A method and apparatus for breast tumor screening based on broadband ultrasound echo signatures
By combining broadband ultrasound echo features, target detection, and statistical learning models, a highly accurate screening method for breast tumors has been achieved, solving the problems of high subjectivity and low accuracy in existing technologies, and providing an efficient and reliable method and device for breast tumor screening.
Patent Information
- Application Number
- CN202210871399.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-07-22
- Publication Date
- 2026-02-06
- Estimated Expiration
- 2042-07-22
AI Technical Summary
In existing technologies, breast cancer screening methods are highly subjective, have a high false positive rate, and are time-consuming and labor-intensive. Furthermore, computer-aided methods have low accuracy and are computationally difficult, while deep learning models are limited by data scale and interpretability.
A screening method using broadband ultrasound echo features, combined with a target detection model and a statistical learning model, is employed. Ultrasound echo radio frequency signals are acquired through multi-angle scanning, time-spectrum features are calculated, lesion locations and types are labeled, and a dataset is trained using deep learning and statistical learning models to generate high-contrast visualized ultrasound images.
It achieves high accuracy and reliability in breast tumor screening, reduces false positives, improves screening efficiency, and provides structural and physical interpretable spectral feature analysis.
Smart Images
Figure CN115227293B_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application relates to an ultrasonic medical imaging technology, in particular to a breast tumor screening method and device based on wideband ultrasonic echo characteristics. BACKGROUND
[0002] Medical ultrasonic examination is a commonly used breast examination method. Professional doctors often complete dynamic examination of suspicious areas by operating an ultrasonic transducer. In clinical practice, the evaluation of breast lesions under ultrasonography needs to comprehensively consider the size, shape, edge, direction and gray scale of the suspicious area. If the doctor considers that the examination area has a canceration probability, further puncture biopsy is needed for confirmation. This screening is subjective and qualitative on the one hand, and has a high false positive rate, resulting in a lot of unnecessary puncture biopsy; on the other hand, this screening is time-consuming and labor-intensive, and requires a lot of experience of the doctor, which hinders the popularization of breast tumor screening.
[0003] At present, the research on computer-aided ultrasonic breast tumor screening algorithm in the academic field mainly has two directions. Some researches start from the frequency domain characteristics or statistical characteristics of the ultrasonic echo radio frequency signal, and quantitatively analyze the breast lesions. Common ultrasonic quantitative characteristics include scattering coefficient, attenuation rate, scattering sub-size, Nakagami distribution model parameters and Homodyned-K distribution model parameters, etc. These artificially designed and extracted parameters can represent the physical characteristics or microscopic structure characteristics of the tissue to a certain extent. However, quantitative ultrasonic research often faces the problems of low accuracy and difficult calculation. Another part of the research starts from the image and visual angle to analyze the breast ultrasonic image. Among them, various deep learning models have achieved good results in the detection and diagnosis of breast ultrasonography. However, the application of deep learning models in ultrasonic images is restricted by many problems such as data size, data quality and weak interpretability. SUMMARY
[0004] The present application aims at the deficiencies of the prior art, and provides a breast tumor screening method based on wideband ultrasonic echo characteristics. The screening method first uses a target detection model to locate and preliminarily classify breast lesions in the ultrasonic image based on the ultrasonic image, then uses a statistical learning model to locally analyze and identify the tissue inside the lesion using the time-frequency spectrum characteristics of the ultrasonic echo radio frequency signal. Finally, different colors are used to mark the lesion types and lesion degrees of the local lesions in the image, so as to realize the preliminary screening of the breast tumor ultrasonic image.
[0005] To achieve the above-mentioned purpose, the present application adopts the following technical solutions:
[0006] A breast tumor screening method based on wideband ultrasonic echo characteristics, comprising the following steps:
[0007] Step S1, ultrasonic signal acquisition: using a wideband or multi-frequency ultrasonic transducer to perform multi-angle scanning on the breast to obtain ultrasonic echo RF signals;
[0008] Step S2, ultrasonic signal labeling: calculating the envelope and time-frequency spectrum of the ultrasonic echo RF signals respectively, performing logarithmic compression on the envelope to obtain an ultrasound image, labeling the tumor lesion position, contour and type on the ultrasound image, constructing a breast tumor ultrasound image dataset, saving the time-frequency spectrum of the labeled position area, extracting the time-frequency spectrum feature parameters, and constructing a breast tumor ultrasound signal time-frequency spectrum dataset;
[0009] Step S3, image dataset training: training the breast tumor ultrasound image dataset through a target detection model to obtain a breast tumor lesion position detection model based on ultrasound images;
[0010] Step S4, time-frequency spectrum dataset training: training the breast tumor ultrasound signal time-frequency spectrum dataset through a statistical learning model to obtain a breast tumor classification model based on time-frequency spectrum feature parameters;
[0011] Step S5, model prediction: using the trained breast tumor lesion position detection model to mark the tumor area in the test set image, and using the breast tumor classification model to predict the tumor type of the marked area signal, and weighting the prediction probability of each model to obtain the final prediction probability value of the test sample;
[0012] Step S6, prediction result visualization: color coding the lesions in the ultrasound image according to the lesion type and prediction probability, generating a high-contrast wideband ultrasound spectrogram image, and realizing high-contrast and high-resolution visualization of breast tumor screening.
[0013] As a preferred, in step S1, a wideband or multi-frequency ultrasonic transducer is used to perform multi-angle scanning on the breast to obtain ultrasonic echo RF signals; the wideband ultrasonic transducer has a wide bandwidth range, containing information of a wider frequency range, which is beneficial to the analysis of signal time-frequency spectrum. The multi-frequency ultrasonic transducer has multiple center frequencies and bandwidth ranges, which are combined into a wide bandwidth range.
[0014] As a preferred, in step S2, Hilbert transform is performed on the ultrasonic signal, and the envelope of each column signal is merged to obtain an ultrasound image, and the breast position, contour and lesion type are labeled on the ultrasound image to establish a breast tumor ultrasound image dataset with each frame of image as a sample and the lesion type as a label.
[0015] As a preferred, in step S2, the signal time-frequency spectrum acquisition method is as follows: using wavelet transform on the signal corresponding to each column of the ultrasound image, calculating the time-frequency spectrum coefficient of the signal at any time under different frequencies. Merging the time-frequency coefficient spectrum of each column of signal, the time-frequency coefficient spectrum of the signal at any position in the image is obtained, and thus the signal time-frequency feature information is obtained.
[0016] As preferred, in step S2, a sliding window grid of a set size is used as a sample unit to calculate mathematical statistics such as time-frequency spectrum mean, variance, kurtosis, skewness, etc. that can be calculated based on time-frequency spectrum coefficients in each sample, and the mathematical statistics are taken as feature parameters, and the tumor category to which the sample belongs is taken as a label to construct a signal time-frequency spectrum dataset. In particular, when the sliding window grid size is 1*1 pixel, the time-frequency spectrum coefficients corresponding to the pixel points can be taken as sample feature parameters to construct the dataset.
[0017] As preferred, in step S3, a breast tumor ultrasound image dataset is trained by a deep learning-based target detection model. First, the ultrasound image dataset is proportionally divided into a training set, a validation set and a test set, then the target detection model is used to cross-train the training set and the validation set to obtain a lesion position detection model, the test set ultrasound image data is predicted by the lesion position detection model, performance indicators such as sensitivity and specificity are calculated, and the parameters are optimized to realize the recognition of the lesion position and type of the test set.
[0018] As preferred, in step S4, a breast tumor ultrasound signal time-frequency spectrum dataset is trained by a statistical learning model. First, the ultrasound signal time-frequency spectrum dataset is proportionally divided into a training set, a validation set and a test set, then the statistical learning model is used to cross-train the training set and the validation set to obtain a breast tumor classification model, the test set ultrasound signal time-frequency spectrum data is predicted by the breast tumor classification model, performance indicators such as sensitivity and specificity are calculated, and the parameters are optimized to realize the recognition of the tumor lesion type and degree of the sample points in the test set.
[0019] As preferred, in step S5, a certain frame of ultrasound image and its corresponding signal time-frequency spectrum in the test set are given, the breast tumor lesion position detection model is used to mark the tumor area in the test set image, and the breast tumor classification model is used to predict the tumor type of the signal time-frequency spectrum corresponding to the marked area. The lesion position detection model can give the tumor area and the category probability in the image, and on this basis, the tumor classification model gives the category probability of the signal time-frequency spectrum of all sample points in the area, and the final prediction probability of all sample points in the tumor area is obtained by weighted calculation.
[0020] As preferred, in step S6, the lesions in the ultrasound image are HSV color coded according to the lesion type and the prediction probability to generate a high-contrast broadband ultrasound spectrogram image, wherein the type of the predicted target area is represented as hue (Hue), the prediction confidence is represented as saturation (Saturation), and the amplitude of the signal is represented as brightness (Value), and finally the high-contrast visualization of breast tumor screening is realized.
[0021] The application also provides a breast tumor screening device based on broadband ultrasound echo features, comprising:
[0022] An ultrasonic signal acquisition module uses a wideband or multi-frequency ultrasonic transducer to perform multi-angle scanning on the breast to obtain ultrasonic echo RF signals;
[0023] An ultrasonic signal labeling module calculates the envelope and time-frequency spectrum of the ultrasonic echo RF signals, respectively, obtains an ultrasonic image by logarithmic compression of the envelope, labels the tumor lesion position, contour and type on the ultrasonic image, constructs a breast tumor ultrasonic image dataset, saves the time-frequency spectrum of the labeled position area, extracts the time-frequency spectrum feature parameters, and constructs a breast tumor ultrasonic signal time-frequency spectrum dataset;
[0024] An ultrasonic dataset training module trains the breast tumor ultrasonic image dataset through a target detection model to obtain a breast tumor lesion position detection model based on ultrasonic images;
[0025] A time-frequency spectrum dataset training module trains the breast tumor ultrasonic signal time-frequency spectrum dataset through a statistical learning model to obtain a breast tumor classification model based on time-frequency spectrum feature parameters;
[0026] A model prediction module uses the trained breast tumor lesion position detection model to mark the tumor area in the test set image, and uses the breast tumor classification model to predict the tumor type of the marked area signal, and weightedly calculates the prediction probability of each model to obtain the final prediction probability value of the test sample;
[0027] A prediction result visualization module color-codes the lesions in the ultrasonic image according to the lesion type and the prediction probability, generates a high-contrast wideband ultrasonic spectrogram image, and realizes high-contrast visualization of breast tumor screening.
[0028] The present application has the following technical effects:
[0029] The present application uses a wideband or multi-frequency ultrasonic transducer to perform multi-angle scanning on a breast to obtain an ultrasonic echo radio frequency signal; the ultrasonic echo radio frequency signal is calculated for an envelope and a time-frequency spectrum respectively, the envelope is logarithmically compressed to obtain an ultrasonic image, a tumor lesion position, contour and type are labeled on the ultrasonic image, a breast tumor ultrasonic image dataset is constructed, and a time-frequency spectrum of a labeled position area is saved, a time-frequency spectrum feature parameter is extracted, and a breast tumor ultrasonic signal time-frequency spectrum dataset is constructed; a breast tumor lesion position detection model based on an ultrasonic image is obtained by training a breast tumor ultrasonic image dataset based on a deep learning target detection model; a breast tumor classification model based on a time-frequency spectrum feature parameter is obtained by training a breast tumor ultrasonic signal time-frequency spectrum dataset based on a statistical learning model; a breast tumor lesion position detection model that has been trained is used to mark a tumor area in a test set image, and a breast tumor classification model is used to predict a tumor type to which a marked area signal belongs, and a final prediction probability value of a test sample is obtained by weighted calculation of respective prediction probabilities of the two models; a lesion in an ultrasonic image is color-coded according to a lesion type and a prediction probability, a high-contrast wideband ultrasonic spectrogram image is generated, and high-contrast visualization of breast tumor screening is realized. The present application proposes a breast tumor screening method combining ultrasonic image features and time-frequency spectrum features of ultrasonic backscattering signals, and adds structural and spectrum feature analysis with physical interpretation on the basis of an intelligent ultrasonic image analysis method, so that an ultrasonic breast tumor screening result is more accurate and reliable. BRIEF DESCRIPTION OF DRAWINGS
[0030] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings needed in the embodiments. Obviously, the drawings in the following description only constitute some embodiments of the present application, and for those skilled in the art, other drawings can also be obtained without creative labor.
[0031] Figure 1 The method flowchart provided by the embodiments of the present application.
[0032] Figure 2 (a) is a normal breast without lesions, (b) is a breast benign tumor, and (c) is a breast malignant tumor ultrasonic image.
[0033] Figure 3 is a time-frequency conversion schematic diagram. (a) is an ultrasonic image obtained after taking an envelope of a radio frequency signal and logarithmic compression; (b) is an image obtained by reconstructing wavelet spectrum coefficients of the radio frequency signal at different frequencies.
[0034] Figure 4is a lesion boundary marking schematic diagram. The gray solid line is a breast tumor region boundary marked by a professional physician, the black dotted line is an outer boundary line obtained by expanding the gray solid line outward, and the white dotted line is an inner boundary line obtained by contracting the gray solid line inward. DETAILED DESCRIPTION
[0035] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by a person of ordinary skill in the art without creative work fall within the protection scope of the present application.
[0036] Embodiment 1
[0037] The present application discloses a breast tumor screening method based on wideband ultrasonic echo features, comprising the following steps:
[0038] Step S1, ultrasonic signal acquisition: using a wideband or multi-frequency ultrasonic transducer to perform multi-angle scanning on the breast to obtain ultrasonic echo radio frequency signals;
[0039] As an embodiment of the present embodiment, a wideband or multi-frequency ultrasonic transducer is used to slowly slide scan the breast region from different directions such as the transverse and longitudinal planes and the cut planes of the patient's lesion, to obtain multiple frames of ultrasonic echo radio frequency signals and save them for offline processing.
[0040] Step S2, ultrasonic signal marking: calculating the envelope and time-frequency spectrum of the ultrasonic echo radio frequency signals respectively, performing logarithmic compression on the envelope to obtain an ultrasonic image, marking the tumor lesion position, contour and type on the ultrasonic image, constructing a breast tumor ultrasonic grayscale image dataset, saving the time-frequency spectrum of the marked position area, extracting time-frequency spectrum feature parameters, and constructing a breast tumor ultrasonic signal time-frequency spectrum dataset;
[0041] As an embodiment of the present embodiment, Hilbert transform and logarithmic compression are used to convert each frame of ultrasonic signal into a visual ultrasonic image form. The implementation is as follows: for each column of ultrasonic radio frequency signals x(t), the following operations are performed:
[0042] Y(t)=20×log 10 (1+H[x(t)])
[0043] Where H[x(t)] is the Hilbert transform of the ultrasonic radio frequency signal x(t), defined as
[0044]
[0045] Y(t) is the envelope of the Hilbert transform of the ultrasonic radio frequency signal x(t) after logarithmic compression, which reflects the amplitude information of the signal, and is reflected as the change of gray scale in the ultrasonic image, as shown in Figure 2 .
[0046] As an embodiment of the present embodiment, for a frame of ultrasonic echo radio frequency signal, the data is composed of a data matrix with the number of sampling points as rows and the number of scan lines as columns. The sampling points correspond to the echo signals of the tissues at different longitudinal depth positions in space. One-dimensional continuous wavelet transform is performed on the column signals to obtain the time-frequency spectrum coefficients of the column signals. The implementation of the continuous one-dimensional wavelet transform is as follows:
[0047]
[0048] wherein CWT{x(t);a,b} represents the continuous wavelet spectrum coefficient of the signal x(t) at scale a and depth b, and the scale is inversely proportional to the frequency; is the wavelet basis. The one-dimensional continuous wavelet transform can be quickly realized in numerical processing analysis software. One-dimensional continuous wavelet transform is performed on each column of the data matrix to expand the data into a three-dimensional matrix, and the data of the matrix is composed of: the number of sampling points x the number of scan lines x the number of scales (frequencies), as shown in Figure 3 . Figure 3 (a) is the gray scale image obtained by taking the envelope of the radio frequency signal and performing logarithmic compression; Figure 3 (b) is the image obtained by reconstructing the wavelet spectrum coefficients corresponding to the radio frequency signal at different frequencies.
[0049] As an embodiment of the present embodiment, as shown in Figure 4 , the breast tumor region is marked in the ultrasonic image, and an ultrasonic gray scale image data set is constructed. The gray solid line is the boundary of the breast tumor region marked by a professional physician, the black dashed line is the outer boundary line obtained by expanding the gray solid line outward, and the white dashed line is the inner boundary line obtained by contracting the gray solid line inward. The signal time-frequency spectrum data for training each tissue will come from within the respective inner boundary contour line, and the background signal time-frequency spectrum data for training each tissue will come from outside the outer boundary of all tissue contour lines; the area between the inner boundary and the outer boundary contour lines can not be involved in the training, and this method is mainly used to reduce the influence of the unclear boundary region on data labeling.
[0050] As an embodiment of the present embodiment, a sliding window grid of a given size is used as a sample unit, and mathematical statistics such as time-frequency spectrum mean, variance, kurtosis, skewness, etc. that can be calculated based on time-frequency spectrum coefficients in each sample and are not completely linearly correlated are calculated, and these mathematical statistics are used as feature parameters, and the tumor category to which the sample belongs is used as a label to construct a signal time-frequency spectrum dataset. In particular, when the size of the sliding window grid is 1*1 pixel, the time-frequency spectrum coefficients corresponding to the pixel points can be used as sample feature parameters to construct the dataset.
[0051] Step S3, image dataset training: training the breast tumor ultrasound image dataset through the target detection model to obtain a breast tumor lesion position detection model based on ultrasound images;
[0052] As an embodiment of the present embodiment, the open source target detection model such as Faster-RCNN, YOLO, etc. is adjusted and optimized according to the data characteristics, and the optimized model is trained on the breast tumor ultrasound image dataset obtained in step S2 to obtain a model that can detect whether there is a tumor lesion in the ultrasound image at a high frame rate, high sensitivity and high accuracy and give the lesion position.
[0053] Step S4, time-frequency spectrum dataset training: training the breast tumor ultrasound signal time-frequency spectrum dataset through a statistical learning model to obtain a breast tumor classification model based on time-frequency spectrum feature parameters;
[0054] As an embodiment of the present embodiment, the open source statistical learning model such as LightGBM integrated learning, feedforward neural network, etc. is adjusted and optimized according to the data characteristics, and the optimized model is trained on the breast tumor ultrasound signal time-frequency spectrum dataset obtained in step S2 to obtain a model that can predict the sample point category and probability inside the lesion based on the local lesion echo signal time-frequency spectrum as a feature.
[0055] Step S5, model prediction: using the trained breast tumor lesion position detection model to mark the tumor area in the test set image, and using the breast tumor classification model to predict the tumor type to which the signal in the marked area belongs, and calculating the prediction probability values of the test sample by weighting the prediction probabilities of the two models respectively;
[0056] As an embodiment of the present embodiment, a certain frame of ultrasound image and its corresponding signal time-frequency spectrum in the test set are given, the breast tumor lesion position detection model is used to mark the tumor area in the test set image, and the breast tumor classification model is used to predict the tumor type to which the signal time-frequency spectrum corresponding to the marked area belongs. The lesion position detection model can give the tumor area in the image and the category probability at the same time, and on this basis, the breast tumor classification model gives the category probability of the signal time-frequency spectrum corresponding to all sample points in the region, and the final prediction probability of all sample points in the tumor region is obtained by weighting calculation.
[0057] Step S6, prediction result visualization: color coding the lesions in the ultrasound image according to the lesion type and the prediction probability, generating a high-contrast broadband ultrasound spectrogram image, realizing high-contrast visualization of breast tumor screening.
[0058] As an embodiment of the present embodiment, the lesions in the ultrasound image are HSV color coded according to the lesion type and the prediction probability, and a high-contrast broadband ultrasound spectrogram image is generated, wherein the type of the predicted target region is represented as hue (Hue), the predicted confidence is represented as saturation (Saturation), and the amplitude of the signal is represented as brightness (Value), and finally high-contrast visualization of breast tumor screening is realized.
[0059] Embodiment 2:
[0060] The present application also provides a breast tumor screening device based on broadband ultrasound echo features, comprising:
[0061] An ultrasound signal acquisition module uses a broadband or multi-frequency ultrasound transducer to perform multi-angle scanning on the breast to obtain an ultrasound echo radio frequency signal;
[0062] An ultrasound signal labeling module calculates the envelope and time-frequency spectrum of the ultrasound echo radio frequency signal, respectively, performs logarithmic compression on the envelope to obtain an ultrasound image, labels the tumor lesion position, contour and type on the ultrasound image, constructs a breast tumor ultrasound image dataset, saves the time-frequency spectrum of the labeled position region, extracts the time-frequency spectrum feature parameters, and constructs a breast tumor ultrasound signal time-frequency spectrum dataset;
[0063] An image dataset training module trains the breast tumor ultrasound image dataset through a target detection model to obtain a breast tumor lesion position detection model based on the ultrasound image;
[0064] A time-frequency spectrum dataset training module trains the breast tumor ultrasound signal time-frequency spectrum dataset through a statistical learning model to obtain a breast tumor classification model based on the time-frequency spectrum feature parameters;
[0065] A model prediction module uses the trained breast tumor lesion position detection model to mark the tumor region in the test set image, and uses the breast tumor classification model to predict the tumor type of the marked region signal, and weightedly calculates the prediction probability of each model to obtain the final prediction probability value of the test sample;
[0066] A prediction result visualization module color codes the lesions in the ultrasound image according to the lesion type and the prediction probability, generates a high-contrast broadband ultrasound spectrogram image, and realizes high-contrast visualization of breast tumor screening.
[0067] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present application, and are not intended to limit the present application; although the present application has been described in detail with reference to the above embodiments, those skilled in the art should understand that the technical solutions recorded in the above embodiments can be modified, or some or all of the technical features can be replaced by equivalents; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope of the technical solutions of the embodiments of the present application.
Claims
1. A breast tumor screening apparatus based on broadband ultrasound echo signatures, characterized by, The method comprises the following steps: An ultrasonic signal acquisition module uses a wideband or multi-frequency ultrasonic transducer to perform multi-angle scanning on a breast to obtain ultrasonic echo radio frequency signals; An ultrasonic signal labeling module calculates the envelope and time-frequency spectrum of the ultrasonic echo radio frequency signals respectively, obtains an ultrasonic image by logarithmic compression of the envelope, labels the tumor lesion position, contour and type on the ultrasonic image, constructs a breast tumor ultrasonic image dataset, saves the time-frequency spectrum of the labeled position area, extracts time-frequency spectrum feature parameters, and constructs a breast tumor ultrasonic signal time-frequency spectrum dataset; Each frame of ultrasonic signal is converted into a visual ultrasonic graph form by using Hilbert transform and logarithmic compression, and each column of ultrasonic radio frequency signals x(t) is operated as follows: ； where H[x(t)] is the Hilbert transform of the ultrasonic radio frequency signal x(t) defined as: ; Y(t) is the envelope obtained by Hilbert transform of the ultrasonic radio frequency signal x(t) and is obtained by logarithmic compression of the envelope; For a frame of ultrasonic echo radio frequency signals, the data is composed of a data matrix with the number of sampling points as rows and the number of scanning lines as columns, and the sampling points correspond to the echo signals of the tissues at different depth positions in space. One-dimensional continuous wavelet transform is performed on the column signals to obtain the time-frequency spectrum coefficients of the column signals. The implementation mode of continuous one-dimensional wavelet transform is as follows: ； wherein represents the continuous wavelet spectrum coefficients of the signal x(t) at scale , depth b, scale being inversely proportional to frequency is the wavelet basis; a one-dimensional continuous wavelet transform is performed for each column of the data matrix, which extends the data into a three-dimensional matrix whose data structure is: number of sampling points x number of scan lines x number of scales or frequencies; The tumor region labeling method of the ultrasonic grayscale image is as follows: the contour of the labeled tumor region forms a closed loop, and the image within the contour line can be used for training of the grayscale image dataset; a small distance is expanded to both sides of the contour line by a self-defined algorithm to obtain an outer boundary contour line and an inner boundary contour line, so as to reduce the influence of the unclear boundary area on data labeling; the signal time-frequency spectrum data for training each type of tissue is obtained from within the respective inner boundary contour line, and the background signal time-frequency spectrum data for training each type of tissue is obtained from outside the outer boundary contour line of all tissue contour lines; the area between the inner boundary and the outer boundary contour lines can not be involved in the training; The extraction method of the time-frequency spectrum feature parameters is as follows: a sliding window grid of a set size is used as a sample unit, mathematical statistics based on the time-frequency spectrum coefficients and not completely linearly related within each sample are calculated, including time-frequency spectrum mean, variance, kurtosis and skewness, the mathematical statistics are taken as feature parameters, the sample belongs to a tumor category is taken as a label, and a signal time-frequency spectrum dataset is constructed; when the size of the sliding window grid is 1*1 pixel, the pixel point corresponding to the time-frequency spectrum coefficient can be taken as a sample feature parameter to construct a dataset; An ultrasonic dataset training module trains the breast tumor ultrasonic image dataset by using a target detection model to obtain a breast tumor lesion position detection model based on ultrasonic images; A time-frequency spectrum dataset training module trains the breast tumor ultrasonic signal time-frequency spectrum dataset by using a statistical learning model to obtain a breast tumor classification model based on time-frequency spectrum feature parameters; A model prediction module uses the trained breast tumor lesion position detection model to mark the tumor area in the test set image, uses the breast tumor classification model to predict the tumor type of the marked area signal, and calculates the prediction probability values of the two models respectively to obtain the final prediction probability value of the test sample. Given a certain frame of ultrasound image and its corresponding signal time-frequency spectrum in the test set, the breast tumor lesion location detection model is used to mark the tumor area in the test set image, and the tumor classification model is used to predict the tumor type of the marked area corresponding to the signal time-frequency spectrum; the lesion location detection model also gives the tumor area in the image and the category probability, on this basis, the breast tumor classification model gives the category probability of the signal time-frequency spectrum corresponding to all sample points in the region, and the final prediction probability of all sample points in the tumor region is obtained by weighted calculation; The prediction result visualization module color encodes the lesions in the ultrasound image according to the lesion type and the prediction probability, generates a high-contrast broadband ultrasound spectrogram image, and realizes the high-contrast visualization of breast tumor screening.
Citation Information
Patent Citations
Dynamic ultrasonic breast nodule real-time segmentation and recognition method based on deep learning
CN111539930A
Biological tissue identification method and device based on ultrasonic echo time-frequency spectrum characteristics
CN114548179A