A pathological section image lesion grading and classification detection method

Abstract

The application discloses a pathological section image lesion grading and classification detection method, and comprises the following steps: step one, using an automatic label labeling method based on normalized mutual information registration to label the cropped image, and generating a data set; step two, improving a DenseNet121 network, and introducing a focal loss function to balance the weights between multiple tissue samples; step three, adding a hybrid attention module between each dense block and conversion layer in the improved DenseNet121, and adding a local supervision function to the loss function of the modified network; step four, combining the two improved network systems in steps two and three to form a lesion grading and classification detection scheme. The application can improve the detection speed and accuracy of lesion images, thereby reducing the cost of manual identification and expanding the application prospect of the medical image classification field.

Description

TECHNICAL FIELD

[0001] The present application belongs to the field of medical image processing, and relates to a medical pathological image classification detection method, in particular to a pathological section image lesion grading classification detection method based on structure optimization and local attention supervision. BACKGROUND

[0002] Pathological diagnosis is an indispensable part of cancer diagnosis and treatment, which has more objective and accurate judgment standard than ultrasound, CT, magnetic resonance and the like. Nowadays, the screening method for cancer tumor is to make pathological sections of patients, and then to scan the sections one by one by pathological experts with rich professional knowledge and experience, and to observe and judge whether the patient's tissue contains lesion tissue by naked eyes. This method is time-consuming, and the pathologist needs to observe each patient's section, while the number of population in China is huge, and the number of doctors with rich experience and professional knowledge is also very small, so it is not possible to guarantee enough time for the pathologist to read the section in the actual medical process, which is one of the reasons for the high missed diagnosis rate of cancer tumor screening in China. In addition, the early lesion characteristics of tumor are not obvious, and the naked eye cannot see the changes of the early lesion characteristics hidden in the tissue, and different experienced doctors have different opinions on the same case, and the patient has not been fundamentally treated in the early stage of virus latency, and has no ability to face the high mortality rate of cancer. Therefore, the current tumor screening method still has many shortcomings, and it is difficult to meet the needs of a large number of patients at present and in the future, and the development of pathology department is also limited. At present, the biggest difficulty of pathological screening based on histomorphology in China is the serious inequality between the huge patient population and the small number of professional doctors. According to the survey, a doctor needs to read more than 700 sections per day on average, and the observation time of the section with complex symptoms needs 10 minutes, and more time is needed for observation and conclusion by multiple professional doctors in the case of complex histological lesion typing, and the computer-aided detection tool needs to be developed urgently.

[0003] With the continuous development of artificial intelligence, artificial intelligence assisted medical treatment has become a general trend, and major medical industries around the world are actively engaged in intelligent diagnosis solutions. Nowadays, deep learning is used to classify medical pathological images, which can be seen everywhere, and more and more practices at home and abroad have proved that deep learning has broad application prospects in the field of image classification. However, when classifying medical pathological images, the pathological images in a window are often not single lesion areas, and are generally accompanied by other normal tissues, and these tissues are not needed for the actual task; in addition, the manual labeling method is difficult to accurately label the classification labels of the lesion area in the medical image, and the attention mechanism can learn the feature difference between the lesion tissue and the normal tissue in histomorphology to focus on the lesion tissue area, so as to achieve the classification of normal tissue and lesion tissue.

[0004] The DenseNet classification model has the reusability of channel features, and can well learn the subtle features in the pathological image. The network model draws on the ideas of ResNet network skip connection and GoogleNet Inception unit structure multi-scale perception, and further innovates the network structure, that is, instead of improving on the basis of deepening the network layer and the network width, starting from the feature map, using fewer parameters to utilize the feature map to the extreme, so as to achieve better classification effect. DenseNet connects all layers in series, so that the input of each layer not only comes from the output of the previous layer, but also comes from the output of all previous layers, which is the basic module of DenseNet network, also known as Dense Block. In order to reduce the complexity of the model, a transition layer (Transition layer) is added in the network. The layer is composed of BN layer, RELU activation function, 1*1 convolution kernel and 2*2 average pooling kernel, and is located between two adjacent Dense Blocks.

[0005] Normalized Mutual Information (NMI) is a measure in information theory for measuring the relationship between two random variables. The greater the mutual information, the greater the degree of dependence between the two random variables. In images, the greater the mutual information, the greater the correlation between the two images, so normalized mutual information can be used to match two images. SUMMARY

[0006] In order to overcome the difficulty and high cost of artificial recognition of different degrees of lesion characteristics, the application provides a pathological section image lesion grading classification detection method based on structure optimization and local attention supervision. The method uses a DenseNet121 model based on structure optimization to classify different tissues existing in a large number of pathological images, wherein the network structure of DenseNet121 is optimized, and a focal loss function is introduced to balance the weights between multiple tissues to improve the recognition rate of each tissue. Then, a DenseNet121 network based on local attention supervision is used to classify normal images and lesion images. By combining the two improved networks, the slice is first classified and the tissue is identified, and then the binary classification task of the lesion image is completed, forming a lesion grading classification detection method, improving the detection rate of the lesion image. Through repeated training of the improved network, the recognition rate of different tissues in the pathological image and the classification accuracy of the lesion image are improved, and the artificial recognition cost is reduced.

[0007] The purpose of the application is achieved by the following technical solutions:

[0008] A pathological section image lesion grading classification detection method, the basic idea of the method is to use different classification networks to realize the two tasks of section tissue recognition and lesion detection, improve the classification performance of the network by improving DenseNet121, and then combine the two classification networks to form a hierarchical classification detection scheme, improve the detection speed and accuracy of the lesion image, thereby reducing the cost of manual identification, expanding the application prospect of the medical image classification field, and specifically comprising the following steps:

[0009] Step one: the high-resolution complete section scanning image is regularly cut, the automatic label labeling method based on normalized mutual information registration is used to label the cut image, and a data set is generated;

[0010] Step two: taking DenseNet121 as a reference network, the DenseNet121 network is improved, the focal loss function is introduced to balance the weights between multiple tissue samples, and the multi-classification task training for tissue recognition is carried out;

[0011] Step three: a mixed attention module is added between each dense block and the conversion layer in the improved DenseNet121 in step two, the spatial and channel features of the pathological image are enhanced, and a local supervision function is added to the loss function of the modified network, which is used for detecting the overall classification performance and local enhancement effect of the network;

[0012] Step four: the improved network based on structure optimization in step two is used to learn the section tissue morphology, the improved network based on local attention supervision in step three is used to learn the lesion tissue characteristics, and the two improved network systems in steps two and three are combined to form a hierarchical classification detection scheme for lesions.

[0013] Compared with the prior art, the present application has the following advantages:

[0014] 1、The automatic label labeling method based on normalized mutual information registration is adopted in the present application, which avoids the complex process of manual identification and labeling of a single image, realizes the labeling of a large number of pathological images, automatically constructs a data set, reduces the cost of manually preparing a data set, and can realize diversified data set generation operation, which can meet the needs of different tasks.

[0015] 2、The DenseNet121 network is improved in the present application, the network structure in DenseNet121 is improved, and the mixed attention module is combined, which enhances the channel feature reuse and the attention degree of DenseNet121 to the image spatial features, improves the recognition rate of the pathological image lesion tissue fine features of DenseNet121, and thus improves the classification accuracy of the network.

[0016] 3、The application adds a local supervision loss function in DenseNet121, locally supervises the DenseNet121 after adding an attention mechanism, further enhances the excitation effect of the attention mechanism in the DenseNet121, makes the network have a higher classification accuracy on the features of the lesion tissue, thereby reduces the misdiagnosis rate and the missed diagnosis rate, and provides more accurate detection results for doctors.

[0017] 4、According to the tumor tissue development process, the application combines two classification network systems that are improved in different ways for the DenseNet121 network, applies them to different classification tasks of the same slice, designs a systematic hierarchical classification detection process, realizes the replacement of the diagnosis process of the slice by the pathologist with artificial intelligence, opens up the application prospect of deep learning in the medical pathology image field, and can realize the detection of the slice lesion at a faster speed and a higher detection rate, thereby improving the survival rate of patients. BRIEF DESCRIPTION OF DRAWINGS

[0018] Figure 1 It is a flowchart of the pathological slice image lesion hierarchical classification detection method based on structure optimization and local attention supervision of the application.

[0019] Figure 2 It is a lesion hierarchical diagnosis strategy diagram based on structure optimization and local attention supervision network.

[0020] Figure 3 It is a four-point matching diagram.

[0021] Figure 4 It is a manually marked original diagram.

[0022] Figure 5 It is a partial network structure diagram after adding a mixed attention module.

[0023] Figure 6 It is a four-corner matching result diagram (part).

[0024] Figure 7 It is a data set visualization diagram, the left diagram is a visualization diagram of slice number S1, and the right diagram is a visualization diagram of slice number S2.

[0025] Figure 8 It is a lesion hierarchical detection result diagram. DETAILED DESCRIPTION

[0026] The technical solutions of the application will be further described below in conjunction with the drawings, but are not limited thereto, and any modification or equivalent replacement to the technical solutions of the application without departing from the spirit and scope of the technical solutions of the application shall be covered in the protection scope of the application.

[0027] The application provides a pathological section image lesion grading and classification detection method based on structure optimization and local attention supervision. First, an automatic label labeling method based on normalized mutual information registration is used to construct a data set, and the image is randomly cropped and rotated as input data. Second, the network structure of the classification network DenseNet121 is optimized, and the focal loss function is combined to balance the weights of multiple classes to realize the multi-classification task. Third, a hybrid attention module is added to the above DenseNet121 network based on structure optimization, and a local supervision function is added to the network loss function to detect the classification results and attention enhanced results of the improved network, realizing the binary classification task of normal images and lesion images. Finally, the two classification task systems are combined to form a lesion grading and classification detection scheme. The improved structure optimization network is used to identify the section tissue, and the improved local attention supervision network is used to identify the lesion image. The pathological image belongs to the tissue category and the classification of whether the image has lesion characteristics is performed. As shown in Figure 1 , there are four steps in total, and the specific steps are as follows:

[0028] Step one: the high-resolution complete section scanning image is regularly cropped, and the cropped image is labeled by using the automatic label labeling method based on normalized mutual information registration to generate a data set. The specific steps are as follows:

[0029] Step one: when training the model, the resolution of the section image is much higher than the resolution of the input image required by the network, so the complete section scanning image that has been manually labeled is first cropped into several sub-images to reduce the resolution of the input image, and then input into the network model for training.

[0030] The cropped sub-image does not have a pre-labeled label. Before labeling the sub-image, the R, G and B channels of the four corner matching method shown in Figure 3 are calculated to calculate the normalized mutual information (NMI) of the R, G and B channels of the four corner points of the sub-image and the R, G and B channels of the sliding window size of the original section scanning image:

[0031]

[0032]

[0033]

[0034]

[0035] Wherein, subscript R represents R channel, NMI (X; Y) is the normalized mutual information of random variable X and random variable Y, H (X) is the information entropy of random variable X, H (Y) is the information entropy of random variable Y, I (X; Y) is the mutual information of random variable X and random variable Y, random variable X is mapped to the subgraph after cutting, random variable Y is mapped to the slice scanning graph original drawing, P (x, y) is the joint probability distribution function of random variable X and random variable Y, P (x), P (y) are the marginal probability distribution function of random variable X and random variable Y respectively.The NMI values of G and B two channels are calculated in the same way:

[0036]

[0037]

[0038] Step one two: in order to speed up the image matching speed and improve the matching accuracy, the dynamic step search method combining breadth and depth is adopted, that is, first breadth search is carried out, search is carried out with larger step, and as many similar matching results as possible are saved in the search process, if the corresponding matching graph cannot be obtained in the search, the step is reduced, and the matching result closest to the shooting graph is manually screened out, and the second depth search is carried out with the result as the original point.

[0039] Step one three: the matching result obtained after search is saved, and the position information (left upper corner and right lower corner two points) of each subgraph on the scanning graph original drawing is saved, and according to the obtained position information and the label information as shown in Figure 4 The label to which the subgraph belongs is obtained; according to the needs of multi-classification and binary classification tasks, all subgraphs are randomly divided into training set, verification set and test set according to the proportion of 3:1:1, and the subgraph visualization information screened on the slice original drawing is generated.

[0040] Step two: taking DenseNet121 as the benchmark network, the DenseNet121 network is improved, the focal loss function is introduced to balance the weight between multiple tissue samples, and the multi-classification task training for tissue identification is carried out.

[0041] Step two one: DenseNet121 is a model with 121 layer network structure in four different network configurations of DenseNet, the improved DenseNet121 is used to realize 1000 kinds of image classification, in order to prevent the occurrence of overfitting problem, the network layer number in DenseNet121 is simplified to a certain extent, the layer number of original dense block (the original layer number of each layer is 6, 12, 24 and 16) is changed to three 6 layer dense blocks, the last dense block is removed, and the last fully connected layer is changed to the size required by multi-classification task.

[0042] Step 22: To avoid interference from small sample images on network weights in multi-class classification tasks, this invention introduces a focal loss function (FL) as an improved loss function for the network:

[0043] FL(p t )=-α t (1-p t ) γ log(p t )

[0044] Among them, FL(p t ) represents parameter p t The focus loss function, α t p is an adjustment factor, which is adjusted according to the actual task requirements. t for:

[0045]

[0046] p is the probability that the image is classified as foreground in the binary classification model, and its value ranges from [0, 1]. γ is the tone factor (1-p). t ) γ The parameter is used to distinguish the difficulty of organizational classification, and its value ranges from [0, 5]. This parameter is adjusted according to the needs of the task in actual operation.

[0047] Step 3: In the improved DenseNet121 from Step 2, a hybrid attention module is added between each dense block and transition layer to enhance the network's weighting of spatial and channel features of lesion regions in pathological images. A local supervision function is also added to the loss function of the modified network to simultaneously detect the network's overall classification performance and local enhancement effect. The specific steps are as follows:

[0048] Step three one: the mixed attention module usually uses the convolutional block attention module (CBAM), which can simultaneously focus on the channel features and spatial features in the classification model, and the mechanism of CBAM is that the feature map is first input into the channel attention module, the channel attention performs maximum pooling and average pooling operations on each spatial position, and the two one-dimensional vectors are input into two weight-shared fully connected layers, the results are added and fused based on pixels, and finally an activation function is used to obtain the channel attention feature map. Then the attention feature map is input into the spatial attention module, the spatial attention module performs maximum pooling and average pooling operations on each channel to obtain two feature maps with unchanged length and width, the two maps are connected by channel, and then input into a convolution layer, and an activation function is used to obtain the spatial feature map output by the spatial attention module, and finally the feature map is multiplied with the original input to obtain the final output feature. The improved network has three dense blocks and three conversion layers, and a mixed attention module is added between each dense block and conversion layer, which does not change the size of the input and output feature maps, as shown in Figure 5 .

[0049] Step three two: the local feature enhancement of the image processed by the conversion layer is enhanced, the feature map of each channel is multiplied with the input original image to obtain a new feature map, and the new feature map and the original feature map are input into the classification network to make a local prediction, and a local supervision loss function is used to detect the local enhancement effect:

[0050] l i =-plog(sigmoid(w i *I))+(1-p)log(1-sigmoid(w i *I))

[0051] Wherein, l i represents the loss of the i-th channel, w i represents the weight of the i-th channel, I is the input original image, p is the probability that the image is judged as foreground in the binary classification model, and the value is between [0, 1], and sigmoid is an activation function:

[0052]

[0053] Wherein, z is the independent variable. The loss function L of the improved local attention supervision network is:

[0054]

[0055] L org =-plog(1-p)+(1-p)logp

[0056] wherein, L org is the original cross-entropy loss function, l j,i represents the loss of the i-th channel after the j-th conversion layer.

[0057] Step four: learn the slice tissue morphology by using the improved network based on structure optimization in step two, learn the lesion tissue characteristics by using the improved network based on local attention supervision in step three, combine the two improved network systems in step two and step three to form a hierarchical classification detection scheme for lesions.

[0058] Different from previous studies, the present application focuses on solving the problem of time-consuming and laborious process of finding lesion tissue under the microscope by pathologists, rather than repeatedly training on the already diagnosed results. In the actual diagnosis process, the doctor will make a global observation through different magnification fields to form a cognition of the slice, and then determine the next screening path, and then make a high-power observation on a certain specific area or tissue to obtain the diagnosis result.

[0059] The present application combines two classification network systems to perform hierarchical classification detection on patient slices: the input slice is used to generate a multi-classification data set for tissue recognition and a two-classification data set for lesion tissue detection according to the method of step one, and then the Figure 2 shown process is used to implement the multi-tissue recognition classification task by the improved network based on structure optimization in step two (complete the full field of view cognition of the slice), and according to the development process of tumor tissue, the easy lesion tissue is selected to be input into the improved network based on local attention supervision in step three, thereby forming hierarchical classification detection, and the other tissue images of the slice where the lesion exists are classified again, and the detection results of each cropped subgraph in the input slice are output.

[0060] Embodiment:

[0061] Perform step one: crop the patient's stomach slice scan image provided by the hospital (since the size of the tissue is different, the total number of pixels in a single slice effective area is generally between 1 billion and 10 billion) into a subgraph with a resolution of 224x224. In order to improve the matching speed and accuracy, the cropped images are stored in the corresponding named subfolder according to the slice number, the cropping batch, and the cropping quadrant. The program reads the pictures under the subfolder and searches them in turn. When the first accurately matched image in the program output folder is output, take this image as the origin, and search for other images in a spiral manner. Set the scaling factor to 0.91 and the step distance to 20. The original image scaling factor is 0.5. The matched result image is named as the original name of the matched image and the corresponding upper left and lower right coordinates on the original image, as shown in Figure 6 According to the matching results obtained from Figure 6 Figure 4 ​The label map is automatically annotated with labels for each subgraph.

[0062] In order to be able to train the model in the future, the visual processing program is used to select the entire image dataset, including but not limited to selecting the background, removing the image intersection, and removing certain labels. Figure 7 The position of the data processed by different selection modes on the original image is shown, the left image selects the background, and the image has intersection, and the right image selects the background, and the image has no intersection. Different processing modes will generate different sample quantities, different organization selections, different region data sets and labels, and generate a visual view of the corresponding data selection. The advantage of this is that different training data can be selected by visualization, and different experiments can be designed for these data to maximize the performance of the classification network.

[0063] Step 2: The improved network structure of DenseNet121 is shown in Table 1.

[0064] Table 1 Improved DenseNet121 network structure

[0065]

[0066] Since the purpose of this experiment is to require the model to accurately identify each tissue in the pathological image, the classification requirement for each tissue is high, therefore, the present application selects the method of converting the multi-classification problem into multiple binary classification problems to evaluate the network of the present application, and uses Macro-Presicion, Macro-Recall and Macro-F1-Score as the evaluation index of the improved network:

[0067]

[0068]

[0069]

[0070] Wherein, k is the number of classification categories, Pre represents the precision (Precision) in binary classification, Re represents the recall (Recall) in binary classification, and F1 represents the F1-Score in binary classification, and the calculation formulas of the precision, recall and F1-Score are as follows:

[0071]

[0072]

[0073]

[0074] The improved network is used in the multi-classification task of organization identification, k = 6, and the adjustment factor a is taken t Take 1, and g takes 0.5. The comparison experimental results of other classification models are shown in Table 2.

[0075] Table 2 Comparison experimental results of multi-classification task

[0076]

[0077] Step three is performed: the improved network has three dense blocks and three conversion layers, and a mixed attention module is added between each dense block and conversion layer, which does not change the size of the input and output feature map. The network result is shown in Table 3.

[0078] Table 3 Network structure of DenseNet121 after adding attention module

[0079]

[0080]

[0081] The improved network is used in the binary classification lesion identification task, and the results are shown in Table 4. The comparison experimental results before and after network improvement are shown in Table 5. The calculation formula of accuracy (Accuracy) is as follows:

[0082]

[0083] Table 4 Comparison experimental results of binary classification task

[0084] Model Precision Recall F1-Score Accuracy The present invention improves the network 0.984 1.000 0.989 0.989 EfficientNet 0.969 0.992 0.980 0.983 ResNet 0.878 0.813 0.844 0.923

[0085] Table 5 Comparison experimental results before and after network improvement

[0086]

[0087]

[0088] Step four is performed: in order to solve the problem that pathologists spend a lot of time and effort in finding pathological tissues under a microscope, the improved network system of step two and step three are combined, the improved network in step two is used to classify the patient's slice according to the tissue morphology, the deep learning model is used to learn the morphology of each tissue, the process of forming the cognition of the tissue in the artificial state is simulated, then according to the consensus of the reading path proposed by different tumor pathologists, for a specific tissue, the improved network in step three is used to further classify the tissue, and the classification result will show whether the tissue is diseased. If there is a lesion, further classification of other tissues in step two is performed using the improved network in step three to check whether there is a virus spread in other tissue images so as to cause tissue lesions, so that the doctor can infer the source of the patient's disease according to the development of the tissue. The present application selects the gastric cancer slice provided by the hospital to perform the experiment, and the experimental results are shown in Table 2, Table 4 and Table 5. Figure 8 The flow generates a report showing the slice number, subgraph number and whether there is a lesion area in the subgraph, and shows the predicted probability value.

[0089] According to the experimental results in Tables 2, 4 and 5, the performance of the DenseNet121 based on structure optimization is better than that of other popular classification networks in the multi-classification task of tissue recognition, and the comprehensive performance of the improved network for lesion area detection can reach more than 98%, and the network after the cascade of the two network systems has an observable detection effect on the lesion detection of patient slices.

Claims

1. A pathological section image lesion grading and classification detection method, characterized by The method comprises the following steps: Step one: the high-resolution complete slice scanning image is regularly cut, the cropped image is labeled by using an automatic label labeling method based on normalized mutual information registration, and a data set is generated; Step two: taking DenseNet121 as a benchmark network, the DenseNet121 network is improved, a focal loss function is introduced to balance the weights between multiple tissue samples, and a multi-classification task training is used for tissue identification; Step three: a mixed attention module is added between each dense block and conversion layer in the improved DenseNet121 in step two, the spatial and channel features of the lesion area in the pathological image are enhanced, and a local supervision function is added to the loss function of the modified network, which is used for detecting the overall classification performance and local enhancement effect of the network at the same time, and the specific steps are as follows: Step three one: the improved network has three dense blocks and three conversion layers, and a mixed attention module is added between each dense block and conversion layer, which does not change the size of the input and output feature maps; Step three two: each channel of the image processed by the conversion layer is strengthened, the feature map of each channel is multiplied with the input original image to obtain a new feature map after local feature enhancement, the new feature map and the original feature map are input into the classification network, a local prediction is made, and a local supervision loss function is used to detect the local enhancement effect: wherein, represents the loss of the i-th channel, represents the weight of the i-th channel, is the input original image, is the probability that the image is judged as foreground in the binary classification model, the loss function of the improved local attention supervision network is: wherein, is the original cross-entropy loss function, represents the loss of the i-th channel after the j-th conversion layer; Step four: the slice tissue morphology is learned by using the improved network based on structure optimization in step two, and the lesion tissue features are learned by using the improved network based on local attention supervision in step three, the two improved network systems in steps two and three are combined to form a hierarchical classification detection scheme for lesions.

2. The pathological section image lesion grading and classification detection method according to claim 1, characterized in that The specific steps of step one are as follows: Step one one: first, cut a complete slice scanning image which has been manually labeled into a plurality of sub-images, reduce the resolution of the input image, and calculate the NMI of the R, G and B channels of the four corner points of the sub-image edge and the R, G and B channels of the sliding window size in the original slice scanning image according to the four corner matching method before labeling the sub-image: where subscript R represents the R channel, is a random variable is a random variable is the normalized mutual information of random variables is a random variable is the information entropy of random variable is a random variable is the information entropy of random variable is a random variable is the mutual information of random variables is a random variable is mapped to the cropped subgraph, and random variable is mapped to the slice scan graph original image, is a random variable is the joint probability distribution function of random variables , , are the marginal probability distribution functions of random variables and random variable ; similarly, the NMI values of the G and B channels are calculated: Step one two: a dynamic step search method combining breadth and depth is adopted, that is, a breadth search is first performed, a search process is saved as much as possible, if the corresponding matching image cannot be obtained in the first search, the step is reduced, the closest matching result to the shooting image is manually selected, and the second depth search is performed with the result as the original point; Step one three: save the matching result obtained after the search, save the position information of each sub-image on the original scanning image, obtain the label to which the sub-image belongs according to the obtained position information and label information, divide all the sub-images into a training set, a verification set and a test set according to the needs of multi-classification and binary classification tasks, and generate the visualization information of the selected sub-images on the original slice image.

3. The pathological section image lesion grading and classification detection method according to claim 1, characterized in that The specific steps of step two are as follows: Step two one: the number of dense blocks in DenseNet121 is changed to three 6-layer dense blocks, the last dense block is removed, and the last fully connected layer is changed to the size required by the multi-classification task; Step two: introduce focal loss function as the loss function of the improved network: wherein, denotes a parameter a focal loss function, is a tuning factor, is: is the probability that an image is classified as foreground in a binary classification model, is a conditioning factor with parameters.

4. The pathological section image lesion grading and classification detection method according to claim 1, characterized in that The where z is an independent variable.

5. The pathological section image lesion grading and classification detection method according to claim 1, characterized in that The specific steps of step four are as follows: The input slice is used to generate the multi-classification data set for tissue identification and the binary classification data set for lesion tissue detection according to the method in step one, and then the improved network based on structure optimization in step two is used to realize the multi-tissue identification classification task, according to the tumor tissue development process, the improved network based on local attention supervision in step three is selected to input the easy lesion tissue, thereby forming hierarchical classification detection, and the other tissue images of the slice with lesions are classified again, and the detection results of each cropped subgraph in the input slice are output.

Images