Ultrasound-metabolic synergistic system for diagnosing SLN metastasis of breast cancer

Abstract

An ultrasound-metabolic synergistic system for diagnosing SLN metastasis of breast cancer. A deep learning model based on grayscale ultrasound is constructed to predict SLN metastasis of breast cancer, and 9 grayscale deep features and 15 grayscale manual features associated with SLN metastasis are obtained. A matrix material based on two-dimensional material MXene / multi-walled carbon nanotubes improves the detection of small metabolic molecules in blood, and metabolomics analysis results show that this has certain potential in detecting SLN metastasis of breast cancer, while revealing metabolic pathways closely associated with metastasis, such as the glucose-alanine cycle and glycolysis, suggesting that alterations in relevant molecular pathways may underlie heterogeneous features in ultrasound radiomics. The combination of ultrasound omics and metabolomics further improves the diagnostic performance for SLN metastasis of breast cancer. Therefore, organic integration of the two serves the diagnosis of SLN metastasis of breast cancer, providing a new strategy and theoretical basis for the accurate diagnosis of SLN metastasis of breast cancer, and possesses practical clinical application values.

Description

System for ultrasound-metabolic collaborative diagnosis of breast cancer SLN metastasis TECHNICAL FIELD

[0001] The present application belongs to the field of medical devices, and particularly relates to a system for ultrasound-metabolic collaborative diagnosis of breast cancer SLN (sentinel lymph node) metastasis, in particular a system for ultrasound-metabolic collaborative diagnosis of breast cancer SLN metastasis based on deep learning. BACKGROUND

[0002] Breast cancer is a common malignant tumor in women, and surgical operation is its main treatment method. Whether to perform axillary lymph node dissection is one of the important clinical decisions. Axillary lymph node dissection significantly increases complications such as limb swelling and movement disorders of patients, therefore, it is of great significance to accurately assess the axillary lymph node metastasis of patients. Sentinel lymph node (SLN) biopsy is currently a routine method for predicting axillary lymph node metastasis of breast cancer in clinical practice. SLN negative exemption axillary lymph node dissection has become a consensus, and 1-2 SLN positive patients who plan to undergo breast conserving surgery and additional postoperative radiotherapy can also not undergo axillary lymph node dissection. However, SLN biopsy has a certain false negative rate (7.8-27.3%), and also causes certain complications such as upper limb movement limitation and sensory disturbance. Therefore, it is urgent to develop a new strategy for accurate diagnosis of SLN metastasis with excellent performance and clinical practicability, and high efficiency, accuracy and non-invasiveness are the keys.

[0003] Therefore, there is a need in the industry for a non-invasive breast cancer SLN metastasis diagnosis method and system. TECHNICAL PROBLEM

[0004] The prior art in the industry needs a non-invasive breast cancer SLN metastasis diagnosis method and system. SUMMARY

[0005] The purpose of the present application is to provide a system for ultrasound-metabolic collaborative diagnosis of breast cancer SLN metastasis to solve the problem of the need for a non-invasive breast cancer SLN metastasis diagnosis method and system in the prior art in the industry.

[0006] The technical scheme of the present application is: a system for ultrasound-metabolic collaborative diagnosis of breast cancer SLN metastasis, comprising:

[0007] A gray-scale ultrasound omics feature extraction device is used to extract omics feature information of gray-scale ultrasound including shape features, intensity features and texture features after processing the breast cancer lesion ultrasound image;

[0008] The Doppler ultrasound feature extraction device is used to extract Doppler ultrasound feature information including blood flow full ratio, peripheral ratio and intramodular ratio after breast cancer lesion image processing.

[0009] The metabolomics target molecule obtaining device is used to screen specific mass-to-charge ratio (m / z) target molecules and find metabolic pathways including glucose-alanine cycle and glycolysis closely related to SLN metastasis.

[0010] The evaluation device further comprises:

[0011] The heterogeneity feature obtaining sub-device is used to analyze the gray-scale ultrasound feature information and the Doppler ultrasound feature information, and select heterogeneity features expressing metastatic and non-metastatic SLN.

[0012] The relationship establishing sub-device is used to establish a correlation between the heterogeneity features and metabolic pathways including glucose-alanine cycle and glycolysis.

[0013] The diagnosis sub-device is used to extract the obtained target tissue based on the correlation relationship, including gray-scale ultrasound feature information, Doppler ultrasound feature information and m / z target molecules, and further predict breast cancer SLN metastasis using the correlation relationship.

[0014] Preferably, the Doppler ultrasound feature extraction device and the gray-scale ultrasound feature extraction device further comprise:

[0015] The ultrasonic diagnostic instrument comprises a probe, and has a Doppler ultrasound function, which is used to scan the target lesion by the probe to obtain data including lesion blood flow, peripheral ratio and intramodular ratio, adjust the sampling frame size to contain the tumor and its surrounding 1-2 cm range, adjust the color gain to detect small blood vessels to obtain the most vascular cross-sectional data of the lesion, and obtain the related ultrasonic features of the lymph node.

[0016] Preferably, the gray-scale ultrasound feature extraction device further comprises:

[0017] The controller further comprises:

[0018] The image cutting and preprocessing module is used to perform image cutting and preprocessing operations on the obtained ultrasonic imaging data.

[0019] The feature extraction and dimension reduction processing module is configured to extract manual features in the ultrasonic imaging data based on a pyradiomics package in a python environment, including morphological features, intensity features, texture features, and wavelet transformed features, and perform dimension reduction on the extracted features to obtain a feature subset with lower dimension and higher resolution as the gray-scale ultrasonic feature information.

[0020] Preferably, the heterogeneity feature obtaining sub-device further comprises:

[0021] The classification prediction unit is configured to train the SLN metastasis prediction model based on the breast cancer lesion ultrasonic image, learn the classification and discrimination result by taking the N0, N+1-2 and N+≥3 as labels, and the screened ultrasonic feature as input, and predict the heterogeneity features of the metastatic and non-metastatic SLN by using the trained SLN metastasis prediction model.

[0022] The relationship establishing sub-device further comprises a metastasis prediction model, which is configured to learn the relationship discrimination result by taking the heterogeneity features of the metastatic and non-metastatic SLN, the glucose-alanine cycle, and the glycolysis-related molecules as input, and predict the relationship between the heterogeneity features and the glucose-alanine cycle and glycolysis-related molecules by using the trained metastasis prediction model.

[0023] Preferably, the gray-scale ultrasonic feature extraction device further comprises extracting breast tumor morphological features, which are further configured as:

[0024] Extracting internal texture features: using a local information enhancement algorithm to extract fine-grained features including texture, calcification, and uniformity, and the probability value is:

[0025] ,

[0026] is a parameter matrix, performing probability distribution normalization processing, optimizing the matrix and the convolution kernel parameters, and minimizing the loss function to obtain the optimal solution of local feature recognition, and the loss function is defined as follows:

[0027] ,

[0028] ,

[0029] is the number of input images, is an indicative function.

[0030] Extracting acoustic features: using gray scale contrast, histogram equalization and other graphics processing methods to extract the echo difference characteristics of the image inside and outside the lesion;

[0031] Extracting tumor size: according to the length and width data of the segmented tumor.

[0032] Preferably, the extraction of the morphological features of the breast tumor further comprises: for the features that cannot be separated alone, joint feature extraction based on text prompts is performed, multiple features are jointly predicted, and for the image-text data set , wherein , is an image-text data set, is the data volume, and a training loss function is used for feature corresponding training:

[0033] .

[0034] The metabolomics target molecule obtaining device further comprises: using data analysis software to process the serum metabolomics spectrum peaks, finding significant difference characteristic peaks, finding significant metabolic molecules through the difference of metabolomics, drawing corresponding heat maps and volcano plots, finding up-regulated and down-regulated factors, and then comparing with a metabolic database to obtain abnormal metabolic pathways of breast cancer SLN metastasis.

[0035] Preferably, the significant metabolic molecules closely related to SLN metastasis include glucose-alanine cycle, glycolysis metabolic pathways and their network relationships.

[0036] Preferably, the system is based on an optimized MXene / MWCNTs matrix material, the MXene / MWCNTs matrix material is prepared by etching Al from Ti2CTi3 to prepare a two-dimensional MXene matrix material by using hydrofluoric acid etching method, including Adv. Funct. Mater. 2021, 31; ACS Nano 2019, 13, 3042; Nat. Commun. 2020, 11, 1, the synthesized MXene material is characterized by scanning electron microscope and energy spectrum analysis; the interface fusion of MWCNTs and MXene is performed by thermal fusion method, and the particle size, ionization performance, thermal conductivity index and hydrophilic and hydrophobic characterization experiments of the MXene / MWCNTs matrix material are performed. BRIEF DESCRIPTION OF DRAWINGS

[0037] Fig. 1 is a schematic diagram of an ultrasound-metabolic cooperative system for diagnosing breast cancer SLN metastasis;

[0038] Fig. 2 is a mechanism block diagram of the system for diagnosing breast cancer SLN metastasis in cooperation with ultrasound-metabolism;

[0039] Fig. 3 is the technical principle of the ultrasound platform establishment and the diagnosis of breast cancer SLN metastasis by ultrasound metabonomics;

[0040] Fig. 4 is the technical principle of the metabonomics platform establishment and the diagnosis of breast cancer SLN metastasis by metabonomics;

[0041] Fig. 5 is a comparison of the diagnostic performance of different models for predicting SLN metastasis (N0 vs. N+), a is the ROC diagnostic performance chart, and b is the confusion matrix chart predicted by the EfficientNet-B7 model;

[0042] Fig. 6 is the Grad-CAM heat map of the EfficientNet-B7 deep learning model, which prompts the artificial intelligence model to predict the possibility of breast cancer SLN metastasis. The gray-scale ultrasound images of breast lumps without SLN metastasis (A1-10) and their feature heat maps (B1-10) are shown. The gray-scale ultrasound images of breast lumps with SLN metastasis (C1-10) and their feature heat maps (D1-10) are shown.

[0043] Fig. 7 is the metabolic fingerprint analysis and typing of breast cancer samples with / without SLN metastasis. a) Metabolic fingerprint spectrum of breast cancer without SLN metastasis b) Metabolic fingerprint spectrum of breast cancer with SLN metastasis c) PCA and d) OPLS-DA for typing analysis of breast cancer with / without SLN metastasis;

[0044] Fig. 8 a) is a heat map; b) is a volcano plot and c) is an AUC analysis for typing analysis of breast cancer with / without SLN metastasis;

[0045] Fig. 9 a) is a network diagram of enriched pathways b) metabolic pathways. DETAILED DESCRIPTION

[0046] I. Brief introduction of the entire invention process of the inventors of our company.

[0047] 1. Ultrasound metabonomics helps diagnose breast cancer SLN metastasis

[0048] Ultrasound is widely used in breast cancer screening and preoperative evaluation due to its convenience and non-radiation. It is also the first choice for axillary lymph node evaluation. However, the diagnostic performance of conventional ultrasound for axillary lymph node metastasis in breast cancer is low, with an area under the curve (AUC) value of 0.585-0.719, which cannot meet the clinical diagnostic needs. Ultrasonicomics can convert ultrasound images into high-dimensional, available quantitative image features through high-throughput data feature extraction algorithms, and use various algorithms for deep analysis of the features, which has great potential in the diagnosis of benign and malignant tumors. With the rapid development of artificial intelligence convolutional neural networks (CNN) and precise quantitative medical imaging technology, image recognition technology and data algorithms are constantly updated, and the mining and analysis of medical image big data are realized, which greatly expands the analysis efficiency of medical images in ultrasonicomics. Previous studies have shown that ultrasonicomics provides an effective imaging technique for the diagnosis of lymph node metastasis in breast cancer. However, ultrasonicomics is heavily dependent on feature extraction algorithms and feature analysis methods. The design of deep networks, the selection of ultrasonicomics features (such as gray scale, blood vessel ratio, and blood vessel number), the feature dimension reduction method, and the selection of feature classifiers are extremely important for diagnostic performance. Studies have shown that relying solely on ultrasound image features is insufficient to achieve accurate diagnosis of breast cancer SLN metastasis, especially for the efficient differentiation of N0 (no SLN metastasis), N

[0049] N +1-2 (1 to 2 SLN metastasis), and N +≥3 (3 or more SLN metastasis), which often requires the assistance of other diagnostic techniques or indicators, such as tumor grading, histological tumor size, lymphatic vessel invasion, Ki-67 or hormone receptor status, but these indicators often require invasive diagnostic procedures. More importantly, deep learning features have the characteristics of a 'black box', with no accurate and complete formulas and definitions, and lack of biological interpretability. Even through the construction of heat maps to perform reverse reasoning on the decision-making of deep learning models, it cannot fully meet the requirements of biological interpretability. Therefore, it is urgent to combine other non-invasive detection methods to improve the diagnostic performance of ultrasonicomics for breast cancer SLN metastasis and improve its clinical interpretability.

[0050] 2. Metabolomics assisted diagnosis of breast cancer SLN metastasis

[0051] In recent years, by analyzing the metabolite profile of breast cancer patients through metabolomics technology, valuable potential markers are screened out, which opens up a new way for the individualized and precise diagnosis and treatment of breast cancer. The occurrence and progression of tumors are closely related to the changes of metabolites. During the malignant transformation of breast cancer, the metabolism of tumor cells undergoes a series of profound reorganization, which plays a key role in the progression of the disease and shows dependence on specific metabolic pathways. As a new discipline in the 'post-genomic' era, metabolomics analyzes the dynamic changes of all endogenous small molecule metabolites in biological systems under physiological or pathological conditions, thereby exploring the relationship between endogenous small molecule substances and the occurrence and development of diseases. At the same time, metabolomics has the technical characteristics of high sensitivity, high resolution and high throughput, as well as the advantages of non-invasive / micro-invasive and easy implementation, and is widely used in tumor marker screening, disease diagnosis, new drug research and development and other fields, showing great development potential. Therefore, the discovery of new specific biomarkers through metabolomics technology can provide a new idea for the precise diagnosis of breast cancer SLN metastasis, and metabolomics diagnosis data mining and identification are the key.

[0052] Among various metabolic molecular detection methods, high-throughput nanoparticle-enhanced laser desorption / ionization mass spectrometry (NPELDI-MS) technology can quickly provide differential information of metabolite levels, and with the help of artificial intelligence, it can analyze the serum metabolic fingerprint of different populations, which is a potential new diagnostic tool in clinical practice. In the field of breast cancer diagnosis and treatment of metabolomics, high-throughput NPELDI-MS performs well in the early diagnosis of breast cancer (AUC is 0.948, 95% CI: 0.922-0.973, accuracy is 88.8%, sensitivity is 88.9%, and specificity is 88.8%), showing a good clinical application prospect. In the field of breast cancer metabolic fingerprint mapping, it is necessary to rely on high specificity and sensitivity. Tandem mass spectrometry can detect 14 methionine-related metabolites such as methionine and S-adenosyl methionine in breast cancer cells at the same time, and further metabolic spectrum analysis is helpful for the study of the pathogenesis of triple-negative breast cancer. In the field of breast cancer treatment target metabolome screening, the applicant analyzed the metabolome and lipidome of 330 triple-negative breast cancer samples and 149 healthy human breast tissues, constructed a complete triple-negative breast cancer metabolome atlas, and found that (n-acetyl aspartic acid is an important pro-tumor metabolite and a potential therapeutic target for high-risk basal cell-like immunosuppressive tumors. In summary, metabolomics has made certain breakthroughs in the diagnosis of breast cancer, metabolic fingerprint spectrum, and therapeutic targets, and has been widely used in the screening of breast cancer, the establishment of metabolic fingerprint library, and the development of individualized treatment plan. However, there is no report on the use of metabolomics technology for the diagnosis of breast cancer lymph node metastasis.

[0053] 3. Ultrasound-metabolic multi-omics combined to improve the diagnostic performance of SLN metastasis in breast cancer

[0054] In recent years, artificial intelligence-based multi-omics analysis methods for predicting breast cancer lymph node metastasis have received extensive attention. With the deepening development of tumor research, the applicant has increasingly realized that tumor metastasis is a complex multi-link process, which depends on itself and is also dependent on the internal and external environment. Similarly, whether breast cancer can occur lymph node metastasis depends on the accumulation of tumor cell mutations and the acquired metastatic advantage, but ultimately whether it can successfully metastasize is determined by multiple factors. Therefore, the combination of multi-dimensional tumor metastasis-related information brings new ideas for the accurate diagnosis of SLN. As we all know, the change of metabolism is in the middle of the biological information flow, between genes, proteins and cells, tissues, and plays a role in the transmission of biological information. Therefore, metabolomics can systematically and comprehensively reflect the abnormalities of breast cancer lymph node metastasis at the biological molecular level and metabolic pathways, especially showing the overall information changes of breast cancer patients, which can make up for the lack of spatial dimension of ultrasound omics. In addition, there may be some potential spatiotemporal relationship between imageomics and metabolomics, for example, some imageomics features with high predictive ability may be related to small molecule metabolites of tumor cells in a certain physiological period, therefore, by exploring the relationship between small molecule metabolites and breast ultrasound images, further biological explanation of ultrasound omics can be provided, more valuable information for clinical treatment can be provided, the development of individualized diagnosis and treatment of ultrasound imaging technology can be promoted, and the popularization of precision medicine can be promoted.

[0055] II. Brief description of the applicant's invention idea

[0056] The applicant constructed a deep learning model based on gray-scale ultrasound to predict breast cancer SLN metastasis, and the preliminary research results showed that the AUC value of the model in predicting SLN metastasis was 0.819, indicating the value of ultrasound omics in predicting breast cancer SLN metastasis, and 9 gray-scale deep features and 15 gray-scale manual features related to SLN metastasis were obtained. The applicant further applied a matrix material based on two-dimensional material MXene / multi-walled carbon nanotubes (MWCNTs) to improve the detection of small molecules in blood metabolism, and the metabolomics analysis results showed that it had certain potential in detecting breast cancer SLN metastasis (AUC value was 0.808), and excavated the glucose-alanine cycle, glycolysis and other metabolic pathways closely related to metastasis, with 6 kinds of metabolic small molecules being extremely strongly correlated with 16 of the above 24 ultrasound omics features and 8 being strongly correlated, showing that the change of related molecular pathways may be the basis of the heterogeneous characteristics of ultrasound omics, and ultrasound omics combined with metabolomics can further improve the diagnostic efficiency of breast cancer SLN metastasis. Therefore, the two are organically integrated to serve breast cancer SLN metastasis diagnosis together, and the internal relationship between ultrasound-metabolomics is preliminarily explored, which will provide new strategies and theoretical basis for the accurate diagnosis of breast cancer SLN metastasis, and has practical clinical application value.

[0057] III. Ultrasound-metabolic collaborative diagnosis system for breast cancer SLN metastasis

[0058] Referring to FIG. 1, it is a schematic diagram of the ultrasound-metabolic collaborative diagnosis system for breast cancer SLN metastasis. It includes:

[0059] The gray-scale ultrasound feature extraction device 11 is used to extract the gray-scale ultrasound feature information including shape features, intensity features and texture features after processing the breast cancer lesion ultrasound image;

[0060] The Doppler ultrasound feature extraction device 12 is used to extract the Doppler ultrasound feature information including blood flow total ratio, peripheral ratio and intramodular ratio after processing the breast cancer lesion image;

[0061] The metabolic target molecule acquisition device 13 is used to screen out specific mass-to-charge ratio m / z target molecules and find metabolic pathways closely related to SLN metastasis including glucose-alanine cycle and glycolysis;

[0062] The evaluation device 14 further includes:

[0063] The heterogeneity feature acquisition sub-device 141 is used to analyze the gray-scale ultrasound feature information and the Doppler ultrasound feature information and select heterogeneity features expressed in metastatic and non-metastatic SLN;

[0064] The relationship establishment sub-device 142 is used to establish a correlation between the heterogeneity features and metabolic pathways including glucose-alanine cycle and glycolysis;

[0065] The diagnosis sub-device 15 is used to extract the gray-scale ultrasound feature information, the Doppler ultrasound feature information and the m / z target molecules based on the correlation between the target tissues, and further predict the breast cancer SLN metastasis using the correlation.

[0066] I. First, introduce the ultrasound platform and the technical path of ultrasound-based diagnosis

[0067] Referring to FIG. 2, the system collects clinical information, blood samples and ultrasound image data of SLN metastasis and non-metastasis breast cancer patients, extracts heterogeneous image features in gray-scale ultrasound and Doppler ultrasound and significant differential biochemical markers in metabolomics, organically integrates the above features and markers, and establishes a prediction model based on deep learning, realizes efficient prediction of the presence or absence of SLN metastasis, and further realizes accurate typing of No, N+1-2 and N+≥3; and through the evaluation of correlation and significance and other elements, the metabolomics and ultrasound omics features are analyzed jointly, the internal relationship between the two is systematically expounded, and the molecular biology basis of the heterogeneous image features of ultrasound omics is explored.

[0068] The Doppler ultrasound feature extraction device and the gray-scale ultrasound feature extraction device further comprise: an ultrasound diagnostic instrument comprising a probe, the ultrasound diagnostic instrument having a Doppler ultrasound function to scan the target lesion by the probe to obtain gray-scale ultrasound, to enable the Doppler ultrasound to obtain data including lesion blood flow, peripheral ratio and intramodular ratio, to adjust the sampling frame size to contain the tumor and its surrounding 1-2 cm range, to adjust the color gain to detect small blood vessels to obtain the most vascular cross-section related data of the lesion, and to obtain the ultrasound features related to the lymph node, and to save the related ultrasound feature images in the ultrasound diagnostic instrument. For example, the ultrasound diagnostic instrument can use Philips Epiq 7 ultrasound diagnostic instrument, 6-12 MHz linear probe, and optimize the imaging parameters such as depth and gain according to the image condition of each patient. The patient lies on the examination bed, and the breast, axillary fossa and supraclavicular fossa are fully exposed. Radial scanning is performed with the nipple as the center to determine the position of the lesion. After the gray-scale ultrasound examination is completed, the Doppler ultrasound is enabled to observe the blood flow of the lesion, the sampling frame size is adjusted to contain the tumor and its surrounding 1-2 cm range, the color gain is adjusted to just detect small blood vessels while suppressing false color, and the most vascular cross-section of the lesion is found. After full breast ultrasound examination, axillary and supraclavicular ultrasound examination is performed and the ultrasound features of the lymph nodes are recorded. The above typical images are stored in the ultrasound diagnostic instrument.

[0069] The Doppler ultrasound feature extraction device and the gray-scale ultrasound feature extraction device are mainly divided in function. In fact, after the above-mentioned ultrasound diagnostic instrument comprising a probe collects ultrasound data, the subsequent data will be sent to a controller with a storage unit to realize the above-mentioned functions. The controller can also have a separate storage unit to store related data.

[0070] The gray-scale ultrasound feature extraction device further comprises:

[0071] The controller further comprises:

[0072] Image cutting and preprocessing module: for image cutting and preprocessing operation on the acquired ultrasonic imaging data. It further includes: proportionally randomly dividing the enrolled patients into training group, validation data group and test data group. The study uses MATLAB software to develop a manual cropping tool, which enables the sonographer to identify the pixel points marking the lesion boundary and collect the coordinates of these points, and through calculation to obtain the minimum circumscribed rectangle enclosing frame of the lesion. The image in the enclosing frame is called the region of interest. To expand the sample size, the study uses geometric method and singular value decomposition method for data augmentation. The acquired ultrasonic imaging data is standardized, normalized and processed to reduce the interference of factors such as high speckle image, fuzzy boundary, low signal-to-noise ratio, low contrast and uneven intensity.

[0073] Feature extraction and dimension reduction processing module: for extracting manual features in ultrasonic imaging data based on pyradiomics package in python environment, including morphological features, intensity features, texture features, and wavelet transformed features. The extracted omics features are dimensionally reduced to obtain a lower-dimensional, higher-resolution feature subset as the omics feature information of gray-scale ultrasound. Specifically, the manual features of gray-scale ultrasound images are extracted based on pyradiomics package in python environment, including morphological features, intensity features, texture features, and wavelet transformed features totaling 2060. In addition, an attention mechanism based on EfficientNetB7 is constructed by pytorch as a deep feature extractor. The network takes EfficientNetB7 as the framework, and performs spatial-channel attention module on the output of the last convolutional layer to strengthen the network feature extraction, followed by global average pooling operation to reduce the risk of overfitting and add a fully connected layer with 256 neurons. Finally, 256 deep learning features are extracted from the fully connected layer. The trained deep network can generate gradient weighted class activation mapping (Grad-CAM) heat map, thereby realizing the visual positioning of the lesion area. The place with larger activation value (i.e. redder color) indicates that the area is more likely to be related to SLN metastasis. In summary, a total of 2316 gray-scale ultrasound omics features are extracted for subsequent analysis.

[0074] The heterogeneity feature obtaining sub-device further includes:

[0075] The classification prediction unit: the SLN metastasis prediction model based on breast cancer lesion ultrasound images is trained, taking N0, N+1-2 and N+≥3 as labels, the screened ultrasound group features as input, and the classification discrimination result as output for learning, and the heterogeneity features of the metastatic and non-metastatic SLN are classified and predicted by the trained SLN metastasis prediction model. The support vector machine (SVM) is selected for classification, which has the advantages of fast convergence speed and high classification accuracy. The SLN metastasis prediction model based on breast cancer lesion ultrasound images is established after ten-fold cross-validation. The performance of the model is measured by AUC, accuracy, sensitivity and specificity.

[0076] One specific implementation example is that the ultrasound group model extracts breast tumor location information and breast tumor morphological features based on the obtained breast gray-scale ultrasound and color Doppler ultrasound images of breast cancer patients, which specifically includes:

[0077] For the ultrasound image, the cutting tool is used to cut off the irrelevant text information in the image, and the obtained ultrasound image data is standardized and normalized to reduce the interference factors of high gray-scale image spots, blurred boundaries, low signal-to-noise ratio, low contrast and uneven intensity;

[0078] The breast tumor location information is extracted, and the foreground feature is first enhanced by the foreground feature optimization, attention mechanism algorithm and context semantic association module:

[0079] ,

[0080] is the i-th feature map, which has an output stride of pixels relative to the input image, and the feature map set is obtained by transverse connection of the top and bottom paths, Γ is the transverse connection realized by the learnable convolution layer, and Ι represents the nearest neighbor up-sampling with a scale factor of 2;

[0081] ,

[0082] ,

[0083] is the feature map obtained by converting through the scale perception projection function , is a learnable parameter of ;

[0084] The scene embedding vector is expressed as follows:

[0085] ,

[0086] is the projection function, the output space is , the variable is the point-wise inner product similarity estimate:

[0087] ,

[0088] ,

[0089] The re-encoder has learnable parameters , The term included is used to weight the re-encoded feature map, and finally the foreground feature map enhanced in foreground-background relationship is obtained :

[0090] .

[0091] The technical scheme of the embodiment combines the class imbalance problem of the tumor region (foreground) and the whole (background) region of the breast tumor ultrasound, designs a foreground feature optimization, attention mechanism algorithm and context semantic correlation module to enhance the target of the foreground feature, and can more accurately and effectively identify the tumor region.

[0092] Specifically, for the collection of ultrasound images: a 10-15MHz linear probe is used, and the imaging parameters such as depth and gain are optimized according to the image condition of each patient; the patient is supine on the examination bed, and the breast, axillary fossa and supraclavicular fossa are fully exposed; radial scanning is performed with the nipple as the center to determine the position of the lesion; after the patient receives full breast ultrasound examination, the same ultrasound physician performs axillary ultrasound examination, and records the suspicious ultrasound features of the axillary lymph nodes; after the ultrasound physician finishes the gray-scale ultrasound examination, the Doppler ultrasound is started to observe the blood flow of the lesion, the sampling frame size is adjusted to include the tumor and its surrounding 1-2cm range, the color gain is adjusted to small blood vessels just being detected, and false color is suppressed, and the section with the most blood vessels of the lesion is searched; the whole scanning process is stored in the ultrasound diagnostic instrument. For ultrasound image cutting and preprocessing: the cutting tool is used to cut off irrelevant text information in the picture; in order to expand the sample size, geometric method and singular value decomposition method are used for data augmentation.

[0093] Further, breast tumor position information extraction is performed, further using the homogeneous positioning method, realizing standardized registration and automatic quantitative statistics of tumor position information of the whole population, using the differential homeomorphism method to realize flexible registration under a unified template, and introducing the Lagrange equation for optimization:

[0094] The picture to be registered is provided and target registration image , the registration calculation formula is:

[0095] ,

[0096] Considering that ultrasound tumor features have low contrast compared to the entire image, we introduced the window sliding multi-head self-attention module (W-MSA) and the shifted multi-head self-attention module (SW-MSA) into the traditional natural language model architecture to achieve efficient detection of location information:

[0097] ,

[0098] and Respectively expressed in The output features of the multi-head self-attention module W-MSA or the shifted multi-head self-attention module SW-MSA in the modules and the multi-layer perception layer module.

[0099] Ultrasound features of breast tumors vary in scale and resolution, creating challenges for large-scale statistical analysis of location information. This embodiment designs a homogeneous localization method to achieve standardized registration and automated quantitative statistics of tumor location information across the entire population.

[0100] Furthermore, the extraction of breast tumor morphological features includes:

[0101] Extract internal texture features: Use local information enhancement algorithm to extract fine-grained features including texture, calcification, and uniformity, and probability value for:

[0102] ,

[0103] is the parameter matrix, Normalize the probability distribution and optimize Matrix and convolution kernel parameters make the loss function Minimize, thus obtaining the optimal solution for local feature recognition. The loss function is defined as follows:

[0104] ,

[0105] ,

[0106] is the number of input images, is an indicative function;

[0107] Extracting acoustic features: using gray scale contrast, histogram equalization and other graphics processing methods to extract the echo difference characteristics of the image inside and outside the lesion;

[0108] Extracting tumor size: according to the length and width data of the segmented tumor.

[0109] Combined with literature knowledge and Chinese breast imaging reporting and data system (BI-RADS), the breast tumor features are effectively extracted, including: tumor internal information (cystic and solid, calcification, echo, uniformity), edge information (edge sharpness, capsule integrity, shadow degree), tumor size (aspect ratio, diameter) and the like. This embodiment gives a specific means for extracting breast tumor morphological features.

[0110] Preferably, the extraction of breast tumor morphological features also includes: joint feature extraction based on text prompts for features that cannot be separated alone, realizing joint prediction of multiple features, for image-text data set , wherein , is an image-text data set, is the data volume, and the training loss function is trained for feature correspondence:

[0111] .

[0112] Further, the extraction of breast tumor morphological features also includes: joint feature extraction based on text prompts for features that cannot be separated alone, realizing joint prediction of multiple features, for image-text data set , wherein , is an image-text data set, is the data volume, and the training loss function is trained for feature correspondence:

[0113] .

[0114] This embodiment is adapted to the application scenario, and the joint feature extraction based on text prompts for features that cannot be separated alone is provided, thereby providing a basis for realizing joint prediction of multiple features.

[0115] Preferably, the ultrasound omics model further comprises a fusion module for globally interacting and fusing breast tumor location information and breast tumor morphological feature information, and the input breast tumor location information and breast tumor morphological feature information are combined into a feature matrix , divided into a fixed-size block sequence , mapped to an embedding vector based on linear projection , and a classification label is introduced , combined into the standard input form of Transformer, position embedding information, which transforms image patches into ordered one-dimensional embeddings and introduces self-attention mechanism:

[0116] ,

[0117] ,

[0118] , , denote query, key and value matrices respectively, denote input sequence vector length, multi-head attention mechanism performs self-attention calculation on input sequence, and multiple outputs are spliced to obtain projection, and the final output vector is calculated;

[0119] Multi-head attention calculation process is as follows:

[0120] ,

[0121] wherein:

[0122] ,

[0123] ,

[0124] Finally, the feature output of Transformer encoder is converted into classification vector corresponding to conversion site:

[0125] ,

[0126] When the response vector is whether lymph node metastasis occurs, the functional variable is position information, and the bounded closed interval is a square integrable function, dimensional scalar variable is morphological feature, is a dimensional sub-vector of , let be the conditional mean of , and the generalized linear model is described by the following formula:

[0127] ,

[0128] and​ For known functions, the slope function in the position information Depends on The value, which allows the position information There is an interactive effect with the morphological characteristics , which is expressed as follows:

[0129] ,

[0130] where is an unknown second derivative continuous function, is a single index weight vector, and are unknown coefficient functions, respectively representing the main effect of the position information and the interactive effect of the two kinds of information; the morphological characteristics are independently extracted by the morphological extraction module; the position information main effect is independently extracted by the position extraction module; and the interactive effect information of the two is independently extracted by the fusion module.

[0131] II. Metabolomics platform establishment and breast cancer SLN metastasis metabolomics diagnosis

[0132] 2.1. The metabolomics platform obtains patient blood sample data in advance. For example, the patient uses BD 6 mL coagulation tube to collect 1 tube of fasting venous blood. After blood sampling, mix gently and avoid hemolysis. The collected blood is placed at room temperature for about 30 minutes, and the blood sample separation is completed within 4 hours: centrifuge at 3000-3500 rpm for 10-15 minutes in a horizontal centrifuge. Use a pipette to aspirate the upper serum and divide it into pre-labeled nuclease-free cryotubes, 1000 uL per tube, and store them in a -80°C freezer to avoid repeated freezing and thawing.

[0133] 2.2. The metabolomics platform obtains 5 MXene / MWCNTs matrix materials

[0134] A1 is prepared by etching MXene matrix material from Ti2CTi3 with hydrofluoric acid (Adv. Funct. Mater. 2021, 31; ACS Nano 2019, 13, 3042; Nat. Commun. 2020, 11, 1), and the synthesized MXene material is characterized by scanning electron microscopy and energy spectrum analysis. The interface fusion method is used to fuse MWCNTs and MXene, and the particle size, ionization performance, thermal conductivity index and hydrophilic and hydrophobic properties of MXene / MWCNTs matrix material are characterized.

[0135] 2.3. Metabolomics testing by the metabolomics platform

[0136] For example, MXene / MWCNTs were used as matrix materials to analyze serum samples from 100 cases of No, 90 cases of N+1-2, and 90 cases of N+³3, a total of 280 cases of breast cancer patients. 0.5 μL of serum sample was placed on a polished steel target plate MTP384, and after air drying, 1 μL of matrix material was added for ionization. The autoflex max (Bruker Daltonics, Bremen, Germany) time-of-flight tandem mass spectrometer was used for sample detection, and the smartbeam-II laser was used to collect mass spectrum spectra of 100-5000 Da in reflection mode at 355 nm with a laser frequency of 1,000 Hz. Each sample was randomly measured 25 times at 20 different grating points.

[0137] 2.4. Metabolic fingerprint analysis of breast cancer SLN metastasis

[0138] The serum metabolomics peaks were studied using data analysis software (such as SIMCA-P) to find significant difference characteristic peaks. By their metabolic differences, significant metabolites were found, and corresponding heat maps and volcano plots (J. Med. Primato1.2022, 1111, 12610) were drawn to find up-regulated and down-regulated factors, and then compared with the metabolic database to find abnormal metabolic pathways of breast cancer SLN metastasis, and to achieve effective identification and diagnosis.

[0139] III. Relationship establishment sub-device

[0140] Relationship establishment sub-device 142: for establishing a correlation between the heterogeneity feature and the metabolic pathway including glucose-alanine cycle, glycolysis.

[0141] In this example, a metastasis prediction model is established, which uses the heterogeneity features of metastatic and non-metastatic SLNs, glucose-alanine cycle, and glycolysis-related molecules as input, and the relationship discrimination result as output for learning. The relationship between the heterogeneity features and the glucose-alanine cycle and glycolysis-related molecules is predicted by the trained metastasis prediction model. The core is to obtain the correlation between metabolic molecules / pathways and ultrasound image features, and to evaluate it by R 2 and p value.

[0142] Application example

[0143] Applicants constructed a deep learning model based on grayscale ultrasound images to predict breast cancer SLN metastasis. We collected 158 ultrasound images of breast lesions (50 cases for training set, 36 cases for validation data set and 72 cases for test data set), of which 79 cases had SLN metastasis and 79 cases had no metastasis. Five mainstream CNNs (Inception-v2, ResNet-50, DensNet-201, ResNeXt-101 and EfficientNet-B7) were adopted as extractors of deep feature extraction modules to evaluate the influence of different CNNs on the prediction performance of the model. Comprehensive analysis showed that EfficientNet-B7 was the model with the best generalization performance and the highest diagnostic efficiency among all networks (Table 1). Experimental results confirmed that the AUC, accuracy, sensitivity and specificity of the model in the validation data set and test data set reached 0.819 and 0.726, 77.78% and 72.22%, 83.33% and 75.00%, 72.22% and 69.44% respectively (Figure 5, indicating that the model has high diagnostic efficiency for breast cancer SLN metastasis.

[0144]

[0145]

[0146] To further improve the prediction performance and visibility of SLN metastasis, applicants further extracted high-dimensional grayscale ultrasound sonographic features, including grayscale manual features and grayscale deep features, in a feature fusion manner (Table 2). Through feature screening and dimension reduction, 9 grayscale deep features and 15 grayscale manual features were finally obtained (Table 3). At the same time, the visualization positioning of the lesion area was realized through Grad-CAM heat map, which provided corresponding visual explanation for the prediction results of EfficientNet-B7 network (Figure 6). Among them, the place with larger activation value (i.e. the color is redder) means that this area is more likely to belong to a valuable feature, i.e. the key attention area for predicting SLN metastasis, while the blue part indicates the key attention area for predicting SLN non-metastasis.

[0147]

[0148]

[0149]

[0150] The applicants further analyzed the metabolic fingerprint of 30 cases of SLN non-metastasis and 33 cases of SLN metastasis, and the typical fingerprint of SLN non-metastasis and SLN metastasis is shown in 7a-b. Further, the applicants used principal component analysis (PCA) and orthogonal partial least squares discriminant analysis (OPLS-DA) to analyze the above samples, and the two types of samples were partially overlapped in the PCA method, and the OPLS-DA discriminant analysis statistical method could clearly distinguish the samples with or without SLN metastasis (Fig. 7c-d), indicating that there may be one or more significantly different metabolites that can be used for the diagnosis of breast cancer SLN metastasis.

[0151] Based on the better classification effect, the applicants analyzed the m / z that played an important role in the heat map and volcano plot, and the analysis results clearly showed the key difference molecules, especially the potential up-regulated and down-regulated molecules (8a-b). Further evaluation of the diagnostic efficiency of breast cancer SLN metastasis showed that the AUC value was 0.808 (Fig. 8c), indicating that this method has good potential in the application of diagnosing breast cancer SLN metastasis.

[0152] Metabolomics further revealed that the glucose-alanine cycle, glycolysis and other metabolic pathways and their network relationships closely related to metastasis had different expressions between metastasis and non-metastasis samples (Fig. 9a-b). The applicants further performed Pearson correlation analysis on the expression frequency of the above 24 ultrasonic features and the expression frequency of the different metabolites, and the results showed that 6 metabolic small molecules were extremely strongly correlated with 16 of the above 24 ultrasonic features, and 8 were strongly correlated (Table 4), suggesting that the change of tumor metabolic pathways affects its growth pattern and morphology, and the external performance is the ultrasonic features of the ultrasonic images. Ultrasonicomics combined with metabolomics can realize its specific diagnosis and prediction.

[0153]

[0154] The diagnostic sub-device is used for extracting the obtained target tissue based on the correlation relationship, including the omics feature information of gray group ultrasound, the Doppler ultrasound omics feature information and the m / z target molecule, and further predicting the breast cancer SLN metastasis by using the correlation relationship. A metastasis prediction model is established, which is used for learning the input of the related molecules of the heterogeneity characteristics of the metastasis and non-metastatic SLN, glucose-alanine cycle and glycolysis, and the output of the relationship discrimination result, and predicting the relationship of the heterogeneity characteristics and the related molecules of the glucose-alanine cycle and glycolysis by using the trained metastasis prediction model.

[0155] The embodiments of the present application are described in detail above with reference to the drawings, but the present application is not limited to the above-described embodiments. Even if various changes are made to the present application, if the changes belong to the scope of the claims of the present application and equivalent technologies thereof, they still fall within the protection scope of the present application.

Claims

1. A system for ultrasound-metabolic synergic diagnosis of metastasis of SLN in breast cancer, characterized by, Comprise: Gray-scale ultrasound omics feature extraction device, used to extract gray-scale ultrasound omics feature information including shape features, intensity features and texture features after breast cancer lesion ultrasound image processing; Doppler ultrasound omics feature extraction device, used to extract Doppler ultrasound omics feature information including blood flow total ratio, peripheral ratio and intramodular ratio after breast cancer lesion image processing; Metabolomics target molecule acquisition device, used to screen specific mass-to-charge ratio m / z target molecules and find metabolic pathways closely related to SLN metastasis including glucose-alanine cycle and glycolysis; Evaluation device: further comprising: Heterogeneity feature acquisition sub-device: for analyzing the gray-scale ultrasound omics feature information and the Doppler ultrasound omics feature information, and selecting heterogeneity features expressing metastatic and non-metastatic SLN; Relationship establishment sub-device: for establishing a correlation between the heterogeneity features and metabolic pathways including glucose-alanine cycle and glycolysis; Diagnosis sub-device: for extracting gray-scale ultrasound omics feature information, Doppler ultrasound omics feature information and m / z target molecules from the target tissue based on the correlation, and further predicting breast cancer SLN metastasis using the correlation.

2. The system of claim 1, wherein, The Doppler ultrasound omics feature extraction device and the gray-scale ultrasound omics feature extraction device further comprise: An ultrasound diagnostic instrument comprising a probe, the ultrasound diagnostic instrument having Doppler ultrasound function, used to scan the target lesion by the probe to obtain gray-scale ultrasound, enable Doppler ultrasound to obtain data including lesion blood flow, peripheral ratio and intramodular ratio, adjust the sampling frame size to include the tumor and its surrounding 1-2 cm range, adjust the color gain to detect small blood vessels to obtain the cross-sectional data with the most lesion blood vessels, and obtain the ultrasound features related to lymph nodes, and save the related ultrasound feature images to the ultrasound diagnostic instrument.

3. The system of claim 2, wherein, The gray-scale ultrasound omics feature extraction device further comprises: A controller, the controller further comprising: Image cutting and preprocessing module: for image cutting and preprocessing of the obtained ultrasound imaging data; Feature extraction and dimension reduction processing module: for extracting manual features in the ultrasound imaging data based on the pyradiomics package in the python environment, including morphological features, intensity features, texture features and wavelet transformed features, and performing dimension reduction operation on the extracted omics features to obtain a feature subset with lower dimension and higher resolution as the gray-scale ultrasound omics feature information.

4. The system of claim 1 or 3, wherein, The heterogeneity feature acquisition sub-device further comprises: Classification prediction unit: for training the SLN metastasis prediction model based on the breast cancer lesion ultrasound image, taking N0, N+1-2 and N+≥3 as labels, the screened ultrasound omics features as input, and the classification discrimination result as output for learning, and classifying and predicting the heterogeneity features expressing metastatic and non-metastatic SLN through the trained SLN metastasis prediction model.

5. The system of claim 4, wherein, The relationship establishing sub-device further comprises an establishment transfer prediction model, which is used for learning the relationship between the heterogeneity features of the transfer and non-transfer SLN, the glucose-alanine cycle and glycolysis as input and the relationship discrimination result as output, and predicting the relationship between the heterogeneity features and the glucose-alanine cycle and glycolysis through the trained establishment transfer prediction model.

6. The system of claim 1 or 3, wherein, The gray-scale ultrasound group feature extraction device further comprises extracting breast tumor morphological features, which are further formulated as: Extracting internal texture features: using local information enhancement algorithm to realize the extraction of fine granularity features including texture, calcification, uniformity, and probability value To be: ， for the parameter matrix, carrying out the probability distribution normalization process, by optimizing matrix and convolution kernel parameters, such that a loss function min, so as to obtain a local feature recognition optimal solution, and a loss function is defined as follows: ， ， For the input image number, To be an indicative function; Extracting echogenic features: using gray-scale contrast, histogram equalization and other graphics processing methods, extracting the echo difference features of the lesion inside and outside the image; Extracting tumor size: the tumor size is calculated according to the length and width data of the segmented tumor.

7. The system of claim 6, wherein, The extraction of morphological features of breast tumors also includes text prompt-based joint feature extraction for features that cannot be separated alone, realizing joint prediction of multiple features, and for a graph-text data set , , is a graph-text data set, is the data amount, and a training loss function is used for feature corresponding training: 。 8. The system of claim 1, wherein, The metabolomics target molecule obtaining device further comprises: using data analysis software to process the serum metabolomics spectrum peaks, find the significantly different characteristic peaks, find the significant metabolic molecules through the difference of metabolomics, draw the corresponding heat map and volcano plot, find the up-regulated and down-regulated factors, and then compare with the metabolic database to obtain the abnormal metabolic pathways of breast cancer SLN metastasis.

9. The system of claim 1, wherein, The significant metabolic molecules closely related to SLN metastasis include the glucose-alanine cycle, glycolysis metabolic pathways and their network relationships.

10. The system of claim 1, wherein, The system is based on an optimized MXene / MWCNTs matrix material, the MXene / MWCNTs matrix material is prepared by etching Al from Ti2CTi3 to prepare a two-dimensional MXene matrix material, including Adv. Funct. Mater. 2021, 31; ACS Nano 2019, 13, 3042; Nat. Commun. 2020, 11, 1, the synthesized MXene material is characterized by scanning electron microscope and energy spectrum analysis; the interface fusion of MWCNTs and MXene is carried out by thermal fusion method, and the particle size, ionization performance, heat conduction index and hydrophilic and hydrophobic properties of the MXene / MWCNTs matrix material are characterized.

Images