module, and superimposes and fuses the multi-scale features with the shallow feature map in the feature preliminary extraction network, so as to achieve sufficient lesion feature extraction, improve the algorithm's recognition accuracy of lesions, and reduce the number of network layers, avoiding the loss of feature information caused by an overly deep network structure. And based on a multi-layer attention mechanism, the lesion features are weighted, which improves the algorithm's ability to detect lesions. At the same time, a trained multi-component segmentation algorithm module based on a Sobel operator is added in parallel after the preliminary feature extraction network to assist in optimizing the loss function. Through multi-level lesion feature extraction, enhancement, and optimization training, the lesion detection effect of the detection algorithm can be effectively improved.

[0006] The objective of the present invention is achieved through the following technical solutions:

[0007] A multi-level lesion detection optimization pathological slice image segmentation detection method comprises the following steps:

[0008] Step 1: Load N pathological images I N , and each image is equipped with its corresponding segmentation label map L N , use the MobileNetV3 lightweight network to obtain the initial extraction feature map Extract using Sobel operator The region edge image H N , and based on H N right Perform edge enhancement to obtain edge enhancement feature map use Train a multi-component segmentation algorithm FCN based on the trained multi-component segmentation algorithm model Segmentation edge enhancement feature map The segmentation result graph W of the multi-component segmentation algorithm based on the Sobel operator is obtained N ;

[0009] Step 2: Based on the preliminary extraction of feature maps The multi-scale lesion feature extraction is performed using the hollow space pyramid pooling module structure to obtain the multi-scale feature map P N , and the shallow feature map I′ in the MobileNetV3 network N Downward skip connection, and multi-scale feature map P N Superimpose and fuse along the channel dimension to obtain the lesion feature map

[0010] Step 3: Based on the lesion feature map The lesion area features are enhanced by connecting the ECA channel attention mechanism module and the SAM spatial attention mechanism module in series. The enhanced feature map is upsampled by bilinear interpolation to obtain the segmentation result of the multi-level lesion detection algorithm based on the multi-layer attention mechanism. N ;

[0011] Step 4: Segmentation result graph W using the multi-component segmentation algorithm based on the Sobel operator N And the segmentation result of the multi-level lesion detection algorithm based on the multi-layer attention mechanism Figure Z N , improve the loss function calculation method, and train the improved pathological section image lesion detection algorithm model to obtain the pathological section image segmentation result map.

[0012] Compared with the prior art, the present invention has the following advantages:

[0013] 1) The present invention adopts the lightweight network MobileNetV3 as the preliminary feature extraction structure, which can effectively reduce the number of parameters and the amount of model calculation while ensuring that the feature extraction effect remains unchanged.

[0014] 2) The present invention adopts a dilated space pyramid pooling module as a deep feature extraction structure. Due to the complex regional feature distribution and high difficulty in accurate segmentation of pathological slice images, the image lesion detection and segmentation task needs to make full use of global and local feature information, and the dilated space pyramid pooling module uses a parallel dilated convolution module structure to extract features from feature maps at multiple scales, which can effectively utilize the multi-scale feature information in the image to improve the lesion recognition accuracy of the detection algorithm. At the same time, the dilated convolution module structure increases the receptive field during the convolution operation, reduces the number of network convolution layers, and can effectively avoid the problem of loss of detail feature information caused by too deep network layers.

[0015] 3) The present invention is based on a multi-layer attention mechanism to perform adaptive lesion area feature enhancement on the feature map. Due to the large amount of feature information in the pathological section image, there are many feature information in the extracted feature map that are not very useful for the lesion detection and segmentation task, which will reduce the lesion detection ability of the detection algorithm. The channel attention mechanism module can extract important feature information for each channel, while the spatial attention mechanism module can extract important feature information in space. Therefore, the multi-layer attention mechanism can comprehensively and effectively enhance the important feature information in the feature map and reduce the proportion of unimportant feature information, thereby enhancing the algorithm model's ability to detect lesions.

[0016] 4) The present invention simultaneously adopts two lesion detection modules: a multi-component segmentation algorithm based on the Sobel operator and a multi-level lesion detection algorithm based on a multi-layer attention mechanism, and optimizes the training loss function calculation formula of the network model, which can effectively improve the training effect of the network and improve the lesion detection and segmentation effect of the improved algorithm of the present invention.

[0017] 5) The present invention improves the accuracy of lesion recognition through a multi-level feature extraction structure, improves the lesion detection ability through a multi-layer attention mechanism module, and improves the loss function, which can overcome the problems of complex feature distribution and redundant feature information of pathological images, making it difficult to accurately detect lesions. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 The present invention is a flowchart of the implementation of the pathological slice image segmentation detection method optimized for multi-level lesion detection.

[0019] Figure 2 This is an example diagram of pathological slice image samples used in training the lesion detection algorithm of the present invention.

[0020] Figure 3 This is an example of the segmentation results of the multi-component segmentation algorithm based on the Sobel operator.

[0021] Figure 4 This is a structural diagram of the pathological section image lesion detection algorithm with multi-level lesion detection optimization adopted in the implementation of the present invention.

[0022] Figure 5 This is a schematic diagram comparing the segmentation results of the pathological image S2_N17_Q1 using the improved algorithm of the present invention and the traditional segmentation algorithm.

[0023] Figure 6 This is a schematic diagram comparing the segmentation results of the pathological image S2_N17_Q2 using the improved algorithm of the present invention and the traditional segmentation algorithm.

[0024] Figure 7 Schematic diagram comparing the segmentation results of the pathological image S2_N17_Q3 using the improved algorithm of the present invention and the traditional segmentation algorithm.

[0025] Figure 8 This is a schematic diagram comparing the segmentation results of the pathological image S2_N17_Q4 using the improved algorithm of the present invention and the traditional segmentation algorithm. DETAILED DESCRIPTION

[0026] The technical solution of the present invention is further described below in conjunction with the accompanying drawings, but is not limited thereto. Any modification or equivalent replacement of the technical solution of the present invention without departing from the spirit and scope of the technical solution of the present invention should be included in the protection scope of the present invention.

[0027] The present invention provides a pathological slice image segmentation detection method with multi-level lesion detection optimization. The method first loads a pathological slice image dataset, uses a MobileNetV3 lightweight network to preliminarily extract features, and uses a multi-component segmentation algorithm based on the Sobel operator to perform image segmentation detection. At the same time, the preliminarily extracted feature map is connected to a dilated space pyramid pooling module to perform multi-scale lesion feature extraction, and the multi-scale feature map is superimposed and fused with a shallow feature map from the MobileNetV3 network along the channel dimension. The superimposed feature map is adaptively enhanced in terms of lesion features based on a multi-layer attention mechanism module, thereby realizing a multi-level lesion detection algorithm based on a multi-layer attention mechanism. Based on the multi-component segmentation algorithm and the multi-level lesion detection algorithm, the loss function calculation method is improved to improve the lesion detection capability and recognition accuracy of the segmentation network, so that a high-precision pathological slice image lesion detection algorithm can be obtained. Figure 1 As shown, it is divided into four steps, the specific steps are as follows:

[0028] Step 1: Load N pathological images I N , and each image is equipped with its corresponding segmentation label map L N , use the MobileNetV3 lightweight network to obtain the initial extraction feature map Extract using Sobel operator The region edge image H N , and based on H N right Perform edge enhancement to obtain edge enhancement feature map use Train a multi-component segmentation algorithm FCN based on the trained multi-component segmentation algorithm model Segmentation edge enhancement feature map The segmentation result graph W of the multi-component segmentation algorithm based on the Sobel operator is obtained N The specific steps are as follows:

[0029] Step 1: Load all images used for pathological slice image lesion segmentation detection algorithm model training I N ={i1,i2,…,i N}, where i n Represents a pathological picture used for model training, that is, the set {i1,i2,…,i N}, n=1,2,…,N; at the same time, I N Each image in has a corresponding segmentation label map L N ={l1,l2,…l N}, where l n Respectively with i n Correspondingly, the number represented by n is in l nwith i n After the image is loaded, the MobileNetV3 lightweight network is used to perform preliminary image feature extraction to obtain a preliminary extracted feature map.

[0030] Step 1 and 2: Get the initial feature map Then, use the Sobel operator to extract The region edge image H N , as follows:

[0031]

[0032] Among them, x and y are the horizontal and vertical coordinates of the pixel points of the feature map respectively; based on the regional edge image H N , for the initial extraction of feature maps Perform edge enhancement to obtain edge enhancement feature map

[0033]

[0034] Step 13: Use Train a constructed multi-component segmentation algorithm FCN to obtain a trained multi-component segmentation algorithm model This trained Connect to the MobileNetV3 network, freeze the model parameters, and then enhance the input edge feature map Perform multi-component segmentation of an image:

[0035]

[0036] Among them, W N This is the segmentation result diagram of the multi-component segmentation algorithm based on the Sobel operator. The tissue area in the image is roughly detected. Indicates the use of a trained algorithm model For the input feature map Perform segmentation detection, the detection result is shown in Figure W N The size is I N Size.

[0037] Step 2: Based on the preliminary extraction of feature maps The multi-scale lesion feature extraction is performed using the hollow space pyramid pooling module structure to obtain the multi-scale feature map P N , and the shallow feature map I′ in the MobileNetV3 network N Downward skip connection, and multi-scale feature map P N Superimpose and fuse along the channel dimension to obtain the lesion feature map The specific steps are as follows:

[0038] Step 21: After the MobileNetV3 network, connect a dilated spatial pyramid pooling module in parallel to further extract the initial feature map Perform multi-scale feature extraction; connect five feature extraction modules in parallel in the dilated spatial pyramid pooling module, namely: a convolution module C, a global average pooling module A and three dilated convolution modules B1(u1), B2(u2), B3(u3), where u1, u2, u3 are the dilated convolution expansion rates of each module; extract the initial feature map Five feature extraction modules are input respectively, and the feature extraction results are:

[0039]

[0040] Based on the feature maps of different scales obtained from the above formula, the feature maps are superimposed along the channel dimension and the number of feature map channels is compressed to obtain a multi-scale feature map P N :

[0041]

[0042] Among them, Cat(…) means superimposing the feature maps along the channel dimension, and Y{…} means compressing the number of channels to X through convolution operation, avoiding feature redundancy caused by too many channels.

[0043] Step 22: Convert the shallow feature map I′ in the MobileNetV3 network N Take it out and compare it with the multi-scale feature map P N Superposition along the channel dimension; in order to avoid the shallow feature map I′ N The feature information in the convolution operation is lost too much, and the feature map with M channels after the deep convolution operation in the MobileNetV3 network is taken out as the shallow feature map I′ N At the same time, in order to avoid the superposition of shallow features after fusion, N The influence is too great, resulting in insufficient lesion feature extraction. N and P N Before superposition, the feature map I′ N The number of channels is compressed as follows:

[0044] M′=αX

[0045] Among them, M′ is the compressed shallow feature map I′ N The number of channels, α is a constant variable, ranging from 0 to 1; the shallow feature map I′ after channel compression N With multi-scale feature map P N The superposition fusion is performed as follows:

[0046]

[0047] in, It is the lesion feature map after the deep features and shallow features are fused. Up(P N ,I′ N ) indicates that P is upsampled by bilinear interpolation. N The feature map size increases to I′ N Size.

[0048] Step 3: Based on the lesion feature map The lesion area features are enhanced by connecting the ECA channel attention mechanism module and the SAM spatial attention mechanism module in series. The enhanced feature map is upsampled by bilinear interpolation to obtain the segmentation result of the multi-level lesion detection algorithm based on the multi-layer attention mechanism. N The specific steps are as follows:

[0049] Step 31: In the ECA module, first calculate the size of the convolution kernel k value of the ECA module's convolution operation. Use the number of channels to calculate the corresponding convolution kernel k value. The feature map with more channels uses a larger k value, and the feature map with fewer channels uses a smaller k value, so as to achieve full multi-channel fusion. First, use the lesion feature map The number of channels is calculated by parameter t:

[0050]

[0051] Among them, ch is the lesion feature map The number of channels, b and g are constant variables, both of which are positive integers. Then, the adaptive convolution kernel k value is calculated using parameter t:

[0052]

[0053] Among them, % represents the remainder operation.

[0054] Step 32: After the adaptive convolution kernel k value is determined, the ECA channel attention mechanism module and the SAM spatial attention mechanism module are used to analyze the lesion feature map. The feature map of the lesion area is enhanced, and the enhanced feature map is upsampled by bilinear interpolation:

[0055]

[0056] Among them, Z N The segmentation result diagram represents the multi-level lesion detection algorithm based on the multi-layer attention mechanism. Represents the lesion feature map based on the channel attention mechanism For feature enhancement, the size of the module’s adaptive convolution kernel is k, A SAM(…) indicates feature enhancement based on the spatial attention mechanism.

[0057] Step 4: Segmentation result graph W using the multi-component segmentation algorithm based on the Sobel operator N And the segmentation result of the multi-level lesion detection algorithm based on the multi-layer attention mechanism Figure Z N , improve the loss function calculation method, and train the improved pathological slice image lesion detection algorithm model to obtain the pathological slice image segmentation result map. The specific steps are as follows:

[0058] Step 41: Based on step 1, the segmentation result graph W of the multi-component segmentation algorithm based on the Sobel operator is obtained. N ; Based on steps 2 and 3, the segmentation result of the multi-level lesion detection algorithm based on the multi-layer attention mechanism is obtained. N ; To optimize the accuracy of the pathological section image lesion detection algorithm model, use W N and Z N Improve the algorithm loss function calculation method. The improved loss function calculation formula is as follows:

[0059]

[0060] Among them, M loss is the loss function of the improved algorithm proposed in this invention; L N is the standard segmentation label map; λ is a constant variable with a value range of 0 to 1; β and γ are category weight parameters used to adjust the segmentation accuracy of a certain category area; when β,γ>1, the number of false negatives in the category can be reduced, and when β,γ<1, the number of false positives in the category can be reduced; in order to reduce the missed diagnosis rate of segmentation detection in the lesion area, the β and γ weight settings of the corresponding category of the lesion area should be greater than 1.

[0061] Step 42: After setting the model structure of the pathological section image lesion detection algorithm proposed by the present invention and optimizing the loss function calculation formula, use the dataset image I N And the corresponding segmentation label map L N , training algorithm model, can realize high-precision pathological section image lesion detection algorithm model, and then obtain the pathological section image segmentation result map.

[0062] Example:

[0063] Execute step 1: Load 2875 pathology images I N , and each image is equipped with its corresponding segmentation label map L N , use the MobileNetV3 lightweight network to obtain the initial extraction feature map Extract using Sobel operator The region edge image H N, and based on H N right Perform edge enhancement to obtain edge enhancement feature map use Train a multi-component segmentation algorithm FCN based on the trained multi-component segmentation algorithm model Segmentation feature map The segmentation result graph W of the multi-component segmentation algorithm based on the Sobel operator is obtained N The specific steps are as follows:

[0064] Step 1: Load all images used for pathological slice image lesion segmentation detection algorithm model training I N ={i1,i2,…,i 2875}, where i n Represents a pathological picture used for model training, that is, the set {i1,i2,…,i 2875}, n=1,2,…,2875. At the same time, I N Each image in has a corresponding segmentation label map L N ={l1,l2,…l 2875}, where l n Respectively with i n Correspondingly, the number represented by n is in l n with i n The pathological images and labeling diagrams are the same. Figure 2 After the image is loaded, the MobileNetV3 network is used to perform preliminary image feature extraction to obtain the feature map

[0065] Step 1 and 2: Get the initial extracted feature map Then, use the Sobel operator to extract The region edge image H N , as follows:

[0066]

[0067] Where x and y are the horizontal and vertical coordinates of the pixel points in the feature map, respectively. N , for the feature map Perform edge enhancement to obtain edge enhancement feature map

[0068]

[0069] Step 13: Use Train a constructed multi-component segmentation algorithm FCN to obtain a trained multi-component segmentation algorithm model Evaluate the multi-component segmentation algorithm model in terms of dice coefficient, pixel accuracy, and weight file size The segmentation accuracy of ,the evaluation results are shown in Table 1.

[0070] Table 1 Segmentation accuracy of multi-component segmentation algorithm model

[0071]

[0072] This trained The model is connected after the MobileNetV3 network, and the model parameters are frozen, and then the input feature map is Perform multi-component segmentation of an image:

[0073]

[0074] Among them, W N This is the segmentation result of the multi-component segmentation algorithm based on the Sobel operator, which roughly detects the tissue area in the image. Indicates the use of a trained algorithm model For the input feature map Perform segmentation detection, the detection result is shown in Figure W N The size is I N The size of the image. The image segmentation result of the multi-component segmentation algorithm based on the Sobel operator is shown in Figure W N Examples include Figure 3 shown.

[0075] Execute step 2: Based on the preliminary extraction of feature maps The multi-scale lesion feature extraction is performed using the hollow space pyramid pooling module structure to obtain the multi-scale feature map P N , and the shallow feature map I′ in the MobileNetV3 network N Downward skip connection, and multi-scale feature map P N Superimpose and fuse along the channel dimension to obtain the lesion feature map The specific steps are as follows:

[0076] Step 21: After the MobileNetV3 network, connect a dilated spatial pyramid pooling module in parallel to further extract the initial feature map Multi-scale feature extraction is performed. Five feature extraction modules are connected in parallel in the dilated spatial pyramid pooling module, namely a convolution module C, a global average pooling module A, and three dilated convolution modules B1 (6), B2 (12), and B3 (18), where 6, 12, and 18 are dilated convolution expansion rate settings. Input 5 feature extraction modules respectively, and the feature extraction results are:

[0077]

[0078] Based on the feature maps of different scales obtained from the above formula, the feature maps are superimposed along the channel dimension and the number of feature map channels is compressed to obtain a multi-scale feature map P N :

[0079]

[0080] Among them, Cat(…) means superimposing the feature map along the channel dimension, and Y{…} means compressing the number of channels to 256 through convolution operation, avoiding feature redundancy caused by too many channels.

[0081] Step 22: Transform the shallow feature map I′ in the MobileNetV3 network N Take it out and compare it with the multi-scale feature map P N Overlay along the channel dimension.

[0082] In order to avoid the shallow feature map I′ N The feature information in the convolution operation is lost too much, so the feature map with 80 channels after the deep convolution operation in the MobileNetV3 network is taken out as the shallow feature map I′ N At the same time, in order to avoid the superposition of shallow features after fusion, N The influence is too great, resulting in insufficient lesion feature extraction. N and P N Before superposition, the feature map I′ N The number of channels is compressed as follows:

[0083] M′=αX

[0084] Among them, M′ is the compressed shallow feature map I′ N The number of channels. α is a constant variable, set α = 0.1875. The shallow feature map I′ after channel compression N With multi-scale feature map P N The superposition fusion is performed as follows:

[0085]

[0086] in, It is the lesion feature map after the deep features and shallow features are fused. Up(P N ,I′ N ) indicates that P is upsampled by bilinear interpolation. N The feature map size increases to I′ N Size.

[0087] Execute step 3: Based on the lesion feature map , the multi-layer attention mechanism module is used to enhance the features of the lesion area, and the enhanced feature map is upsampled by bilinear interpolation to obtain the segmentation result of the multi-level lesion detection algorithm based on the multi-layer attention mechanism. N The specific steps are as follows:

[0088] Step 31: In the multi-layer attention mechanism module, the lesion feature is enhanced by connecting the ECA channel attention mechanism module and the SAM spatial attention mechanism module in series. In the ECA module, first calculate the size of the convolution kernel k value of the ECA module's convolution operation. Use the number of channels to calculate the corresponding convolution kernel k value. The feature map with more channels uses a larger k value, and the feature map with fewer channels uses a smaller k value, so as to achieve full multi-channel fusion. First, use the lesion feature map The number of channels is calculated by parameter t:

[0089]

[0090] Among them, ch is the lesion feature map The number of channels, b and g are constant variables, set b = 1, g = 2. Then, use the parameter t to calculate the adaptive convolution kernel k value:

[0091]

[0092] Among them, % represents the remainder operation.

[0093] Step 32: After the adaptive convolution kernel k value is determined, the ECA channel attention mechanism module and the SAM spatial attention mechanism module are used to analyze the lesion feature map. The features of the lesion area are enhanced, and the enhanced feature map is upsampled by bilinear interpolation:

[0094]

[0095] Among them, Z N A diagram showing the segmentation results of a multi-level lesion detection algorithm based on a multi-layer attention mechanism. Represents the lesion feature map based on the channel attention mechanism For feature enhancement, the size of the module’s adaptive convolution kernel is k. SAM (…) indicates feature enhancement based on the spatial attention mechanism.

[0096] Execute step 4: Segmentation result image W of multi-component segmentation algorithm based on Sobel operator N And the segmentation result of the multi-level lesion detection algorithm based on the multi-layer attention mechanism Figure Z N , improve the loss function calculation method, and train the improved pathological section image lesion detection algorithm model. The specific steps are as follows:

[0097] Step 41: Based on the execution of step 1, the segmentation result graph W of the multi-component segmentation algorithm based on the Sobel operator algorithm is obtained. N Based on the execution of steps 2 and 3, the segmentation result of the multi-level lesion detection algorithm based on the multi-layer attention mechanism is obtained. N In order to optimize the accuracy of the pathological section image lesion detection algorithm model, W N and Z N Improve the algorithm loss function calculation method. The improved loss function calculation formula is as follows:

[0098]

[0099] Among them, M loss It is the loss function of the improved algorithm proposed in this invention. N is the standard segmentation label map. λ is a constant variable, set λ = 0.5. β and γ are class weight parameters, set β = 1.6, γ = 1.6. The structure diagram of the pathological slice image lesion detection algorithm for multi-level lesion detection optimization proposed by the present invention is as follows: Figure 4 shown.

[0100] Step 42: After setting the model structure of the pathological section image lesion detection algorithm proposed by the present invention and optimizing the loss function calculation formula, use the dataset image I N And the corresponding segmentation label map L N , training algorithm model. The schematic diagrams comparing the segmentation results of the improved algorithm proposed in the present invention and the traditional segmentation algorithm for pathological images S2_N17_Q1, S2_N17_Q2, S2_N17_Q3 and S2_N17_Q4 are shown in the figure. Figure 5 , Figure 6 , Figure 7 and Figure 8 The comparative evaluation results of the improved algorithm of the present invention and the traditional segmentation algorithm are shown in Table 2.

[0101] Table 2 Comparative evaluation results of the improved algorithm of the present invention and the traditional segmentation algorithm

[0102]

[0103] It can be found from Table 2 that the pathological section image lesion detection algorithm with multi-level lesion detection optimization proposed in the present invention is superior to the traditional image segmentation algorithm in terms of dice coefficient, pixel accuracy and weight file size. Therefore, through the above experimental verification, the following conclusion can be drawn: the pathological section image lesion detection algorithm with multi-level lesion detection optimization proposed in the present invention can effectively improve the lesion detection ability and recognition accuracy of the lesion detection algorithm when the pathological image feature distribution is complex and the feature information is redundant.

Claims

1. A multi-level lesion detection optimization pathological slice image segmentation detection method, characterized in that The method comprises the following steps: Step 1: Load N pathological images I N , and each image is equipped with its corresponding segmentation label map L N , use the MobileNetV3 lightweight network to obtain the initial extraction feature map Extract using Sobel operator The region edge image H N , and based on H N right Perform edge enhancement to obtain edge enhancement feature map use Train a multi-component segmentation algorithm FCN based on the trained multi-component segmentation algorithm model Segmentation edge enhancement feature map The segmentation result graph W of the multi-component segmentation algorithm based on the Sobel operator is obtained N ; Step 2: Based on the preliminary extraction of feature maps The multi-scale lesion feature extraction is performed using the hollow space pyramid pooling module structure to obtain the multi-scale feature map P N , and the shallow feature map I′ in the MobileNetV3 network N Downward skip connection, and multi-scale feature map P N Superimpose and fuse along the channel dimension to obtain the lesion feature map Step 3: Based on the lesion feature map The lesion area features are enhanced by connecting the ECA channel attention mechanism module and the SAM spatial attention mechanism module in series. The enhanced feature map is upsampled by bilinear interpolation to obtain the segmentation result of the multi-level lesion detection algorithm based on the multi-layer attention mechanism. N ; Step 4: Segmentation result graph W using the multi-component segmentation algorithm based on the Sobel operator N And the segmentation result of the multi-level lesion detection algorithm based on the multi-layer attention mechanism Figure Z N , improve the loss function calculation method, and train the improved pathological section image lesion detection algorithm model to obtain the pathological section image segmentation result map.

2. The multi-level lesion detection optimized pathological slice image segmentation detection method according to claim 1 is characterized in that The specific steps of step one are as follows: Step 1: Load all images used for pathological section image lesion segmentation detection algorithm model training I N ={i1,i2,…,i N }, where i n Represents a pathological picture used for model training, that is, the set {i1,i2,…,i N }, n=1,2,…,N; at the same time, I N Each image in has a corresponding segmentation label map L N ={l1,l2,…l N }, where l n Respectively with i n Correspondingly, the number represented by n is in l n with i n After the image is loaded, the MobileNetV3 lightweight network is used to perform preliminary image feature extraction to obtain a preliminary extracted feature map. Step 1 and 2: Get the initial feature map Then, use the Sobel operator to extract The region edge image H N , based on the region edge image H N , for the initial extraction of feature maps Perform edge enhancement to obtain edge enhancement feature map Step 13: Use Train a constructed multi-component segmentation algorithm FCN to obtain a trained multi-component segmentation algorithm model This trained Connect to the MobileNetV3 network, freeze the model parameters, and then enhance the input edge feature map Perform multi-component segmentation of the image and obtain the segmentation result graph W of the multi-component segmentation algorithm based on the Sobel operator N .

3. The multi-level lesion detection optimized pathological slice image segmentation detection method according to claim 2 is characterized in that The H N As shown below: Among them, x and y are the horizontal and vertical coordinates of the pixel points of the feature map respectively; The edge enhancement feature map As shown below: The W N As shown below: Among them, W N This is the segmentation result diagram of the multi-component segmentation algorithm based on the Sobel operator. Indicates the use of a trained algorithm model For the input feature map Perform segmentation detection.

4. The pathological slice image segmentation detection method for multi-level lesion detection optimization according to claim 1 is characterized in that The specific steps of step 2 are as follows: Step 21: After the MobileNetV3 network, connect a dilated spatial pyramid pooling module in parallel to further extract the initial feature map Perform multi-scale feature extraction; connect five feature extraction modules in parallel in the dilated spatial pyramid pooling module, namely: a convolution module C, a global average pooling module A and three dilated convolution modules B1(u1), B2(u2), B3(u3), where u1, u2, u3 are the dilated convolution expansion rates of each module; extract the initial feature map The five feature extraction modules are input respectively, and based on the obtained feature maps of different scales, the feature maps are superimposed along the channel dimension, and the number of feature map channels is compressed to obtain a multi-scale feature map P N ; Step 22: Convert the shallow feature map I′ in the MobileNetV3 network N Take it out and compare it with the multi-scale feature map P N Superimpose along the channel dimension; extract the feature map with the number of channels M after the deep convolution operation in the MobileNetV3 network as the shallow feature map I′ N ; At the same time, in I′ N and P N Before superposition, the feature map I′ N The number of channels is compressed; the shallow feature map I′ after channel compression N With multi-scale feature map P N Perform superposition fusion to obtain the lesion feature map 5. The multi-level lesion detection optimized pathological slice image segmentation detection method according to claim 4 is characterized in that The multi-scale feature map P N As shown below: Among them, Cat(…) means superimposing the feature map along the channel dimension, and Y{…} means compressing the number of channels to X through convolution operation; The compression formula is as follows: M′=αX Among them, M′ is the compressed shallow feature map I′ N The number of channels, α is a constant variable; The lesion feature map As shown below: in, It is the lesion feature map after the deep features and shallow features are fused. Up(P N ,I′ N ) indicates that P is upsampled by bilinear interpolation. N The feature map size increases to I′ N Size.

6. The multi-level lesion detection optimized pathological slice image segmentation detection method according to claim 1, characterized in that The specific steps of step three are as follows: Step 31: In the ECA module, calculate the size of the adaptive convolution kernel k value of the convolution operation of the ECA module; Step 32: After the adaptive convolution kernel k value is determined, the ECA channel attention mechanism module and the SAM spatial attention mechanism module are used to analyze the lesion feature map. The lesion area features are enhanced, and the enhanced feature map is upsampled by bilinear interpolation to obtain the segmentation result map Z of the multi-level lesion detection algorithm based on the multi-layer attention mechanism. N .

7. The pathological slice image segmentation detection method for multi-level lesion detection optimization according to claim 6 is characterized in that The calculation method of the adaptive convolution kernel k value of the convolution operation is as follows: First, use the lesion feature map The number of channels is calculated by parameter t: Among them, ch is the lesion characteristic map The number of channels, b and g are constant variables; Then, the adaptive convolution kernel k value is calculated using the parameter t: Among them, % represents the remainder operation.

8. The multi-level lesion detection optimized pathological slice image segmentation detection method according to claim 6, characterized in that The segmentation result of the multi-level lesion detection algorithm based on the multi-layer attention mechanism is shown in Figure Z N As shown below: Among them, Z N The segmentation result diagram represents the multi-level lesion detection algorithm based on the multi-layer attention mechanism. Represents the lesion feature map based on the channel attention mechanism For feature enhancement, the size of the module’s adaptive convolution kernel is k, A SAM (…) indicates feature enhancement based on the spatial attention mechanism.

9. The multi-level lesion detection optimized pathological slice image segmentation detection method according to claim 1, characterized in that The specific steps of step 4 are as follows: Step 41: To optimize the accuracy of the pathological slice image lesion detection algorithm model, use W N and Z N Improve the algorithm loss function calculation method. The improved loss function calculation formula is as follows: M loss =λloss(L N ,W N )+loss(L N ,Z N ), Among them, M loss is the loss function of the improved algorithm; L N is the standard segmentation label map; λ is a constant variable; β and γ are class weight parameters; Step 42: After setting up the improved pathological slice image lesion detection algorithm model structure and optimizing the loss function calculation formula, use the dataset image I N And the corresponding segmentation label map L N , train the algorithm model, realize the high-precision pathological section image lesion detection algorithm model, and then obtain the pathological section image segmentation result map.

Images