Deep learning-based cervical immunohistochemical lesion grading method, system and device

Abstract

This invention provides a deep learning-based method, system, and device for grading cervical immunohistochemical lesions, relating to the interdisciplinary field of deep learning and medicine. Through data acquisition and preprocessing, digital images of cervical immunohistochemical slides are obtained, acquiring the RGB matrix and corresponding image data. Deep learning is used to obtain corresponding data features, which are then fused using a Transformer model. A reference feature library is constructed, and error thresholds for different CIN levels are preset. The total error between the input sample and the corresponding reference feature in each sub-region is calculated to determine the lesion level of each sub-region. Evaluation is then performed to obtain the grading result for the current region to be graded. By introducing multiple modal information such as RGB image features and image spatial features, combined with a dynamic attention mechanism and a multi-scale error evaluation strategy, automated and refined grading of cervical lesions is achieved.

Description

Technical Field

[0001] This invention relates to the interdisciplinary field of deep learning and medicine, and in particular to a method, system and device for grading cervical immunohistochemical lesions based on deep learning. Background Technology

[0002] Cervical cancer is one of the most common and deadliest malignant tumors among women worldwide. Its development typically involves a progressive process from cervical intraepithelial neoplasia (CIN) to invasive carcinoma. Early detection and accurate grading of CIN lesions are crucial for clinical intervention and prevention of cervical cancer.

[0003] In clinical pathological diagnosis, immunohistochemical staining techniques are widely used to assist in assessing the severity of cervical lesions. Among them, the p16 / Ki67 double staining method has become an important tool due to its good specificity and sensitivity for high-grade lesions (such as CIN2 and above). However, traditional immunohistochemical image analysis relies on manual slide reading, which has problems such as high subjectivity, low efficiency, and poor consistency, especially in primary healthcare institutions.

[0004] In recent years, artificial intelligence and deep learning technologies have made significant progress in the field of medical image recognition and analysis, providing new approaches to solving the aforementioned problems. The application of multimodal fusion methods, such as convolutional neural networks (CNNs), Transformers, and point cloud processing, has made it possible to automatically identify lesion regions in tissue slices, extract key features, and perform accurate grading.

[0005] Current research has attempted to classify cervical lesions using single-modal models (such as those based solely on RGB images), but these models generally suffer from the following limitations: 1. Insufficient information utilization: They ignore three-dimensional structural information such as cell spatial distribution and nuclear morphology; 2. Insufficient generalization ability: Due to factors such as staining differences and different image resolutions, the models have poor robustness; 3. Weak interpretability: Most models are "black box" mechanisms, which are difficult to meet the diagnostic criteria required by clinicians.

[0006] Therefore, there is an urgent need for an intelligent diagnostic system that integrates multiple image features, has high-precision grading capabilities, and is highly interpretable, in order to improve the automation level and reliability of cervical lesion diagnosis. Summary of the Invention

[0007] The main objective of this invention is to provide a method, system, and device for grading cervical immunohistochemical lesions based on deep learning, thereby solving at least one technical problem involved in the prior art.

[0008] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is: a deep learning-based method for grading cervical immunohistochemical lesions, comprising the following steps: Data acquisition and preprocessing: Multi-source data acquisition and preprocessing were performed on cervical pathological immunohistochemical sections to obtain high-resolution digital images. After image denoising, color correction and other preprocessing steps, the corresponding RGB pixel matrix was extracted. The RGB matrix was then converted into image data to form an initial region to be analyzed containing tissue morphology information. The initial region was then divided into several sub-regions, and the corresponding RGB matrix and image feature data of each sub-region were extracted as the basis for subsequent classification and grading analysis. Multimodal feature extraction employs an EfficientNet-B4 network to extract features from the RGB matrix, and a ResNet-50 network combined with a Feature Pyramid Network (FPN) to extract spatial features from the image. The two feature types are then fused using a Transformer model, and an attention mechanism is used to dynamically adjust the weights of each feature type to obtain the fused feature representation. ; The weight matrix W is dynamically adjusted through an attention mechanism. Error calculation and hierarchical decision-making: A reference feature library is constructed, and RGB error thresholds, image feature error thresholds, and point cloud feature error thresholds are preset for different CIN levels. The total error between the input sample and the corresponding reference feature in each sub-region is calculated, and the formula for calculating the total error is as follows: ; In the formula, and These are the corresponding weight coefficients; Based on multi-scale error analysis and a weighted voting mechanism, the lesion level of each sub-region is determined as CIN1, CIN2, CIN3, or requires review. The grading rules are as follows: ; Obtain the classification results, acquire the classification judgment results of each sub-region, evaluate them, obtain the classification results of the current region to be classified, and perform visualization processing.

[0009] In the preferred scheme, the classification result of each sub-region is obtained, evaluated, and the current classification result includes: The ratio of the number of grades in each sub-region to the total number of grades in the statistical sub-regions, with the grade containing the largest ratio being the overall lesion grade of the current region to be graded.

[0010] In a preferred embodiment, the data preprocessing includes magnifying the digital image of the cervical immunohistochemical section and performing p16 / Ki67 double staining; After standardized staining, the Macenko method was used to eliminate staining bias. Multiple scale patches are then extracted and combined with nuclear density maps to generate the final input data.

[0011] In a preferred embodiment, the visualization process includes providing a heatmap to display key decision-making areas, a radar chart of feature error distribution to show multi-dimensional error comparisons, and marking of three-dimensional abnormal areas.

[0012] In a preferred embodiment, the multi-scale patches extraction includes extracting image regions at multiple magnification levels, including three scales: 0.5x, 1.0x, and 2.0x.

[0013] In the preferred embodiment, the feature fusion module adopts a Transformer architecture and dynamically adjusts the weight allocation of each modality feature through an attention mechanism; The dynamic weight allocation module includes a neural network submodule, which consists of a fully connected layer and a Softmax activation function, used to calculate the weighting coefficients for different feature modes.

[0014] In the preferred embodiment, the error threshold setting includes RGB error thresholds and image feature error thresholds for different lesion levels, with the error thresholds decreasing for each level, and CIN3 setting the minimum error tolerance.

[0015] In the preferred embodiment, the total error calculation formula includes learnable parameters α, β, and γ, which correspond to the weighting coefficients of RGB, image features, and point cloud features, respectively, and α=0.4 and β=0.6.

[0016] A deep learning-based cervical immunohistochemical lesion grading system includes: The data acquisition and preprocessing module is used to acquire and preprocess multi-source data from cervical pathological immunohistochemical sections to obtain high-resolution digital images. After completing preprocessing steps such as image denoising and color correction, the corresponding RGB pixel matrix is ​​extracted. The RGB matrix is ​​then converted into image data to form an initial region to be analyzed containing tissue morphology information. The initial region is then divided into several sub-regions, and the corresponding RGB matrix and image feature data in each sub-region are extracted as the basis for subsequent classification and grading analysis. The multimodal feature extraction module uses the EfficientNet-B4 network to extract features from the RGB matrix, ResNet-50 combined with the Feature Pyramid Network (FPN) to extract spatial features from the image, and PointNet++ to extract structural features from the point cloud data. The two types of features are fused through the Transformer model, and the weight allocation of each feature type is dynamically adjusted using an attention mechanism to obtain the fused feature representation. The error calculation and hierarchical decision module is used to construct a reference feature library and preset RGB error thresholds, image feature error thresholds, and point cloud feature error thresholds for different CIN levels. It calculates the total error between the input sample and the corresponding reference feature, where the formula for calculating the total error is as follows: ; In the formula, and The values ​​in the middle are the corresponding weight coefficients; Based on multi-scale error analysis and weighted voting mechanism, the lesion grade is determined as CIN1, CIN2, CIN3 or requires review. The grading rules are as follows: ; The grading results module is used to obtain the grading determination results for each sub-region, evaluate them, obtain the current grading results, and perform visualization processing.

[0017] An electronic device, comprising a memory and a processor; The memory is used to store computer programs; The processor is configured to implement the deep learning-based cervical immunohistochemical lesion grading method when executing the computer program.

[0018] This invention proposes a deep learning-based multimodal fusion cervical immunohistochemical lesion grading technology. Through data acquisition and preprocessing, digital images of cervical immunohistochemical slides are obtained, and an RGB matrix is ​​acquired. The RGB matrix is ​​then transformed to obtain corresponding image data and point cloud data. An EfficientNet-B4 network is used to extract features from the RGB matrix, and a ResNet-50 network combined with a Feature Pyramid Network (FPN) is used to extract spatial features from the image. The two feature sets are fused using a Transformer model. A reference feature library is then constructed, and RGB error thresholds and image feature error thresholds for different CIN levels are preset. The total error between the input sample and the corresponding reference feature in each sub-region is calculated. Based on multi-scale error analysis and a weighted voting mechanism, the lesion level of each sub-region is determined and evaluated to obtain the grading result for the current region to be graded. By introducing multiple modal information such as RGB image features and image spatial features, combined with a dynamic attention mechanism and a multi-scale error evaluation strategy, automated and refined grading of cervical lesions is achieved, providing a visual interpretation output. This technology has significant clinical application value and promising prospects for technological promotion.

[0019] Technical effects: 1. Improve diagnostic accuracy: Through multimodal collaboration and the interaction between the RGB matrix and image features, the limitations of single-modality approaches are overcome, thereby improving the accuracy of the data.

[0020] 2. Enhanced small lesion detection capability: The introduction of multi-scale patch extraction and progressive grading strategy enables the model to capture tiny lesion areas at different magnification levels, improving the sensitivity to early or localized lesions. Combined with point cloud structural features, it enhances the ability to perceive image information such as cell nucleus density and arrangement, which helps to identify small lesion areas with irregular shapes.

[0021] 3. The model's robustness and generalization ability have been improved. A reference feature library for pathologically confirmed samples has been constructed, and differential error thresholds have been set for each level to ensure clear and controllable grading boundaries. A complete AI solution for the automated analysis of cervical immunohistochemical images is provided, suitable for large-scale screening scenarios. Attached Figure Description

[0022] The present invention will be further described below with reference to the accompanying drawings and embodiments: Figure 1 This is a flowchart of the lesion grading method of the present invention; Figure 2 This is a schematic diagram illustrating the overall execution of the lesion grading method of the present invention; Figure 3 This is a panoramic view of the cell sample to be tested in this invention; Figure 4 This invention provides a lesion grading map; Figure 5 This is the original diagram of the invention, as well as the effect diagrams of positive cell identification, basal layer identification, and distance measurement for obtaining lesion grading. Detailed Implementation

[0023] Example 1 like Figure 1-5 As shown, a deep learning-based method for grading cervical immunohistochemical lesions includes the following steps: S1: Data Acquisition and Preprocessing. Multi-source data acquisition and preprocessing are performed on cervical pathological immunohistochemical sections to obtain high-resolution digital images. After image denoising, color correction and other preprocessing steps, the corresponding RGB pixel matrix is ​​extracted. The RGB matrix is ​​then converted into image data to form an initial region to be analyzed containing tissue morphology information. The initial region is divided into several sub-regions, and the corresponding RGB matrix and image feature data in each sub-region are extracted as the basis for subsequent classification and grading analysis.

[0024] S2: Multimodal feature extraction. The EfficientNet-B4 network is used to extract features from the RGB matrix, and the ResNet-50 network combined with the Feature Pyramid Network (FPN) is used to extract spatial features from the image. The two feature types are then fused using a Transformer model, and an attention mechanism is used to dynamically adjust the weight distribution of each feature type to obtain the fused feature representation. ; The weight matrix W is dynamically adjusted through an attention mechanism.

[0025] S3: Error Calculation and Hierarchical Decision Making. A reference feature library is constructed, and RGB error thresholds and image feature error thresholds for different CIN levels are preset. The total error between the input sample and the corresponding reference feature in each sub-region is calculated using the following formula: ; In the formula, , These are the RGB matrix eigenvectors, the image space eigenvectors, and... , The values ​​in are the weight coefficients of the corresponding RGB matrix eigenvectors and image space eigenvectors, respectively. This is a vector concatenation operation.

[0026] Based on multi-scale error analysis and a weighted voting mechanism, the lesion level of each sub-region is determined as CIN1, CIN2, CIN3, or requires review. The grading rules are as follows: ; S4: Obtain the grading results, acquire the grading judgment results of each sub-region, evaluate them, obtain the grading results of the current region to be graded, and perform visualization processing.

[0027] In this embodiment, data acquisition and preprocessing are performed, followed by multimodal feature extraction using a deep learning model and feature fusion, which improves the accuracy of the data. Then, error calculation and hierarchical decision-making are conducted to construct a reference feature library. RGB error thresholds and image feature error thresholds for different CIN levels are preset, and the total error between the input sample and the corresponding reference feature in each sub-region is calculated. This dynamic error threshold mechanism improves the stability of fuzzy lesion classification and increases the overall classification accuracy. Through the deep integration of deep learning and multimodal feature analysis, the accuracy, efficiency, and standardization of cervical lesion grading are significantly improved, demonstrating high clinical practical value.

[0028] The theoretical basis of this embodiment is that the RGB matrix is ​​the most original pixel-level representation of an image, containing color and texture information and providing color distribution information (such as immunohistochemical staining intensity).

[0029] Image features are typically high-level semantic features extracted from the RGB matrix through convolution operations, providing information on organizational structure and morphology.

[0030] Furthermore, point cloud features can be fused, which are derived from RGB-D data or 3D structural information reconstructed from RGB images, providing 3D spatial distribution and topological relationships.

[0031] Structural features are extracted from point cloud data using PointNet++; the three types of features are fused using a Transformer model, and the weights of each feature type are dynamically adjusted using an attention mechanism to obtain the fused feature representation. ; The weight matrix W is dynamically adjusted through an attention mechanism.

[0032] In the preferred scheme, the classification result of each sub-region is obtained, evaluated, and the current classification result includes: The ratio of the number of grades in each sub-region to the total number of grades in the statistical sub-regions, with the grade containing the largest ratio being the overall lesion grade of the current region to be graded.

[0033] In the preferred embodiment, data preprocessing includes magnifying the digital images of cervical immunohistochemical sections and performing p16 / Ki67 double staining; After standardized staining, the Macenko method was used to eliminate staining bias. Multiple scale patches are then extracted and combined with nuclear density maps to generate the final input data.

[0034] In the preferred embodiment, visualization processing includes providing a heatmap to display key decision-making areas, a radar chart of feature error distribution to show multi-dimensional error comparisons, and marking of three-dimensional abnormal areas.

[0035] In this embodiment, heatmap generation includes defining the output of the last convolutional layer of the CNN, obtaining the category score gradient, and thus generating a heatmap to highlight key lesion areas of interest to the model. Three-dimensional abnormal region labeling uses image feature data and the corresponding prediction error distribution, combined with a set error threshold, to obtain an abnormal point set, helping doctors understand abnormal changes in the spatial distribution of cell nuclei. The feature error distribution radar chart can visually demonstrate the degree of matching between RGB and image features.

[0036] like Figure 3 and Figure 4 As shown, this is a real-world use case in this embodiment (CIN3 count / total number of suspicious areas = 0.707, very serious level).

[0037] In the preferred scheme, multi-scale patch extraction includes extracting image regions at multiple magnification levels, including 0.5x, 1.0x, and 2.0x. The weighted average expression for the errors at different magnification levels is as follows: ; In the formula, Es is the error value calculated at the magnification factor s, and ws=[0.2,0.5,0.3] is the weight of each scale.

[0038] In the preferred scheme, the feature fusion module adopts the Transformer architecture and dynamically adjusts the weight allocation of each modality feature through an attention mechanism; The dynamic weight allocation module contains a neural network submodule, which consists of fully connected layers and a Softmax activation function, used to calculate the weighting coefficients for different feature modes.

[0039] In the preferred embodiment, the error threshold settings include RGB error thresholds and image feature error thresholds for different lesion levels, with the error thresholds decreasing for each level, and CIN3 setting the minimum error tolerance.

[0040] In this embodiment, the reference feature library is constructed as follows: Collect 1000 pathologically confirmed samples (CIN1: 5000, CIN2: 3000, CIN3: 2000). Calculate the mean of the features for each class of samples as a reference feature: .

[0041] In this embodiment, the error threshold used after calculation is shown in Table 1.

[0042] Based on theoretical and empirical values, α=0.4, β=0.35, and γ=0.25 are learnable parameters.

[0043] Table 1 Dynamic Error Threshold Table

[0044] In the preferred embodiment, the total error calculation formula includes learnable parameters α and β, which correspond to the weighting coefficients of RGB and image features, respectively, and α=0.4 and β=0.6.

[0045] For each candidate region, RGB matrix feature enhancement is performed, using the following formula: ; in, ⊕ indicates element-wise multiplication, and ⊕ indicates feature concatenation; the enhanced RGB matrix is ​​processed using a pyramid pooling module: regional feature encoding is adopted, and the regional features are input into the classification head.

[0046] Furthermore, the weight matrix W is dynamically adjusted through an attention mechanism, and the output of the dynamic weight network is: ; Specifically, it can be elaborated as follows: ; ; The above formula outputs three types of weight values, which are normalized and then used for weighted fusion of the three modal features.

[0047] like Figure 5 As shown, the images are the original images and the images after lesion grading in this embodiment. It can be seen that the technical solution of this embodiment effectively identifies positive cells, basal layer, and distance measurement, thereby achieving accurate grading.

[0048] In this embodiment, differentiated error tolerance and multi-scale dynamic error thresholds are set for different magnifications; multiple data features are integrated into cervical pathology analysis; adaptive weight allocation is adopted, and the contribution of each modality feature is dynamically adjusted through an attention mechanism; and a progressive grading strategy is adopted, first coarse segmentation and then fine classification, which improves the detection rate of small lesions.

[0049] Example 2 To further illustrate with reference to Example 1, a deep learning-based cervical immunohistochemical lesion grading system is provided, comprising: The data acquisition and preprocessing module is used to acquire and preprocess multi-source data from cervical pathological immunohistochemical sections to obtain high-resolution digital images. After completing preprocessing steps such as image denoising and color correction, the corresponding RGB pixel matrix is ​​extracted. The RGB matrix is ​​then converted into image data to form an initial region to be analyzed containing tissue morphology information. The initial region is then divided into several sub-regions, and the corresponding RGB matrix and image feature data in each sub-region are extracted as the basis for subsequent classification and grading analysis. The multimodal feature extraction module uses the EfficientNet-B4 network to extract features from the RGB matrix and the ResNet-50 combined with the Feature Pyramid Network (FPN) to extract spatial features from the image. The two types of features are fused through the Transformer model, and the attention mechanism is used to dynamically adjust the weight distribution of each feature type to obtain the fused feature representation.

[0050] The error calculation and hierarchical decision module is used to construct a reference feature library and preset RGB error thresholds and image feature error thresholds for different CIN levels, and calculate the total error between the input sample and the corresponding reference feature.

[0051] Based on multi-scale error analysis and weighted voting mechanism, the lesion grade was determined to be CIN1, CIN2, CIN3 or required to be reviewed.

[0052] The grading results module is used to obtain the grading determination results for each sub-region, evaluate them, obtain the current grading results, and perform visualization processing.

[0053] An electronic device including a memory and a processor.

[0054] Memory is used to store computer programs.

[0055] A processor, used to implement a deep learning-based cervical immunohistochemical lesion grading method as described in Example 1 when executing a computer program.

[0056] The above embodiments are merely preferred technical solutions of the present invention and should not be considered as limitations on the present invention. The scope of protection of the present invention should be limited to the technical solutions described in the claims, including equivalent substitutions of the technical features described in the claims. That is, equivalent substitutions and improvements within this scope are also within the scope of protection of the present invention.

Claims

1. A deep learning-based method for grading cervical immunohistochemical lesions, characterized in that, Includes the following steps: Data acquisition and preprocessing: Multi-source data acquisition and preprocessing were performed on cervical pathological immunohistochemical sections to obtain high-resolution digital images. After image denoising, color correction and other preprocessing steps, the corresponding RGB pixel matrix was extracted. The RGB matrix was then converted into image data to form an initial region to be analyzed containing tissue morphology information. The initial region was then divided into several sub-regions, and the corresponding RGB matrix and image feature data of each sub-region were extracted. Multimodal feature extraction employs an EfficientNet-B4 network to extract features from the RGB matrix, and a ResNet-50 network combined with a Feature Pyramid Network (FPN) to extract spatial features from the image. The two feature types are then fused using a Transformer model, and an attention mechanism is used to dynamically adjust the weights of each feature type to obtain the fused feature representation. ； in, , These are the RGB matrix eigenvectors and the image space eigenvectors, respectively. This is a vector concatenation operation; Features are fused using a Transformer model; Error calculation and hierarchical decision-making: A reference feature library is constructed, and RGB error thresholds and image feature error thresholds at different CIN levels are preset. The total error between the input sample and the corresponding reference feature in each sub-region is calculated, and the formula for calculating the total error is as follows: ； In the formula, and The values ​​in are the weight coefficients of the corresponding RGB matrix feature vector and image space feature vector, respectively. Based on multi-scale error analysis and a weighted voting mechanism, the lesion level of each sub-region is determined as CIN1, CIN2, CIN3, or requires review. The grading rules are as follows: ； In the formula, The lesion grade, This represents the final error after multi-scale fusion. Obtain the classification results, acquire the classification judgment results of each sub-region, evaluate them, obtain the classification results of the current region to be classified, and perform visualization processing.

2. The deep learning-based cervical immunohistochemical lesion grading method according to claim 1, characterized in that, Obtain the classification result for each sub-region, evaluate it, and obtain the current classification result, which specifically includes: The ratio of the number of grades in each sub-region to the total number of grades in the statistical sub-regions, with the grade containing the largest ratio being the overall lesion grade of the current region to be graded.

3. The deep learning-based cervical immunohistochemical lesion grading method according to claim 1, characterized in that, The data preprocessing includes magnifying the digital images of cervical immunohistochemical sections and performing p16 and Ki67 double staining. After standardized staining, the Macenko method was used to eliminate staining bias. Multiple scale patches are then extracted and combined with nuclear density maps to generate the final input data.

4. The deep learning-based cervical immunohistochemical lesion grading method according to claim 1, characterized in that, The visualization process includes providing heatmaps to display key decision-making areas, radar charts to show multi-dimensional error comparisons of feature error distribution, and marking of three-dimensional abnormal areas.

5. A deep learning-based cervical immunohistochemical lesion grading method according to claim 3, characterized in that, Multi-scale patches extraction involves extracting image regions at multiple magnification levels, including 0.5x, 1.0x, and 2.0x.

6. The deep learning-based cervical immunohistochemical lesion grading method according to claim 1, characterized in that, The two features are fused through a Transformer model. The Transformer architecture is used, and the weight distribution of each modality feature is dynamically adjusted through an attention mechanism. The dynamic weight allocation includes a neural network submodule consisting of fully connected layers and a Softmax activation function, which is used to calculate the weighting coefficients for different feature modes.

7. The deep learning-based cervical immunohistochemical lesion grading method according to claim 1, characterized in that, Error threshold settings include RGB error thresholds and image feature error thresholds for different lesion levels, with each level having a decreasing error threshold. CIN3 sets the minimum error tolerance.

8. The deep learning-based cervical immunohistochemical lesion grading method according to claim 1, characterized in that, The total error calculation formula includes learnable parameters α and β, where α and β correspond to the weight coefficients of the RGB matrix feature vector and the image space feature vector, respectively, and α = 0.4 and β = 0.

6.

9. A deep learning-based cervical immunohistochemical lesion grading system, characterized in that, include: The data acquisition and preprocessing module is used to acquire and preprocess multi-source data from cervical pathological immunohistochemical sections to obtain high-resolution digital images. After completing preprocessing steps such as image denoising and color correction, the corresponding RGB pixel matrix is ​​extracted. Then, the RGB matrix is ​​converted into image data to form an initial region to be analyzed containing tissue morphology information. The current initial region is divided into several sub-regions, and the corresponding RGB matrix and image feature data in each sub-region are extracted respectively. The multimodal feature extraction module uses the EfficientNet-B4 network to extract features from the RGB matrix and the ResNet-50 network combined with the Feature Pyramid Network (FPN) to extract spatial features from the image. The two feature types are then fused using a Transformer model, and an attention mechanism is used to dynamically adjust the weights of each feature type to obtain the fused feature representation. ； in, , These are the RGB matrix eigenvectors and the image space eigenvectors, respectively. This is a vector concatenation operation; Features are fused using a Transformer model; The error calculation and hierarchical decision module is used to construct a reference feature library and preset RGB error thresholds and image feature error thresholds for different CIN levels. It calculates the total error between the input sample and the corresponding reference feature, where the formula for calculating the total error is as follows: ； In the formula, and The values ​​in are the weight coefficients of the corresponding RGB matrix eigenvectors and image space eigenvectors, respectively. Based on multi-scale error analysis and weighted voting mechanism, the lesion grade is determined as CIN1, CIN2, CIN3 or requires review. The grading rules are as follows: ； In the formula, The lesion grade, This represents the final error after multi-scale fusion. The grading results module is used to obtain the grading determination results for each sub-region, evaluate them, obtain the current grading results, and perform visualization processing.

10. An electronic device, characterized in that, Including memory and processor; The memory is used to store computer programs; The processor is configured to, when executing the computer program, implement a deep learning-based cervical immunohistochemical lesion grading method as described in any one of claims 1 to 8.

Images