AI auxiliary screening method based on mammary gland molybdenum target image

Abstract

The invention discloses an AI auxiliary screening method based on a mammary gland molybdenum target image, and relates to the technical field of image data identification processing, and the method comprises the following steps: S1, image collection: collecting mammary gland molybdenum target image data of a patient, carrying out the standardization processing, removing the noise in the image, standardizing the gray value, and enhancing the local contrast, and obtaining a structured image data set. According to the AI auxiliary screening method based on the mammary gland molybdenum target image, multiple machine learning algorithms are fused for focus classification, so that the classification accuracy is improved; multiple classification models are constructed in an integrated learning mode, prediction results are fused through a weighted average mechanism and a voting mechanism, model prediction confidence is calculated for weighted fusion, the advantages of all the models are brought into full play, and the limitation of a single model is reduced; the screening accuracy and efficiency of the mammary gland molybdenum target image are improved, the focus detection and classification difficulty is reduced, and the condition that the model has more false negative or false positive results is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of image data recognition and processing, and specifically to an AI-assisted screening method based on mammary gland mammography images. Background Art

[0002] Breast diseases pose a huge threat to women's health. The incidence of breast cancer has been increasing year by year. Early diagnosis is of great significance to improving patients' survival rate and quality of life. Mammography is a commonly used screening method that can clearly present the internal structure of the breast and the characteristics of lesions. However, traditional manual reading of films has many limitations. On the one hand, radiologists are prone to strong subjectivity in reading films due to visual fatigue and differences in experience levels, and may miss subtle lesions, such as misjudging tiny calcifications. On the other hand, manual reading of films is labor-intensive and has low diagnostic efficiency during large-scale screening, resulting in long waiting times for patients and affected diagnostic accuracy.

[0003] In recent years, artificial intelligence image recognition technology has developed rapidly. Technologies such as convolutional neural networks have excellent capabilities in processing image data and can automatically learn image features. At the same time, a large amount of accumulated mammography data provides rich materials for AI model training, which helps the model learn different lesion feature patterns and improve generalization ability. However, the existing AI-assisted mammography screening methods also have shortcomings. The accuracy of the model is restricted by many factors, such as differences in imaging equipment and shooting conditions, especially low-quality images produced by old equipment in primary medical institutions, such as many noise points and low contrast. In addition, the lesions of breast cancer are diverse in morphology, and their size, shape, density and other characteristics vary, which further increases the difficulty of lesion detection and classification. The model will have more false negative or false positive results. Therefore, how to automatically analyze and diagnose mammography images, identify lesions of different shapes, sizes and densities, so as to improve the screening accuracy and efficiency of mammography images, reduce the difficulty of lesion detection and classification, and improve the situation where the model has more false negative or false positive results is the problem we want to solve. To this end, an AI-assisted screening method based on mammography is proposed. Summary of the invention

[0004] In order to solve the above technical problems, the present invention is implemented by the following technical solutions: an AI-assisted screening method based on mammography, comprising the following steps:

[0005] S1: Image acquisition: Collect the patient's breast mammography image data and perform standardization processing to remove noise in the image, standardize the gray value and enhance the local contrast to obtain a structured image data set;

[0006] S2: Feature extraction: Perform multi-scale and multi-feature analysis on standardized image data sets to extract lesion features, including the size, shape, density, and edge contour features of the lesions;

[0007] S3: Abnormality detection: Use the trained deep learning model to compare with normal image features, identify abnormal lesion areas in lesion features, determine whether there are suspected breast cancer lesions, and generate annotation information. At the same time, generate breast lesions based on the annotation information, where the annotation information includes lesion location, size, and density;

[0008] S4: Lesion classification: Identify and classify the lesion types of breast lesions through a variety of machine learning algorithms, identify the categories to which the breast lesions belong, and obtain lesion classification results. The lesion types include benign lesions, malignant lesions, and suspected lesions that cannot be classified;

[0009] S5: Diagnostic report generation: Generate a diagnostic report based on the lesion classification results. The diagnostic report is used to provide diagnostic conclusions and examination recommendations. The diagnostic report includes lesion description and treatment recommendations.

[0010] Preferably, the noise removal process in the image is as follows:

[0011] Based on the collected mammography data, the distribution characteristics and intensity changes of the noise are observed, the noise is classified, and the noise type is distinguished as quantum noise or electronic noise. The noise type is obtained, and the noise in the image is removed based on the noise type. The process is as follows:

[0012] If quantum noise exists, a noise threshold is set, and when the change in pixel value in the local area exceeds the noise threshold, adaptive filtering is started;

[0013] If electronic noise exists, then if the imaging device allows, collect a dark field image, and subtract the pixel value corresponding to the dark field image from the original image to eliminate the influence of electronic noise and complete the correction;

[0014] Among them, the distribution characteristics include uniform distribution or local concentration.

[0015] Preferably, for the standardization of grayscale values, the process is as follows:

[0016] Since different imaging devices and imaging conditions may lead to differences in the range and distribution of grayscale values, they need to be standardized to a unified range of 0-255; analyze the grayscale histogram of mammography images to determine the distribution range and distribution characteristics of grayscale values ​​in the image;

[0017] Use linear grayscale transformation or nonlinear grayscale transformation method, and select the corresponding mapping function according to the distribution of grayscale histogram. Standardize the grayscale value, the process is as follows:

[0018] If the grayscale value is evenly distributed, then a linear grayscale transformation is used: a linear mapping function y=ax+b is used, where x is the original grayscale value, y is the standardized grayscale value, and a and b are coefficients calculated based on the target grayscale range;

[0019] If the grayscale value distribution is uneven, then use nonlinear grayscale transformation: use the logarithmic transformation y=c*log(1+x) in the nonlinear transformation function; where c is the adjustment coefficient to make the grayscale value distribution uniform and improve the image contrast;

[0020] Among them, the gray value distribution is uneven: there are many low gray value or high gray value pixels concentrated.

[0021] Preferably, the process of enhancing local contrast is as follows:

[0022] Based on local statistical information, a local contrast stretching enhancement algorithm is used. According to the gray value distribution in the local neighborhood of each pixel in the image, a window between 3x3 and 11x11 is selected to stretch the pixel value to enhance the local contrast. The process is as follows:

[0023] Calculate the minimum and maximum values ​​in the neighborhood of each pixel, and then convert the gray value x of the pixel into Perform the transformation, set the target grayscale range to 0-255, and enhance the grayscale value contrast in the local area.

[0024] Preferably, step S2 comprises:

[0025] S21: Multi-scale decomposition: Based on wavelet transform, the image dataset is decomposed into sub-images of different scales and directions by three-layer wavelet decomposition. In each layer of decomposition, the image dataset is decomposed into a low-frequency sub-band and three high-frequency sub-bands.

[0026] S22: Scale-related feature calculation:

[0027] Compute lesion-related features on sub-images of different scales:

[0028] In the low-frequency subband of wavelet decomposition, the Otsu threshold method is used to preliminarily segment the lesion area, calculate the number of pixels, and obtain the segmentation result. Then, the lesion boundary is located and refined in the high-frequency subband to complete the correction of the segmentation result and obtain the size characteristics of the lesion.

[0029] In the high-frequency subbands of different layers of wavelet decomposition, the edge contour of the lesion is calculated, and then its shape descriptor is calculated. The changes in the shape characteristics of the lesion at different scales are compared to determine whether the lesion has a regular shape or whether there are irregular features such as lobulation and burrs, and then the shape characteristics of the lesion are obtained;

[0030] Based on the multi-layer sub-images decomposed by wavelet, the threshold segmentation method is used to obtain the lesion area on images of different scales, and then the average grayscale value of the pixels in the lesion area is calculated to obtain the quantized lesion density characteristics; the Canny operator in the edge detection operator is used to calculate the amplitude and direction of the image gradient to obtain the edge contour characteristics of the lesion; the size characteristics, shape characteristics, density characteristics and edge contour characteristics of the lesion obtained are combined to form the lesion characteristics.

[0031] Preferably, in step S3, the process of generating breast lesions is as follows:

[0032] S31: Comparison between the trained deep learning model and normal image features: Using the trained deep learning model, the collected and standardized mammography image data are compared with the normal mammography image features; wherein the mammography image data includes lesion features;

[0033] S32: Identify abnormal lesion areas: The deep learning model compares the mammography data to find areas with different characteristics from normal mammography images and marks these areas as abnormal lesion areas.

[0034] S33: Determine suspected breast cancer lesions: Determine whether there are suspected breast cancer lesions in the area based on the characteristics of the identified abnormal lesion area;

[0035] S34: Generate annotation information: If there is a suspected breast cancer lesion, generate annotation information including the location, size, and density of the lesion, clarify the location coordinates of the lesion in the image, measure and record the size of the lesion, and calculate the average grayscale value of the pixels in the lesion area to obtain density information;

[0036] S35: Generate breast lesions based on the annotation information: According to the annotation information, determine the corresponding abnormal area in the image as a breast lesion, and display the annotation information of the lesion on the image.

[0037] Preferably, in step S33, the process of determining whether there is a suspected breast cancer lesion in the region is as follows:

[0038] S331: Data training: Using annotated mammography data as training set, the convolutional neural network (CNN) structure is used to train the network to identify lesion areas; by allowing the model to learn the characteristic patterns of many known lesion areas, the model is able to recognize the characteristics of lesions; the known lesion areas include benign and malignant lesions;

[0039] S332: Abnormality discrimination: Based on the trained convolutional neural network, feature matching is performed on each pixel of the mammography image data, and the pixels in the mammography image data are compared and analyzed one by one with the lesion feature patterns learned by the model;

[0040] Identify areas with similar features to known lesions. When the pixel features of a certain area show a high degree of similarity with the lesion features in the training set, the area is determined to be a suspected lesion area.

[0041] Generate a confidence score. The model generates a confidence score for each suspected lesion area based on the degree of feature matching. The score is used to evaluate the reliability of the detection results. The higher the confidence score, the greater the possibility that the area is a lesion.

[0042] The preset confidence threshold is used to determine whether a suspected breast cancer lesion is present. If the confidence score of a suspected lesion area is higher than the preset confidence threshold, the area is considered to contain a suspected breast cancer lesion. Otherwise, if the confidence score is lower than the confidence threshold, the area is considered normal.

[0043] Preferably, in step S4, the process of obtaining the lesion classification result is as follows:

[0044] S41: Classification model establishment: Multiple classification models were constructed using ensemble learning methods, including convolutional neural network (CNN), random forest, and support vector machine (SVM). The classification models were trained using mammary mammography data with lesion types labeled to enable the models to learn image feature patterns of different types of lesions. Lesion types included benign, malignant, and suspected.

[0045] S42: Multi-model fusion step:

[0046] S421: Fusion of prediction results based on different classification models:

[0047] A weighted average mechanism is used to assign different weights to each model according to its accuracy and recall during training and validation. If the CNN model has a high accuracy rate in classifying benign lesions on the validation set, it will be given a relatively high weight in the weighted average.

[0048] At the same time, a voting mechanism is used. Each model votes on the category to which the breast lesion belongs. The category with the most votes is used as the preliminary fusion result. If three models classify a breast lesion, two of which vote as benign and one vote as malignant, then the preliminary fusion result is benign.

[0049] S422: Calculate the prediction confidence of each model and weightedly fuse the results of multiple models to obtain the final classification result:

[0050] Through the calculation method within the model, the confidence of each model in the prediction result is obtained; then these confidence results are weighted and fused to obtain a more accurate and reliable fused lesion classification result;

[0051] Among them, the calculation methods within the three models are: CNN is based on the probability output layer, random forest is based on the distribution ratio of samples in each decision tree, and SVM is based on the distance to the hyperplane.

[0052] Preferably, in step S41, the process of establishing the three classification models is as follows:

[0053] For the convolutional neural network (CNN), its structural design is based on the learning requirements of breast lesion image features. It extracts complex features in the image through multiple convolutional layers and pooling layers to obtain the neural network prediction results.

[0054] For random forests, multiple decision trees are constructed, and each decision tree is used to vote on the classification results of the data to determine the final classification and obtain the random forest prediction results. When constructing a random forest, a feature subset and a data subset are randomly selected to train each decision tree to improve the generalization ability of the model.

[0055] For the support vector machine (SVM), the benign, malignant, and suspected lesion data are separated by finding the optimal hyperplane, and the different categories of lesion data are mapped to a high-dimensional space through the kernel function of the support vector machine to obtain the vector machine prediction results, so as to better handle nonlinear classification problems.

[0056] Preferably, step S42 further includes: optimizing the classification accuracy, the process is as follows:

[0057] Tune the hyperparameters of the classification model through cross-validation: divide the training data into multiple subsets, use k-fold cross-validation, select one of the subsets as the validation set each time, and the remaining subsets as the training set, train the model and evaluate the performance on the validation set to obtain the evaluation structure;

[0058] Adjust the model's hyperparameters based on the evaluation results, including adjusting the convolution kernel size and learning rate in CNN; adjusting the number of trees and tree depth in random forest; adjusting the penalty parameter C and kernel function parameters in SVM.

[0059] The present invention provides an AI-assisted screening method based on mammography, which has the following beneficial effects:

[0060] This AI-assisted screening method based on mammography can extract information from multiple aspects such as the size, shape, density, and edge contour features of lesions through multi-scale and multi-feature analysis, providing rich basis for accurate judgment; it combines the above features to comprehensively and accurately describe the lesion situation; and uses a fusion of multiple machine learning algorithms for lesion classification to improve classification accuracy; it uses an integrated learning method to build multiple classification models based on convolutional neural networks CNN, random forests, and support vector machines SVM, and fuses the prediction results through a weighted average mechanism and a voting mechanism, and calculates the model prediction confidence for weighted fusion, giving full play to the advantages of each model and reducing the limitations of a single model; it improves the screening accuracy and efficiency of mammography, reduces the difficulty of lesion detection and classification, and improves the situation where the model has a large number of false negative or false positive results. BRIEF DESCRIPTION OF THE DRAWINGS

[0061] Figure 1 The present invention is a flowchart of an AI-assisted screening method based on mammography. DETAILED DESCRIPTION

[0062] The present invention is further described in detail below in conjunction with the accompanying drawings and specific embodiments. The embodiments of the present invention are provided for the purpose of illustration and description, and are not intended to be exhaustive or to limit the present invention to the disclosed forms. Many modifications and variations will be apparent to those of ordinary skill in the art. The embodiments are selected and described in order to better illustrate the principles and practical applications of the present invention, and to enable those of ordinary skill in the art to understand the present invention and thereby design various embodiments with various modifications suitable for specific uses.

[0063] like Figure 1 As shown, the present invention provides a technical solution: an AI-assisted screening method based on mammography, comprising the following steps:

[0064] S1: Image acquisition: Collect the patient's breast mammography image data and perform standardization processing to remove noise in the image, standardize the gray value and enhance the local contrast to obtain a structured image data set;

[0065] S2: Feature extraction: Perform multi-scale and multi-feature analysis on standardized image data sets to extract lesion features, including the size, shape, density, and edge contour features of the lesions;

[0066] S3: Abnormality detection: Use the trained deep learning model to compare with normal image features, identify abnormal lesion areas in lesion features, determine whether there are suspected breast cancer lesions, and generate annotation information. At the same time, generate breast lesions based on the annotation information, where the annotation information includes lesion location, size, and density;

[0067] S4: Lesion classification: Use a variety of machine learning algorithms to identify and classify the types of breast lesions, identify the categories to which the breast lesions belong, and obtain lesion classification results. Lesion types include benign lesions, malignant lesions, and suspected lesions that cannot be classified;

[0068] S5: Diagnostic report generation: Generate a diagnostic report based on the lesion classification results. The diagnostic report is used to provide diagnostic conclusions and examination recommendations. The diagnostic report includes lesion description and treatment recommendations.

[0069] It should be further explained that, in the specific implementation process, multi-scale and multi-feature analysis can extract information from multiple aspects of the size, shape, density, and edge contour features of the lesions, providing rich basis for accurate judgment; the above features are combined to comprehensively and accurately describe the lesion situation; and a variety of machine learning algorithms are fused for lesion classification to improve classification accuracy; an integrated learning method is used to construct convolutional neural networks CNN, random forests, and support vector machines SVM, and the prediction results are fused through a weighted average mechanism and a voting mechanism, and the model prediction confidence is calculated for weighted fusion, giving full play to the advantages of each model and reducing the limitations of a single model.

[0070] The generated diagnostic report contains lesion descriptions and treatment recommendations, providing doctors with detailed references to help them make more accurate diagnostic decisions, especially in early breast cancer screening, which helps to increase the early diagnosis rate and improve patient prognosis. It can identify benign lesions, malignant lesions, and suspected lesions that cannot be classified, providing references in different directions for further clinical diagnosis and treatment, and helping to rationally arrange medical resources and the patient's subsequent diagnosis and treatment process.

[0071] The process of removing noise from an image is as follows:

[0072] Based on the collected mammography data, the distribution characteristics and intensity changes of the noise are observed, the noise is classified, and the noise type is distinguished as quantum noise or electronic noise. The noise type is obtained, and the noise in the image is removed based on the noise type. The process is as follows:

[0073] If quantum noise exists, a noise threshold is set, and when the change in pixel value in the local area exceeds the noise threshold, adaptive filtering is started;

[0074] If electronic noise exists, then if the imaging device allows, collect a dark field image, and subtract the pixel value corresponding to the dark field image from the original image to eliminate the influence of electronic noise and complete the correction;

[0075] Among them, the distribution characteristics include uniform distribution or local concentration;

[0076] It should be further explained that, in the specific implementation process, the dark field image is an image collected under the same imaging conditions without X-ray irradiation; among them, the quantum noise is adaptively filtered: an adaptive median filter is used to adjust the filter window size according to the statistical characteristics of the local area of ​​the image; in areas with more noise, the window is correspondingly enlarged to remove the noise; in areas with rich image details, the window is reduced to protect the details.

[0077] The standardized processing flow can remove image noise, standardize grayscale values ​​and enhance local contrast, making image data more suitable for subsequent analysis and improving processing efficiency. For example, for different types of noise: quantum noise and electronic noise, different removal methods are used, and appropriate transformation methods are selected for grayscale values ​​according to the histogram distribution. Local contrast is enhanced by using a local contrast stretching enhancement algorithm. These preprocessing operations help to conduct subsequent feature extraction and analysis more quickly and accurately. The entire process from image acquisition to diagnostic report generation is automated, reducing manual intervention and improving screening efficiency. It can quickly process a large amount of breast mammography image data and is suitable for large-scale population screening.

[0078] For the standardization of grayscale values, the process is as follows:

[0079] Since different imaging devices and imaging conditions may lead to differences in the range and distribution of grayscale values, they need to be standardized to a unified range of 0-255; analyze the grayscale histogram of mammography images to determine the distribution range and distribution characteristics of grayscale values ​​in the image;

[0080] Use linear grayscale transformation or nonlinear grayscale transformation method, and select the corresponding mapping function according to the distribution of grayscale histogram. Standardize the grayscale value, the process is as follows:

[0081] If the grayscale value is evenly distributed, then a linear grayscale transformation is used: a linear mapping function y=ax+b is used, where x is the original grayscale value, y is the standardized grayscale value, and a and b are coefficients calculated based on the target grayscale range;

[0082] If the grayscale value distribution is uneven, then use nonlinear grayscale transformation: use the logarithmic transformation y=c*log(1+x) in the nonlinear transformation function; where c is the adjustment coefficient to make the grayscale value distribution uniform and improve the image contrast;

[0083] Among them, the gray value distribution is uneven: there are many low gray value or high gray value pixels concentrated;

[0084] It should be further explained that, in the specific implementation process, if the grayscale value range of the original image is 0-4095 and the target range is 0-255, for linear mapping, a=255 / 4095 and b=0 are calculated; for logarithmic transformation, the c value can be adjusted so that the transformed grayscale value can be reasonably distributed in the range of 0-255, thereby enhancing the contrast between different tissue areas in the image and facilitating subsequent feature extraction and analysis.

[0085] The process of enhancing local contrast is as follows:

[0086] Based on local statistical information, a local contrast stretching enhancement algorithm is used. According to the gray value distribution in the local neighborhood of each pixel in the image, a window between 3x3 and 11x11 is selected to stretch the pixel value to enhance the local contrast. The process is as follows:

[0087] Calculate the minimum and maximum values ​​in the neighborhood of each pixel, and then convert the gray value x of the pixel into Perform transformation, set the target grayscale range to 0-255, and enhance the grayscale value contrast in the local area;

[0088] It should be further explained that, in the specific implementation process, during the enhancement of local contrast, considering the delicate texture of breast tissue and the relatively low contrast between the lesion area and the surrounding tissue, an appropriate neighborhood window size can be selected according to the size of the breast tissue and the lesion, so as to enhance the contrast without destroying the overall structure of the image.

[0089] Step S2 includes:

[0090] S21: Multi-scale decomposition: Based on wavelet transform, the image dataset is decomposed into sub-images of different scales and directions by three-layer wavelet decomposition. In each layer of decomposition, the image dataset is decomposed into a low-frequency sub-band and three high-frequency sub-bands.

[0091] It should be further explained that, in the specific implementation process, the low-frequency sub-band is an approximate image, which contains the overall information of the image, and its resolution gradually decreases with the increase of the number of decomposition layers; the high-frequency sub-band is a detail image, which reflects the details and edge information of the image in different directions;

[0092] S22: Scale-related feature calculation:

[0093] Compute lesion-related features on sub-images of different scales:

[0094] In the low-frequency subband of wavelet decomposition, the Otsu threshold method is used to preliminarily segment the lesion area, calculate the number of pixels, and obtain the segmentation result. Then, the lesion boundary is located and refined in the high-frequency subband to complete the correction of the segmentation result and obtain the size characteristics of the lesion.

[0095] In the high-frequency subbands of different layers of wavelet decomposition, the edge contour of the lesion is calculated, and then its shape descriptor is calculated. The changes in the shape characteristics of the lesion at different scales are compared to determine whether the lesion has a regular shape or whether there are irregular features such as lobulation and burrs, and then the shape characteristics of the lesion are obtained;

[0096] It should be further explained that, in the specific implementation process, for the shape features of the lesions, the shape descriptors of the lesions at different scales are calculated, and the process is as follows: wherein the shape descriptors include circularity and rectangularity:

[0097] The calculation formula of circularity is: Among them, Area is the area of ​​the lesion, and Perimeter is the perimeter of the lesion. The closer the value is to 1, the closer the lesion is to a circle.

[0098] The rectangularity is expressed as the ratio of the area of ​​the lesion to the area of ​​the smallest circumscribed rectangle, reflecting the regularity of the shape of the lesion;

[0099] In multi-scale analysis, the changes of these shape descriptors at different scales are observed to analyze the stability and characteristic of the lesion shape;

[0100] The gray level co-occurrence matrix GLCM is used to extract the texture features of the lesions on multi-scale images:

[0101] In the sub-images of different layers of wavelet decomposition, the statistics of GLCM are calculated on each scale of the image, including energy, contrast, correlation, and entropy;

[0102] By observing the changing rules of texture features at different scales, we can find that: at low scales, it reflects the overall texture trend of the lesion, while at high scales, it reflects the subtle texture changes inside the lesion;

[0103] Among them, energy reflects the uniformity of image grayscale distribution and texture coarseness, contrast reflects the contrast of local grayscale changes in the image, correlation describes the linear relationship between pixel grayscale values ​​in the image, and entropy represents the randomness of image texture;

[0104] Analyzing the scale changes of these texture features can help distinguish different types of lesions, including that the texture features of benign and malignant lesions may have different manifestations at different scales;

[0105] Based on the multi-layer sub-images decomposed by wavelet, the threshold segmentation method is used to obtain the lesion area on images of different scales, and then the average grayscale value of the pixels in the lesion area is calculated to obtain the quantized lesion density characteristics; the Canny operator in the edge detection operator is used to calculate the amplitude and direction of the image gradient to obtain the edge contour characteristics of the lesion; the size characteristics, shape characteristics, density characteristics and edge contour characteristics of the lesion obtained are combined to form the lesion characteristics.

[0106] In step S3, the process of generating breast lesions is as follows:

[0107] S31: Comparison between the trained deep learning model and normal image features: Using the trained deep learning model, the collected and standardized mammography image data are compared with the normal mammography image features; wherein the mammography image data includes lesion features;

[0108] S32: Identify abnormal lesion areas: The deep learning model compares the mammography data to find areas with different characteristics from normal mammography images and marks these areas as abnormal lesion areas.

[0109] S33: Determine suspected breast cancer lesions: Determine whether there are suspected breast cancer lesions in the area based on the characteristics of the identified abnormal lesion area;

[0110] S34: Generate annotation information: If there is a suspected breast cancer lesion, generate annotation information including the location, size, and density of the lesion, clarify the location coordinates of the lesion in the image, measure and record the size of the lesion, and calculate the average grayscale value of the pixels in the lesion area to obtain density information;

[0111] S35: Generate breast lesions based on the annotation information: According to the annotation information, determine the corresponding abnormal area in the image as a breast lesion, and display the annotation information of the lesion on the image.

[0112] In step S33, the process of determining whether there is a suspected breast cancer lesion in the region is as follows:

[0113] S331: Data training: Using annotated mammography data as training set, the convolutional neural network (CNN) structure is used to train the network to identify lesion areas; by allowing the model to learn the characteristic patterns of many known lesion areas, the model is able to recognize the characteristics of lesions; the known lesion areas include benign and malignant lesions;

[0114] S332: Abnormality discrimination: Based on the trained convolutional neural network, feature matching is performed on each pixel of the mammography image data, and the pixels in the mammography image data are compared and analyzed one by one with the lesion feature patterns learned by the model;

[0115] Identify areas with similar features to known lesions. When the pixel features of a certain area show a high degree of similarity with the lesion features in the training set, the area is determined to be a suspected lesion area.

[0116] Generate a confidence score. The model generates a confidence score for each suspected lesion area based on the degree of feature matching. The score is used to evaluate the reliability of the detection results. The higher the confidence score, the greater the possibility that the area is a lesion.

[0117] The preset confidence threshold is used to determine whether a suspected breast cancer lesion exists. If the confidence score of a suspected lesion area is higher than the preset confidence threshold, it is considered that there is a suspected breast cancer lesion in the area. On the contrary, if it is lower than the confidence threshold, the area is considered normal.

[0118] It should be further explained that in the specific implementation process, the model optimization steps are: combining transfer learning and reinforcement learning to optimize the parameters of the convolutional neural network, so that the model can still maintain a high detection accuracy under small sample data conditions, thereby more accurately judging suspected breast cancer lesions; introducing knowledge and experience from other related fields or similar tasks with a large amount of data through transfer learning to help the model learn faster and better; using reinforcement learning to give reward or penalty feedback based on the model's judgment results, guiding the model to continuously optimize its own judgment logic and parameter settings, so as to improve the accuracy of judging suspected breast cancer lesions.

[0119] The deep learning model can effectively identify abnormal lesion areas by comparing them with normal image features. Its data training process uses labeled breast mammography image data to enable the model to learn the characteristic patterns of many known lesion areas, thereby having the ability to accurately identify lesion characteristics. When judging suspected breast cancer lesions, it generates a confidence score by matching the features of each pixel, and judges based on the preset threshold to improve detection reliability.

[0120] In step S4, the process of obtaining the lesion classification result is as follows:

[0121] S41: Classification model establishment: Multiple classification models were constructed using ensemble learning methods, including convolutional neural network (CNN), random forest, and support vector machine (SVM). The classification models were trained using mammary mammography data with lesion types labeled to enable the models to learn image feature patterns of different types of lesions. Lesion types included benign, malignant, and suspected.

[0122] S42: Multi-model fusion step:

[0123] S421: Fusion of prediction results based on different classification models:

[0124] A weighted average mechanism is used to assign different weights to each model according to its accuracy and recall during training and validation. If the CNN model has a high accuracy rate in classifying benign lesions on the validation set, it will be given a relatively high weight in the weighted average.

[0125] At the same time, a voting mechanism is used. Each model votes on the category to which the breast lesion belongs. The category with the most votes is used as the preliminary fusion result. If three models classify a breast lesion, two of which vote as benign and one vote as malignant, then the preliminary fusion result is benign.

[0126] S422: Calculate the prediction confidence of each model and weightedly fuse the results of multiple models to obtain the final classification result:

[0127] Through the calculation method within the model, the confidence of each model in the prediction result is obtained; then these confidence results are weighted and fused to obtain a more accurate and reliable fused lesion classification result;

[0128] Among them, the calculation methods within the three models are: CNN is based on the probability output layer, random forest is based on the distribution ratio of samples in each decision tree, and SVM is based on the distance to the hyperplane.

[0129] In step S41, the process of establishing the three classification models is as follows:

[0130] For the convolutional neural network (CNN), its structural design is based on the learning requirements of breast lesion image features. It extracts complex features in the image through multiple convolutional layers and pooling layers to obtain the neural network prediction results.

[0131] For random forests, multiple decision trees are constructed, and each decision tree is used to vote on the classification results of the data to determine the final classification and obtain the random forest prediction results. When constructing a random forest, a feature subset and a data subset are randomly selected to train each decision tree to improve the generalization ability of the model.

[0132] For the support vector machine (SVM), the benign, malignant, and suspected lesion data are separated by finding the optimal hyperplane, and the lesion data of different categories are mapped to a high-dimensional space through the kernel function of the support vector machine to obtain the prediction results of the vector machine, so as to better handle nonlinear classification problems;

[0133] It should be further explained that in the specific implementation process, the convolutional neural network CNN obtains the prediction result:

[0134] The preprocessed mammography data is input into the CNN model. The mammography data first passes through the convolution layer. The convolution kernel slides on the image to extract features of the local area, detect the edge features and texture features of the lesion, and generate a feature map.

[0135] Next, these feature maps are downsampled through the pooling layer to reduce the data dimension and the amount of calculation while retaining important feature information; then, through the alternating processing of multiple convolutional layers and pooling layers, more abstract and advanced features are gradually extracted, which are recorded as advanced features;

[0136] Finally, the extracted advanced features are mapped to the output nodes of each category through the fully connected layer to obtain the output value. Each node corresponds to a category, and the output value represents the probability estimate of the image belonging to that category. Among them, the categories are benign, malignant, and suspected lesions.

[0137] The output of the fully connected layer is normalized using the softmax function to convert the output value into a probability distribution. For example, for a breast lesion, after softmax processing, the result obtained is [0.1 (benign probability), 0.8 (malignant probability), 0.1 (suspected lesion probability)], indicating that the lesion is judged to be malignant. This probability distribution is the prediction result of the CNN model for the breast lesion.

[0138] The process of random forest obtaining prediction results:

[0139] Building decision trees and data sampling: Random forests are composed of multiple decision trees. When building each decision tree, 7% to 80% of the samples are randomly extracted with replacement from the training data set as the training data for the decision tree.

[0140] At the same time, when each node splits, some features are randomly selected for evaluation instead of using all features; in this way, each decision tree has a certain degree of randomness, thereby increasing the diversity of the model;

[0141] Decision tree training and voting: Each decision tree is trained based on its own training data to learn how to classify samples into benign, malignant, or suspected lesions based on features; including the decision tree constructing classification rules based on the size, shape, and density features of the lesions;

[0142] When a new breast lesion is predicted, the lesion data is input into each decision tree in the random forest, and each decision tree votes on the classification based on the rules it has learned; including one decision tree voting for benign and another voting for malignant;

[0143] Determine the final prediction result: Finally, the voting results of all decision trees are counted, and the category with the most votes is the prediction result of the random forest for the breast lesion. For example, if there are 50 decision trees, 30 of which vote for benign, 15 vote for malignant, and 5 vote for suspected lesions, then the prediction result of the random forest is benign.

[0144] The process of obtaining prediction results by support vector machine SVM:

[0145] Data mapping and hyperplane determination:

[0146] SVM first selects a linear kernel as the kernel function, represents the mammography data as a feature vector, and then maps it to a high-dimensional space;

[0147] In this high-dimensional space, SVM searches for an optimal hyperplane to separate benign, malignant, and suspected lesion samples, maximizing the distance between samples of different categories and the hyperplane. This process is achieved by solving an optimization problem to determine the parameters of the hyperplane.

[0148] Prediction classification: When predicting a new breast lesion, its feature vector is mapped to a high-dimensional space through the same kernel function;

[0149] Then the distance from the mapped vector to the hyperplane is calculated. If the distance is on one side of the hyperplane and is greater than the set plane threshold, it is determined to be the category corresponding to that side; that is, if it is on the benign side and the distance is far, it is determined to be benign; if the distance is close to the boundary area, it indicates a suspected lesion.

[0150] Step S42 also includes: optimization of classification accuracy, the process is as follows:

[0151] Tune the hyperparameters of the classification model through cross-validation: divide the training data into multiple subsets, use k-fold cross-validation, select one of the subsets as the validation set each time, and the remaining subsets as the training set, train the model and evaluate the performance on the validation set to obtain the evaluation structure;

[0152] Adjust the model's hyperparameters based on the evaluation results, including adjusting the convolution kernel size and learning rate in CNN; adjusting the number of trees and tree depth in random forest; adjusting the penalty parameter C and kernel function parameters in SVM.

[0153] Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field and related fields without creative work should fall within the scope of protection of the present invention. The structures, devices and operating methods not specifically described and explained in the present invention are implemented according to the conventional means in the field unless otherwise specified and limited.

Claims

1. An AI-assisted screening method based on mammography, characterized in that: The following steps are involved: S1: Image acquisition: Collect the patient's breast mammography image data and perform standardization processing to remove noise in the image, standardize the gray value and enhance the local contrast to obtain a structured image data set; S2: Feature extraction: Perform multi-scale and multi-feature analysis on standardized image data sets to extract lesion features, including the size, shape, density, and edge contour features of the lesions; S3: Abnormality detection: Use the trained deep learning model to compare with normal image features, identify abnormal lesion areas in lesion features, determine whether there are suspected breast cancer lesions, and generate annotation information. At the same time, generate breast lesions based on the annotation information, where the annotation information includes lesion location, size, and density; S4: Lesion classification: Identify and classify the lesion types of breast lesions through a variety of machine learning algorithms, identify the categories to which the breast lesions belong, and obtain lesion classification results. The lesion types include benign lesions, malignant lesions, and suspected lesions that cannot be classified; S5: Diagnostic report generation: Generate a diagnostic report based on the lesion classification results. The diagnostic report includes lesion description and treatment recommendations.

2. The AI-assisted screening method based on mammography according to claim 1, characterized in that: The process of removing noise from an image is as follows: Based on the collected mammography data, the distribution characteristics and intensity changes of the noise are observed, the noise is classified, and the noise type is distinguished as quantum noise or electronic noise. The noise type is obtained, and the noise in the image is removed based on the noise type. The process is as follows: If quantum noise exists, a noise threshold is set, and when the change in pixel value in the local area exceeds the noise threshold, adaptive filtering is started; If electronic noise exists, then if the imaging device allows, collect a dark field image, and subtract the pixel value corresponding to the dark field image from the original image to eliminate the influence of electronic noise and complete the correction; Among them, the distribution characteristics include uniform distribution or local concentration.

3. The AI-assisted screening method based on mammography according to claim 2, characterized in that: For the standardization of grayscale values, the process is as follows: Analyze the grayscale histogram of mammography images to determine the distribution range and characteristics of grayscale values ​​in the image; A linear grayscale transformation or a nonlinear grayscale transformation method is used to select a corresponding mapping function according to the distribution of the grayscale histogram.

4. The AI-assisted screening method based on mammography according to claim 3, characterized in that: The process of enhancing local contrast is as follows: Based on local statistical information, a local contrast stretching enhancement algorithm is used. According to the gray value distribution in the local neighborhood of each pixel in the image, a window between 3x3 and 11x11 is selected to stretch the pixel value to enhance the local contrast. The process is as follows: Calculate the minimum and maximum values ​​in the neighborhood of each pixel, and then convert the gray value x of the pixel into Perform the transformation, set the target grayscale range to 0-255, and enhance the grayscale value contrast in the local area.

5. The AI-assisted screening method based on mammography according to claim 4, characterized in that: Step S2 includes: S21: Multi-scale decomposition: Based on wavelet transform, the image dataset is decomposed into sub-images of different scales and directions by three-layer wavelet decomposition. In each layer of decomposition, the image dataset is decomposed into a low-frequency sub-band and three high-frequency sub-bands. S22: Scale-related feature calculation: Compute lesion-related features on sub-images of different scales: In the low-frequency subband of wavelet decomposition, the Otsu threshold method is used to preliminarily segment the lesion area, calculate the number of pixels, and obtain the segmentation result. Then, the lesion boundary is located and refined in the high-frequency subband to complete the correction of the segmentation result and obtain the size characteristics of the lesion. In the high-frequency subbands of different layers of wavelet decomposition, the edge contour of the lesion is calculated, and then its shape descriptor is calculated. The changes in the shape characteristics of the lesion at different scales are compared to determine whether the lesion has a regular shape or whether there are irregular features such as lobulation and burrs, and then the shape characteristics of the lesion are obtained; Based on the multi-layer sub-images decomposed by wavelet, the threshold segmentation method is used to obtain the lesion area on images of different scales, and then the average grayscale value of the pixels in the lesion area is calculated to obtain the quantized lesion density characteristics; the Canny operator in the edge detection operator is used to calculate the amplitude and direction of the image gradient to obtain the edge contour characteristics of the lesion; the size characteristics, shape characteristics, density characteristics and edge contour characteristics of the lesion obtained are combined to form the lesion characteristics.

6. The AI-assisted screening method based on mammography according to claim 5, characterized in that: In step S3, the process of generating breast lesions is as follows: S31: Comparison between the trained deep learning model and normal image features: Using the trained deep learning model, the collected and standardized mammography image data are compared with the normal mammography image features; wherein the mammography image data includes lesion features; S32: Identify abnormal lesion areas: The deep learning model compares the mammography data to find areas with different characteristics from normal mammography images and marks these areas as abnormal lesion areas. S33: Determine suspected breast cancer lesions: Determine whether there are suspected breast cancer lesions in the area based on the characteristics of the identified abnormal lesion area; S34: Generate annotation information: If there is a suspected breast cancer lesion, generate annotation information including the location, size, and density of the lesion, clarify the location coordinates of the lesion in the image, measure and record the size of the lesion, and calculate the average grayscale value of the pixels in the lesion area to obtain density information; S35: Generate breast lesions based on the annotation information: According to the annotation information, determine the corresponding abnormal area in the image as a breast lesion, and display the annotation information of the lesion on the image.

7. The AI-assisted screening method based on mammography according to claim 6, characterized in that: In step S33, the process of determining whether there is a suspected breast cancer lesion in the region is as follows: S331: Data training: Using annotated mammography data as training set, the convolutional neural network (CNN) structure is used to train the network to identify lesion areas; by allowing the model to learn the characteristic patterns of many known lesion areas, the model is able to recognize the characteristics of lesions; the known lesion areas include benign and malignant lesions; S332: Abnormality discrimination: Based on the trained convolutional neural network, feature matching is performed on each pixel of the mammography image data, and the pixels in the mammography image data are compared and analyzed one by one with the lesion feature patterns learned by the model; Identify areas with similar features to known lesions. When the pixel features of a certain area show a high degree of similarity with the lesion features in the training set, the area is determined to be a suspected lesion area. Generate a confidence score. The model generates a confidence score for each suspected lesion area based on the degree of feature matching. The score is used to evaluate the reliability of the detection results. The higher the confidence score, the greater the possibility that the area is a lesion. The preset confidence threshold is used to determine whether a suspected breast cancer lesion is present. If the confidence score of a suspected lesion area is higher than the preset confidence threshold, the area is considered to contain a suspected breast cancer lesion. Otherwise, if the confidence score is lower than the confidence threshold, the area is considered normal.

8. The AI-assisted screening method based on mammography according to claim 7, characterized in that: In step S4, the process of obtaining the lesion classification result is as follows: S41: Classification model establishment: Multiple classification models were constructed using ensemble learning methods, including convolutional neural network (CNN), random forest, and support vector machine (SVM). The classification models were trained using mammary mammography data with lesion types labeled to enable the models to learn image feature patterns of different types of lesions. Lesion types included benign, malignant, and suspected. S42: Multi-model fusion step: S421: Fusion of prediction results based on different classification models: A weighted average mechanism is used to assign different weights to each model according to its accuracy and recall during training and validation. At the same time, a voting mechanism is used, where each model votes on the category to which the breast lesion belongs, and the category with the most votes is used as the preliminary fusion result; S422: Calculate the prediction confidence of each model and weightedly fuse the results of multiple models to obtain the final classification result: The confidence level of each model in the prediction result is obtained through the calculation method within the model. Then, these confidence level results are weighted and fused to obtain the fused lesion classification result. Among them, the calculation methods within the three models are: CNN is based on the probability output layer, random forest is based on the distribution ratio of samples in each decision tree, and SVM is based on the distance to the hyperplane.

9. The AI-assisted screening method based on mammography according to claim 8, characterized in that: In step S41, the process of establishing the three classification models is as follows: For the convolutional neural network (CNN), its structural design is based on the learning requirements of breast lesion image features. It extracts complex features in the image through multiple convolutional layers and pooling layers to obtain the neural network prediction results. For random forests, multiple decision trees are constructed, and each decision tree is used to vote on the classification results of the data to determine the final classification and obtain the random forest prediction results; For the support vector machine (SVM), the benign, malignant, and suspected lesion data are separated by finding the optimal hyperplane, and the different categories of lesion data are mapped to a high-dimensional space through the kernel function of the support vector machine to obtain the vector machine prediction results.

10. The AI-assisted screening method based on mammography according to claim 9, characterized in that: Step S42 also includes: optimization of classification accuracy, the process is as follows: Tune the hyperparameters of the classification model through cross-validation: divide the training data into multiple subsets, use k-fold cross-validation, select one of the subsets as the validation set each time, and the remaining subsets as the training set, train the model and evaluate the performance on the validation set to obtain the evaluation structure; Adjust the model's hyperparameters based on the evaluation results, including adjusting the convolution kernel size and learning rate in CNN; adjusting the number of trees and tree depth in random forest; adjusting the penalty parameter C and kernel function parameters in SVM.

Images