An identification, analysis, construction and image processing method for special pathological images of the lungs

Abstract

The present invention provides a method for recognizing, analyzing, constructing and processing special pathological images of the lungs, which is characterized in that it includes developing an identification model STASNet for the spread of cancer cells in air spaces in lung cancer based on digitized scanned lung tumor section images. Based on this model, the following objectives can be achieved: assisting pathologists to identify STAS lesions around lung pathological sections in real time during the diagnosis and treatment process, improving the efficiency of clinical diagnosis of STAS, and reducing misdiagnosis and missed diagnosis; on the basis of classifying the presence or absence of STAS, further assisting clinicians to complete semi-quantitative analysis of STAS, indicating the prognosis of patients, and helping doctors formulate personalized diagnosis and treatment plans; as an organic part of the microscope computer imaging system, embedding a microscope image acquisition module to identify STAS lesions in lung cancer images in real time.

Description

Technical Field

[0001] The present invention relates to the technical field of pathological image processing, and specifically to a method for recognizing, analyzing, constructing and processing special pathological images of the lungs. Background Art

[0002] Lung cancer, as a global public health problem, ranks second in incidence and has the highest mortality rate among all cancers, accounting for 18% of cases. Over the past three decades, the overall five-year survival rate of lung cancer has been consistently below 20%. Among them, the main histological subtype of lung cancer is non-small cell lung cancer, and lung adenocarcinoma is the most common pathological subtype. In the past, the patterns of lung cancer metastasis were hematogenous spread, lymph node spread, or direct metastasis. In 2015, Kadota et al. first proposed the concept of spread through air space (STAS), which was defined as tumor cells spreading into the airspace of the lung parenchyma outside the edge of the tumor mass. In the same year, the World Health Organization (WHO) incorporated the STAS concept into the new invasion pattern of lung adenocarcinoma and refined its pathological classification into: micropapillary clusters, solid tumor nests, and single free tumor cells.

[0003] STAS has a relatively high incidence in lung adenocarcinoma. The reported incidence rates in previous studies varied from 14.8% to 60.5%, and it is commonly found in lung adenocarcinomas mainly composed of solid components and micropapillary components. It has also been reported in other cancer types, mainly in invasive carcinomas, and is less common in microinvasive and in-situ carcinomas. Multiple studies have reported the negative impact of STAS on prognosis. The presence of STAS is associated with high recurrence and poor prognosis. For patients with similar cancer stages, the presence of STAS often indicates a worse prognosis and a higher likelihood of recurrence and metastasis. In the evaluation of early lung cancer, the impact of STAS on prognosis has more important guiding significance. As a high-risk factor, STAS has been incorporated into the considerations of many medical centers for whether to perform postoperative adjuvant treatment in patients with early lung adenocarcinoma. Some studies have also pointed out that the farther the distance between STAS and the boundary of the main tumor mass, the worse the prognosis, and the more the number of STAS, the worse the survival. Therefore, a clear diagnosis of STAS lesions helps clinical staff better formulate personalized and precise treatment plans.

[0004] Currently, pathological diagnosis mainly relies on pathologists' examination of slides under a microscope. It is very difficult to quantitatively evaluate airspace dissemination lesions from an overall perspective in this way. At the same time, the clinical attention to STAS is insufficient. Pathologists' interpretation of STAS in early-stage lung adenocarcinoma surgery patients is highly subjective, with only moderate inter-observer agreement among doctors. Moreover, artifacts can interfere with the interpretation process. At the same time, the evaluation results of STAS lack guiding significance for clinical practice. For the formulation of clinical diagnosis and treatment strategies, although a large number of studies have demonstrated the impact of different resection margins and surgical methods on the prognosis of STAS-positive cases, the experience in dealing with STAS-positive cases is still extremely lacking, and its guiding role in the adjuvant treatment of early-stage surgical patients is still unclear. Some studies have used deep learning models based on CT data to predict the preoperative STAS status of patients, which has greatly helped in the personalized formulation of patients' surgical plans. However, the prediction is still based on the presence or absence of STAS and does not involve quantitative analysis of STAS, which is obviously very challenging for CT data. Some studies have also explored the impact of the relationship between STAS and the spatial location of the primary tumor on tumor prognosis, and the results show that the farther the spatial distance between STAS and the primary tumor, the worse the prognosis, but no further quantitative analysis has been carried out. If a method can be proposed to more accurately identify and evaluate STAS in lung tumors and further semi-quantitatively analyze STAS in combination with other risk factors such as spatial location, it will reduce the diagnostic burden of pathologists and enhance the guiding significance for clinical decision-making, enabling patients to receive more personalized diagnosis and treatment.

[0005] In the past few years, the emergence of digital pathology and the application of whole slide images (WSIs) in the field of pathological diagnosis have made it possible to use artificial intelligence (AI) and deep learning (DL) models to diagnose pathological sections. A well-trained DL model can extract high-dimensional features and mine abstract underlying information from the input raw data, showing good performance in the field of pathological image analysis. Morphological features that are difficult to identify even by the human eye can be detected. Some previous studies have also reported that deep learning models perform well in tasks such as tumor region segmentation, prognosis prediction, tumor microenvironment characterization, and driver mutation prediction. Due to the unique robustness of deep learning models, a well-trained DL-based pathological diagnosis model can provide relatively stable diagnostic references for pathologists, reducing the diagnostic burden of pathologists and reducing inter-observer differences in clinical diagnostic practice. Due to the robustness of DL models, the introduction of an AI-assisted pathological diagnosis system can significantly reduce the burden on pathologists and is conducive to accurate and objective diagnosis. Summary of the Invention

[0006] The object of the present invention is to propose a recognition, analysis, construction and image processing method for special pathological images of the lungs, so as to solve the problems in the prior art that there is a lack of means for accurately identifying and evaluating STAS of lung tumors, and it is difficult to further perform semi-quantitative analysis of STAS in combination with other risk factors such as spatial position, increasing the burden on pathologists for diagnosis.

[0007] To achieve the above object, the present invention provides the following technical solutions:

[0008] A recognition, analysis, construction and image processing method for special pathological images of the lungs, characterized in that it includes a pathological section processing module, a STASNet model and a STAS scoring module. After giving a HE pathological section of lung adenocarcinoma and marking its tumor boundary, the pathological section processing module can automatically cut the normal lung tissue outside the tumor boundary in real time, and then the STASNet model identifies the STAS lesions therein to generate a prediction image, and finally calculates the STAS score through the STAS scoring module.

[0009] The pathological section processing module scans the HE section into a digital pathological section image by using a commercial scanner for the HE section, and determines the tumor boundary based on the marking by the pathologist;

[0010] The scanned digital pathological section image is segmented into local image regions with a size of 256 pixels * 256 pixels by using a Python script, and is divided into tumor images and non-tumor region images based on the tumor boundary marking. For non-tumor images, if the white pixels therein have R, G, B values greater than 220 and the proportion is greater than 90%, then the map will be put into the background group; the remaining images containing non-tumor main region tissues will be constructed into a data set and used for model construction and recognition in subsequent modules.

[0011] After the STASNet model receives the image obtained by the pathological section processing module as the input image, the STASNet model introduces a convolutional neural network framework, called the STASNet framework, for module construction, and completes the training and evaluation of the weights of the modules in the tissue image data set.

[0012] The STASNet model is optimized and constructed based on the MobileNet V3 framework, and mainly includes the following key components: convolutional layer, pooling layer, activation function and fully connected layer; the basic components include convolutional layer, pooling layer, fully connected layer, activation function, loss function, optimizer, etc., and the specific details are as follows:

[0013] Convolutional Layer:

[0014] The convolutional layer is the core of feature extraction in the CNN framework. The convolutional layer applies a convolutional kernel with specific features to perform matrix operations on local regions of the image in sequence. The finally obtained new matrix is regarded as a feature map that contains certain features of the original image. A single convolutional layer includes several convolutional kernels for extracting different features of the input image matrix, and finally outputs several feature maps containing different features to the pooling layer or convolutional layer network of the next layer.

[0015] Pooling Layer:

[0016] Since the number of parameters in the CNN framework is too large, the model introduces a pooling layer to reduce the number of parameters. The pooling layer simplifies the input feature map and obtains representative feature values of the region through calculations on local regions of the Feature Map. Common pooling methods include max pooling and mean pooling.

[0017] Fully Connected Layer:

[0018] The fully connected layer is generally connected at the end of the entire CNN to play a role in classifying data. After several convolutional and pooling layers, the multi-dimensional abstract features extracted by the model are comprehensively evaluated and output as specific feature tags. Several fully connected layers form a fully connected neural network. Among them, the fully connected layer that receives input features is the input layer, and the fully connected layer that finally outputs several preset classification tags is the output layer. All layers between the input layer and the output layer are hidden fully connected layers.

[0019] Activation Function:

[0020] During the model training process, parameters continuously flow from the input end to the output end. During this process, the model introduces an activation function to evaluate whether the input result will continue to propagate to the subsequent model.

[0021] MobileNetV3 uses an activation function called Hard Swish, which is a variant of the rectified linear unit ReLU.

[0022] The mathematical expression of the Hard Swish function is as follows:

[0023]

[0024] Among them, ReLU6 is a truncated ReLU function, and its definition is as follows:

[0025]

[0026] Loss Function and Optimizer:

[0027] During the training process of the model, after the input image passes through the feature extraction and classification of the model, the deep learning model introduces a loss function to calculate the gap between the classification result and the true result; after calculating the loss value, the model will apply an optimizer to update and optimize the parameters of each node inside the model in reverse, so as to ensure that a lower loss value can be obtained in the next training;

[0028] Apply the cross-entropy function as the loss function and apply the Adam optimizer to optimize the internal parameters of the model. The formula is as follows:

[0029]

[0030] where xi is a discrete variable, which represents the predicted value of each class result in the prediction result in this study;

[0031] Bneck Block:

[0032] Based on the stacking of multiple said convolutional layers and multiple said pooling layers, the core module Bneck Block of MobileNet V3 can be further constructed. Different from the traditional convolutional layer, Bneck Block separates the traditional convolutional operation into depth-wise convolution and 1*1 point convolution, greatly reducing the number of parameters and computational complexity, and maintaining relatively good performance while improving the model efficiency.

[0033] Bneck Block also enhances the distinction of the importance weights of the model input features by adding a Squeeze-and-Excitation attention mechanism module, thereby improving the model performance.

[0034] The Bneck Block constructed by the present invention includes a group of depth-wise convolutions, a group of attention modules SE Block, and a group of 1*1 point convolutions;

[0035] Overall view of the model:

[0036] Based on the above components, the present invention finally constructs the STASNet model; the overall prediction process is:

[0037] The input image is processed into a matrix of size 256*256*3 and sequentially processed in 15 groups of Bneck Block matrices, and finally a matrix of size 8*8*960 is generated, and then transformed into an output matrix of size 1*1*960 through a pooling layer, and through a fully connected neural network including three layers of fully connected layers, a two-dimensional result, namely STAS and non-SATS, is finally output.

[0038] The training and evaluation of the weights of the module specifically include the following contents:

[0039] Apply the obtained image patches with real class labels processed by the pathological section processing module, make predictions in the STASNet model, calculate the loss function based on the prediction results of the STASNet model and the real results of the image patches, apply an optimizer to optimize the weights, and reuse the optimized model for prediction. By continuously iterating the node weights of the model, a well-trained STASNet framework is obtained;

[0040] The well-trained STASNet framework of the weights can be used to predict STAS lesions in the workflow. The specific workflow is as follows:

[0041] The STASNet model receives the image patches obtained after being processed by the pathological section processing module and is processed by the well-trained STASNet framework. The final output result is "is a STAS lesion / is not a STAS lesion".

[0042] After being processed by the pathological section processing module and the STASNet model, non-tumor image patches with STAS interpretation results or tumor image patches with tumor markers are generated. All image patches are further evaluated in the STAS scoring module, and the STAS semi-quantification of the whole section is completed;

[0043] Specifically, it includes two steps: spatial position calculation and STAS semi-quantification evaluation;

[0044] S1. The spatial position calculation module: The image patches obtained after being processed by the pathological section processing module contain the original spatial coordinates. Based on the original spatial coordinates of the image patches and the tumor boundary coordinates, the spatial position calculation module uses Python scripts to calculate the vertical distances of all image patches to the original spatial coordinates;

[0045] S2. The STAS semi-quantification evaluation module: After the spatial position obtained by being processed by the pathological section processing module and the STAS prediction probability of the image patches obtained by being processed by the STASNet model, the STAS semi-quantification evaluation module further selects the 10 image patches with the highest STAS prediction probability and uses them to calculate the semi-quantification score Top 10Score. The specific formula is as follows:

[0046] Top10Score = ∑ Top10tiles AI score of STAS tile * Tiledistance. Compared with the prior art, the beneficial effects of the present invention are as follows:

[0047] Based on the digitized scanned lung tumor section images, the present invention develops an identification model STASNet for the airspace dissemination lesions of lung cancer. After giving a HE pathological section of lung adenocarcinoma and marking its tumor boundary, the present invention can automatically cut the normal lung tissue outside the tumor boundary in real time, identify the STAS lesions therein to generate a prediction image, and further calculate the STAS score, which can assist pathologists to improve the efficiency of reading films and at the same time prompt the prognosis of patients. It has guiding significance for assisting clinicians in treatment. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] Figure 1 It is a schematic diagram of Bneck Block;

[0049] Figure 2 It is a schematic diagram of the deep learning model of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0050] To clarify the technical problems, technical solutions, implementation processes and performance demonstrations, the present invention will be further described in detail below with reference to the embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention. Various exemplary embodiments, features and aspects of the present disclosure will be described in detail below with reference to the drawings. The same reference numerals in the drawings denote elements with the same or similar functions. Although various aspects of the embodiments are shown in the drawings, the drawings do not have to be drawn to scale unless otherwise specified.

[0051] The special term "exemplary" herein means "serving as an example, embodiment or illustration". Any embodiment described as "exemplary" herein does not have to be construed as superior to or better than other embodiments.

[0052] In addition, in order to better illustrate the present disclosure, numerous specific details are given in the following detailed description. Those skilled in the art should understand that the present disclosure can be implemented without some specific details. In some instances, methods, means, elements and circuits well known to those skilled in the art are not described in detail so as to highlight the gist of the present disclosure.

[0053] Embodiment 1

[0054] As Figure 1 and Figure 2As shown, a method for the recognition, analysis, construction and image processing of special pathological images of the lungs includes a pathological section processing module, a STASNet model and a STAS scoring module. After a HE pathological section of lung adenocarcinoma is given and its tumor boundary is marked, the pathological section processing module can automatically cut the normal lung tissue outside the tumor boundary in real time. Then, the STASNet model identifies the STAS lesions therein to generate a prediction image. Finally, the STAS scoring module calculates the STAS score, which can assist pathologists in improving the efficiency of reading slides and at the same time indicate the prognosis of the patient.

[0055] For the HE section, the pathological section processing module uses a commercial scanner (e.g., Jiangfeng Biology Co., Ltd., Yuyao, Zhejiang, KFBIO, KF-PRO-400 scanner) to scan the HE section into a digital pathological section image (Whole Slide Image, WSI), and determines the tumor boundary based on the markings of pathologists.

[0056] The scanned digital pathological section image is segmented into local image regions of size 256 pixels * 256 pixels using a Python script, and is divided into tumor images and non-tumor region images based on the tumor boundary markings. For non-tumor images, if the proportion of white pixels (R, G, B values greater than 220) is greater than 90%, the map will be placed in the background group; the remaining images containing non-tumor main region tissues will be used to construct a dataset for the model construction and recognition of subsequent modules.

[0057] After the STASNet model receives the image obtained by the pathological section processing module as the input image, the STASNet model introduces a convolutional neural network framework for module construction, and completes the training and evaluation of the module weights in the tissue image dataset.

[0058] The STASNet model is optimized and constructed based on the MobileNet V3 framework, and mainly includes the following key components: convolutional layer, pooling layer, activation function and fully connected layer; the basic components include convolutional layer, pooling layer, fully connected layer, activation function, loss function, optimizer, etc. The specific details are as follows:

[0059] Convolutional Layer:

[0060] The convolutional layer is the core of feature extraction in the CNN framework. The convolutional layer applies convolutional kernels with specific features to perform matrix operations on local regions of the image in sequence. The new matrix obtained finally is regarded as a feature map that contains certain features of the original image. A single convolutional layer includes several convolutional kernels for extracting different features of the input image matrix, and finally outputs several feature maps containing different features to the next layer of the network;

[0061] Pooling Layer:

[0062] Due to the excessive number of parameters in the CNN framework, the model introduces a pooling layer to reduce the number of parameters. The pooling layer simplifies the input feature map and obtains representative feature values for local regions of the Feature Map through calculations on local regions. The introduction of the pooling layer reduces the number of parameters calculated during model training, reduces the impact of secondary features on the model, and improves the stability of the CNN, which is of great significance for preventing overfitting of the model. Commonly used pooling methods include max pooling and mean pooling;

[0063] Fully Connected Layer (FC):

[0064] The fully connected layer is generally connected at the end of the entire CNN to play a role in classifying data. After several convolutional and pooling layers, the multi-dimensional abstract features extracted by the model are comprehensively evaluated and output as specific feature labels. Several fully connected layers form a fully connected neural network. The fully connected layer that receives the input features is the input layer, and the fully connected layer that finally outputs several preset classification labels is the output layer. All the six layers between the input layer and the output layer are hidden fully connected layers. In the entire fully connected neural network, all neurons receive the output parameters of all neurons in the previous layer and comprehensively output them to all neurons in the next layer, so it is called a "fully connected" network;

[0065] Activation Function

[0066] During the model training process, parameters continuously flow from the input end to the output end. During this process, the model introduces an activation function to evaluate whether the input result will continue to propagate to the subsequent model. Such functions are usually non-linear, breaking the characteristic that the model has been performing linear calculations in the convolutional layer and the pooling layer, enabling the model to fit non-linear parameter situations, and at the same time solving the problem of gradient disappearance in the model;

[0067] MobileNetV3 uses an activation function called "Hard Swish", which is a variant of the rectified linear unit (ReLU). The main feature of Hard Swish is that it is computationally lighter and suitable for efficient computing on mobile devices.

[0068] The mathematical expression of the Hard Swish function is as follows:

[0069]

[0070] Among them, ReLU6 is a truncated ReLU function, and its definition is as follows:

[0071]

[0072] Loss Function and Optimizer:

[0073] During the training process of the model, after the input image passes through the feature extraction and classification of the model, the deep learning model introduces a loss function to calculate the gap between the classification result and the true result. After calculating the loss value, the model will apply an optimizer to update and optimize the parameters of each node inside the model in the reverse direction, so as to ensure that a lower loss value can be obtained during the next training, that is, a predicted value closer to the true result, so as to achieve the purpose of "machine learning";

[0074] This application uses the cross-entropy function as the loss function and the Adam optimizer to optimize the internal parameters of the model. The formula is as follows:

[0075]

[0076] Among them, xi is a discrete variable, which represents the predicted value of each class result in the prediction result in this study;

[0077] Bneck Module (Bneck Block):

[0078] Based on the components mentioned above, the core module Bneck Block of MobileNet V3 can be further constructed. Different from the traditional convolutional layer, Bneck Block separates the traditional convolutional operation into depth-wise convolution and 1*1 point convolution, greatly reducing the number of parameters and computational complexity, and maintaining relatively good performance while improving the model efficiency.

[0079] The Bneck Block also enhances the discrimination of the importance weights of the model input features by adding a Squeeze-and-Excitation (SE Block) attention mechanism module, thereby improving the model performance.

[0080] The Bneck Block constructed in the present invention includes a group of depthwise convolutions, a group of attention modules (SE Block), and a group of 1×1 pointwise convolutions. The specific structure is shown in the appendix Figure 1 ;

[0081] Overall view of the model:

[0082] Based on the above components, the present invention finally constructs the STASNet model. The specific structure is shown in the appendix Figure 2 The overall prediction process is as follows:

[0083] The input image is processed into a matrix with a size of 256×256×3 and sequentially processed in 15 groups of Bneck Block matrices, and finally a matrix with a size of 8×8×960 is generated, which is then transformed into an output matrix with a size of 1×1×960 through a pooling layer, and through a fully connected neural network including three fully connected layers, a two-dimensional result, i.e., [STAS, non - SATS], is finally output.

[0084] The training and evaluation of the weights of the said module specifically include the following content:

[0085] The image patches with real class labels obtained by applying the pathological section processing module are processed, predicted in the STASNet model, and the loss function is calculated based on the prediction results of the STASNet model and the real results of the image patches. The optimizer is used for weight optimization, and the optimized model is reused for prediction. By continuously iterating the node weights of the model, a well - trained STASNet framework is obtained;

[0086] The well - trained STASNet framework of the weights can be used to predict STAS lesions in the workflow. The specific workflow is as follows:

[0087] The STASNet model receives the image patches obtained after being processed by the pathological section processing module and is processed by the well - trained STASNet framework, and the final output result is "is a STAS lesion / is not a STAS lesion"

[0088] After being processed by the "pathological section processing module" and the "STASNet model", non - tumor image patches with STAS interpretation results or tumor image patches with tumor markers are generated. All the image patches are further evaluated in the STAS scoring module, and the STAS semi - quantification of the whole section is completed;

[0089] Specifically, it includes two steps: spatial position calculation and STAS semi-quantitative evaluation;

[0090] S1. The spatial position calculation module: After obtaining image patches processed by the pathological section processing module, which will contain the original spatial coordinates, based on the original spatial coordinates of the image patches and the tumor boundary coordinates, the spatial position calculation module uses a Python script to calculate the vertical distance from all image patches to the original spatial coordinates;

[0091] S2. The STAS semi-quantitative evaluation module: After the spatial position obtained by processing through the pathological section processing module and the STAS prediction probability of the image patches obtained by processing through the STASNet model, the STAS semi-quantitative evaluation module further selects the 10 image patches with the highest STAS prediction probability and uses them to calculate the semi-quantitative score Top 10Score; The specific formula is as follows:

[0092] Top10Score = ∑ Top10tiles AI score of STAS tile * Til distance.

[0093] Based on the digitized scanned lung tumor section images, the present invention develops an identification model STASNet for the spread of airspace dissemination lesions in lung cancer. Based on this model, it can assist pathologists in real-time identifying STAS lesions around lung pathological sections during the diagnosis and treatment process, improving the efficiency of clinical diagnosis of STAS and reducing misdiagnosis and missed diagnosis; On the basis of classifying the presence or absence of STAS, it further assists clinicians in completing the semi-quantitative analysis of STAS, indicating the prognosis of patients, and helping doctors formulate personalized diagnosis and treatment plans; As an organic part of the microscope computer imaging system, it is embedded in the microscope image acquisition module to real-time identify STAS lesions in lung cancer images; As an organic component module of the large-scale pathological image recognition model, it generates a structured pathological report to improve the standardization of clinical diagnosis and treatment.

[0094] The above shows and describes the basic principles, main features and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited by the above embodiments. The above embodiments and the descriptions in the specification are only preferred examples of the present invention and do not limit the present invention. Without departing from the spirit and scope of the present invention, the present invention will have various changes and improvements, and these changes and improvements all fall within the scope of the present invention claimed. The scope of protection claimed by the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for recognizing, analyzing, constructing and processing special lung pathological images, characterized in that: The invention comprises a pathological slice processing module, a STASNet model and a STAS scoring module. After a HE pathological slice of lung adenocarcinoma is given and its tumor boundary is annotated, the pathological slice processing module can automatically cut the normal lung tissue outside the tumor boundary in real time, and then the STASNet model identifies the STAS lesions therein to generate a predicted image, and finally the STAS score is calculated by the STAS scoring module; the pathological slice processing module uses a commercial scanner to scan the HE slice into a digital pathological slice image for the HE slice, and determines the tumor boundary based on the annotation of the pathologist; The scanned digital pathology slice images are segmented into local image areas of 256 pixels * 256 pixels using Python scripts, and are divided into tumor images and non-tumor area images based on tumor boundary annotation. For non-tumor images, if the white pixels have R, G, B values ​​greater than 220 and a proportion greater than 90%, they will be placed in the background group; the remaining images containing non-tumor main area tissues will construct a data set and be used for model construction and recognition of subsequent modules; After the STASNet model receives the image obtained by the pathological slice processing module as an input image, the STASNet model introduces a convolutional neural network framework, called the STASNet framework, to construct the module, and completes the training and evaluation of the module weights in the tissue image data set; the STASNet model is optimized and constructed based on the MobileNet V3 framework, and mainly includes the following key components: convolution layer, pooling layer, activation function and fully connected layer; basic components include convolution layer, pooling layer, fully connected layer, activation function, loss function, optimizer, etc., and the specific details are as follows: Convolutional Layer: The convolutional layer is the core of the CNN framework feature extraction. The convolutional layer uses a convolutional kernel with specific features to perform matrix operations on the local area of ​​the image in sequence. The new matrix finally obtained is regarded as a feature map containing a certain type of feature of the original image. A single convolutional layer includes several convolution kernels for extracting different features of the input image matrix, and finally outputs several feature maps containing different features to the pooling layer or convolutional layer network of the next layer. Pooling Layer: Since the number of parameters in the CNN framework is too large, the model introduces a pooling layer to reduce the number of parameters; the pooling layer simplifies the input feature map and obtains the representative feature value of the area by calculating the local area of ​​the Feature Map; the commonly used pooling methods are Max Pooling and Mean Pooling; Fully Connected Layer: The fully connected layer is generally connected to the end of the entire CNN to classify the data. After passing through several convolution and pooling layers, the multi-dimensional abstract features extracted by the model are comprehensively evaluated and output as specific feature tags; several fully connected layers form a fully connected neural network, in which the fully connected layer that accepts input features becomes the input layer, and the fully connected layer that finally outputs several preset classification labels is the output layer. All layers between the input layer and the output layer are hidden fully connected layers; Activation Function: During model training, parameters continuously flow from the input to the output. During this process, the model introduces an activation function to evaluate whether the input result will continue to propagate to subsequent models. MobileNetV3 uses an activation function called Hard Swish, which is a variant of the rectified linear unit ReLU; The mathematical expression of the Hard Swish function is as follows: ； Among them, ReLU6 is a truncated ReLU function, which is defined as follows: ； Loss Function and Optimizer: During the model training process, after the input image is extracted and classified by the model, the deep learning model introduces a loss function to calculate the difference between the classification result and the actual result. After calculating the loss value, the model will apply the optimizer to reversely update and optimize the parameters of each node inside the model, so as to ensure that a lower loss value can be obtained in the next training. The cross entropy function is used as the loss function, and the Adam optimizer is used to optimize the internal parameters of the model; the formula is as follows: ； Where xi is a discrete variable, which represents the predicted value of each type of result in the prediction results in this study; Bneck Block: Based on the stacking of multiple convolutional layers and multiple pooling layers, the core module Bneck Block of MobileNet V3 can be constructed; Different from the traditional convolutional layer, Bneck Block greatly reduces the number of parameters and computational complexity by separating the traditional convolution operation into depth-wise convolution and 1*1 point-wise convolution, thereby improving the model efficiency while maintaining relatively good performance; Bneck Block also enhances the importance weight distinction of model input features by adding a Squeeze-and-Excitation attention mechanism module, and the memory cloth improves model performance; The constructed Backeck Block includes a set of depth-wise convolutions, a set of attention modules SE Block and a set of 1*1 point convolutions; Overall view of the model: Based on the above components, the STASNet model was finally constructed; the overall prediction process is: The input image is processed into a matrix of size 256*256*3 and processed sequentially in 15 groups of Bneck Block matrices to finally generate a matrix of size 8*8*960, which is then transformed into an output matrix of size 1*1*960 through a pooling layer and passed through a fully connected neural network including three fully connected layers to finally output a two-dimensional result, namely [STAS, non-STAS].

2. The method for recognizing, analyzing, constructing and processing special lung pathological images according to claim 1, characterized in that: The training and evaluation of the module weights specifically include the following: The image blocks with real category labels obtained by processing the pathological slice processing module are predicted in the STASNet model, and the loss function is calculated based on the prediction results of the STASNet model and the real results of the image blocks, and the optimizer is used to optimize the weights, and the optimized model is reused for prediction, and the well-trained STASNet framework is obtained by continuously iterating the model node weights; The STASNet framework with perfect weight training can be used to predict STAS lesions in the workflow. The specific workflow is as follows: The STASNet model receives the image blocks obtained after being processed by the pathological slice processing module, and processes them through the trained and perfected STASNet framework, and the final output result is [STAS, non-STAS].

3. The method for recognizing, analyzing, constructing and processing special lung pathological images according to claim 1, characterized in that: After being processed by the pathology slice processing module and the STASNet model, non-tumor image blocks with STAS interpretation results or tumor image blocks with tumor markers are generated. All image blocks are evaluated in the STAS scoring module, and STAS semi-quantification of the entire slice is completed; Specifically, it includes two steps: spatial position calculation and STAS semi-quantitative evaluation; S1, the spatial position calculation module: the image blocks obtained by the pathological slice processing module will contain original spatial coordinates. Based on the original spatial coordinates of the image blocks and the tumor boundary coordinates, the spatial position calculation module uses a Python script to calculate the vertical distances of all image blocks to the original spatial coordinates; S2, the STAS semi-quantitative evaluation module: after the spatial position obtained by the pathological slice processing module and the STAS prediction probability of the image block obtained by the STASNet model, the STAS semi-quantitative evaluation module selects the 10 image blocks with the highest STAS prediction probability and uses them to calculate the semi-quantitative score Top 10 Score; the specific formula is as follows: 。

Images