Detection method of radiosensitizing drugs for gastrointestinal tumors based on spectral detection technology

Abstract

The present invention relates to the field of spectral analysis technology, and proposes a method for detecting gastrointestinal tumor radiotherapy sensitizing drugs based on spectral detection technology, including: collecting spectral curves of a number of normal sensitizing drugs and sensitizing drugs containing impurities; classifying all spectral curves to obtain base classifiers and a number of spectral categories; obtaining a number of spectral clusters of each spectral category; obtaining a number of spectral feature vectors and spectral feature values ​​of each spectral cluster of each spectral category; matching clusters of different spectral categories to obtain a number of matching cluster pairs; obtaining a number of difference cluster pairs, and the difference coefficient of each spectral curve in each spectral cluster in the difference cluster pairs; obtaining an integrated classifier for distinguishing spectral curves; and performing quality detection on the sensitizing drugs to be tested by using the integrated classifier. The present invention aims to solve the problem that when performing impurity detection on sensitizing drugs, the classification effect is poor due to the similarity of spectral curves, which affects the detection results.

Description

Technical Field

[0001] The present invention relates to the technical field of spectral analysis, and in particular to a method for detecting gastrointestinal tumor radiotherapy sensitizing drugs based on spectral detection technology. Background Art

[0002] The quality inspection of sensitizing drugs for gastrointestinal tumor radiotherapy can ensure that the activity and performance of sensitizing drugs meet the standards, thereby effectively enhancing the therapeutic effect of gastrointestinal tumor radiotherapy; only through strict quality control can patients receive effective treatment and improve the success rate of treatment. Usually, spectral analysis is used for quality inspection of sensitizing drugs. When the quality of sensitizing drugs is unqualified, they will contain impurities, that is, non-standard chemical components, and contain a large number of standard chemical components, resulting in a certain similarity between the spectral curve of the sensitizing drugs containing impurities and the spectral curve of normal sensitizing drugs; usually, for such spectral curves with large similarity, the difference that only occupies a small part cannot be well extracted, resulting in poor differentiation between sensitizing drugs containing impurities and normal sensitizing drugs, resulting in poor quality inspection results of sensitizing drugs based on similarity classification of spectral curves based on existing algorithms. Summary of the invention

[0003] The present invention provides a method for detecting gastrointestinal tumor radiotherapy sensitizing drugs based on spectral detection technology to solve the problem that the classification effect is poor and the detection result is affected by the similarity of spectral curves when performing impurity detection on sensitizing drugs in the prior art. The technical solution adopted is as follows:

[0004] One embodiment of the present invention provides a method for detecting gastrointestinal tumor radiosensitization drugs based on spectral detection technology, the method comprising the following steps:

[0005] Collecting spectral curves of a number of normal sensitizing drugs and sensitizing drugs containing impurities, and recording them as a number of qualified spectral curves and a number of impurity spectral curves respectively;

[0006] Classify all spectral curves to obtain base classifiers and several spectral categories; cluster the spectral curves under each spectral category according to the similarity of the spectral curves to obtain several spectral clusters of each spectral category; obtain several spectral feature vectors and spectral feature values ​​of each spectral cluster of each spectral category according to the spectral curves in each spectral cluster;

[0007] According to the spectral feature vectors and spectral feature values ​​of the spectral clusters, clusters of different spectral categories are matched to obtain a number of matching cluster pairs; according to the spectral feature vectors and spectral feature values ​​of the spectral clusters in the matching cluster pairs, a number of difference cluster pairs and the difference coefficient of each spectral curve in each spectral cluster in the difference cluster pairs are obtained; based on the difference coefficient, the weight of the base classifier is updated to obtain an integrated classifier for distinguishing spectral curves;

[0008] The spectral curve of the sensitizing drug to be tested is obtained, and the quality of the sensitizing drug to be tested is tested according to the spectral curve of the sensitizing drug to be tested and through an integrated classifier.

[0009] Optionally, the obtaining of the base classifier and the plurality of spectral categories comprises a specific method as follows:

[0010] The decision tree is used as the type of base classifier in adaboost, and all spectral curves are classified by the base classifier to obtain a number of spectral curves belonging to positive samples and a number of spectral curves belonging to negative samples;

[0011] All qualified spectral curves in the positive sample are grouped into a qualified positive sample category, and all impurity spectral curves in the positive sample are grouped into an unqualified positive sample category; all qualified spectral curves in the negative sample are grouped into a qualified negative sample category, and all impurity spectral curves in the negative sample are grouped into an unqualified negative sample category;

[0012] The positive sample qualified category, the positive sample unqualified category, the negative sample qualified category and the negative sample unqualified category are all recorded as spectral categories.

[0013] Optionally, the obtaining of a plurality of spectrum clusters of each spectrum category includes the following specific methods:

[0014] For any spectral category, the spectral category contains several spectral curves, and the cosine similarity is calculated for any two spectral curves. The difference obtained by subtracting the cosine similarity from 1 is taken as the difference distance between the two spectral curves; all spectral curves of the spectral category are density clustered, and the distance measurement uses the difference distance between the spectral curves to obtain several clusters, which are recorded as several spectral clusters of the spectral category.

[0015] Optionally, the method of obtaining a plurality of spectral feature vectors and spectral feature values ​​of each spectral category and each spectral cluster includes the following specific methods:

[0016] For any spectral cluster, each spectral curve in the spectral cluster is converted into a vector form to obtain a spectral vector of each spectral curve, and the spectral vectors are arranged row by row to obtain a spectral matrix of the spectral cluster; the spectral matrix is ​​decomposed by SVD to obtain a left singular matrix and a right singular matrix;

[0017] According to the right singular matrix obtained from the spectral matrix of the spectral cluster, a plurality of spectral feature vectors and spectral feature values ​​of the spectral cluster are obtained.

[0018] Optionally, the obtaining of the plurality of spectral feature vectors and spectral feature values ​​of the spectral cluster includes the following specific methods:

[0019] Each column element in the right singular matrix is ​​extracted as a right singular vector, all right singular vectors are taken as eigenvectors, recorded as several spectral eigenvectors of the spectral cluster, and the eigenvalue corresponding to each spectral eigenvector is taken as several eigenvalues ​​of the spectral cluster.

[0020] Optionally, the obtaining of a plurality of matching cluster pairs includes the following specific methods:

[0021] The qualified category of positive samples and the qualified category of negative samples are taken as the qualified sample group. For the qualified sample group, all the spectral feature vectors of any spectral cluster in the qualified category of positive samples are taken as the left nodes of the cluster bipartite graph, and all the spectral feature vectors of any spectral cluster in the qualified category of negative samples are taken as the right nodes of the cluster bipartite graph. The cosine similarity of the spectral feature vectors corresponding to the nodes on the left and right sides of the cluster bipartite graph is taken as the edge value between the nodes on the left and right sides, and the cluster bipartite graph is subjected to KM matching, and the obtained KM matching value is taken as the matching degree of the two spectral clusters.

[0022] According to the matching degree of the two spectral clusters, several matching cluster pairs are obtained through KM matching.

[0023] Optionally, the method of obtaining a plurality of matching cluster pairs through KM matching includes:

[0024] All spectral clusters in the qualified category of positive samples are taken as the left nodes of the category bipartite graph, and all spectral clusters in the qualified category of negative samples are taken as the right nodes of the category bipartite graph. The matching degree of the spectral clusters corresponding to the nodes on the left and right sides of the category bipartite graph is taken as the edge value between the nodes on the left and right sides. KM matching is performed on the category bipartite graph to obtain several matching spectral cluster pairs, which are recorded as several matching cluster pairs.

[0025] Optionally, the obtaining of a plurality of difference cluster pairs and a difference coefficient of each spectral curve in each spectral cluster in the difference cluster pairs includes the following specific methods:

[0026] For any matching cluster pair, the cluster bipartite graph corresponding to the matching cluster pair obtains several matching spectral feature vector pairs during the KM matching process, and the spectral feature values ​​corresponding to the matching spectral feature vectors are taken as a matching feature value pair to obtain several matching feature value pairs of the matching cluster pair; for any matching feature value pair, the absolute value of the difference between the two spectral feature values ​​in the matching feature value pair is calculated and recorded as the matching difference of the matching feature value pair;

[0027] Obtain the matching difference of each matching feature value pair of the matching cluster pair, calculate the mean of all matching differences, record the matching feature value pairs whose matching differences are greater than the mean as the difference feature value pairs of the matching cluster pair; record the mean of all difference feature value pairs as the feature difference of the matching cluster pair;

[0028] According to a number of matching feature value pairs of each matching cluster pair, a matching difference of each matching feature value pair, a number of difference feature value pairs and feature differences, a difference coefficient of each spectral curve in the difference cluster pair and each spectral cluster in the difference cluster pair is obtained.

[0029] Optionally, the specific method of obtaining the difference cluster pair and the difference coefficient of each spectral curve in each spectral cluster in the difference cluster pair includes:

[0030] For any matching cluster pair, the ratio of the feature difference of the matching cluster pair to the maximum value of the matching differences of all matching feature value pairs of all matching cluster pairs is recorded as the difference coefficient of the matching cluster pair;

[0031] If the difference coefficient of the matching cluster pair is greater than the difference threshold, the matching cluster pair is recorded as a difference cluster pair, and the difference coefficient of the difference cluster pair is used as the difference coefficient of each spectral curve in the two spectral clusters in the difference cluster pair.

[0032] Optionally, the method of updating the weights of the base classifiers based on the difference coefficients to obtain an integrated classifier for distinguishing spectral curves includes the following specific methods:

[0033] In the process of updating the weights of the base classifier, for any spectral curve of a spectral cluster belonging to a difference category pair, if the spectral curve belongs to a positive sample qualified category or a positive sample unqualified category, the difference coefficient is added to the negative exponential part of the natural constant in the calculation formula for updating the weights;

[0034] If the spectral curve belongs to the negative sample qualified category or the negative sample unqualified category, the difference coefficient is added to the exponential part of the natural constant in the calculation formula for updating the weight;

[0035] After the base classifier is updated, all spectral curves are continuously classified to obtain the difference cluster pairs and the difference coefficients of several spectral curves in the spectral clusters. The updated base classifier is continuously updated, and the iterative update is performed in this way. The final classifier is recorded as an integrated classifier for distinguishing spectral curves.

[0036] The beneficial effects of the present invention are as follows: the present invention obtains spectral curves of normal sensitizing drugs and sensitizing drugs containing impurities, and classifies them based on a base classifier to obtain several spectral categories, and clusters them to obtain spectral clusters, extracts spectral feature vectors and spectral feature values ​​of spectral clusters, and then performs matching analysis between spectral clusters based on spectral curves of the same type, thereby quantifying difference cluster pairs and obtaining difference coefficients of spectral curves, iteratively updating to obtain an integrated classifier, and performing detection through the integrated classifier, and finally completing quality detection of the sensitizing drugs to be tested, thereby realizing quality detection of sensitizing drugs for gastrointestinal tumor radiotherapy through spectral detection. Among them, qualified spectral curves and impurity spectral curves are classified based on the base classifier to obtain several spectral categories, and spectral curves in the same spectral category are clustered to obtain spectral clusters and perform singular value decomposition, so as to obtain spectral feature vectors and spectral feature values ​​of spectral clusters, which provide a basis for subsequent matching based on spectral feature vectors and spectral feature values, so as to match spectral clusters of different spectral categories and extract features that cause misclassification in negative sample categories; by matching spectral clusters corresponding to spectral curves of different spectral categories but the same type, matching cluster pairs are obtained, and the difference features and common features that cause classification errors are extracted through the matching cluster pairs; then, the correlation quantification of the difference features between spectral clusters is performed based on the matching cluster pairs, so as to reflect the proportion of the difference features between spectral clusters in the difference cluster pairs, and then obtain the difference coefficient, which can numerically reflect the difference features that cause classification errors; the base classifier is updated based on the difference coefficient, and finally an integrated classifier is obtained, and the integrated classifier can solve the problem of easy classification errors when the impurity spectral curve and the qualified spectral curve are similar to each other to the greatest extent. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0038] Figure 1 A schematic flow chart of a method for detecting gastrointestinal tumor radiosensitizing drugs based on spectral detection technology provided by one embodiment of the present invention. DETAILED DESCRIPTION

[0039] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0040] See also Figure 1 , which shows a flow chart of a method for detecting gastrointestinal tumor radiosensitization drugs based on spectral detection technology provided by an embodiment of the present invention, the method comprising the following steps:

[0041] Step S001, collecting spectral curves of a number of normal sensitizing drugs and sensitizing drugs containing impurities, and recording them as a number of qualified spectral curves and a number of impurity spectral curves respectively.

[0042] The purpose of this embodiment is to obtain a spectral curve of the sensitizing drug by a spectrometer, and to perform quality detection of the sensitizing drug to be tested by analyzing the spectral curve. Therefore, it is necessary to first collect spectral curves of several normal sensitizing drugs and sensitizing drugs containing impurities.

[0043] Specifically, in this embodiment, 500 spectral curves of normal sensitizing drugs and 100 spectral curves of sensitizing drugs containing impurities are collected by a spectrometer; the spectral curves of normal sensitizing drugs are recorded as qualified spectral curves, and several qualified spectral curves are obtained; the spectral curves of sensitizing drugs containing impurities are recorded as impurity spectral curves, and several impurity spectral curves are obtained.

[0044] Step S002: classify all spectral curves to obtain base classifiers and several spectral categories; cluster the spectral curves under each spectral category according to the similarity of the spectral curves to obtain several spectral clusters of each spectral category; and obtain several spectral feature vectors and spectral feature values ​​of each spectral cluster of each spectral category according to the spectral curves in each spectral cluster.

[0045] It should be noted that by comparing the correctly classified 0-1 (indicating that the current classifier has learned the distinguishing features) and the incorrectly classified 0-1 (no distinguishing features have been learned) samples of each base classifier, when the sample weights are updated next time, the more obvious the features are in the correctly classified 0-1 samples, and the less obvious the features are in the incorrectly classified 0-1 samples. The positive samples (correctly classified samples, i.e., 1 samples) and negative samples (incorrectly classified samples, i.e., 0 samples) corresponding to the same features, that is, samples with the same features but large differences in the features, are obtained. As a result, the degree of reduction in the attention paid to the positive samples that meet the conditions is reduced, and the degree of attention paid to the negative samples that meet the conditions is further increased. This allows the final integrated classifier to achieve a higher recognition accuracy for the current situation, that is, when the impurity spectral curve and the qualified spectral curve are relatively similar.

[0046] Preferably, in one embodiment of the present invention, all spectral curves are classified to obtain a base classifier and several spectral categories, including the following specific method:

[0047] The decision tree is used as the type of base classifier in adaboost, and all spectral curves are classified by the base classifier to obtain several spectral curves belonging to positive samples and several spectral curves belonging to negative samples; the positive sample is a correctly classified spectral curve, all qualified spectral curves in the positive sample are combined into a positive sample qualified category, and all impurity spectral curves in the positive sample are combined into a positive sample unqualified category; the negative sample is an incorrectly classified spectral curve, all qualified spectral curves in the negative sample are combined into a negative sample qualified category, and all impurity spectral curves in the negative sample are combined into a negative sample unqualified category. It should be noted that the classification error means that the spectral curve itself is a qualified spectral curve but is mistakenly classified as an impurity spectral curve; or the spectral curve itself is an impurity spectral curve but is mistakenly classified as a qualified spectral curve; adaboost and the base classifier are both existing algorithms, and will not be repeated in this embodiment; the positive sample qualified category, the positive sample unqualified category, the negative sample qualified category and the negative sample unqualified category are all recorded as spectral categories, and a total of four spectral categories are obtained.

[0048] What needs to be explained is that it is necessary to extract the difference features of positive and negative samples, that is, the positive samples (the positive samples contain qualified spectral curves and impurity spectral curves) are classified correctly because some features of the qualified spectral curves in the positive samples are more obvious and are better learned by the classifier, while the misclassified negative samples (for example, qualified spectral curves are mistakenly classified as impurity spectral curves, or impurity spectral curves are mistakenly classified as qualified spectral curves) have weaker feature intensity on this feature, that is, this feature is less obvious.

[0049] Preferably, in one embodiment of the present invention, the spectral curves under each spectral category are clustered according to the similarity of the spectral curves to obtain a number of spectral clusters for each spectral category, including the specific method of:

[0050] For any spectral category, the spectral category contains several spectral curves, and the cosine similarity is calculated for any two spectral curves, and the difference obtained by subtracting the cosine similarity from 1 is used as the difference distance between the two spectral curves; all spectral curves of the spectral category are density clustered, and the density clustering in this embodiment adopts DBSCAN clustering, and the distance metric adopts the difference distance between the spectral curves to obtain several clusters, which are recorded as several spectral clusters of the spectral category; according to the above method, the spectral curves of each spectral category are density clustered respectively to obtain several spectral clusters of each spectral category.

[0051] It should be noted that, for a spectral category, it is necessary to analyze the common features of the spectral curves therein, so as to adjust the base classifier based on the common features, so that after the sample weights are updated, the classification accuracy can be improved based on the extracted common features; therefore, it is necessary to perform cluster analysis on the spectral curves in the same spectral category and extract common features based on spectral clusters.

[0052] Preferably, in one embodiment of the present invention, a number of spectral feature vectors and spectral feature values ​​of each spectral cluster of each spectral category are obtained according to the spectral curve in each spectral cluster, including the specific method of:

[0053] For any spectral cluster, each spectral curve in the spectral cluster is converted into vector form to obtain the spectral vector of each spectral curve. The spectral vectors are arranged row by row to obtain the spectral matrix of the spectral cluster. Any row of elements in the spectral matrix from left to right is a spectral vector of a spectral curve. The spectral matrix is ​​decomposed by SVD to obtain the left singular matrix and the right singular matrix.

[0054] It should be noted that SVD decomposition is singular value decomposition, which is an existing method and will not be described in detail in this embodiment; each column element in the left singular matrix is ​​a left singular vector, which reflects the changes in the column space of the spectral matrix; each column element in the right singular matrix is ​​a right singular vector, which reflects the changes in the row space of the spectral matrix; and one row element in the spectral matrix is ​​a spectral vector, so it is necessary to extract the right singular vector as a feature vector to reflect the characteristics of the spectral vector.

[0055] Specifically, extract each column element in the right singular matrix as a right singular vector, take all right singular vectors as eigenvectors, record them as a number of spectral eigenvectors of the spectral cluster, and take the eigenvalues ​​corresponding to each spectral eigenvector as a number of eigenvalues ​​of the spectral cluster; it should be noted that the process of obtaining eigenvectors and their corresponding eigenvalues ​​through SVD decomposition is an existing process in singular value decomposition, which will not be repeated in this embodiment; obtain a number of spectral eigenvectors and spectral eigenvalues ​​of each spectral cluster according to the above method.

[0056] At this point, by classifying the qualified spectral curves and the impurity spectral curves based on the base classifier, several spectral categories are obtained, and the spectral curves in the same spectral category are clustered to obtain spectral clusters and perform singular value decomposition, thereby obtaining the spectral feature vectors and spectral feature values ​​of the spectral clusters, which provide a basis for subsequent matching based on spectral feature vectors and spectral feature values, thereby matching spectral clusters of different spectral categories and extracting the features that cause misclassification in the negative sample category.

[0057] Step S003, according to the spectral feature vectors and spectral feature values ​​of the spectral clusters, clusters of different spectral categories are matched to obtain a number of matching cluster pairs; according to the spectral feature vectors and spectral feature values ​​of the spectral clusters in the matching cluster pairs, a number of difference cluster pairs and the difference coefficient of each spectral curve in each spectral cluster in the difference cluster pairs are obtained; based on the difference coefficient, the weight of the base classifier is updated to obtain an integrated classifier for distinguishing spectral curves.

[0058] Preferably, in one embodiment of the present invention, clusters of different spectral categories are matched according to the spectral feature vectors and spectral feature values ​​of the spectral clusters to obtain a number of matching cluster pairs, including the specific method of:

[0059] The qualified category of positive samples and the qualified category of negative samples are taken as the qualified sample group, and the unqualified category of positive samples and the unqualified category of negative samples are taken as the unqualified sample group, thus obtaining two sample groups; taking the qualified sample group as an example, all spectral feature vectors of any spectral cluster in the qualified category of positive samples are taken as the left nodes of the cluster bipartite graph, and all spectral feature vectors of any spectral cluster in the qualified category of negative samples are taken as the right nodes of the cluster bipartite graph, and the cosine similarity of the spectral feature vectors corresponding to the nodes on the left and right sides of the cluster bipartite graph is taken as the edge value between the nodes on the left and right sides, and the cluster bipartite graph is subjected to KM matching, and the obtained KM matching value is taken as the matching degree of the two spectral clusters; according to the above method, the matching degree of each spectral cluster in the qualified category of positive samples and each spectral cluster in the qualified category of negative samples is obtained.

[0060] Furthermore, all spectral clusters in the qualified category of positive samples are used as the left nodes of the category bipartite graph, and all spectral clusters in the qualified category of negative samples are used as the right nodes of the category bipartite graph. The matching degree of the spectral clusters corresponding to the nodes on the left and right sides of the category bipartite graph is used as the edge value between the left and right nodes. KM matching is performed on the category bipartite graph to obtain several matching spectral cluster pairs, which are recorded as several matching cluster pairs, and then several matching cluster pairs of the qualified sample group are obtained; several matching cluster pairs of the unqualified sample group are obtained according to the above method; it should be noted that a matching cluster pair contains two spectral clusters of different spectral categories in the same sample group.

[0061] It should be noted that by dividing the spectral categories into sample groups and matching the spectral clusters of different spectral categories based on the sample groups, the obtained matching cluster pairs can show common characteristics and difference characteristics, and the difference characteristics are the characteristics that are more likely to cause misclassification, providing a basis for the subsequent acquisition of difference cluster pairs.

[0062] Preferably, in one embodiment of the present invention, a plurality of difference cluster pairs and a difference coefficient of each spectral curve in each spectral cluster in the difference cluster pairs are obtained according to the spectral feature vectors and spectral feature values ​​of the spectral clusters in the matching cluster pairs, including the specific method of:

[0063] For any matching cluster pair, the cluster bipartite graph corresponding to the matching cluster pair obtains several matching spectral feature vector pairs during the KM matching process, and the spectral feature values ​​corresponding to each of the matching spectral feature vectors are taken as a matching feature value pair, thus obtaining several matching feature value pairs of the matching cluster pair; for any matching feature value pair, the absolute value of the difference between the two spectral feature values ​​in the matching feature value pair is calculated, and recorded as the matching difference of the matching feature value pair; the matching difference of each matching feature value pair of the matching cluster pair is obtained according to the above method, the mean of all matching differences is calculated, and the matching feature value pairs whose matching differences are greater than the mean are recorded as the difference feature value pairs of the matching cluster pair; the mean of all difference feature value pairs is recorded as the feature difference of the matching cluster pair; according to the above method, several matching feature value pairs, matching differences of each matching feature value pair, several difference feature value pairs and feature differences are obtained for each matching cluster pair.

[0064] Furthermore, for any matching cluster pair, the ratio of the feature difference of the matching cluster pair to the maximum value of the matching differences of all matching feature value pairs of all matching cluster pairs is recorded as the difference coefficient of the matching cluster pair; a difference threshold is preset, and the difference threshold in this embodiment is described as 0.7; if the difference coefficient of the matching cluster pair is greater than the difference threshold, the matching cluster pair is recorded as a difference cluster pair, and the difference coefficient of the difference cluster pair is used as the difference coefficient of each spectral curve in the two spectral clusters in the difference cluster pair; if the difference coefficient of the matching cluster pair is less than or equal to the difference threshold, it is not regarded as a difference category pair; then several difference category pairs and the difference coefficient of each spectral curve in each spectral cluster in the difference cluster pair are obtained.

[0065] It should be noted that the difference characteristics between the spectral clusters in the difference cluster pairs are more obvious, and the corresponding difference coefficients are larger. Therefore, it is necessary to make adjustments based on the difference coefficients when updating the weights to make the difference feature judgment more accurate and thus improve the accuracy of the classification effect.

[0066] Preferably, in one embodiment of the present invention, the weights of the base classifiers are updated based on the difference coefficients to obtain an integrated classifier for distinguishing spectral curves, including the following specific methods:

[0067] In the process of updating the weights of the base classifier, for any spectral curve belonging to a difference category pair of the spectral cluster to which it belongs, that is, a spectral curve with a difference coefficient, if the spectral curve belongs to a qualified positive sample category or an unqualified positive sample category, that is, the spectral curve is correctly classified as a positive sample, then the negative exponential part of the natural constant in the calculation formula for updating the weight is added with the difference coefficient, that is, the negative exponential part of the natural constant in the calculation formula is increased, then the exponential part is reduced, and the updated weight is reduced; if the spectral curve belongs to a qualified negative sample category or an unqualified negative sample category, that is, the spectral curve is incorrectly classified as a negative sample, then the exponential part of the natural constant in the calculation formula for updating the weight is added with the difference coefficient, that is, the exponential part of the natural constant in the calculation formula is increased, and the updated weight is increased; wherein the calculation formula for updating the weight is the prior art in the adaboost algorithm, and this embodiment will not be repeated here.

[0068] Furthermore, the base classifier is updated according to the above method. After the update is completed, all spectral curves are classified according to the above method to re-obtain the difference coefficients of the difference cluster pairs and several spectral curves in the spectral clusters therein, and the updated base classifier is continued to be updated. The iterative update is performed in this way, and the final classifier is recorded as an integrated classifier for distinguishing spectral curves.

[0069] At this point, by matching the spectral clusters corresponding to spectral curves of different spectral categories but the same type, matching cluster pairs are obtained, and the difference features and common features that lead to classification errors are extracted through the matching cluster pairs; then, the correlation quantification of the difference features between the spectral clusters is performed based on the matching cluster pairs, thereby reflecting the proportion of the difference features between the spectral clusters in the difference cluster pairs, and then the difference coefficient is obtained, which can numerically reflect the difference features that lead to classification errors; the base classifier is updated based on the difference coefficient, and finally an integrated classifier is obtained, which can solve the problem of easy classification errors when the impurity spectral curve and the qualified spectral curve are similar to each other.

[0070] Step S004: obtaining the spectral curve of the sensitizing drug to be tested, and performing quality detection on the sensitizing drug to be tested according to the spectral curve of the sensitizing drug to be tested and through an integrated classifier.

[0071] Preferably, in one embodiment of the present invention, the spectral curve of the sensitizing drug to be tested is obtained, and the quality of the sensitizing drug to be tested is detected according to the spectral curve of the sensitizing drug to be tested and through an integrated classifier, and the specific method includes:

[0072] After obtaining the integrated classifier, the sensitizing drug to be tested is recorded as the sensitizing drug to be tested, and the spectral curve of the sensitizing drug to be tested is obtained through a spectrometer; the spectral curve of the sensitizing drug to be tested is input into the integrated classifier, and the test result is output. If the test result is 1, the sensitizing drug to be tested is a normal sensitizing drug; if the test result is 0, the sensitizing drug to be tested contains impurities and the quality of the sensitizing drug to be tested is unqualified.

[0073] It should be noted that the final integrated classifier can effectively reduce the probability of misclassification of whether the sensitizing drug corresponding to the spectral curve contains impurities. At the same time, the output detection result is 1 corresponding to the spectral curve of the normal sensitizing drug, and the detection result is 0 corresponding to the spectral curve containing impurities. Therefore, the spectral curve of the sensitizing drug to be tested is obtained and tested by the integrated classifier, and finally the quality inspection of the sensitizing drug to be tested is completed, thereby realizing the quality inspection of sensitizing drugs for gastrointestinal tumor radiotherapy through spectral detection.

[0074] At this point, this embodiment is completed.

[0075] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A method for detecting gastrointestinal tumor radiosensitization drugs based on spectral detection technology, characterized in that: The method comprises the following steps: Collecting spectral curves of a number of normal sensitizing drugs and sensitizing drugs containing impurities, and recording them as a number of qualified spectral curves and a number of impurity spectral curves respectively; Classify all spectral curves to obtain base classifiers and several spectral categories; cluster the spectral curves under each spectral category according to the similarity of the spectral curves to obtain several spectral clusters of each spectral category; obtain several spectral feature vectors and spectral feature values ​​of each spectral cluster of each spectral category according to the spectral curves in each spectral cluster; According to the spectral feature vectors and spectral feature values ​​of the spectral clusters, clusters of different spectral categories are matched to obtain a number of matching cluster pairs; according to the spectral feature vectors and spectral feature values ​​of the spectral clusters in the matching cluster pairs, a number of difference cluster pairs and the difference coefficient of each spectral curve in each spectral cluster in the difference cluster pairs are obtained; based on the difference coefficient, the weight of the base classifier is updated to obtain an integrated classifier for distinguishing spectral curves; Acquire the spectral curve of the sensitizing drug to be tested, and perform quality detection on the sensitizing drug to be tested according to the spectral curve of the sensitizing drug to be tested and through an integrated classifier; The specific method of obtaining a plurality of matching cluster pairs includes: The qualified category of positive samples and the qualified category of negative samples are taken as the qualified sample group. For the qualified sample group, all the spectral feature vectors of any spectral cluster in the qualified category of positive samples are taken as the left nodes of the cluster bipartite graph, and all the spectral feature vectors of any spectral cluster in the qualified category of negative samples are taken as the right nodes of the cluster bipartite graph. The cosine similarity of the spectral feature vectors corresponding to the nodes on the left and right sides of the cluster bipartite graph is taken as the edge value between the nodes on the left and right sides, and the cluster bipartite graph is subjected to KM matching, and the obtained KM matching value is taken as the matching degree of the two spectral clusters. According to the matching degree of the two spectral clusters, a number of matching cluster pairs are obtained through KM matching; The specific method of obtaining a plurality of difference cluster pairs and a difference coefficient of each spectral curve in each spectral cluster in the difference cluster pairs includes: For any matching cluster pair, the cluster bipartite graph corresponding to the matching cluster pair obtains several matching spectral feature vector pairs during the KM matching process, and the spectral feature values ​​corresponding to the matching spectral feature vectors are taken as a matching feature value pair to obtain several matching feature value pairs of the matching cluster pair; for any matching feature value pair, the absolute value of the difference between the two spectral feature values ​​in the matching feature value pair is calculated and recorded as the matching difference of the matching feature value pair; Obtain the matching difference of each matching feature value pair of the matching cluster pair, calculate the mean of all matching differences, record the matching feature value pairs whose matching differences are greater than the mean as the difference feature value pairs of the matching cluster pair; record the mean of all difference feature value pairs as the feature difference of the matching cluster pair; According to a number of matching feature value pairs of each matching cluster pair, a matching difference of each matching feature value pair, a number of difference feature value pairs and feature differences, a difference coefficient of each spectral curve in the difference cluster pair and each spectral cluster in the difference cluster pair is obtained.

2. The method for detecting gastrointestinal tumor radiosensitization drugs based on spectral detection technology according to claim 1, characterized in that: The specific method of obtaining the base classifier and the plurality of spectral categories includes: The decision tree is used as the type of base classifier in adaboost, and all spectral curves are classified by the base classifier to obtain a number of spectral curves belonging to positive samples and a number of spectral curves belonging to negative samples; All qualified spectral curves in the positive sample are grouped into a qualified positive sample category, and all impurity spectral curves in the positive sample are grouped into an unqualified positive sample category; all qualified spectral curves in the negative sample are grouped into a qualified negative sample category, and all impurity spectral curves in the negative sample are grouped into an unqualified negative sample category; The positive sample qualified category, the positive sample unqualified category, the negative sample qualified category and the negative sample unqualified category are all recorded as spectral categories.

3. The method for detecting gastrointestinal tumor radiosensitization drugs based on spectral detection technology according to claim 1, characterized in that: The specific method of obtaining a plurality of spectral clusters of each spectral category includes: For any spectral category, the spectral category contains several spectral curves, and the cosine similarity is calculated for any two spectral curves. The difference obtained by subtracting the cosine similarity from 1 is taken as the difference distance between the two spectral curves; all spectral curves of the spectral category are density clustered, and the distance measurement uses the difference distance between the spectral curves to obtain several clusters, which are recorded as several spectral clusters of the spectral category.

4. The method for detecting gastrointestinal tumor radiosensitization drugs based on spectral detection technology according to claim 1, characterized in that: The specific method of obtaining a plurality of spectral feature vectors and spectral feature values ​​of each spectral category and each spectral cluster includes: For any spectral cluster, each spectral curve in the spectral cluster is converted into a vector form to obtain a spectral vector of each spectral curve, and the spectral vectors are arranged row by row to obtain a spectral matrix of the spectral cluster; the spectral matrix is ​​decomposed by SVD to obtain a left singular matrix and a right singular matrix; According to the right singular matrix obtained from the spectral matrix of the spectral cluster, a plurality of spectral feature vectors and spectral feature values ​​of the spectral cluster are obtained.

5. The method for detecting gastrointestinal tumor radiosensitization drugs based on spectral detection technology according to claim 4, characterized in that: The specific method of obtaining the plurality of spectral feature vectors and spectral feature values ​​of the spectral cluster includes: Each column element in the right singular matrix is ​​extracted as a right singular vector, all right singular vectors are taken as eigenvectors, recorded as several spectral eigenvectors of the spectral cluster, and the eigenvalue corresponding to each spectral eigenvector is taken as several eigenvalues ​​of the spectral cluster.

6. The method for detecting gastrointestinal tumor radiosensitization drugs based on spectral detection technology according to claim 1, characterized in that: The specific method of obtaining a plurality of matching cluster pairs through KM matching is as follows: All spectral clusters in the qualified category of positive samples are taken as the left nodes of the category bipartite graph, and all spectral clusters in the qualified category of negative samples are taken as the right nodes of the category bipartite graph. The matching degree of the spectral clusters corresponding to the nodes on the left and right sides of the category bipartite graph is taken as the edge value between the nodes on the left and right sides. KM matching is performed on the category bipartite graph to obtain several matching spectral cluster pairs, which are recorded as several matching cluster pairs.

7. The method for detecting gastrointestinal tumor radiosensitization drugs based on spectral detection technology according to claim 1, characterized in that: The specific method of obtaining the difference cluster pairs and the difference coefficient of each spectral curve in each spectral cluster in the difference cluster pairs includes: For any matching cluster pair, the ratio of the feature difference of the matching cluster pair to the maximum value of the matching differences of all matching feature value pairs of all matching cluster pairs is recorded as the difference coefficient of the matching cluster pair; If the difference coefficient of the matching cluster pair is greater than the difference threshold, the matching cluster pair is recorded as a difference cluster pair, and the difference coefficient of the difference cluster pair is used as the difference coefficient of each spectral curve in the two spectral clusters in the difference cluster pair.

8. The method for detecting gastrointestinal tumor radiosensitization drugs based on spectral detection technology according to claim 1, characterized in that: The method of updating the weight of the base classifier based on the difference coefficient to obtain an integrated classifier for distinguishing spectral curves includes the following specific methods: In the process of updating the weights of the base classifier, for any spectral curve of a spectral cluster belonging to a difference category pair, if the spectral curve belongs to a positive sample qualified category or a positive sample unqualified category, the difference coefficient is added to the negative exponential part of the natural constant in the calculation formula for updating the weights; If the spectral curve belongs to the negative sample qualified category or the negative sample unqualified category, the difference coefficient is added to the exponential part of the natural constant in the calculation formula for updating the weight; After the base classifier is updated, all spectral curves are continuously classified to obtain the difference cluster pairs and the difference coefficients of several spectral curves in the spectral clusters. The updated base classifier is continuously updated, and the iterative update is performed in this way. The final classifier is recorded as an integrated classifier for distinguishing spectral curves.

Images