An online classification method and system for breast cancer based on artificial intelligence
By preprocessing and building a lightweight network model in the cloud service center, and combining fuzzy ensemble theory to classify breast cancer images, the problems of breast cancer image transmission time and delay in diagnosis and treatment are solved, and efficient and accurate online diagnosis of breast cancer is achieved. Ensure that patients in remote areas receive timely medical support.
Patent Information
- Application Number
- CN202411209582.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-30
- Publication Date
- 2024-11-22
- Estimated Expiration
- 2044-08-30
AI Technical Summary
The large size of breast cancer pathological image files makes it take a long time to transmit, and professional doctors cannot diagnose and treat online in time, making it difficult for patients in remote areas to obtain timely medical support.
Using the online classification method of breast cancer based on artificial intelligence, features are extracted through cloud service center preprocessing and convolutional neural network, lightweight network EfficientNet-b4 model is constructed and feature fusion layer is added, and label fuzzy processing is combined with fuzzy set theory. Ensure the accuracy and security of image classification.
It significantly shortens the image transmission time, improves the efficiency and accuracy of breast cancer diagnosis, ensures that patients in remote areas can obtain the same level of medical services as cities, and improves the accuracy of the classification model through timed updates.
Smart Images

Figure CN118711838B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of neural network technology, and in particular to an artificial intelligence-based online breast cancer classification method and system. Background Art
[0002] For fatal diseases such as cancer, early diagnosis is crucial to saving patients' lives. Although online diagnosis and treatment systems are currently in use, breast cancer pathology image files are usually large in size, which makes the transmission process time-consuming. In addition, even if the images can be successfully transmitted, professional doctors may not be able to diagnose and treat online at any time, and this waiting may miss the best time for treatment for patients. Therefore, it is particularly important to use advanced information and communication technologies to support breast cancer diagnosis in remote areas.
[0003] With the rapid development of 5G and artificial intelligence technologies, high-quality remote electronic health diagnosis is also developing. The high speed and low latency characteristics of 5G networks significantly shorten the image transmission time, while the deep learning capabilities of artificial intelligence can assist doctors in rapid and accurate classification. These technological advances not only improve diagnostic efficiency, but also bring timely medical support to patients in remote areas, ensuring that they can receive the same level of medical services as in cities. Summary of the invention
[0004] In view of the deficiencies in the prior art, the present invention provides an artificial intelligence-based online breast cancer classification method and system, which aims to solve the problems in the background technology.
[0005] To achieve the above object, the present invention provides the following technical solution: an online classification method for breast cancer based on artificial intelligence, comprising the following steps:
[0006] Step S1: Obtain breast cancer diagnosis and treatment data from different central hospitals and transmit them to the cloud service center; the diagnosis and treatment data includes diagnosis and treatment images and diagnosis and treatment results of the diagnosis and treatment images;
[0007] Step S2: pre-processing the breast cancer diagnosis and treatment data of the central hospital through the cloud service center;
[0008] Step S3: using a convolutional neural network to extract features from the pre-processed breast cancer diagnosis and treatment data of the central hospital;
[0009] Step S4: construct a classification model, use the features extracted in step S3 to train the classification model, and encrypt the weights of the trained classification model through the cloud service center to obtain the optimal classification model, and transmit the optimal classification model to the edge server;
[0010] The classification model is composed of the convolutional neural network in step S3 and the lightweight network EfficientNet-b4, and a feature fusion layer is added to the lightweight network EfficientNet-b4, and the output number of the last connection layer of the lightweight network EfficientNet-b4 is set;
[0011] Step S5: The collaborative hospital uploads the medical images to the edge server, blurs the labels of the medical images uploaded by the collaborative hospital, and then classifies the medical images with blurred labels through the optimal classification model.
[0012] Furthermore, a feature fusion layer is added to the lightweight network EfficientNet-b4, wherein the lightweight network EfficientNet-b4 includes 7 MBConv layers, and the features extracted by the first 3 MBConv layers are used as shallow features, and the features extracted by the last 4 MBConv layers are used as deep features; first, the shallow features are fused with the deep features using skip connections, and the features are added by adding Add layers between different layers of the lightweight network EfficientNet-b4, and the deep features are upsampled by the UpSampling2D function; secondly, a feature pyramid network is applied to fuse the shallow features layer by layer on the basis of the deep features, and the features between two adjacent layers are subjected to 1×1 convolution and 3×3 convolution by the Conv2D convolution function; finally, the output of the last fully connected layer of the lightweight network EfficientNet-b4 is changed to 3, and output by the softmax function.
[0013] Furthermore, the specific process of fuzzifying the labels of the medical images uploaded by the cooperative demand hospitals is as follows: the classification labels of the medical images uploaded by the cooperative demand hospitals are fuzzified by using fuzzy set theory, that is, the fuzzy membership degree is introduced to represent the fuzziness degree of each classification label;
[0014] Suppose the diagnostic images to be classified, i.e. the diagnostic images of the cooperative demand hospital, are , Classification label collection Contains three classification labels: benign, malignant and healthy. Each classification label ; Define a fuzzy membership function To express belong The degree of blur, The value range is between [0,1], and the fuzzy membership function It is expressed as:
[0015] ;
[0016] In the formula, Yes Assessment for Characteristic function of the correlation; and is the parameter for adjusting the fuzzy membership curve; represents the natural logarithm;
[0017] Diagnostic images to be classified The fuzzy label vector is expressed as:
[0018] ;
[0019] In the formula, represents the regional fuzzy score; represents the fuzzy score of the first region; represents the fuzzy score of the second region; Indicates The fuzzy score of a region.
[0020] Furthermore, it is assumed that there are multiple central hospitals , Indicates Central hospitals, each of which generates breast cancer diagnosis and treatment data of , Indicates the first breast cancer diagnosis and treatment data, the cloud service center is , the edge server is , Indicates Edge servers.
[0021] Furthermore, the specific process of step S3 is: using a convolutional neural network to extract features of the standardized diagnostic and treatment images; the convolutional neural network automatically learns and extracts high-level features of the standardized diagnostic and treatment images through multi-layer convolution operations and pooling operations; the calculation formula of the convolution operation is:
[0022] ;
[0023] In the formula, is the convolution kernel weight; is bias; represents the diagnostic image after convolution processing; It represents the quotient of the normalized diagnostic image and the convolution kernel width; Represents the quotient of the standardized diagnostic image and the height of the convolution kernel; is the original image pixel value, i.e., the unprocessed diagnosis and treatment image of the central hospital; is the convolution kernel bias; Diagnosis and treatment images for the central hospital;
[0024] The calculation formula for the pooling operation is:
[0025] ;
[0026] In the formula, represents the diagnostic image after pooling; It represents the maximum value in the neighborhood of the i-th pixel on the left, i-th pixel on the right, j-th pixel above, and j-th pixel below the pixel in the diagnosis and treatment image; Represents the diagnostic image after multi-layer convolution processing.
[0027] Furthermore, the specific process of step S4 is: the features of breast cancer diagnosis and treatment data of different central hospitals are divided into a training set and a test set, the training set is used to train the classification model, and the test set is used to test the classification model; in the process of training the classification model, the weights pre-trained by the pre-trained ImageNet model are used as the initial weights of the classification model, which are expressed as:
[0028] ;
[0029] In the formula, is the initial weight of the classification model; Represents the pre-trained weights of the pre-trained ImageNet model; represents the gradient; represents breast cancer diagnosis and treatment data;
[0030] After obtaining the initial weight of the classification model, the initial weight of the classification model is encrypted using the RC4 stream cipher, and the encrypted hash value is calculated using the SHA-1 algorithm. After the calculation is completed, the initial weight of the classification model is uploaded to the edge server;
[0031] After receiving the initial weights of the classification model, the edge server performs verification. If the verification is successful, the classification model is decrypted and loaded using the initial weights of the classification model. If the verification fails, a request is made to download the initial weights of the classification model again.
[0032] Furthermore, the specific process of the encrypted trained classification model in step S5 classifying the diagnosis and treatment images uploaded by the cooperative demand hospital is as follows: the cooperative demand hospital transmits the patient's diagnosis and treatment images to the edge server, and the classification model in the edge server classifies the patient's diagnosis and treatment images transmitted by the cooperative demand hospital, which is expressed as:
[0033] ;
[0034] In the formula, Indicates The first hospital in need of collaboration Diagnostic images; Represents the classification model pair in the edge server The prediction results; Represents the classification model in the edge server.
[0035] Furthermore, the specific process of step S2 is: firstly, the diagnosis and treatment images in the breast cancer diagnosis and treatment data of the central hospital are processed into 512 512, filtering, image enhancement, and data enhancement are performed on the resized diagnostic and treatment images in sequence, and finally, the diagnostic and treatment images after data enhancement are standardized;
[0036] The specific process of filtering, image enhancement, and data enhancement of the resized diagnostic image is as follows:
[0037] Gaussian filtering and median filtering are used to reduce random noise on the diagnosis and treatment images of the central hospital. The calculation formula of Gaussian filtering is:
[0038] ;
[0039] In the formula, is the original image pixel value, i.e., the unprocessed diagnosis and treatment image of the central hospital; Diagnosis and treatment images for the central hospital; is the standard deviation; is the filter window size; represents Gaussian filtering operation; Indicates pixel coordinates; Indicates that the filter window is The position in the axis direction; Indicates that the filter window is The position on the axis; represents the natural logarithm;
[0040] The calculation formula of median filtering is:
[0041] ;
[0042] In the formula, Represents the median value of the pixel intensity within a specified neighborhood; Represents the median filter operation;
[0043] The filtered diagnostic images were processed using two image enhancement techniques: histogram equalization and adaptive histogram equalization. The calculation formula for histogram equalization is:
[0044] ;
[0045] In the formula, is the number of gray levels; is the total number of pixels of the filtered diagnostic image; Is grayscale The number of pixels; Represents the value of the pixel point with x and y coordinates after histogram equalization;
[0046] The threshold segmentation of the diagnostic image after histogram equalization is expressed as:
[0047] ;
[0048] In the formula, is the selected threshold;
[0049] Using data enhancement technology, rotation and translation transformations are performed on the enhanced diagnostic images;
[0050] The rotation transformation is expressed as:
[0051] ;
[0052] In the formula, is the rotation matrix, is the rotation angle; represents the diagnosis and treatment image after rotation transformation;
[0053] The translation transformation is expressed as:
[0054] ;
[0055] In the formula, represents the diagnostic image after translation transformation; is the translation matrix; Indicates the diagnostic image after image enhancement The distance the axis moves; Indicates the diagnostic image after image enhancement The distance the axis moves;
[0056] The specific process of standardizing the diagnostic and treatment images after data enhancement is as follows: normalizing the diagnostic and treatment images after data enhancement, which is expressed as:
[0057] ;
[0058] In the formula, represents the normalized diagnosis and treatment image; Indicates the minimum pixel value in the diagnosis and treatment image; Indicates the maximum value of pixels in the diagnostic image.
[0059] Furthermore, breast cancer diagnosis and treatment data is obtained from the central hospital at regular intervals to perform real-time training on the classification model to update the initial weights of the classification model, and the updated initial weights of the classification model are uploaded to the edge server to replace the initial weights of the original classification model.
[0060] A system for implementing an artificial intelligence-based online classification method for breast cancer, comprising:
[0061] The data acquisition module is used to obtain breast cancer diagnosis and treatment data from different central hospitals and transmit it to the cloud service center;
[0062] The preprocessing module is used to preprocess the breast cancer diagnosis and treatment data of the central hospital through the cloud service center;
[0063] A feature extraction module is used to extract features from the pre-processed breast cancer diagnosis and treatment data of the central hospital using a convolutional neural network;
[0064] A construction module is used to construct a classification model, train the classification model using the extracted features, encrypt the weights of the trained classification model through the cloud service center, obtain the optimal classification model, and transmit the optimal classification model to the edge server;
[0065] The classification module is used to upload the medical images of the collaborative demand hospital to the edge server, blur the labels of the medical images uploaded by the collaborative demand hospital, and then classify the medical images with blurred labels through the optimal classification model.
[0066] Compared with the existing technology, the present invention has the following beneficial effects:
[0067] (1) The present invention encrypts the initial weights of the classification model by using the RC4 stream cipher and calculates the encrypted hash value using the SHA-1 algorithm. After the calculation is completed, the initial weights of the classification model are uploaded to the edge server to ensure the security of the classification model. The initial weights of the classification model are updated by regularly obtaining breast cancer diagnosis and treatment data from the central hospital for real-time training of the classification model, thereby continuously improving the accuracy of the classification model.
[0068] (2) The present invention adopts a lightweight network EfficientNet-b4 and adds a feature fusion layer to the lightweight network EfficientNet-b4 to perform feature extraction, which can make the classification of diagnostic images more accurate; by using a label fuzzy method to process the images to be diagnosed, the score weights of local areas of the images to be diagnosed can be dynamically adjusted, the attention of areas irrelevant to the classification and areas with unclear case characteristics can be reduced, and the attention of regional features related to and obvious lesion characteristics can be increased. BRIEF DESCRIPTION OF THE DRAWINGS
[0069] Figure 1 It is a flow chart of the steps of the present invention. DETAILED DESCRIPTION
[0070] like Figure 1 As shown, the present invention provides a technical solution: an online classification method for breast cancer based on artificial intelligence, comprising the following steps:
[0071] Step S1: Obtain breast cancer diagnosis and treatment data from different central hospitals and transmit them to the cloud service center; the diagnosis and treatment data includes diagnosis and treatment images and the diagnosis and treatment results of the diagnosis and treatment images, and the diagnosis and treatment results include three classification labels: benign, malignant, and healthy.
[0072] Step S2: Preprocessing the breast cancer diagnosis and treatment data of the central hospital through the cloud service center.
[0073] Step S3: Use a convolutional neural network (CNN) to extract features from the pre-processed breast cancer diagnosis and treatment data of the central hospital.
[0074] Step S4: construct a classification model, use the features extracted in step S3 to train the classification model, and encrypt the weights of the trained classification model through the cloud service center to obtain the optimal classification model, and transmit the optimal classification model to the edge server;
[0075] The classification model is composed of the convolutional neural network CNN and the lightweight network EfficientNet-b4 in step S3, and a feature fusion layer is added to the lightweight network EfficientNet-b4, and the output of the last connection layer of the lightweight network EfficientNet-b4 is set to 3.
[0076] Step S5: The collaborative hospital uploads the medical images to the edge server, blurs the labels of the medical images uploaded by the collaborative hospital, and then classifies the medical images with blurred labels through the optimal classification model.
[0077] Among them, the central hospital is a Grade 3A hospital, and the cooperative demand hospital is a village hospital.
[0078] Among them, there are a number of central hospitals , Indicates Central hospitals, each of which generates breast cancer diagnosis and treatment data of , Indicates the first breast cancer diagnosis and treatment data, the cloud service center is , the edge server is , Indicates Edge servers.
[0079] Among them, since there are some differences in the image features obtained from different central hospitals, image processing is required to process the images into pictures of the same specifications; the specific process of step S2 is: in order to facilitate the classification model analysis, the diagnosis and treatment images in the breast cancer diagnosis and treatment data of the central hospital are first processed into 512 The size of the image is 512; secondly, in order to improve the generalization performance of the classification model during training, the diagnosis and treatment images of the central hospital are filtered, image enhanced, and data enhanced in turn. Finally, they are standardized to reduce the computational complexity of the classification model.
[0080] Specifically, in order to enhance the diversity of diagnosis and treatment images, Gaussian filtering and median filtering are used to reduce random noise on the diagnosis and treatment images of the central hospital; the calculation formula of Gaussian filtering is:
[0081] (1);
[0082] In the formula, is the original image pixel value (unprocessed diagnosis and treatment image of the central hospital); Diagnosis and treatment images for the central hospital; is the standard deviation; is the filter window size; represents Gaussian filtering operation; Indicates pixel coordinates; Indicates that the filter window is The position in the axis direction; Indicates that the filter window is The position on the axis; Represents the natural logarithm.
[0083] The calculation formula of median filtering is:
[0084] (2);
[0085] In the formula, Represents the median value of the pixel intensity within a specified neighborhood; Represents a median filter operation.
[0086] Image enhancement techniques, such as histogram equalization and color-adaptive histogram equalization (CLAHE), are used to process the filtered diagnostic images to improve the image contrast and make the lesion area more obvious.
[0087] The calculation formula for histogram equalization is:
[0088] (3);
[0089] In the formula, is the number of gray levels; is the total number of pixels of the filtered diagnostic image; Is grayscale The number of pixels; Represents the value of the pixel at (x, y) coordinates after histogram equalization.
[0090] The diagnostic image after histogram equalization is segmented to extract the region of interest (ROI), including the tumor or lesion area; the segmentation method uses threshold segmentation, which is expressed as:
[0091] (4);
[0092] In the formula, is the selected threshold, which is set to 0.6 in this embodiment.
[0093] In addition, data enhancement technology is also used to generate more diverse training samples by performing random transformations on the enhanced diagnostic images, such as rotation and translation, to improve the generalization ability of the model.
[0094] The rotation transformation is expressed as:
[0095] (5);
[0096] In the formula, is the rotation matrix, is the rotation angle; Represents the diagnosis and treatment image after rotation transformation.
[0097] The translation transformation is expressed as:
[0098] (6);
[0099] In the formula, represents the diagnostic image after translation transformation; is the translation matrix; Indicates the diagnostic image after image enhancement The distance the axis moves; Indicates the diagnostic image after image enhancement The distance the axis moves.
[0100] Finally, the diagnostic images after data enhancement are normalized to eliminate the brightness and contrast differences between different diagnostic images and ensure image consistency. Specifically, the diagnostic images after data enhancement are normalized, which is expressed as:
[0101] (7);
[0102] In the formula, represents the normalized diagnosis and treatment image; Indicates the minimum pixel value in the diagnosis and treatment image; Indicates the maximum value of pixels in the diagnostic image.
[0103] The specific process of step S3 is as follows: using a convolutional neural network (CNN) to extract features of the standardized diagnostic images; the convolutional neural network (CNN) automatically learns and extracts high-level features of the standardized diagnostic images through multi-layer convolution operations and pooling operations; the calculation formula of the convolutional neural network (CNN) convolution operation is:
[0104] (8);
[0105] In the formula, is the convolution kernel weight; is the convolution kernel bias; represents the diagnostic image after convolution processing; It represents the quotient of the normalized diagnostic image and the convolution kernel width; Represents the quotient of the standardized diagnostic image and the convolution kernel height.
[0106] The calculation formula of the convolutional neural network (CNN) pooling operation is:
[0107] (9);
[0108] In the formula, represents the diagnostic image after pooling; It represents the maximum value in the neighborhood of the i-th pixel on the left, i-th pixel on the right, j-th pixel above, and j-th pixel below the pixel in the diagnosis and treatment image; Represents the diagnostic image after multi-layer convolution processing.
[0109] In order to improve the performance of breast cancer image classification, the present invention adds a feature fusion layer to the lightweight network EfficientNet-b4, wherein the lightweight network EfficientNet-b4 includes 7 MBConv layers (inverted linear bottleneck layers with depthwise separable convolutions), and the features extracted by the first 3 MBConv layers are used as shallow features, and the features extracted by the last 4 MBConv layers are used as deep features. First, the shallow features are fused with the deep features using Skip Connections. Add layers are added between different layers of the lightweight network EfficientNet-b4 to add features (add the features of the 1st and 4th layers, add the features of the 2nd and 5th layers first, and add the features of the 3rd and 6th layers). The deep features are upsampled using the UpSampling2D function provided by pytorch to make the deep and shallow features have the same shape. Secondly, the feature pyramid network (FPN) is applied to fuse the shallow features layer by layer on the basis of the deep features. The Conv2D convolution function provided by pytorch is used to perform 1×1 convolution and 3×3 convolution on the features between two adjacent layers to fuse the features between the two adjacent layers to participate in the operation of the next layer and adjust the number of channels. The combination of these two methods can enhance the model's ability to capture multi-scale features, thereby improving the accuracy of breast cancer image classification. Finally, in order to adapt to the breast cancer classification task focused on by the present invention, the output of the last fully connected layer of the lightweight network EfficientNet-b4 is changed to 3 and output through the softmax function.
[0110] The specific process of step S4 is as follows: the features of breast cancer diagnosis and treatment data from different central hospitals are divided into training sets and test sets at a ratio of 8:2, the training set is used to train the classification model, and the test set is used to test the classification model. In the process of training the classification model, the weights pre-trained by the pre-trained ImageNet model are used as the initial weights of the classification model;
[0111] (10);
[0112] In the formula, is the initial weight of the classification model; Represents the pre-trained weights of the pre-trained ImageNet model; represents the gradient; Represents breast cancer diagnosis and treatment data.
[0113] After obtaining the initial weight of the classification model, the initial weight of the classification model is encrypted using the RC4 stream cipher, and the encrypted hash value is calculated using the SHA-1 algorithm. After the calculation is completed, the initial weight of the classification model is uploaded to the edge server.
[0114] After receiving the initial weights of the classification model, the edge server performs verification. If the verification is successful, the classification model is decrypted and loaded using the initial weights of the classification model. If the verification fails, a request is made to download the initial weights of the classification model again.
[0115] Among them, new patients will be diagnosed continuously in the central hospital, and new breast cancer diagnosis and treatment data will be continuously generated; therefore, in order to improve the accuracy of the classification model, breast cancer diagnosis and treatment data are regularly obtained from the central hospital to perform real-time training on the classification model to update the initial weights of the classification model, and the updated initial weights of the classification model are uploaded to the edge server to replace the initial weights of the original classification model.
[0116] The specific process of the encrypted trained classification model in step S5 classifying the diagnosis and treatment images uploaded by the cooperative demand hospital is as follows: the cooperative demand hospital transmits the patient's diagnosis and treatment images to the edge server, and the classification model in the edge server classifies the patient's diagnosis and treatment images transmitted by the cooperative demand hospital;
[0117] (11);
[0118] In the formula, Indicates The first hospital in need of collaboration Diagnostic images; Represents the classification model pair in the edge server The prediction results; Represents the classification model in the edge server.
[0119] Among them, in the daily diagnosis scenarios of hospitals, some difficult and complicated cases often occur. Due to their complexity and variability, these cases make it difficult for classification models to accurately classify them. For difficult and complicated cases that appear in the classification of diagnostic images in breast cancer diagnosis and treatment data, the label fuzzification method is used to fuzzify the classification labels of diagnostic images uploaded by collaborative demand hospitals, aiming to improve the flexibility and accuracy of classification. Specifically, in the classification of diagnostic images, for some cases with unclear features of diagnostic images, the fuzzy set theory is used to fuzzy the classification labels of diagnostic images, that is, the fuzzy membership degree is introduced to represent the degree of fuzziness of each classification label.
[0120] Assume that the diagnostic images to be classified (diagnostic images of cooperative demand hospitals) are , Classification label collection It contains three classification labels: benign, malignant and healthy. Each classification label ; Define a fuzzy membership function To express belong The degree of blur, The value range is between [0,1], and the fuzzy membership function It can be expressed as:
[0121] (12);
[0122] In the formula, Yes Assessment for Characteristic function of the correlation; and is the parameter for adjusting the fuzzy membership curve.
[0123] Diagnostic images to be classified The fuzzy label vector is expressed as:
[0124] (13);
[0125] In the formula, represents the regional fuzzy score; represents the fuzzy score of the first region; represents the fuzzy score of the second region; Indicates The fuzzy score of a region.
[0126] In practical applications, when the fuzzy membership of the diagnostic image to be classified is evenly distributed among multiple classification labels, it means that the diagnostic image to be classified is difficult to classify clearly and belongs to a difficult and complicated disease. At this time, further manual classification or other auxiliary examinations are required.
[0127] A system for implementing an artificial intelligence-based online classification method for breast cancer, comprising: a data acquisition module, used to acquire breast cancer diagnosis and treatment data of different central hospitals and transmit the data to a cloud service center; a preprocessing module, used to preprocess the breast cancer diagnosis and treatment data of the central hospital through the cloud service center; a feature extraction module, used to extract features from the preprocessed breast cancer diagnosis and treatment data of the central hospital using a convolutional neural network; a construction module, used to construct a classification model, train the classification model using the extracted features, encrypt the weights of the trained classification model through the cloud service center to obtain an optimal classification model, and transmit the optimal classification model to an edge server; a classification module, used to upload diagnosis and treatment images of a cooperative demand hospital to an edge server, perform label fuzzification on the diagnosis and treatment images uploaded by the cooperative demand hospital, and then classify the diagnosis and treatment images with blurred labels using the optimal classification model.
[0128] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.
Claims
1. An online classification method for breast cancer based on artificial intelligence, It is characterized in that The steps include: Step S1: Obtain breast cancer diagnosis and treatment data from different central hospitals and transmit them to the cloud service center; the diagnosis and treatment data includes diagnosis and treatment images and diagnosis and treatment results of the diagnosis and treatment images; Step S2: pre-processing the breast cancer diagnosis and treatment data of the central hospital through the cloud service center; Step S3: using a convolutional neural network to extract features from the pre-processed breast cancer diagnosis and treatment data of the central hospital; Step S4: construct a classification model, use the features extracted in step S3 to train the classification model, and encrypt the weights of the trained classification model through the cloud service center to obtain the optimal classification model, and transmit the optimal classification model to the edge server; The classification model is composed of the convolutional neural network in step S3 and the lightweight network EfficientNet-b4, and a feature fusion layer is added to the lightweight network EfficientNet-b4, and the output number of the last connection layer of the lightweight network EfficientNet-b4 is set; Step S5: the cooperative demand hospital uploads the diagnosis and treatment images to the edge server, blurs the labels of the diagnosis and treatment images uploaded by the cooperative demand hospital, and then classifies the diagnosis and treatment images with blurred labels through the optimal classification model; A feature fusion layer is added to the lightweight network EfficientNet-b4, where the lightweight network EfficientNet-b4 includes 7 MBConv layers, and the features extracted by the first 3 MBConv layers are used as shallow features, and the features extracted by the last 4 MBConv layers are used as deep features; first, the shallow features are fused with the deep features using skip connections, and the features are summed by adding Add layers between different layers of the lightweight network EfficientNet-b4, and the deep features are upsampled by the UpSampling2D function; secondly, a feature pyramid network is applied to fuse the shallow features layer by layer on the basis of the deep features, and the features between two adjacent layers are subjected to 1×1 convolution and 3×3 convolution by the Conv2D convolution function; finally, the output of the last fully connected layer of the lightweight network EfficientNet-b4 is changed to 3, and output by the softmax function; The specific process of label fuzzification of the medical images uploaded by the cooperative demand hospitals is as follows: the classification labels of the medical images uploaded by the cooperative demand hospitals are fuzzified by using fuzzy set theory, that is, fuzzy membership is introduced to represent the degree of fuzziness of each classification label; Suppose the diagnostic images to be classified, i.e. the diagnostic images of the cooperative demand hospital, are , Classification label collection Contains three classification labels: benign, malignant and healthy. Each classification label ; Define a fuzzy membership function To express belong The degree of blur, The value range is between [0,1], and the fuzzy membership function It is expressed as: ; In the formula, Yes Assessment for Characteristic function of the correlation; and is the parameter for adjusting the fuzzy membership curve; represents the natural logarithm; Diagnostic images to be classified The fuzzy label vector is expressed as: ; In the formula, represents the regional fuzzy score; represents the fuzzy score of the first region; represents the fuzzy score of the second region; Indicates Fuzzy score of the region; The specific process of step S4 is: the features of breast cancer diagnosis and treatment data from different central hospitals are divided into training sets and test sets, the training set is used to train the classification model, and the test set is used to test the classification model; in the process of training the classification model, the weights pre-trained by the pre-trained ImageNet model are used as the initial weights of the classification model, which are expressed as: ; In the formula, is the initial weight of the classification model; Represents the pre-trained weights of the pre-trained ImageNet model; represents the gradient; represents breast cancer diagnosis and treatment data; After obtaining the initial weight of the classification model, the initial weight of the classification model is encrypted using the RC4 stream cipher, and the encrypted hash value is calculated using the SHA-1 algorithm. After the calculation is completed, the initial weight of the classification model is uploaded to the edge server; After receiving the initial weights of the classification model, the edge server verifies it. If the verification succeeds, the classification model is decrypted and loaded using the initial weights of the classification model. If the verification fails, the initial weights of the classification model are downloaded again. The specific process of the encrypted trained classification model in step S5 classifying the diagnosis and treatment images uploaded by the cooperative demand hospital is as follows: the cooperative demand hospital transmits the patient's diagnosis and treatment images to the edge server, and the classification model in the edge server classifies the patient's diagnosis and treatment images transmitted by the cooperative demand hospital, which is expressed as: ; In the formula, Indicates The first hospital in need of collaboration Diagnostic images; Represents the classification model pair in the edge server The prediction results; Represents the classification model in the edge server.
2. The method for online classification of breast cancer based on artificial intelligence according to claim 1, Features: Set up a total of multiple central hospitals , Indicates Central hospitals, each of which generates breast cancer diagnosis and treatment data of , Indicates the first breast cancer diagnosis and treatment data, the cloud service center is , the edge server is , Indicates Edge servers.
3. The method for online classification of breast cancer based on artificial intelligence according to claim 2, Features: The specific process of step S3 is: using a convolutional neural network to extract features of the standardized diagnostic and treatment images; the convolutional neural network automatically learns and extracts high-level features of the standardized diagnostic and treatment images through multi-layer convolution operations and pooling operations; the calculation formula of the convolution operation is: ; In the formula, is the convolution kernel weight; is bias; represents the diagnostic image after convolution processing; It represents the quotient of the normalized diagnostic image and the width of the convolution kernel; Represents the quotient of the standardized diagnostic image and the height of the convolution kernel; is the original image pixel value, i.e., the unprocessed diagnosis and treatment image of the central hospital; is the convolution kernel bias; Diagnosis and treatment images for the central hospital; The calculation formula for the pooling operation is: ; In the formula, represents the diagnostic image after pooling; It represents the maximum value in the neighborhood of the i-th pixel on the left, i-th pixel on the right, j-th pixel above, and j-th pixel below the pixel in the diagnosis and treatment image; Represents the diagnostic image after multi-layer convolution processing.
4. The method for online classification of breast cancer based on artificial intelligence according to claim 3, Features: The specific process of step S2 is: firstly, the diagnosis and treatment images in the breast cancer diagnosis and treatment data of the central hospital are processed into 512 512, filtering, image enhancement, and data enhancement are performed on the resized diagnostic and treatment images in sequence, and finally, the diagnostic and treatment images after data enhancement are standardized; The specific process of filtering, image enhancement, and data enhancement of the resized diagnostic image is as follows: Gaussian filtering and median filtering are used to reduce random noise on the diagnosis and treatment images of the central hospital. The calculation formula of Gaussian filtering is: ; In the formula, is the original image pixel value, i.e., the unprocessed diagnosis and treatment image of the central hospital; Diagnosis and treatment images for the central hospital; is the standard deviation; is the filter window size; represents Gaussian filtering operation; Indicates pixel coordinates; Indicates that the filter window is The position in the axis direction; Indicates that the filter window is The position on the axis; represents the natural logarithm; The calculation formula of median filtering is: ; In the formula, Represents the median value of the pixel intensity within a specified neighborhood; Represents the median filter operation; The filtered diagnostic images were processed using two image enhancement techniques: histogram equalization and adaptive histogram equalization. The calculation formula for histogram equalization is: ; In the formula, is the number of gray levels; is the total number of pixels of the filtered diagnostic image; Is grayscale The number of pixels; Represents the value of the pixel point with x and y coordinates after histogram equalization; The threshold segmentation of the diagnostic image after histogram equalization is expressed as: ; In the formula, is the selected threshold; Using data enhancement technology, rotation and translation transformations are performed on the enhanced diagnostic images; The rotation transformation is expressed as: ; In the formula, is the rotation matrix, is the rotation angle; represents the diagnosis and treatment image after rotation transformation; The translation transformation is expressed as: ; In the formula, represents the diagnostic image after translation transformation; is the translation matrix; Indicates the diagnostic image after image enhancement The distance the axis moves; Indicates the diagnostic image after image enhancement The distance the axis moves; The specific process of standardizing the diagnostic and treatment images after data enhancement is as follows: normalizing the diagnostic and treatment images after data enhancement, which is expressed as: ; In the formula, represents the normalized diagnosis and treatment image; Indicates the minimum pixel value in the diagnosis and treatment image; Indicates the maximum value of pixels in the diagnostic image.
5. The method for online classification of breast cancer based on artificial intelligence according to claim 4, Features: Breast cancer diagnosis and treatment data is obtained from the central hospital regularly to perform real-time training on the classification model to update the initial weights of the classification model. The updated initial weights of the classification model are uploaded to the edge server to replace the initial weights of the original classification model.
6. A system for implementing the artificial intelligence-based online breast cancer classification method according to any one of claims 1 to 5, It is characterized in that include: The data acquisition module is used to obtain breast cancer diagnosis and treatment data from different central hospitals and transmit it to the cloud service center; The preprocessing module is used to preprocess the breast cancer diagnosis and treatment data of the central hospital through the cloud service center; A feature extraction module is used to extract features from the pre-processed breast cancer diagnosis and treatment data of the central hospital using a convolutional neural network; A construction module is used to construct a classification model, train the classification model using the extracted features, encrypt the weights of the trained classification model through the cloud service center, obtain the optimal classification model, and transmit the optimal classification model to the edge server; The classification module is used to upload the medical images of the collaborative demand hospital to the edge server, blur the labels of the medical images uploaded by the collaborative demand hospital, and then classify the medical images with blurred labels through the optimal classification model.
