Skin lesion detection method based on neural network

Abstract

The invention provides a skin lesion detection method based on a neural network, and relates to the field of skin lesion detection, and the method comprises the following steps: S1, data preparation: obtaining a skin lesion image # imgabs0 #, and marking a lesion region in the skin lesion image # imgabs1 #; and S2, model training and optimization: establishing a fusion detection network model # imgabs2 #, inputting the skin lesion image # imgabs3 # into the fusion detection network model # imgabs4 # for training, outputting to obtain a detection result, and optimizing the fusion detection network model # imgabs5 # based on the detection result. Through the method, the following beneficial effects can be achieved: 1, the region extraction network is introduced, and the original lesion picture is preprocessed; secondly, a fusion detection network module is adopted, and a new method is provided for detecting skin lesions more comprehensively and accurately; and thirdly, by constructing a two-stage AI architecture, layering can be carried out according to confidence through sample extraction, high and low confidence sample characteristics are considered, precise navigation is optimized for the model, sample deviation is calibrated by experts, the labeling precision is improved from multiple dimensions, and precise judgment of the model is facilitated.

Description

Technical Field

[0001] The present invention relates to the field of skin lesion detection, and particularly to a skin lesion detection method based on a neural network. Background Art

[0002] Skin lesion detection is an important medical task, usually used for early detection and diagnosis of skin diseases. Traditional skin lesion detection methods mainly rely on the experience of professional doctors, and this method has subjectivity and error. In recent years, skin lesion detection methods based on neural networks have made significant progress in the field of computer-aided medical diagnosis. Through deep learning techniques, especially convolutional neural networks, these methods can efficiently detect skin lesions. With the support of large-scale datasets and accurate annotations, the use of image enhancement techniques improves the robustness of the model. The application of these comprehensive techniques has brought improvements in the accuracy, automation, and real-time performance of skin lesion detection, providing better services for future medical clinical practice.

[0003] In the prior art, the Chinese patent with the publication number CN109493342B discloses "a method for classifying skin disease picture lesion types based on deep learning". This method is designed based on the existing ResNet50 and InceptionV3 networks, and there are certain limitations in the model algorithm, and it cannot well adapt to other types of skin lesions or perform well when facing a wider dataset. Secondly, this method uses specific lesion pictures in training and tuning, resulting in poor performance of the model when facing different types or unseen skin lesions. Summary of the Invention

[0004] Aiming at the deficiencies of the above-mentioned prior art, the technical problem to be solved by the present invention is to provide a skin lesion detection method based on a neural network with high accuracy and strong generalization ability.

[0005] To solve the above technical problem, the technical solution adopted by the present invention is: The present invention provides a skin lesion detection method based on a neural network, including the following steps: S1. Prepare data: Obtain skin lesion images , and label the lesion areas in the skin lesion images ; S2. Model training and optimization: Establish a fusion detection network model , and input the skin lesion images into the fusion detection network model for training, and output the detection results, and optimize the fusion detection network model based on the detection results ; S3. Model testing: Establish a two-level intelligent processing framework, and combine and utilize the fusion detection network model Detect the skin lesion category to obtain the prediction result .

[0006] In the preferred solution, step S2 specifically includes the following steps: S21. Obtain the feature extraction module and the feature reconstruction module, send the skin lesion image to the feature extraction module for processing to obtain the first feature map, and send the first feature map to the feature reconstruction module for processing to obtain the second feature map; S22. The fusion detection network model includes an original feature extraction layer, a fusion feature extraction layer, a fully connected layer, and a softmax layer; Send the skin lesion image to the original feature extraction layer to obtain the original feature map, send the second feature map to the fusion feature extraction layer to obtain the fusion feature map, splice the original feature map and the fusion feature map and send them to the fully connected layer and the softmax layer to obtain the detection result, label and classify the detection result to obtain the classified and labeled sample ; S23. Establish a cloud database, and transmit the classified and labeled samples and the intermediate data and results during the training process of the fusion detection network model to the cloud database for storage; S24. Use the loss function to calculate the difference between the prediction result of the fusion detection network model and the true label , and update the parameters of the fusion detection network model through the backpropagation algorithm to minimize the loss function and optimize the fusion detection network model .

[0007] In the preferred solution, the loss function specifically used in step S24 is: ; wherein, represents the number of lesion categories, represents the th element of the true label, represents the th element of the predicted label of the fusion detection network model , represents the number of samples.

[0008] In the preferred solution, step S3 further includes the following steps: Divide the two-level intelligent processing framework into an intelligent terminal and a central cloud, and interact through the intelligent terminal and the central cloud to detect the skin lesion category; The intelligent terminal is used for the target image of the skin lesion part collected Perform a preliminary detection using the trained fusion detection network model to quickly process the image and obtain a preliminary prediction result and confidence , and dynamically adjust the processing mode of the target image according to the confidence . Upload the target image with a confidence lower than the set threshold to the central cloud for review; wherein the threshold is divided into a local processing threshold and a cloud processing threshold . . .

[0009] In a preferred solution, step S3 specifically includes the following steps: S31. Collect the target image of the skin lesion site through an image acquisition device ; S32. Preprocess the collected target image , including cropping, scaling, and normalization, to obtain the preprocessed target image , and perform image enhancement processing on the preprocessed target image ; S33. Model prediction: Input the preprocessed target image into the trained fusion detection network model and output the prediction result of the target image of the skin lesion site ; ; S34. Sample classification and intelligent extraction: Classify the prediction result to the cloud database according to the confidence ; wherein, when the confidence is, the prediction result is classified into the high-confidence group ; when the confidence is, the prediction result is classified into the low-confidence group .

[0010] In a preferred solution, in step S33, the specific calculation steps of the model prediction are as follows: The intelligent terminal pre-loads the trained fusion detection network model , and when the preprocessed target image is input, use the trained fusion detection network model to process the preprocessed target image Perform feature extraction to obtain the target feature vector ; Let the set of classified and labeled samples in the cloud database be , and the corresponding set of classified and labeled features be , that is ; Calculate the Euclidean distance between the target feature vector and the th classified and labeled feature vector as follows: ; Among them, represents the dimension of the feature vector, represents the index variable for the dimension of the feature vector; Calculate the minimum value of the Euclidean distance between the target image and all classified and labeled samples , and convert the minimum value to the confidence level through inversion as follows: ; Among them, is the number of classified and labeled images in the cloud database; Set the threshold to , and determine the processing mode of the target image based on the comparison result between the confidence level; The processing mode , including the basic processing mode , the advanced processing mode , and the collaborative processing mode ; Set the range of the local processing threshold to 90%, and set the range of the cloud processing threshold to 80%; The formula for the processing mode is as follows: ; Among them, when , that is , then the decision is to directly process on the intelligent terminal and determine the final conclusion; When , that is , then the decision is that the intelligent terminal will send the corresponding target image Uploaded to the central cloud for processing; When , that is , the decision is that the intelligent terminal uploads the difficult-to-recognize part and the remaining part that has been recognized and labeled locally, and the central cloud conducts in-depth analysis on the unrecognizable part and returns the processing result to the intelligent terminal.

[0011] In the preferred solution, in step S34, the specific calculation steps of sample classification and intelligent extraction are as follows: Set the sample ratio extracted from the cloud database for quality control and model optimization to ; Set the initial value of the sample ratio to 5%, and the initial value can be adaptively adjusted dynamically; Set the total number of classified and labeled images in the cloud database to , then the total number of samples extracted . Each time of extraction, according to the number of samples in the current low-confidence group and high-confidence group , dynamically allocate the number of samples extracted from the low-confidence group and high-confidence group ; The high-confidence group adopts a stratified sampling strategy, and stratifies the samples in the high-confidence group according to the key features of skin lesions, and selects lesion type, size, shape, color, and boundary clarity as the stratification dimensions; Set the total number of samples in the high-confidence group to . The number of samples extracted from this layer is: , and it is necessary to satisfy . The fixed ratio total upper limit is ; Within each stratum, assign a unique index ) to the samples; Among them, is the number of samples in this stratum, use a random number generator to generate random integers between 1 and , is the proportion of the number of samples in each stratum in the high-confidence group ; Merge the samples extracted from each stratum to obtain the sample set extracted from the high-confidence group ; The low-confidence group adopts a dynamic sampling ratio adjustment strategy for dynamic sampling, and adjusts the sampling ratio according to the change of sample distribution;​ Let the original sampling ratio be , and the adjusted sampling ratio be: , and ; Among them, is the adjustment factor, is the ratio change; Let the total number of samples in the low confidence group be , the number of samples extracted , assign a unique index to each sample in the low confidence group, use a random number generator to generate random integers between 1 and , and select the samples with index as the sample set extracted from the low confidence group ; Merge the sample set extracted from the high confidence group and the sample set extracted from the low confidence group to obtain the total sample set for verification and model training as: , and upload the total sample set to the central cloud.

[0012] In the preferred solution, the following steps are further included: S4. Sample verification and correction: Given that the prediction result of the fusion detection network model for the extracted samples is: ; Execute the verification program for the extracted total sample set , input the true label annotated by experts, and compare the true label annotated by experts with the prediction result in the total sample set one by one, accurately classify the skin lesions in the samples and evaluate the lesion stage, obtain the analysis result, and determine the set of samples with incorrect prediction results to be corrected as: ; Among them, is the number of error samples. For each sample , the expert determines the true class label and the accurate lesion stage; When the prediction result is inconsistent with the analysis result, execute the correction operation. For each error sample ​Make corrections to obtain the corrected samples, and divide the correction process into category correction and conclusion correction; If the correction process is category correction, then correct the skin pathology category of the misclassified samples in the prediction result ; If the correction process is conclusion correction, then correct the skin lesion conclusion of the misclassified samples in the prediction result ;

[0013] In the preferred solution, the following steps are further included: S5. Model retraining and updating: Use the true results corrected by experts to retrain the fusion detection network model and train the parameters of the fusion detection network model using the cross-entropy loss function and the gradient descent algorithm , obtain the update, and after the training is completed, obtain a new fusion detection network model , and evaluate the new fusion detection network model . The evaluation metrics cover classification performance metrics such as accuracy, recall, precision, and F1 value. Determine whether to update the new fusion detection network model to the smart terminal according to the evaluation metrics.

[0014] In the preferred solution, the following steps are further included in step S5: Let the evaluation metric set be: , and define a model update decision function according to the evaluation metric set ; Set that when the accuracy rate increases by more than 5% and the number of model parameters increases by no more than 10%; If , then determine to update the new fusion detection network model to the smart terminal. Conversely, if , it means not to update; Let the set of smart terminal devices be ; Divide the smart terminals into a test group and a control group ; Update the new fusion detection network model to the test group and automatically perform performance evaluation to obtain the evaluation results. Feed the evaluation results back to the central cloud and compare with the control group to obtain the comparison results. Determine the deployment plan according to the comparison results and continue the next round of sample extraction, verification, correction, and model training and updating cycle.

[0015] The present invention provides a skin lesion detection method based on a neural network, and the following beneficial effects can be achieved through the above method: First, a region extraction network is introduced to preprocess the original lesion images, and this module is expected to make significant progress in improving the accuracy and effectiveness of the lesion region. Second, a fusion detection network module is adopted, which provides a new method for more comprehensively and accurately detecting skin lesions. The design of this module is expected to achieve significant effects in fusing multi-source information and enhancing the model performance. By training the fusion detection network model, it is expected to improve the discriminative ability of the model for skin lesions, making it more adaptable to diversity and complexity. Third, by constructing a two-level AI architecture, it is possible to stratify by confidence through sample extraction, taking into account the characteristics of both high-confidence and low-confidence samples, accurately guiding the model optimization, calibrating the sample deviation by experts, improving the annotation accuracy from multiple dimensions, assisting the model in accurate discrimination, adapting to other types of skin lesions or facing a wider dataset, avoiding a large amount of duplicate or unnecessary data processing in the central cloud, thereby reducing the waste of computing resources, and accurately discriminating in the face of different types or unseen skin lesions. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] The present invention will be further described below with reference to the drawings and embodiments: Figure 1 It is a schematic flowchart of a skin lesion detection method based on a neural network provided by an embodiment of the present application; Figure 2 It is a schematic diagram of a region extraction network module of a skin lesion detection method based on a neural network provided by an embodiment of the present application; Figure 3 It is a schematic diagram of a fusion detection network module of a skin lesion detection method based on a neural network provided by an embodiment of the present application; DETAILED DESCRIPTION OF THE EMBODIMENTS To better understand the purpose, structure, and function of the present invention, the embodiments in the present application and the features in the embodiments can be combined with each other without conflict. The present invention will be described in detail below with reference to the drawings and embodiments.

[0017] Embodiment 1 As Figures 1 to 3 shown, this embodiment provides a skin lesion detection method based on a neural network, including the following steps: S1. The specific steps of data preparation are as follows: S11. Collect image data: Collect a large number of images containing skin lesions, covering different types of skin lesions (such as moles, melanoma, basal cell carcinoma, etc.), different skin color populations, different age groups, and different lesion stages, etc., to ensure the diversity and representativeness of the data.

[0018] Let the collected skin lesion images be ; S12. Data annotation: The lesion areas in the skin lesion images are annotated by professional dermatologists or trained personnel to clarify the benign and malignant categories of each lesion, providing accurate label information for the training of the fusion detection network model .

[0019] S2. The specific steps for training the fusion detection network model are as follows: S21. Preprocessing of the region extraction network module Feature extraction: The skin lesion images are fed into the feature extraction module. First, they pass through a 1×1 ordinary convolution layer, a normalization layer, a Relu activation function, and a 3×3 max pooling layer to obtain an extracted feature map. Then, the extracted feature map is fed into a 3×3 dilated convolution layer, a normalization layer, a Relu activation function, and a 2×2 max pooling layer to obtain a first feature map. This process helps to extract key features from the original skin lesion images to distinguish normal and lesion areas.

[0020] Feature reconstruction: The first feature map is fed into the feature reconstruction module. First, it passes through a 3×3 transposed convolution layer, a normalization layer, a Relu activation function, and a 3×3 average pooling layer to obtain a reconstructed feature map. Then, the reconstructed feature map is fed into a 5×5 dilated convolution layer, a normalization layer, a Relu activation function, and a 2×2 average pooling layer to obtain a second feature map, thereby optimizing and reconstructing the features and improving the characterization quality of the lesion area.

[0021] S22. Processing of the fusion detection network module Original feature extraction Shallow channel feature extraction: The skin lesion images are fed into the shallow channel module, which consists of a 1×1 ordinary convolution layer, a 3×3 ordinary convolution layer, an Elu activation function, a shallow channel attention mechanism, a 5×5 ordinary convolution layer, and a 3×3 global pooling layer. The calculation formula for the shallow channel attention mechanism is: ; The shallow channel feature map is calculated; where represents the element at the i-th row and j-th column of the input feature map of the shallow channel attention mechanism, represents the element at the i-th row and j-th column of the output feature map, , , , are learnable parameters, represents the height of the input feature map, represents the width of the input feature map, represents the Relu activation function, represents the Sigmoid activation function. This process focuses on extracting local features.

[0022] Deep channel feature extraction: Skin lesion image is fed into the deep channel module, which consists of a 3×3 ordinary convolutional layer, a 5×5 ordinary convolutional layer, a Sigmoid activation function, a deep channel attention mechanism, a 7×7 depthwise separable convolutional layer, and a 5×5 adaptive average pooling layer. The calculation formula of the deep channel attention mechanism is: ; The deep channel feature map is calculated; Among them, represents the element at the i-th row and j-th column of the input feature map of the deep channel attention mechanism, represents the element at the i-th row and j-th column of the output feature map, , , , , , are learnable parameters, the element at the k-th channel of the input feature map at the i-th row and j-th column, represents the Relu activation function, represents the Sigmoid activation function, C represents the number of channels of the input feature map, and the deep channel module focuses on extracting more global and abstract features.

[0023] Feature map concatenation: Concatenate the shallow channel feature map and the deep channel feature map to obtain the original feature map to obtain a richer feature expression and improve the fusion detection network model performance.

[0024] Fusion feature extraction The second feature map is fed into the fusion feature extraction layer, first passing through two 3×3 ordinary convolutional layers, a Relu activation function, and a normalization layer.

[0025] Then it passes through the first residual layer, which consists of a 3×3 ordinary convolutional layer, a 1×1 ordinary convolutional layer, an upsampling layer, and a normalization layer, helping to alleviate the gradient vanishing problem and retain detailed information.

[0026] Then it passes through the second residual layer, which consists of a 3×3 ordinary convolutional layer, a 5×5 ordinary convolutional layer, an upsampling layer, and a normalization layer, enhancing the network receptive field.

[0027] Finally, the calculation formula of the high-resolution enhancement layer is: ; Calculate the fused feature map; Among them, represents the pixel value of the output feature map at position and channel ; represents the pixel value of the input feature map at position and channel ; represents the weight of the convolution kernel, represents the pixel value of the input feature map at position and channel ; and respectively represent the offsets of the convolution kernel in the horizontal and vertical directions, represents the channel, represents the number of channels of the input feature map, and respectively represent the radii of the convolution kernel in the horizontal and vertical directions, improving the resolution of the output feature map, accurately locating and depicting the lesion area.

[0028] Classification decision: Concatenate the original feature map and the fused feature map and send them into the fully connected layer and the softmax layer to obtain the detection result, that is, the probability distribution of whether the skin lesion is benign or malignant, and label and classify the detection result to obtain the classified and labeled samples .

[0029] Establish a cloud database to store the classified and labeled samples and the fused detection network model and the intermediate data and results during the training process, and adopt distributed storage technology to ensure the high availability and fast read and write capabilities of the data.

[0030] Among them, each classified and labeled sample carries detailed annotation information, including the true category of the skin lesion, lesion characteristics such as lesion type, size, shape, color, boundary clarity, etc., and the interpretation of the lesion stage.

[0031] S3. Loss function calculation and model optimization The loss function adopted is: ; Among them, represents the number of lesion categories, represents the th element of the true label, represents the th element of the predicted label of the fused detection network model ; represents the number of samples. During the training process, the difference between the prediction result and the true label of the fusion detection network model is calculated according to the loss function, and the parameters of the fusion detection network model are updated through the backpropagation algorithm to minimize the loss function and continuously optimize the fusion detection network model to improve its discrimination ability for skin lesions

[0032] S4. Model testing: Establish a two-level intelligent processing framework and use the trained fusion detection network model to detect the skin lesion category ; Among them, the two-level intelligent processing framework is divided into an intelligent terminal and a central cloud. The skin lesion category is detected through the interaction between the intelligent terminal and the central cloud The intelligent terminal is used to perform preliminary detection on the target image of the skin lesion site and quickly process the image using the trained fusion detection network model to obtain a preliminary prediction result and confidence ; according to the confidence dynamically adjust the processing mode of the target image , and upload the target image with a confidence lower than the set threshold to the central cloud for review. The specific steps are as follows : S41. Collect the target image of the skin lesion site through one of a professional dermatoscope or a high-resolution camera to ensure that the image is clear, complete, and contains sufficient lesion information

[0033] S42. Image preprocessing: Preprocess the collected target image , including operations such as cropping, scaling, and normalization, to make its size and format meet the requirements of the trained fusion detection network model to obtain the preprocessed target image . At the same time, the preprocessed target image can be subjected to image enhancement processing, including rotation, flipping, brightness adjustment, etc., to increase data diversity and improve the generalization ability of the fusion detection network model

[0034] S43. Model prediction: Input the preprocessed target image into the trained fusion detection network model . The model will output the prediction result of the target image of the skin lesion site , and the specific steps are as follows :​​​​​ S431. Local Model Loading and Feature Extraction: The intelligent terminal pre-loads the trained fusion detection network model. When the pre-processed target image is input, the trained fusion detection network model is used to perform feature extraction on the pre-processed target image to obtain the target feature vector. ; Let the set of classified and labeled samples in the cloud database be , and the corresponding set of classified and labeled features be , that is ; S432. Euclidean Distance and Confidence Calculation: Calculate the Euclidean distance between the target feature vector and the th classified and labeled feature vector as follows: ; where represents the dimension of the feature vector, which is used to measure the dimension size of the feature space that reflects the difference between two feature vectors. represents the number of extracted features, and these features together form a -dimensional feature space to describe the characteristics of the image. is an index variable representing the dimension of the feature vector, used to traverse each dimension of the feature vector, taking values from 1 to in sequence. When calculating the Euclidean distance, for each dimension , the element of the target feature vector on this dimension is calculated with the element of the classified and labeled feature vector on the same dimension to obtain the square of the difference . That is, through the traversal of , the difference contributions of the two feature vectors on each dimension can be calculated in sequence, and these differences are squared and accumulated. This enables the calculation of the Euclidean distance to cover all dimensions of the feature vector, thus accurately reflecting the overall similarity or difference between the two feature vectors. Further, when , the square of the difference on the first dimension is calculated, when calculating the square of the difference on the second dimension, and so on, until all dimensions are calculated, and finally the Euclidean distance between the two feature vectors is obtained.

[0035] Based on the Euclidean distance Convert to confidence by inversion The formula is as follows: ; Among them, is the number of classified and labeled images in the cloud database. First, calculate the target image and all classified and labeled samples The minimum value of the Euclidean distance in , and then convert it to confidence by inversion , and then convert it to a percentage form, indicating the target image and the classified and labeled samples The similarity between them. The higher the confidence , the more similar the target image is to the classified and labeled samples , and the higher the confidence of the fusion detection network model in the prediction result .

[0036] S433. Threshold setting and decision-making: Set the threshold to , and determine the processing mode of the target image according to the comparison result between the confidence and the threshold ; ; The processing mode , including the basic processing mode , the advanced processing mode and the collaborative processing mode ; Among them, the threshold is divided into the local processing threshold and the cloud processing threshold ; Specifically, the range of the local processing threshold is set to 90%, and the range of the cloud processing threshold is set to 80%; The formula of the processing mode is as follows: ; Among them, when , that is , the decision is to directly process on the intelligent terminal and determine the final conclusion; Specifically, the smart terminal not only provides a preliminary diagnosis result but also generates a detailed conclusion report. The report content includes the type prediction of lesions such as benign moles and malignant melanomas, the description of lesion characteristics such as size, shape, and color distribution, the comparative analysis with common skin lesions, and corresponding health suggestions, including the time interval for regular reexaminations and precautions in daily life; The conclusion report is presented to patients and doctors in a form with both pictures and texts, enabling them to more intuitively understand the lesion situation. At the same time, the smart terminal uploads the conclusion report and relevant image data to the cloud database for storage, facilitating subsequent query and comparative analysis; When , that is , the decision is that the smart terminal will upload the corresponding target image to the central cloud for processing; Specifically, the smart terminal further preprocesses the image to reduce the data transmission volume and improve the cloud processing efficiency. Further, lossy compression is performed on the image, but it is ensured that the quality of the compressed image can still meet the requirements of the central cloud algorithm. Breakpoint resume and asynchronous transmission technologies are adopted to ensure the reliable transmission of data in an unstable network environment. At the same time, when the smart terminal uploads data, it attaches a detailed record of the preprocessing process and the intermediate results of the fusion detection network model for detection, so that the central cloud can better understand the data source and the preliminary analysis situation, and speed up the cloud processing speed. After receiving complex case data, the central cloud starts a multi-algorithm collaborative review mode. In addition to using a high-accuracy algorithm to perform a detailed analysis of the image, it also organizes multiple deep learning models to simultaneously verify the case and fuse and compare the results. At the same time, the central cloud uses an expert system to review and correct the verification results.

[0037] When , that is , the decision is that the smart terminal will upload the difficult-to-identify part and the remaining part that has been identified and marked locally. The central cloud performs in-depth analysis on the part that cannot be identified and returns the processing result to the smart terminal.

[0038] Specifically, when the smart terminal uploads the part that cannot be identified and the remaining part that has been identified and marked locally, it makes more detailed markings on the identified part. The marking content includes feature points such as the bounding box coordinates of the lesion area, the center point of the color abnormal area suspected of being a lesion, and the turning points of the irregular shape, as well as the preliminary feature description obtained according to the local model. The fusion detection network module is used to perform feature reconstruction and enhancement processing on the difficult case image, highlighting the areas and features where lesions may exist, and improving the recognition accuracy of the central cloud for difficult cases; When the central cloud performs in-depth analysis on difficult cases, it uses multi-modal data fusion technology to comprehensively verify the cases. After the central cloud completes the analysis of difficult cases, it not only returns the diagnosis results, but also provides detailed analysis reports and suggestions. The report includes the confirmed type of the lesion, the comparative analysis with similar cases in the cloud database, the detection results and weights of different algorithms, treatment suggestions, and subsequent monitoring plans. At the same time, according to the characteristics of difficult cases, the central cloud provides targeted model update suggestions for intelligent terminals, such as adding training data for specific types of lesions or adjusting certain parameters of the model, to improve the processing ability of intelligent terminals for similar difficult cases.

[0039] S44. Sample Classification and Intelligent Extraction: According to the confidence level Classify the prediction results into the cloud database; Among them, when the confidence level , then classify the prediction results into the high-confidence group ; When the confidence level , then classify the prediction results into the low-confidence group ; Set the proportion of samples extracted from the cloud database for quality control and model optimization to be ; Let the initial value of the sample proportion be 5%, and the initial value can be adaptively and dynamically adjusted; Let the total number of classified and labeled images in the cloud database be , then the total number of extracted samples . Each time of extraction, according to the number of samples in the current low-confidence group and high-confidence group , dynamically allocate the number of samples extracted from the low-confidence group and high-confidence group . The specific formula steps are as follows: S441. High-confidence group Adopt a stratified sampling strategy, and stratify the samples in the high-confidence group according to the key features of skin lesions, and select the lesion type, size, shape, color, and boundary clarity as the stratification dimensions; The lesion types are classified into different categories such as moles, melanoma, basal cell carcinoma, and squamous cell carcinoma; according to the lesion size, they are divided into different size ranges of small, medium, and large. Further, the small size can be defined as the proportion of the lesion area in the total image area being less than 10%, the medium size being 10% to 30%, and the large size being greater than 30%; according to the shape, they are divided into circular, oval, and irregular shapes, which can be determined by calculating the shape parameters of the lesion area through an image analysis algorithm; according to the color, they are divided into pigmented types, i.e., darker colors, and hypopigmented types, i.e., lighter colors; classified by the boundary clarity as clear, i.e., the boundary of the lesion is distinct from the surrounding normal skin, and blurred, i.e., the boundary is not clear and the transition to the surrounding tissue is smooth; Calculation and sampling of the number of samples in each layer: Calculate the proportion of the number of samples in each stratification in the high-confidence group as: ) Among them, is the total number of stratifications; Let the total number of samples in the high-confidence group be , and the number of samples drawn from this layer is: , ensuring that the number of samples drawn is adapted to the proportion of the samples in the high-confidence group in the total samples, and it is necessary to satisfy , and the fixed proportion total upper limit is .

[0040] Within each stratification, assign a unique index to the samples ) Among them, is the number of samples in this stratification, and use a random number generator to generate random integers between 1 and , and select the samples with the index as the samples drawn from this stratification.

[0041] Combine the samples drawn from each stratification to obtain the final sample set drawn from the high-confidence group .

[0042] S442. For the low-confidence group Dynamic sampling adopts a dynamic adjustment of the sampling ratio strategy, and adjusts the sampling ratio according to the change in the sample distribution; Regularly analyze the distribution changes of the samples in the low-confidence group in different feature dimensions, count the proportion of the number of samples with different lesion types, lesion sizes, shapes, colors, etc., and when it is found that the change in the sample proportion in a certain feature dimension exceeds the set threshold , then adjust the sampling ratio of the samples related to this feature dimension;​ Let the original sampling ratio be , and the adjusted sampling ratio is: , and ; Among them, is the adjustment factor, which is determined according to the actual situation, is the proportion change amount, and it is also ensured that is within a reasonable range; Furthermore, if the proportion of a certain rare lesion type in the new data in the low-confidence group increases to 15% and the change exceeds 20%; Let , then the formula is: , and to prevent sampling overfitting.

[0043] According to the adjusted sampling ratio , samples are drawn from the low-confidence group ; Let the total number of samples in the low-confidence group be , the number of samples drawn , ensure that the total number of samples drawn meets the set ratio, assign a unique index to each sample in the low-confidence group, use a random number generator to generate random integers between 1 and , and select the samples with index as the final sample set drawn from the low-confidence group .

[0044] S443. Sample merging and uploading: Merge the sample set drawn from the high-confidence group and the sample set drawn from the low-confidence group to obtain the total sample set for verification and model training as: ; And upload the total sample set to the central cloud for subsequent verification and processing.

[0045] S5. Sample verification and correction S51. Sample verification: Given that the prediction result of the fusion detection network model for the drawn samples is: ; Execute the verification program for the drawn total sample set , input the true labels annotated by experts, and the true labels annotated by experts​ Compare one by one with the prediction results in the total sample set to accurately classify the skin lesions in the samples and evaluate the lesion stage, obtain the analysis results, and determine the prediction results according to the analysis results The set of samples to be corrected with incorrect predictions is ; where is the number of incorrect samples. For each sample , the expert determines the true class label and the accurate lesion stage. S52. Prediction result correction: When the prediction result

[0046] is inconsistent with the analysis result, perform the correction operation. For each incorrect sample perform the correction to obtain the corrected sample, and the correction process is divided into class correction and conclusion correction; Class correction: Correct the skin pathology class of the incorrect samples in the prediction result ; Specifically, the expert will comprehensively consider various features of the sample image (such as the color, shape, size, texture, etc. of the lesion), the clinical symptoms of the patient (such as whether there is itching, pain, bleeding, etc.), and the comparative analysis with known typical cases and other factors to determine the correct skin pathology class. For example, if the sample image shows an irregular lesion boundary, uneven color, and relatively deep depth, combined with the patient's symptom of rapid growth of the recent lesion, the expert may correct the class originally predicted as a benign nevus to the class of malignant melanoma.

[0047] Conclusion correction: Correct the skin lesion conclusion of the incorrect samples in the prediction result ; Specifically, the expert will make a comprehensive judgment based on the performance characteristics such as the change in the lesion size, the evolution of the morphology, and the change in the relationship with the surrounding tissues of the lesion in the image, the duration of the patient's symptoms, the trend of symptom changes, and other disease course development situations, as well as the relevant clinical examination results such as dermoscopy examination and pathological biopsy results to correct the lesion conclusion. If the lesion shows an increase in area, ulceration, and the patient's symptoms worsen in the image, the expert may correct the sample originally predicted as stage I of the skin lesion to stage II of the skin lesion.

[0048] S6. Model retraining and updating: Use the true results corrected by the expert to retrain the fusion detection network model and continuously train the fusion detection network model using the cross-entropy loss function and the gradient descent algorithm After the training is completed, the new model is updated to the intelligent terminal to achieve continuous improvement of the model and improve the accuracy of subsequent detections, as follows: S61. Model retraining S611. Determine the corrected sample set as: ; Among them, is the number of corrected samples, and its corresponding feature vector set is ; That is, the true result corrected by the expert is: ; Specifically, the feature vector is obtained by using the same feature extraction method as in step S2 when training the model, that is, using the feature extraction part in the previously mentioned fusion detection network model to process the sample image.

[0049] S612. Set the parameters of the fusion detection network model as: ; S613. Define the loss function as: ; Among them, is a loss function that can be selected to suit the skin lesion classification task; For classification problems, the cross-entropy loss function is used, that is, the formula is expressed as follows: ; Among them, for different classification categories, represents the one-hot encoding of the true class label of the sample , is the class probability vector predicted by the model.

[0050] S614. Use the gradient descent algorithm to train the parameters of the fusion detection network model , and the specific steps are as follows: The parameter update formula is: ; Among them, is the learning rate, and its value directly affects the speed and effect of model convergence. If the value is too large, the model may skip the optimal solution during training and cannot converge; if the value is too small, the training process will be extremely slow, consuming too much computing resources and time, while is the gradient of the loss function at the current parameter . The gradient indicates the direction in which the loss function drops fastest under the current parameter state and is the key guidance for model parameter update; When calculating the gradient, the gradients of the parameters of different layers in the fusion detection network model are calculated separately according to the chain rule; For the convolution layer output, furthermore, for the gradient of the convolution layer output with respect to the convolution kernel it can be calculated by the backpropagation algorithm, and then according to the gradient of the loss function with respect to the prediction result it is calculated through the chain rule as: ; Similarly, the gradients of other parameters can be calculated, thus realizing the update of the parameters of the fusion detection network model ; During the training process, in order to further improve the training efficiency and stability, we adopt optimization strategies such as mini-batch gradient descent. The total sample set

[0051] is divided into several mini-batches. Let the mini-batch size be then the total sample set can be divided into mini-batches. Each time, one of these mini-batches is selected, and the gradient is calculated using the mini-batch samples and the parameters are updated. Compared with traditional batch gradient descent, mini-batch gradient descent has a smaller computational amount in each iteration and can complete a parameter update faster, enabling the fusion detection network model to explore the parameter space faster during the training process, while avoiding the problem of excessive gradient noise caused by stochastic gradient descent, that is, calculating the gradient using only one sample each time, ensuring a certain degree of stability.

[0052]

[0053] where is the regularization coefficient; When calculating the gradient, for the L2 regularization term the derivative with respect to the parameter is This derivative will participate in the total gradient calculation, thereby constraining the parameter size during parameter update and avoiding overfitting of the training data by the fusion detection network model

[0053] S62, Model Evaluation and Update Decision After training is completed, conduct a comprehensive evaluation of the new fusion detection network model to evaluate classification performance metrics such as accuracy, recall, precision, and F1 value, as well as metrics such as model complexity and model stability. The specific steps are as follows: Evaluate accuracy: ; Among them, is an indicator function. When , , otherwise it is 0; Evaluate recall: ; Among them, is the number of true positives, is the number of false negatives. The true positives and false negatives can be defined according to the specific lesion category. For the malignant lesion category, represents the number of samples that are actually malignant and are predicted by the model as malignant, represents the number of samples that are actually malignant but are predicted by the model as benign; Evaluate precision: ; Among them, is the number of false positives. For the malignant lesion category, represents the number of samples that are actually benign but are predicted by the model as malignant; Evaluate F1 value: ; Let the evaluation index set be: .

[0054] S622. Define the model update decision function according to the evaluation index set , and determine whether to update the new fusion detection network model to the intelligent terminal; Set that when the accuracy rate increases by more than 5% and the number of parameters of the new fusion detection network model increases by no more than 10%, that is, when the accuracy rate increases by more than 5%, it means that the new fusion detection network model has a significant improvement in classification performance. In an implementable manner, the accuracy rate of the original fusion detection network model is 80%. After training and updating, the accuracy rate of the new fusion detection network model reaches 85% or more as above, indicating that the new fusion detection network model can classify skin lesions more accurately and can provide more reliable diagnostic assistance in practical applications.

[0055] Meanwhile, it is required that the number of model parameters increases by no more than 10%. This is to control the complexity of the model. Because if the number of model parameters increases significantly, overfitting may occur, that is, the model performs well on the training data but its performance deteriorates on new data in actual applications. Moreover, too many parameters will also increase the storage and computing burden on intelligent terminal devices. In an implementable manner, the original model has 100 parameters, and the number of parameters of the updated model does not exceed 110; Only when these two conditions are met simultaneously, , it means that the new fusion detection network model is updated to the intelligent terminal, On the contrary, if the accuracy does not increase by more than 5%, or the number of model parameters increases by more than 10%, then , it means not to update. This is because if the improvement in accuracy is not obvious, the new fusion detection network model after update may not bring actual performance improvement. And if the model complexity grows too fast, it may bring potential risks such as overfitting or device performance problems. Therefore, the model is not updated at this time.

[0056] S63. Model Update and Continuous Optimization S631. If and it is determined to update the new fusion detection network model to the intelligent terminal, a gray release strategy is adopted as follows: Let the set of intelligent terminal devices be ; The devices are divided into a test group and a control group ; First, the new model is deployed to the devices in the test group for piloting. Observe the performance of the new fusion detection network model in the actual application scenario, including whether there are compatibility issues and the processing ability for skin lesion images collected in different environments, and compare with the control group to obtain a comparison result. Determine the deployment plan according to the comparison result. If the pilot on the devices in the test group is successful, that is, the new fusion detection network model performs well in various evaluation indicators and no obvious problems occur, gradually expand the update scope and deploy the new fusion detection network model to more intelligent terminal devices.

[0057] S632. When the intelligent terminal is updated to the new fusion detection network model After that, it automatically conducts a self - detection and performance evaluation once, and feeds back the update results to the central cloud, including information such as the accuracy rate and processing time of the updated model in actual use, so that the central cloud can monitor and manage the update process; Then it continues the next round of sample extraction, verification, correction, and model training update loop, continuously optimizing the model to improve its accuracy, reliability, and generalization ability in skin lesion detection, avoiding the occurrence of overfitting. At the same time, as the model is continuously updated, more accurately labeled samples can be gradually accumulated, further enriching the cloud database, providing a more solid data foundation for subsequent model training and optimization, forming a virtuous cycle optimization process, and continuously improving the performance of the entire skin lesion detection system.

[0058] Embodiment 2 Further illustrating in combination with Embodiment 1, the embodiment of the present application further provides an electronic device, which may include: A memory, a processor, and a computer program stored on the memory and executable on the processor.

[0059] When the processor executes the program, it implements the skin lesion detection method based on neural network provided in the above - mentioned embodiment.

[0060] Furthermore, the electronic device further includes: A communication interface for communication between the memory and the processor.

[0061] The memory is used to store a computer program executable on the processor.

[0062] The memory may include a high - speed RAM memory, and may also include non - volatile memory, such as at least one disk memory.

[0063] If the memory, the processor, and the communication interface are implemented independently, the communication interface, the memory, and the processor can be interconnected through a bus and complete communication with each other. The bus can be an Industry Standard Architecture (ISA) bus, a Peripheral Component Interconnect (PCI) bus, or an Extended Industry Standard Architecture (EISA) bus, etc. The bus can be divided into an address bus, a data bus, a control bus, etc.

[0064] The processor may include one or more processing units. For example, the processor may include an application processor (AP), an Application Specific Integrated Circuit (ASIC), a modem processor, a Central Processing Unit (CPU), an image signal processor (ISP), a controller, a memory, a video codec, a digital signal processor (DSP), a baseband processor, and / or a neural-network processing unit (NPU), etc. Among them, different processing units may be independent devices or integrated in one or more processors. Among them, the controller may be the nerve center and command center. The controller can generate operation control signals according to the instruction operation code and timing signal to complete the control of fetching and executing instructions. A memory may also be provided in the processor for storing instructions and data. In some embodiments, the memory in the processor is a cache memory. This memory can save the instructions or data that the processor has just used or recycled. If the processor needs to use the instruction or data again, it can be directly called from the memory. This avoids repeated accesses, reduces the waiting time of the processor, and thus improves the efficiency of the system.

[0065] A visualization module for displaying images, videos, etc. The visualization module may include a display panel, which may be a liquid crystal display (LCD), an organic light-emitting diode (OLED), an active-matrix organic light-emitting diode (AMOLED), a flexible light-emitting diode (FLED), a Miniled, a MicroLed, a Micro-oLed, a quantum dot light-emitting diode (QLED), etc.

[0066] Optionally, in specific implementation, if the memory, the processor, and the communication interface are integrated on a single chip, the memory, the processor, and the communication interface can complete communication with each other through an internal interface.

[0067] An embodiment of the present application further provides a computer-readable storage medium, on which a computer program is stored, and when the program is executed by a processor, the above-mentioned neural network-based skin lesion detection method is implemented.

[0068] An embodiment of the present application further provides a computer program product, the computer program can run computer instructions, and when the computer instructions are executed by a processor, the above-mentioned neural network-based skin lesion detection method is implemented.

[0069] The logic and / or steps represented in the flowchart or described in other ways herein, for example, can be considered as a definite sequence list of executable instructions for implementing logical functions, and can be specifically implemented in any computer-readable medium for use by an instruction execution system, apparatus, or device (such as a computer-based system, a system including a processor, or other systems that can fetch instructions from the instruction execution system, apparatus, or device and execute the instructions), or in combination with these instruction execution systems, apparatus, or devices. For the purposes of this specification, a "computer-readable medium" can be any device that can contain, store, communicate, propagate, or transport a program for use by or in connection with an instruction execution system, apparatus, or device. More specific examples (non-exhaustive list) of computer-readable media include the following: an electrical connection part (electronic device) having one or N wirings, a portable computer disk cartridge (magnetic device), a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber device, and a portable compact disc read-only memory (CDROM). Additionally, the computer-readable medium can even be paper or other suitable media on which the program can be printed, because the program can be obtained electronically by optically scanning the paper or other media, followed by editing, interpretation, or otherwise processing as appropriate, and then stored in a computer memory.

[0070] The above-mentioned storage medium can be a read-only memory, a disk, or an optical disc, etc. Although the embodiments of the present application have been shown and described above, it can be understood that the above embodiments are exemplary and should not be construed as limiting the present application. Those of ordinary skill in the art can make changes, modifications, substitutions, and variations to the above embodiments within the scope of the present application.

Claims

1. A method for detecting skin lesions based on a neural network, characterized in that: The following steps are involved: S1. Prepare data: Obtain skin lesion images , for skin lesion images The lesion area is marked in the figure; S2. Model training and optimization: Establishing a fusion detection network model , and the skin lesion image Input to the fusion detection network model Conduct training, output the detection results, and optimize the fusion detection network model based on the detection results ; S3. Model testing: Establish a two-level intelligent processing framework and combine it with the fusion detection network model Detect skin lesion categories and get prediction results .

2. The neural network-based skin lesion detection method according to claim 1, characterized in that: Step S2 specifically includes the following steps: S21, obtain a feature extraction module and a feature reconstruction module, and transform the skin lesion image Send the first feature map to the feature reconstruction module for processing to obtain a second feature map; S22. Fusion detection network model Includes original feature extraction layer, fusion feature extraction layer, fully connected layer and softmax layer; Image the skin lesions The original feature map is sent to the original feature extraction layer to obtain the original feature map, the second feature map is sent to the fusion feature extraction layer to obtain the fusion feature map, the original feature map and the fusion feature map are spliced ​​and sent to the full connection layer and the softmax layer to obtain the detection result, the detection result is labeled and classified to obtain the classified labeled sample ; S23. Establish a cloud database to store classified and labeled samples And the fusion detection network model The intermediate data and results during the training process are transmitted to the cloud database for storage; S24. Using loss function to calculate fusion detection network model Prediction results and true labels The difference between them is calculated and the fusion detection network model is updated through the back propagation algorithm. Parameters to minimize the loss function and optimize the fusion detection network model .

3. The neural network-based skin lesion detection method according to claim 2, characterized in that: The loss function specifically used in step S24 is: ； in, represents the number of lesion categories, The true label elements, Represents the fusion detection network model The predicted label of elements, Indicates the sample size.

4. The neural network-based skin lesion detection method according to claim 1, characterized in that: Step S3 also includes the following steps: dividing the two-level intelligent processing framework into an intelligent terminal and a central cloud, and detecting the category of skin lesions through interaction between the intelligent terminal and the central cloud; The intelligent terminal is used to collect the target image of the skin lesion Perform preliminary detection and use the trained fusion detection network model Process the image quickly to get preliminary prediction results and confidence , according to the confidence Dynamically adjust the target image Processing mode , the confidence Below the set threshold The target image Upload to the center’s cloud for review; The threshold is Divide into local processing thresholds and cloud processing thresholds .

5. The neural network-based skin lesion detection method according to claim 4, characterized in that: Step S3 specifically includes the following steps: S31. Capturing a target image of a skin lesion site by an image acquisition device ; S32, the collected target image Perform preprocessing, including cropping, scaling, and normalization, to obtain the preprocessed target image , and the preprocessed target image Perform image enhancement processing; S33, model prediction: the preprocessed target image Input to the trained fusion detection network model and output the target image of the skin lesion area The prediction results ; S34, sample classification and intelligent extraction: based on confidence The predicted results Classify to cloud database; Among them, when the confidence , then the prediction result Classify into high confidence group ; When confidence , then the prediction result Classified into low confidence group .

6. The neural network-based skin lesion detection method according to claim 5, characterized in that: In step S33, the specific calculation steps of the model prediction are as follows: the intelligent terminal preloads the trained fusion detection network model , when the preprocessed target image After input, use the trained fusion detection network model After preprocessing, the target image Perform feature extraction to obtain the target feature vector ; Suppose there are classified and labeled samples in the cloud database The collection is , and its corresponding classified and labeled feature set is ,Right now ; Calculate the target feature vector With Classified and labeled feature vectors The Euclidean distance formula between is as follows: ; in, represents the dimension of the feature vector, Variables represented as indices of the feature vector dimensions; Calculate the target image With all classified and labeled samples Minimum value of the Euclidean distance , and according to the minimum value Convert to confidence by inverting the value The formula is as follows: ； in, is the number of classified and annotated images in the cloud database; Set the threshold to , according to the confidence With threshold The comparison result determines the target image Processing mode ; Processing Mode , including the base processing mode , Advanced Processing Mode and collaborative processing mode ; Local processing threshold The range is set to 90%, and the cloud processing threshold The range is set to 80%; Processing Mode The formula is as follows: ； Among them, when ,Right now , then the decision is to directly process and determine the final conclusion on the intelligent terminal; when ,Right now , then the decision is that the intelligent terminal will send the corresponding target image Upload to the central cloud for processing; when ,Right now , the decision is that the smart terminal uploads the difficult-to-recognize parts and the remaining parts that have been identified and marked locally, and the central cloud performs in-depth analysis on the unrecognizable parts and returns the processing results to the smart terminal.

7. The neural network-based skin lesion detection method according to claim 5 or 6, characterized in that: In step S34, the specific calculation steps of sample classification and intelligent extraction are as follows: Set the sample ratio extracted from the cloud database for quality control and model optimization to ; Assume sample proportion The initial value is 5%, and the initial value can be adaptively adjusted dynamically; Assume that the total number of classified and labeled images in the cloud database is , then the total number of samples drawn , each time a sample is drawn, according to the current low confidence group and high confidence group The number of samples in the low confidence group is dynamically allocated and high confidence group The number of samples drawn from High confidence group A stratified sampling strategy was used to select high-confidence groups based on key features of skin lesions. The samples in were stratified, and the lesion type, size, shape, color, and boundary clarity were selected as stratification dimensions; Set high confidence group The total number of samples 6 in , the number of samples drawn from this layer is: , and must meet , the total upper limit of the fixed ratio is ; Within each stratum, samples are assigned unique indices ); in, The number of samples in this stratum is generated using a random number generator. From 1 to A random integer between , For each stratum in the high confidence group The proportion of samples in The samples extracted from each stratum are combined to obtain the samples from the high confidence group. The sample set extracted from ; Low confidence group Dynamic sampling adopts a strategy of dynamically adjusting the sampling ratio, adjusting the sampling ratio according to changes in sample distribution; Assume the original sampling ratio is , the adjusted sampling ratio is: ,and ; in, is the adjustment factor, is the proportional change; Set low confidence group The total number of samples in is , the number of samples drawn , assign a unique index to each sample in the low confidence group , using a random number generator From 1 to A random integer between , select index as The samples are taken as the sample set drawn from the low confidence group ; From the high confidence group The sample set extracted from and from the low confidence group The sample set extracted from After merging, the total sample set used for verification and model training is: , the total sample set Upload to the central cloud.

8. The neural network-based skin lesion detection method according to claim 7, characterized in that: The following steps are also included: S4. Sample verification and correction: known fusion detection network model The prediction results for the extracted samples are: ; Perform the total sample set drawn The verification procedure uses the real labels annotated by experts as input , the true labels annotated by experts With the total sample set The prediction results in Compare one by one, accurately classify and stage the skin lesions in the samples, obtain the analysis results, and determine the prediction results based on the analysis results The set of incorrect samples to be corrected for: ; in, is the number of error samples, for each sample , experts determine the true category label and accurate lesion staging; When the prediction results When the analysis result is inconsistent, a correction operation is performed for each erroneous sample. Correction is performed to obtain a corrected sample, and the correction process is divided into category correction and conclusion correction; The correction process is category correction, then the prediction result Correction was performed on the skin pathology category of the erroneous samples; The correction process is the conclusion correction, then the prediction result Correction was performed on the conclusion of skin lesions in the erroneous samples.

9. The neural network-based skin lesion detection method according to claim 8, characterized in that: The following steps are also included: S5. Model retraining and updating: using the real results corrected by experts Fusion Detection Network Model Retrain and use the cross entropy loss function and gradient descent algorithm to train the fusion detection network model Parameters , get updated, after training is completed, get a new fusion detection network model , and the new fusion detection network model Evaluation: the evaluation indicators include classification performance indicators such as accuracy, recall, precision and F1 value, and the new fusion detection network model is determined based on the evaluation indicators. Update to smart terminal.

10. The neural network-based skin lesion detection method according to claim 9, characterized in that: Step S5 also includes the following steps: Assume that the evaluation index set is: , according to the evaluation index set Define the model update decision function ; It is set that when the accuracy rate increases by more than 5% and the number of model parameters increases by no more than 10%; like , then determine the new fusion detection network model Update to smart terminal, otherwise, if Indicates no update; Assume that the set of intelligent terminal devices is ; Divide smart terminals into test groups and control group ; The new fusion detection network model Update to test group And automatically perform performance evaluation, get the evaluation results, feed the evaluation results back to the central cloud, and compare them with the control group Compare and obtain the comparison results, determine the deployment plan based on the comparison results, and continue the next round of sample extraction, verification, correction and model training update cycle.

Images