Mammary gland sentinel node body surface noninvasive positioning and property detection method and device and medium

Abstract

The invention discloses a mammary gland sentinel node body surface non-invasive positioning and property detection method, a mammary gland sentinel node body surface non-invasive positioning and property detection device and a medium, relates to the technical field of biomedical engineering and spectral analysis, and synchronously realizes non-invasive rapid positioning and accurate intelligent property judgment of mammary gland sentinel node through collecting a Raman spectrum of a body surface through an optical fiber for the first time. According to the technical scheme, the method is characterized in that a representative Raman spectrum is selected by fusing multiple strategies to construct a reference set, ridge regression is introduced to ensure stable solution, accurate and self-adaptive correction of the scattering effect of the complex biological sample is achieved, the problem of poor adaptability caused by single benchmark in a traditional method is effectively solved, and the method is suitable for large-scale popularization and application. And the generalization capability and robustness of the scattering correction model are obviously improved. Then, by dynamically learning the specialities of different expert models under different sample characteristics, a proper expert weight is adaptively allocated to each test sample, so that the complementary advantages of multiple algorithms are intelligently fused, and the body surface lymph node property detection precision and reliability are remarkably improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of biomedical engineering and spectral analysis, more particularly, it relates to a breast sentinel lymph node body surface non-invasive positioning and property detection method, device and medium. BACKGROUND

[0002] The breast sentinel lymph node is the first station hub of breast cancer lymph node metastasis, and whether it occurs metastasis is the decisive basis for tumor TNM staging, determining the scope of surgery, developing adjuvant therapy and evaluating the prognosis of patients. Breast cancer lymph node positioning and property detection can accurately diagnose the disease, guide surgery and treatment plan and judge prognosis.

[0003] The existing clinical methods for lymph node positioning and property detection include sentinel lymph node biopsy method and intraoperative frozen section pathological method, both of which are invasive, irreversible damage to the human body, and have the risk of false negative (missed metastasis). Intraoperative frozen section pathological method needs pathological results within 30 minutes, and may miss small metastasis or small lesions in lymph nodes; the clinical ultrasound positioning combined with fine needle aspiration method is also a invasive detection, and depends on the experience of the operator, which is difficult to find small metastasis lesions with small volume and no obvious abnormal shape, resulting in missed detection, and the display effect of deep lymph nodes is poor; the clinical sentinel lymph node positioning method also includes nuclide lymph imaging, injecting radioactive nuclide markers, and displaying lymphatic drainage pathway and suspicious lymph nodes through imaging. This method has ionizing radiation, requires special equipment and radioactive drug qualification, and has limited resolution for small metastasis. It can be seen that the current clinical methods for sentinel lymph node positioning and property detection have the disadvantages of trauma, irreversible damage, low detection accuracy, easy missed detection of small metastasis, and dependence on doctor's experience. Therefore, there is an urgent need for non-traumatic, high-precision, real-time, intelligent and safe breast sentinel lymph node body surface non-invasive positioning and property detection method and device, which has great clinical significance for realizing individualized precise treatment of breast cancer, avoiding unnecessary lymph node dissection and its related complications, and ultimately improving the quality of life of patients.

[0004] As a non-invasive molecular vibration spectroscopy technology, fiber Raman spectroscopy technology can non-destructively obtain the fingerprint information of key biological molecules such as nucleic acids, proteins and lipids in biological tissues. The hardware device has the characteristics of small size, low cost and convenient operation. This technology provides an ideal solution for breast sentinel lymph node body surface high-precision intelligent detection due to its unique advantages of non-invasiveness, real-time, safety and early molecular diagnosis potential. This technology does not require complex pathological section preparation, can directly collect Raman spectrum of lymph nodes through the skin on the body surface, and realizes positioning and property diagnosis of sentinel lymph nodes through intelligent algorithm program, replacing artificial subjective interpretation, and is more likely to develop a new outpatient or preoperative diagnosis technology, showing great clinical application value.

[0005] When a laser irradiates biological tissue on the body surface, photons interact with the tissue, generating Raman spectra that reflect molecular fingerprints. However, the tissue's physical properties (such as particle size and uneven distribution) and surface contours also produce strong physical scattering effects. For fiber optic miniature Raman spectrometers, the simplified equipment results in a relatively low signal-to-noise ratio (SNR) for the acquired biological tissue Raman spectra. The raw Raman spectra contain strong scattering effects introduced by the tissue's physical properties, leading to systematic differences in spectral intensity, increasing intra-class dispersion of the data, and severely affecting the accuracy of subsequent data processing models. Therefore, this invention proposes a non-invasive method, device, and medium for the localization and property detection of sentinel lymph nodes on the body surface of the breast. Summary of the Invention

[0006] The above-mentioned technical objective of the present invention is achieved through the following technical solution: The first aspect of this invention provides a non-invasive method for locating and detecting the properties of sentinel lymph nodes in the breast, comprising the following steps: Includes the following steps: Raman spectra of the skin near the sentinel lymph node to be tested, Raman spectra of multiple tested sentinel lymph nodes, and labels of the properties of multiple tested sentinel lymph nodes for each category were obtained. The tracer Raman spectrum was collected, and the surface Raman spectrum of the sentinel lymph node to be tested and the tracer Raman spectrum were input into the spectral matching algorithm to locate the breast sentinel lymph node. The surface Raman spectra of the pre-measured sentinel lymph nodes and the surface Raman spectra of the sentinel lymph nodes to be measured were pre-processed with smoothing and baseline correction, and then scattering correction was performed to obtain the corrected dataset of the measured spectra and the spectra to be measured. The calibrated measured spectral dataset is divided into a training set and a validation set. An expert model is constructed, and the expert model is trained using the training set to output a class probability vector. The validation set is used to construct a gated training set, and then the gated model is trained using this set to obtain the finally trained gated model. After locating the sentinel lymph nodes in the breast, the corrected spectrum to be measured is input into the trained gating model, which outputs a soft membership matrix as the confidence weight of each expert model. The weights of each expert model are multiplied by the probability vector output by that expert model and then added together to obtain the fused total output probability. The final predicted category is the category corresponding to the highest probability in the total output probability vector.

[0007] A second aspect of the present invention also provides an apparatus / device / system for non-invasive localization and property detection of sentinel lymph nodes in the breast, comprising a memory, a processor, and a computer program stored in the memory, characterized in that the processor executes the computer program to implement the steps of the above method.

[0008] A third aspect of the present invention also provides a computer-readable storage medium having a computer program / instructions stored thereon, which, when executed by a processor, implement the steps of the above-described method.

[0009] A fourth aspect of the present invention also provides a computer program product, including a computer program / instructions that, when executed by a processor, implement the steps of the above-described method.

[0010] In summary, the present invention has the following beneficial effects: 1. Achieve non-invasive, rapid, and precise surface localization of breast sentinel lymph nodes, completely overcoming the limitations of traditional methods. Through specific enrichment of tracers such as indocyanine green, methylene blue, patent blue, isosulfur blue, mitoxantrone, or radionuclides, combined with Raman spectroscopy matching technology, the location of sentinel lymph nodes can be pinpointed solely through surface scanning without puncture or surgical sampling. This process causes no tissue damage, completely avoiding irreversible harm from invasive procedures, and significantly improving patient comfort and safety. Furthermore, the localization process does not rely on operator experience, effectively resolving issues of missed detections and misjudgments in clinical localization, and providing a precise target area for subsequent qualitative testing.

[0011] 2. After lymph node localization, the nature of the lymph nodes is precisely and intelligently detected simultaneously using the acquired surface Raman spectroscopy. This offers significant advantages such as real-time detection, non-invasiveness, and painlessness, solving the problems of current ultrasound methods that rely on operator experience, making it difficult to detect small metastatic lesions with no obvious morphological abnormalities, leading to missed detections, and providing poor visualization of deep lymph nodes. It also avoids the invasiveness, at least 30-minute waiting time, and irreversible damage associated with traditional biopsies or intraoperative frozen section pathology. This invention is particularly suitable for outpatient and preoperative diagnostic scenarios, providing doctors with accurate information on the benign or malignant nature of lymph nodes, assisting in making precise surgical decisions, and offering a powerful new solution for the early detection and preoperative assessment of lymph node metastasis.

[0012] 3. The raw Raman spectra contain strong scattering effects introduced by tissue physical properties, leading to systematic differences in spectral intensity, increasing the intra-class dispersion of the data, and affecting the accuracy of subsequent data processing models. Existing scattering correction algorithms typically rely on a single reference spectrum, which is difficult to adapt to the inherent strong heterogeneity of biological tissue Raman spectra, easily resulting in under-correction or over-correction. To address this problem, this invention proposes a correction method based on multi-reference spectral ridge regression. This method selects representative spectra to construct a reference set by fusing multiple strategies and introduces ridge regression to ensure stable solutions, thereby achieving accurate and adaptive correction of scattering effects in complex biological samples, effectively overcoming the poor adaptability problem caused by the single reference in traditional methods.

[0013] 4. Due to its simplified equipment, the fiber optic miniature Raman spectrometer acquires biological tissue Raman spectra with a low signal-to-noise ratio. When using multiple reference spectra for regression analysis, directly applying the least squares method is susceptible to noise, leading to unstable calibration results. To address this issue, this invention introduces a ridge regression regularization term into the regression model for parameter solving. This mechanism, by constraining the norm of the solution vector, effectively suppresses the instability in parameter estimation caused by collinearity between reference spectra and noise interference, avoids overfitting, and significantly improves the generalization ability and robustness of the scattering correction model.

[0014] 5. Even after correction, spectral data still exhibits characteristics such as high dimensionality, small sample size, and nonlinearity. Furthermore, different classification models perform differently under varying data characteristics, making it difficult for a single model to guarantee globally optimal performance. To address this issue, this invention dynamically learns the expertise of different expert models under different sample features. This mechanism adaptively assigns appropriate expert weights to each test sample, intelligently integrating the complementary advantages of multiple algorithms and significantly improving the classification accuracy and decision reliability of the integrated system. Attached Figure Description

[0015] Figure 1 This is a typical ICG-stained lymph node Raman spectrum in this invention; Figure 2 This refers to the ICG Raman spectrum used in this invention; Figure 3 This is a flowchart of a non-invasive method for locating and detecting the properties of sentinel lymph nodes in the breast, as described in Examples 1 and 2. Figure 4 This is the overall system framework diagram of the present invention; Figure 5 This is a flowchart of a non-invasive method for locating and detecting the properties of sentinel lymph nodes in the breast, as described in Embodiment 3 of the present invention. Detailed Implementation

[0016] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0017] Specifically, using, such as Figure 4 The non-invasive localization and property detection system for breast sentinel lymph nodes shown implements the method of this application. The non-invasive detection system includes: a control and storage unit, a spectral acquisition unit, and a spectral analysis and processing unit.

[0018] The control and storage unit includes a processor, memory, serial communication module, control program, key control input, and display device. The control program instructs the processor to execute spectral acquisition commands, calls the spectral acquisition unit via serial communication to acquire multi-band mixed spectra, and saves the acquired spectra in memory. The acquired multi-band mixed spectra are displayed on the display device, and the key control input allows adjustment of input parameters. The spectral analysis and processing unit contains algorithms for spectral matching, smoothing and baseline correction preprocessing, spectral scattering correction, and sentinel lymph node property detection. Based on these algorithms, it provides analysis results, which are stored in memory. The control program calls the algorithms to execute spectral analysis and displays the results on the display device in the control and storage unit. The spectral acquisition unit can be a miniature spectrometer or other instruments capable of spectral acquisition. In this embodiment, a 785nm excitation source is used, and the Y-shaped optical fiber includes the optical fiber and a probe. The spectrometer and laser are connected via the Y-shaped optical fiber. The excitation light emitted by the laser is reflected by the tissue surface and received by the spectrometer, thereby acquiring the multi-band mixed spectrum of the tissue.

[0019] Example 1: A non-invasive method for locating and detecting the properties of sentinel lymph nodes in the breast, such as... Figure 1 As shown, it includes the following steps: Step 1: Inject a tracer such as indocyanine green, methylene blue, patent blue, isosulfur blue, mitoxantrone, or a radionuclide into the areola area intradermally. When inserting the needle, tilt the bevel upwards at a 5° angle into the skin. Inject 0.2-0.3 ml of tracer at each injection point to form a wheal. Two to three injection points can be selected, avoiding leakage of the injection solution. After injection, gently massage the injection site to allow the tracer to drain through the lymphatic system to the sentinel lymph nodes and accumulate.

[0020] Step 2: Place the fiber optic probe of the fiber optic miniature Raman spectrometer close to and scan the skin surface near the sentinel lymph nodes of the breast. Obtain the Raman spectra of the skin surface near the sentinel lymph nodes to be tested, the surface Raman spectra of multiple tested sentinel lymph nodes, and the properties of multiple tested sentinel lymph nodes for various categories. Also, collect Raman spectra of tracers such as indocyanine green, methylene blue, patent blue, isothiocyanate, mitoxantrone, or radionuclides.

[0021] Step 3: The surface Raman spectrum of the sentinel lymph node to be tested is matched with the Raman spectrum of tracers such as indocyanine green, methylene blue, patent blue, isosulfur blue, mitoxantrone, or radionuclides using a spectral matching algorithm. If the match is successful, the tested site is the sentinel lymph node, thus realizing the surface localization of the sentinel lymph node.

[0022] Step 4: Perform smoothing and baseline correction preprocessing on the measured sentinel lymph node surface Raman spectra and the sentinel lymph node surface Raman spectra to be measured.

[0023] Step 5: Construct a system using three methods that includes... K Reference spectral matrix of representative spectra R ① Calculate the average spectrum of the entire spectral dataset. : In the formula, M This represents the total number of spectra. For the first i Spectral vectors.

[0024] ② Select the one with the smallest Euclidean distance from the average spectrum. Spectrum. A measure of the measured spectrum. With mean spectrum The formula for distance is: Calculate the distances for all spectra. Then, select the one with the smallest distance. Adding the spectrum to the reference spectrum set .

[0025] ③ Principal Component Analysis (PCA) based method. This method is applied to the mean-centered spectral dataset. Perform PCA and project the centralized image onto the front. L Calculate the score matrix along each principal component direction. T : In the formula, This is the eigenvector matrix obtained by the PCA process.

[0026] Select the principal component direction that contributes the most to the explanation of variance, ranked by score. Spectrum Add to the reference spectrum set.

[0027] Finally, the spectra selected by the three methods are combined to form the final reference spectral matrix. Step 6: For the measured spectrum to be corrected Using a reference spectral matrix R Ridge regression modeling is performed, and the model is... In the formula, The scale factor weights to be estimated are: This is the offset. Regularization constraints are applied by introducing ridge regression: In the formula, This is the regularization coefficient.

[0028] For ease of solution, an extended matrix is ​​defined. and parameter vector Then the objective function can be simplified to: Its closed solution is: In the formula, I It is an identity matrix.

[0029] Step 7: Calculate the sum of squared residuals for individual regression of each reference spectrum. : In the formula, k As an index of the reference spectrum, each reference spectrum used for regression will obtain a set of estimators. Then, the weight of each reference spectrum is determined based on the residuals: And perform weight normalization: Step 8: Utilize weights For each set of parameters obtained in step 6 and A weighted average is then performed to obtain the final scattering correction parameters. and : Step 9: Calculate the corrected spectrum using the final parameters: Thus, Raman spectra that have eliminated most of the scattering effects and mainly retain the chemical composition information were obtained.

[0030] Step 10: Divide the pre-treated sentinel lymph node surface Raman spectra stained with tracers such as indocyanine green, methylene blue, patent blue, isosulfur blue, mitoxantrone, or radionuclides into a training set. and verification set Based on the pathological results, the corresponding categories were marked, and the lymph nodes were divided into normal lymph nodes, inflammatory lymph nodes, and metastatic lymph nodes.

[0031] Step 11: From the 12 initial models—Support Vector Machine, Gaussian Process Classifier, Random Forest, Extreme Gradient Boosting, Convolutional Neural Network, Logistic Regression, Naive Bayes, Linear Discriminant Analysis, Kernel Extreme Learning Machine, K-Nearest Neighbors, Adaptive Augmentation, and Multilayer Perceptron—three models are selected to construct an expert model system. Ultimately, Kernel Extreme Learning Machine (KELM), Random Forest (RF), and Gaussian Process Classifier (GPC) are chosen as the foundational expert models. (In the training set...) The following are the training methods for the three expert models: KELM, RF, and GPC, which are designated as expert 1, expert 2, and expert 3, respectively. Each model will train on the following samples. x Each will output a normalized class probability vector. , indicating the probability that the expert model predicts the sample belongs to each category: In the formula, Indicates the first m Expert model prediction samples x Belongs to the C The probability of the category.

[0032] Step 12: Utilize the validation set Determine the "Best Expert" label. z Construct a dataset for training the gating network. The rules are as follows: ①If only one expert makes a correct prediction, then the expert's number becomes the new label for the sample in the validation set. ②If multiple experts make correct predictions, select the one that appears earlier in the priority order or the one with the highest probability of predicting the correct label. ③ If no expert makes a correct prediction, then select the one with the highest probability of predicting the correct label category.

[0033] Finally, construct the gating training set. .

[0034] In the formula, To verify the performance of a single Raman spectrum sample after scattering correction, z The label for this sample is "Best Expert".

[0035] Using the constructed gating training set Train the gating model.

[0036] Step 13: Construct a Gaussian kernel similarity matrix for all samples in the gated training set. This is used to measure the structural relationships between samples. The first sample and the first The similarity calculation rules between samples are as follows: In the formula, and The first The training sample and the first The feature vectors of each training sample The bandwidth parameter of the Gaussian kernel function controls the rate of similarity decay. The smaller the value, the faster the similarity decays and the more localized the neighborhood becomes.

[0037] The first can be calculated based on the Gaussian kernel similarity matrix. i Laplace score of wavenumber feature points: In the formula, Indicates the first The training sample at the th ... Intensity at each wavenumber point Indicates the first The training sample at the th ... Intensity at each wavenumber point For all training samples at the 1st The average intensity at each wavenumber point.

[0038] Step 14: For the sentinel lymph node sample to be tested Calculate its relationship with all training samples The weighted Laplace cosine similarity is calculated according to the following rules: In the formula, This indicates that the sentinel lymph node sample to be tested was in the first... The intensity at each wavenumber point.

[0039] Selection of sentinel lymph node samples for testing The highest cosine similarity The training samples constitute a set For each predicted category, calculate the fuzzy membership degree of that category: In the formula, For the sample class currently used to calculate fuzzy membership, For fuzzy kernel bandwidth parameters, The function for labeling sample categories follows these rules: Step 15: Further, the boundary concept of rough sets is introduced to calculate the sentinel lymph node samples to be tested. The confidence level for each category is determined according to the following rules: In the formula, Belonging to the same category The collection of all training samples, To adjust the power exponent parameter for similarity decay.

[0040] The gating model training is now complete.

[0041] Step 16: Locate the sentinel lymph nodes of the subject using steps 1-3, and use the surface Raman spectrum of the sentinel lymph nodes after smoothing, baseline correction, and scattering correction preprocessing. The trained gating model is used to predict its soft membership matrix. As the confidence weights for each expert model: In the formula, Indicates the first j The test sample pair m The weights of an expert model, which satisfy: Step 17: Multiply the weights of each expert model by the probability vector output by that expert model, then sum the results term by term to obtain the total output probability after fusion. The final predicted category is the category corresponding to the highest probability in the total output probability vector: This completes the non-invasive surface characterization of the sentinel lymph nodes in the breast.

[0042] Example 2: A non-invasive method for locating and detecting the properties of sentinel lymph nodes in the breast, such as... Figure 1 As shown, it includes the following steps: Step 1: Inject indocyanine green (ICG) intradermally into the areola area. When inserting the needle, tilt the bevel upwards at a 5° angle into the skin. Inject 0.2-0.3 ml of ICG at each injection point to form a wheal. Two to three injection points can be selected to avoid leakage. After injection, gently massage the injection site to allow the tracer to reach and accumulate in the sentinel lymph nodes via lymphatic drainage.

[0043] Step 2: Place the fiber optic probe of the fiber optic miniature Raman spectrometer close to and scan the skin surface near the sentinel lymph nodes of the breast. Acquire the Raman spectra of the skin surface near the sentinel lymph node to be tested, the surface Raman spectra of multiple tested sentinel lymph nodes, and the labels of the properties of multiple tested sentinel lymph nodes for each category; the original spectral range covers 0-3012 cm⁻¹. -1 A total of 1038 data points were collected. One Raman spectrum of indocyanine green liquid was also acquired. Typical Raman spectra of lymph nodes after ICG staining are shown below. Figure 1 ICG Raman spectra can be found Figure 2 .

[0044] Step 3: The surface Raman spectrum of the sentinel lymph node to be tested is matched with the indocyanine green Raman spectrum using a spectral matching algorithm. If the match is successful, the tested site is the sentinel lymph node, thus realizing the surface localization of the sentinel lymph node.

[0045] Step 4: Perform smoothing and baseline correction preprocessing on the measured sentinel lymph node surface Raman spectra and the sentinel lymph node surface Raman spectra to be measured.

[0046] Step 5: Construct a system using multiple strategies. K Reference spectral matrix of representative spectra R The specific implementation is as follows: First, calculate the average spectrum of the entire spectral dataset as a baseline reference: Next, calculate the Euclidean distance between each spectrum and the average spectrum, and select the spectrum with the smallest scattering: Sort by distance from smallest to largest, then select the first... Spectral composition .

[0047] Then, based on principal component analysis, the mean-centered spectral dataset was analyzed. Perform PCA and project the centralized image onto the front. L Calculate the score matrix along each principal component direction. T : In the formula, This is the eigenvector matrix obtained by the PCA process.

[0048] Finally, the top-scoring principal component directions that contribute the most to the variance explanation are selected. Spectrum Add to the reference spectrum set.

[0049] Finally, the spectra selected by the three strategies are combined to form a reference spectral matrix. .

[0050] Step 6: For the measured spectrum to be corrected Using a reference spectral matrix R Ridge regression modeling is performed, and the model is... In the formula, The scale factor weights to be estimated are: This is the offset. Regularization constraints are applied by introducing ridge regression: In the formula, This is the regularization coefficient.

[0051] For ease of solution, an extended matrix is ​​defined. and parameter vector Then the objective function can be simplified to: Its closed solution is: In the formula, I It is an identity matrix.

[0052] Step 7: Calculate the sum of squared residuals for individual regression of each reference spectrum. : In the formula, k As an index of the reference spectrum, each reference spectrum used for regression will obtain a set of estimators. Then, the weight of each reference spectrum is determined based on the residuals: And perform weight normalization: Step 8: Utilize weights For each set of parameters obtained in step 2 and A weighted average is then performed to obtain the final scattering correction parameters. and : Step 9: Calculate the corrected spectrum using the final parameters: Thus, Raman spectra that have eliminated most of the scattering effects and mainly retain the chemical composition information were obtained.

[0053] Step 10: Divide the pre-treated indocyanine green stained sentinel lymph node surface Raman spectra into a training set. and verification set The data was then categorized based on pathological results, into normal lymph nodes, inflammatory lymph nodes, and metastatic lymph nodes. Specifically, after indocyanine green staining, there were 201 normal lymph node spectra, 198 inflammatory lymph node spectra, and 206 metastatic lymph node spectra. The corrected spectral dataset was then randomly divided into training and validation sets in an 8:2 ratio. The specific division is as follows: Training set: 485 spectra, including 161 normal lymph node spectra, 158 inflammatory lymph node spectra, and 166 metastatic lymph node spectra.

[0054] Validation set: 120 spectra, including 40 normal lymph node spectra, 40 inflammatory lymph node spectra, and 40 metastatic lymph node spectra.

[0055] Step 11: From the 12 initial models—Support Vector Machine, Gaussian Process Classifier, Random Forest, Extreme Gradient Boosting, Convolutional Neural Network, Logistic Regression, Naive Bayes, Linear Discriminant Analysis, Kernel Extreme Learning Machine, K-Nearest Neighbors, Adaptive Augmentation, and Multilayer Perceptron—three models are selected to construct an expert model system. Ultimately, Kernel Extreme Learning Machine (KELM), Random Forest (RF), and Gaussian Process Classifier (GPC) are chosen as the foundational expert models. (In the training set...) The following are the training methods for the three expert models: KELM, RF, and GPC, which are designated as expert 1, expert 2, and expert 3, respectively. Each model will train on the following samples. x Each will output a normalized class probability vector. , indicating the probability that the expert model predicts the sample belongs to each category: In the formula, Indicates the first m Expert model prediction samples x Belongs to the C The probability of the category.

[0056] Step 12: Utilize the validation set Determine the "Best Expert" label. zConstruct a dataset for training the gating network. The rules are as follows: ①If only one expert makes a correct prediction, then the expert's number becomes the new label for the sample in the validation set. ②If multiple experts make correct predictions, select the one that appears earlier in the priority order or the one with the highest probability of predicting the correct label. ③ If no expert makes a correct prediction, then select the one with the highest probability of predicting the correct label category.

[0057] Finally, construct the gating training set. .

[0058] In the formula, To verify the performance of a single Raman spectrum sample after scattering correction, z The label for this sample is "Best Expert".

[0059] Through this process, 120 validation set samples were assigned the corresponding "best expert" label, forming a gated training set.

[0060] Using the constructed gating training set Train the gating model.

[0061] Step 13: Construct a Gaussian kernel similarity matrix for all samples in the gated training set. This is used to measure the structural relationships between samples. The first sample and the first The similarity calculation rules between samples are as follows: In the formula, and The first The training sample and the first The feature vectors of each training sample The bandwidth parameter of the Gaussian kernel function controls the rate of similarity decay. The smaller the value, the faster the similarity decays and the more localized the neighborhood becomes.

[0062] The first can be calculated based on the Gaussian kernel similarity matrix. i Laplace score of wavenumber feature points: In the formula, Indicates the first The training sample at the th ... Intensity at each wavenumber point Indicates the first The training sample at the th ... Intensity at each wavenumber point For all training samples at the 1st The average intensity at each wavenumber point.

[0063] Step 14: For the sentinel lymph node sample to be tested Calculate its relationship with all training samples The weighted Laplace cosine similarity is calculated according to the following rules: In the formula, Indicates the sample to be predicted is in the th... The intensity at each wavenumber point.

[0064] Selection of sentinel lymph node samples for testing The highest cosine similarity The training samples constitute a set For each predicted category, calculate the fuzzy membership degree of that category: In the formula, For the sample class currently used to calculate fuzzy membership, For fuzzy kernel bandwidth parameters, The function for labeling sample categories follows these rules: Step 15: Further, the boundary concept of rough sets is introduced to calculate the sentinel lymph node samples to be tested. The confidence level for each category is determined according to the following rules: In the formula, Belonging to the same category The collection of all training samples, To adjust the power exponent parameter for similarity decay.

[0065] The gating model training is now complete.

[0066] Step 16: Sentinel lymph nodes were located using steps 1-3. The Raman spectra of the sentinel lymph nodes, preprocessed with smoothing, baseline correction, and scattering correction, were used as the test set, including 21 normal breast lymph node spectra, 18 inflammatory lymph node spectra, and 20 metastatic lymph node spectra. A trained gating model was used to predict their soft membership matrices. As the confidence weights for each expert model: In the formula, Indicates the first j The test sample pair m The weights of an expert model, which satisfy: Step 17: Multiply the weights of each expert model by the probability vector output by that expert model, then sum the results term by term to obtain the total output probability after fusion. The final predicted category is the category corresponding to the highest probability in the total output probability vector: This completes the non-invasive surface characterization of the sentinel lymph nodes in the breast.

[0067] Experimental results: In terms of localization, the ICG spectral matching method of this invention can stably and accurately identify the corresponding surface region of sentinel lymph nodes, providing reliable target points for subsequent spectral acquisition and property analysis. Regarding property discrimination, this experiment uses the test set from step 13 to test the proposed method. Five sets of comparative experiments were conducted using Support Vector Machine (SVM), Adaptive Augmentation (Adaboost), Kernel Extreme Learning Machine (KELM), Random Forest (RF), and Gaussian Process Classifier (GPC), respectively. The original Raman spectral dataset without scattering correction was used for testing. The training set consisted of 485 spectra, including 161 normal lymph node spectra, 158 inflammatory lymph node spectra, and 166 metastatic lymph node spectra. The test set included 21 normal breast lymph node spectra, 18 inflammatory lymph node spectra, and 20 metastatic lymph node spectra. The prediction results for the test set are shown in Table 1. The comparison shows that the method proposed in this invention outperforms other methods in terms of recognition accuracy.

[0068] Table 1. Comparison of recognition results of several pattern recognition algorithms and the method of this invention.

[0069] In summary, the complete solution proposed in this invention, which integrates multi-reference spectral ridge regression scattering correction, spectral matching localization, and gating expert integration algorithms, not only achieves accurate surface localization of breast sentinel lymph nodes but also demonstrates excellent performance in property detection tasks, providing effective technical support for real-time, non-invasive intraoperative diagnosis in clinical practice.

[0070] Example 3: A non-invasive method for locating and detecting the properties of sentinel lymph nodes in the breast, such as... Figure 5 As shown, it includes the following steps: Includes the following steps: S100. Obtain the Raman spectra of the skin surface near the sentinel lymph node to be tested, the surface Raman spectra of multiple tested sentinel lymph nodes, and the labels of the properties of multiple tested sentinel lymph nodes for each category; S200. And acquire tracer Raman spectra, input the surface Raman spectrum of the sentinel lymph node to be tested and the tracer Raman spectrum into the spectral matching algorithm to locate the breast sentinel lymph node; S300. The surface Raman spectra of the pre-measured sentinel lymph nodes and the surface Raman spectra of the sentinel lymph nodes to be measured are pre-processed with smoothing and baseline correction, and then scattering correction is performed to obtain the corrected measured spectrum dataset and the spectrum to be measured. S400. Divide the calibrated measured spectral dataset into a training set and a validation set, construct an expert model, train the expert model using the training set, and output the class probability vector; use the validation set to construct a gated training set, and then use it to train the gated model, finally obtaining the trained gated model; S500. After locating the sentinel lymph nodes in the breast, the corrected spectrum to be measured is input into the trained gating model, and the soft membership matrix is ​​output as the confidence weight of each expert model. The weight of each expert model is multiplied by the probability vector output by that expert model and then added item by item to obtain the fused total output probability. Finally, the predicted category is the category corresponding to the highest probability in the total output probability vector.

[0071] In step S100 of this embodiment, acquiring the Raman spectrum set of the previously measured sentinel lymph node and the Raman spectrum of the sentinel lymph node to be measured involves: intradermal injection of tracers such as indocyanine green, methylene blue, patent blue, isosulfur blue, mitoxantrone, or radionuclides into the areola area. The needle is inserted at a 5° angle with the bevel facing upwards. 0.2-0.3 ml of tracer is injected at each injection point to form a wheal. Two to three injection points can be selected to avoid leakage. After injection, the injection site is gently massaged to allow the tracer to reach and accumulate in the sentinel lymph node via lymphatic drainage. The fiber optic probe of the fiber optic miniature Raman spectrometer is placed close to and scans the skin surface near the breast sentinel lymph node to obtain the original Raman spectrum dataset, and Raman spectra of tracers such as indocyanine green, methylene blue, patent blue, isosulfur blue, mitoxantrone, or radionuclides are acquired.

[0072] In step S100 of this embodiment, the step of inputting the surface Raman spectrum of the sentinel lymph node to be tested and the tracer Raman spectrum into the spectral matching algorithm to locate the breast sentinel lymph node is as follows: the spectral matching algorithm is used to perform spectral matching between the surface Raman spectrum of the sentinel lymph node to be tested and the tracer Raman spectrum. If the matching is successful, the tested site is the sentinel lymph node, thereby realizing the surface localization of the sentinel lymph node.

[0073] In step S200 of this embodiment, the scattering correction includes: constructing a reference spectral matrix using the preprocessed measured sentinel lymph node surface Raman spectrum set and the sentinel lymph node surface Raman spectrum to be measured; using ridge regression modeling to eliminate physical scattering interference; and achieving adaptive correction through weighted average parameters.

[0074] The preprocessed set of measured sentinel lymph node surface Raman spectra and the Raman spectra of the sentinel lymph node to be measured are used to construct a reference spectral matrix. This includes calculating the average spectrum, the spectrum with the minimum Euclidean distance, and the spectrum with the maximum PCA variance contribution, and then merging them to obtain the reference spectral matrix.

[0075] The expert model is constructed by selecting two or more from the following algorithms: Support Vector Machine, Gaussian Process Classifier, Random Forest Algorithm, Extreme Gradient Boosting Algorithm, Convolutional Neural Network, Logistic Regression Algorithm, Naive Bayes Classification Algorithm, Linear Discriminant Analysis, Kernel Extreme Learning Machine, K Nearest Neighbor Algorithm, Adaptive Augmentation Algorithm, and Multilayer Perceptron.

[0076] The present invention also provides an apparatus / device / system for a non-invasive method for locating and detecting the properties of sentinel lymph nodes on the surface of the breast, comprising a memory, a processor, and a computer program stored in the memory, characterized in that the processor executes the computer program to implement the steps of the above method.

[0077] The present invention also provides a computer-readable storage medium having a computer program / instructions stored thereon, which, when executed by a processor, implement the steps of the above-described method.

[0078] The present invention also provides a computer program product, including a computer program / instructions that, when executed by a processor, implement the steps of the above-described method.

[0079] Finally, it should be noted that the above descriptions are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A non-invasive method for locating and detecting the properties of sentinel lymph nodes in the breast, characterized by: Includes the following steps: Raman spectra of the skin near the sentinel lymph node to be tested, Raman spectra of multiple tested sentinel lymph nodes, and labels of the properties of multiple tested sentinel lymph nodes for each category were obtained. The tracer Raman spectrum was collected, and the surface Raman spectrum of the sentinel lymph node to be tested and the tracer Raman spectrum were input into the spectral matching algorithm to locate the breast sentinel lymph node. The surface Raman spectra of the pre-measured sentinel lymph nodes and the surface Raman spectra of the sentinel lymph nodes to be measured were pre-processed with smoothing and baseline correction, and then scattering correction was performed to obtain the corrected dataset of the measured spectra and the spectra to be measured. The corrected measured spectral dataset is divided into a training set and a validation set. An expert model is constructed, and the expert model is trained using the training set to output a class probability vector. Using the validation set, a gating training set is constructed, and then the gating model is trained using this set, finally obtaining the trained gating model; After locating the sentinel lymph nodes in the breast, the corrected spectrum to be measured is input into the trained gating model, which outputs a soft membership matrix as the confidence weight of each expert model. The weights of each expert model are multiplied by the probability vector output by that expert model and then added together to obtain the fused total output probability. The final predicted category is the category corresponding to the highest probability in the total output probability vector.

2. The method for non-invasive localization and property detection of breast sentinel lymph nodes according to claim 1, characterized in that: The process of obtaining the measured surface Raman spectrum set of sentinel lymph nodes and the surface Raman spectrum set of sentinel lymph nodes to be tested includes: injecting a tracer intradermally into the areola area, gently massaging the injection site after injection to allow the tracer to reach and accumulate in the sentinel lymph nodes through lymphatic drainage, scanning and collecting the data using a probe to obtain the measured surface Raman spectrum set of sentinel lymph nodes and the surface Raman spectrum set of sentinel lymph nodes to be tested.

3. The method for non-invasive localization and property detection of breast sentinel lymph nodes according to claim 1, characterized in that: The step of inputting the surface Raman spectrum of the sentinel lymph node to be tested and the tracer Raman spectrum into the spectral matching algorithm to locate the sentinel lymph node in the breast is as follows: the spectral matching algorithm is used to perform spectral matching between the surface Raman spectrum of the sentinel lymph node to be tested and the tracer Raman spectrum. If the matching is successful, the tested site is the sentinel lymph node, thereby realizing the surface localization of the sentinel lymph node.

4. The method for non-invasive localization and property detection of breast sentinel lymph nodes according to claim 1, characterized in that: The scattering correction includes: constructing a reference spectral matrix using the preprocessed measured sentinel lymph node surface Raman spectrum set and the sampled sentinel lymph node surface Raman spectrum; using ridge regression modeling to eliminate physical scattering interference; and achieving adaptive correction through weighted average parameters.

5. The method for non-invasive localization and property detection of breast sentinel lymph nodes according to claim 4, characterized in that: The preprocessed set of measured sentinel lymph node surface Raman spectra and the Raman spectra of the sentinel lymph node to be measured are used to construct a reference spectral matrix, including calculating the average spectrum, the spectrum with the minimum Euclidean distance, and the spectrum with the maximum PCA variance contribution, and then merging them to obtain the reference spectral matrix.

6. The method for non-invasive localization and property detection of breast sentinel lymph nodes according to claim 1, characterized in that: The expert model is constructed by selecting two or more from the following algorithms: Support Vector Machine, Gaussian Process Classifier, Random Forest Algorithm, Extreme Gradient Boosting Algorithm, Convolutional Neural Network, Logistic Regression Algorithm, Naive Bayes Classification Algorithm, Linear Discriminant Analysis, Kernel Extreme Learning Machine, K Nearest Neighbor Algorithm, Adaptive Augmentation Algorithm, and Multilayer Perceptron.

7. A device / equipment / system for non-invasive localization and characterization of sentinel lymph nodes in the breast, comprising a memory, a processor, and a computer program stored in the memory, characterized in that, The processor executes the computer program to implement the steps of the method according to any one of claims 1-6.

8. A computer-readable storage medium having a computer program / instructions stored thereon, characterized in that, When the computer program / instructions are executed by the processor, they implement the steps of the method described in any one of claims 1-6.

9. A computer program product comprising a computer program / instructions, characterized in that, When the computer program / instructions are executed by the processor, they implement the steps of the method described in any one of claims 1-6.

Images