A tumor cell section image recognition method, medium and electronic device

Abstract

The present application relates to a kind of tumor cell section image recognition method, medium and electronic equipment, using hyperspectral imaging system to shoot tumor cell section image, obtain the sample image data of tumor cell and its corresponding spectral information, since tumor cell and normal cell have certain difference in protein, nucleic acid, moisture, structure etc., leading to its spectral curve difference.Through the analysis of the size, shape and spectral curve of cell, accurately determine whether there is tumor cell in cell section, this process does not need artificial manual screening pathological image, effectively improve the image screening efficiency, avoid the problem of artificial missed detection error, simultaneously, hyperspectral image data include spectral information, effectively retain the characteristic information of tumor cell, facilitate using spectral information to identify tumor cell, improve resolution accuracy.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of hyperspectral images, in particular to a tumor cell section image recognition method, medium and electronic device. BACKGROUND

[0002] Malignant tumor has high incidence and high recurrence rate, and needs long-term and regular examination. The traditional tissue biopsy method needs to observe the cell section under a microscope, and since many medical images are generated every day, it is inevitable that there will be problems such as low efficiency and missed detection caused by human factors. At present, clinical diagnosis mainly relies on non-invasive imaging examinations such as color ultrasound, magnetic resonance imaging (MRI), computed tomography imaging (CT), and invasive endoscopic examination. In recent years, some scholars have used machine learning methods to identify and classify the above-mentioned tumor medical images, and have achieved certain results, but the above-mentioned image acquisition waveband is limited, and some characteristic information is inevitably lost. SUMMARY

[0003] In view of the above problems, the present application provides a tumor cell section image recognition method, medium and electronic device, which solves the problems of low efficiency of artificial screening of existing pathological images, and loss of characteristic information caused by limited acquisition waveband in machine learning method for identifying tumor medical images.

[0004] To achieve the above-mentioned purpose, in a first aspect, the present application provides a tumor cell section image recognition method, comprising the following steps:

[0005] Obtain sample image data of tumor cells and perform morphological processing thereon to extract a cell image to be analyzed, the sample image data being obtained by a hyperspectral imaging system, the cell image to be analyzed including any one or more of malignant tumor cell images, benign cell images and inflammatory cell images;

[0006] Perform spectral feature extraction on the cell image to be analyzed to obtain a spectral feature vector corresponding to the cell image to be analyzed;

[0007] Perform color matrix feature extraction on the cell image to be analyzed using the extracted spectral feature vector to obtain a color matrix feature vector corresponding to the cell image to be analyzed;

[0008] Fuse the spectral feature vector and the color feature vector to obtain a fused feature vector;

[0009] Classify the fused feature vector using an SVM classifier to obtain a cell image classification result;

[0010] Perform probability calculation and probability calibration based on the cell image classification result to output a cancer probability corresponding to the current tumor cell section image.

[0011] In some embodiments, the SVM-RFE model classifier is used for spectral feature extraction of the cell image to be analyzed, which includes the following steps:

[0012] The number of categories of the cell image to be analyzed is obtained, and the number of SVM-RFE model classifiers is determined according to the number of categories of the cell image to be analyzed. The number of SVM-RFE model classifiers is obtained according to formula (1), and formula (1) is as follows:

[0013] v = N(N-1) / 2,

[0014] where v is the number of SVM-RFE model classifiers, and v is a positive integer;

[0015] The cell image to be analyzed is input into the plurality of SVM-RFE models according to the two-by-two combination of categories, respectively, to obtain a plurality of ranking criterion scores;

[0016] The plurality of ranking criterion scores are summed to obtain a total ranking criterion score;

[0017] The spectral features are removed according to the total ranking criterion score, and the smallest spectral feature is iteratively removed until a preset condition is met, to obtain a final spectral feature set.

[0018] In some embodiments, the specific calculation process of the SVM-RFE model classifier is as follows:

[0019] Let the training sample be: where x i ∈R D , v i ∈{1,2,3}, N is the sample size, x i is the i-th sample, and D is the original sample feature number, i.e., the original spectral band number;

[0020] Initialize the original feature set: S = {1,2,…,D}, and the feature ranking set: R = [];

[0021] Generate three training samples, and find the two-by-two combination of different categories in the training sample T to obtain the final training sample: T j (j = 1,2,3);

[0022] Train the SVM classifier with T j respectively, and obtain the weight ω j (j = 1,2,3) respectively;

[0023] Calculate the ranking criterion score: Find the feature with the smallest ranking criterion score: f = argmin(c k ), and update the feature set R = [f,R];

[0024] Removing the minimum eigenvalues in S obtains a new feature set: S = S [1:f-1, f+1:lengh(S);

[0025] Looping until the number of features in S is k;

[0026] Obtaining a feature ranking list R, i.e. the order of features being eliminated.

[0027] In some embodiments, in the feature ranking list R, each to-be-analyzed cell image corresponds to a k-dimensional spectral feature.

[0028] In some embodiments, performing color matrix feature extraction on the to-be-analyzed cell image using the spectral feature comprises:

[0029] Determining the color components in the color matrix according to the spectral wavelengths corresponding to the spectral feature;

[0030] Extracting color information of low-order moments in the color matrix corresponding to the to-be-analyzed cell image;

[0031] Obtaining a color matrix feature vector according to the color information.

[0032] In some embodiments, the low-order moments include first-order moments, second-order moments, and third-order moments, and the extracting of the color information of the low-order moments in the color matrix corresponding to the to-be-analyzed cell image comprises:

[0033] Letting the k spectral wavelengths corresponding to the k-dimensional spectral feature be the k color components of the color matrix, M representing the number of pixels in the to-be-analyzed cell image, the first-order moment μ i is defined by formula (2), which is as follows:

[0034]

[0035] The second-order moment σ i is defined by formula (3), which is as follows:

[0036]

[0037] The third-order moment s i is defined by formula (4), which is as follows:

[0038]

[0039] wherein μ represents the mean value of all pixels on the i-th color channel, σ represents the standard deviation of all pixels on the i-th color channel, s represents the cubic root of the skewness of all pixels on the i-th channel, and p i,j represents the pixel value of the j-th pixel in the i-th color component (i = 1, 2, …, k);

[0040] The color matrix eigenvector is obtained by formula (5) as follows:

[0041] F color = [μ1, σ1, s1, …, μ k , σ k , s k ].

[0042] In some embodiments, classifying the fused feature vector using the SVM classifier comprises:

[0043] Three SVM classifiers are trained, and the constraint condition of the hyperplane objective function of the SVM classifiers is represented by formula (6) as follows:

[0044]

[0045] subject to y i [(wx i )+b]≥1-ζ i (i = 1, 2, …, N),

[0046] ζ i ≥ 0

[0047] where x i is each sample feature, y i is its label; N is the number of samples; w is the hyperplane normal vector, ζ i is a slack variable, which can allow the data points to deviate from the hyperplane to a certain extent, and C is a penalty factor, representing the importance of the loss brought by outliers.

[0048] In some embodiments, the probability calculation and the probability calibration are calculated using an isotonic regression algorithm, and the calculation process of the isotonic regression algorithm is represented by formula (7) as follows:

[0049]

[0050] where y i is the true label of sample i, is the output probability of sample i after calibration.

[0051] In a second aspect, the present application further provides a computer readable storage medium, which stores computer program instructions, and the computer program instructions, when executed by a processor, implement the method in the first aspect.

[0052] In a third aspect, the present application further provides an electronic device, which comprises a memory and a processor, and the memory is used to store one or more computer program instructions, wherein the one or more computer program instructions are executed by the processor to implement the method in the first aspect.

[0053] Differing from the prior art, the above technical solution uses a hyperspectral imaging system to shoot tumor cell slice images, obtains sample image data of tumor cells and spectrum information corresponding thereto, and since tumor cells and normal cells differ in protein, nucleic acid, moisture, structure, etc., the spectrum curves thereof differ. By analyzing the size, shape and spectrum curve of the cells, it is possible to accurately determine whether tumor cells exist in the cell slice, this process does not need manual screening of pathological images, effectively improves the image screening efficiency, avoids the problem of manual missed detection and wrong detection, meanwhile, the hyperspectral image data contains spectrum information, effectively retains the characteristic information of tumor cells, facilitates identification of tumor cells using spectrum information, and improves the correct identification rate.

[0054] The above invention content related description is only a summary of the technical solution of the present application, in order to enable those skilled in the art to more clearly understand the technical solution of the present application, and then can be implemented according to the content of the description and the drawings, and in order to let the above-mentioned purpose and other purposes, characteristics and advantages of the present application can be more easily understood, the following is described in combination with the specific embodiments of the present application and the drawings. BRIEF DESCRIPTION OF DRAWINGS

[0055] The drawings are only used to show the principles, implementation modes, applications, characteristics and effects of the specific embodiments of the present application and other related contents, and cannot be considered as a limitation of the present application.

[0056] In the drawings of the specification:

[0057] Figure 1 The flow chart of the tumor cell slice image recognition method according to a specific embodiment of the present application is shown;

[0058] Figure 2 The step chart of the tumor cell slice image recognition method according to the first exemplary embodiment of the present application is shown;

[0059] Figure 3 The step chart of the tumor cell slice image recognition method according to the second exemplary embodiment of the present application is shown;

[0060] Figure 4 The schematic diagram of the electronic device for tumor cell slice image recognition according to a specific embodiment of the present application is shown.

[0061] The reference signs involved in the above drawings are explained as follows: 1, electronic device, 11, memory, 12, processor. DETAILED DESCRIPTION

[0062] In order to explain possible application scenarios, technical principles, specific implementation schemes, and the purposes and effects of the present application in detail, the following will be described in detail in combination with the specific embodiments listed and the accompanying drawings. The embodiments described herein are only used to more clearly illustrate the technical solutions of the present application, and therefore only serve as examples, and cannot limit the protection scope of the present application.

[0063] In this paper, the term "embodiment" means that the specific features, structures or characteristics described in combination with the embodiment can be included in at least one embodiment of the present application. The term "embodiment" appearing at various places in the specification does not necessarily refer to the same embodiment, and does not particularly limit its independence or association with other embodiments. In principle, in the present application, as long as there is no technical contradiction or conflict, each technical feature mentioned in each embodiment can be combined in any way to form a corresponding implementable technical solution.

[0064] Unless otherwise defined, the meanings of the technical terms used herein are the same as those commonly understood by those skilled in the art to which the present application belongs; the use of related terms herein is only for the purpose of describing specific embodiments, and is not intended to limit the present application.

[0065] In the description of the present application, the phrase "and / or" is a description of the logical relationship between objects, which means that there can be three relationships, for example, A and / or B, which means that there are three cases: A exists, B exists, and A and B exist at the same time. In addition, the character " / " in this paper generally represents that the associated objects before and after are a kind of "or" logical relationship.

[0066] In the present application, terms such as "first" and "second" are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual quantity, primary and secondary or order relationship between the entities or operations.

[0067] Without more limitations, in the present application, the "includes", "contains", "has" or other similar open expressions used in the sentence are intended to cover non-exclusive inclusion, and these expressions do not exclude the presence of other elements in the process, method or product including the described elements, so that the process, method or product including a series of elements can not only include those limited elements, but also include other elements not explicitly listed, or also include elements inherent to such process, method or product.

[0068] As the same as the understanding in the "Examination Guidelines", in the present application, "greater than", "less than", "exceed" and the like are understood as not including the number; "above", "below", "within" and the like are understood as including the number. In addition, in the description of the embodiments of the present application, the meaning of "a plurality of" is two or more (including two), and similar expressions related to "a plurality of" are also understood in this way, for example, "a plurality of groups", "a plurality of times" and the like, unless otherwise explicitly and specifically limited.

[0069] Please refer to Figure 1 and Figure 2 In a first aspect, the present embodiment provides a tumor cell section image recognition method, comprising the following steps:

[0070] S11, obtain sample image data of tumor cells and perform morphological processing thereon to extract a to-be-analyzed cell image, the sample image data is obtained by a hyperspectral imaging system, and the to-be-analyzed cell image includes any one or more of a malignant tumor cell image, a benign cell image and an inflammatory cell image;

[0071] S12, perform spectral feature extraction on the to-be-analyzed cell image to obtain a spectral feature vector corresponding to the to-be-analyzed cell image;

[0072] S13, perform color matrix feature extraction on the to-be-analyzed cell image by using the extracted spectral feature vector to obtain a color matrix feature vector corresponding to the to-be-analyzed cell image;

[0073] S14, fuse the spectral feature vector and the color feature vector to obtain a fused feature vector;

[0074] S15, classify the fused feature vector by using an SVM classifier to obtain a cell image classification result;

[0075] S16, perform probability calculation and probability calibration based on the cell image classification result to output a cancer probability corresponding to the current tumor cell section image.

[0076] In the embodiment, the sample image data of tumor cells is obtained by a hyperspectral imaging system, the image collected by the hyperspectral imaging system has the characteristic of integration of image and spectrum, for example, the hyperspectral imaging system can simultaneously collect the hyperspectral information and image information of tumor cells, the hyperspectral information of tumor cells is analyzed by means of spectral curve to obtain the specific distribution range of tumor cells. Therefore, the sample image data collected by the hyperspectral imaging system combined with spectral curve analysis can quickly and accurately find the state of tumor cells in the slice sample without complex pre-experiment operation, saving manpower and resources. The commonly used non-invasive image such as color ultrasound, CT and MRI is used to extract and classify blood cells and tumors, which is a modeling analysis of image information of sample images. The hyperspectral imaging technology not only uses the image information of the sample, but also combines the spectral information of the sample, and the sample information obtained is more comprehensive.

[0077] The morphological processing specifically includes the following steps: threshold segmentation by OTSU maximum between-class variance method to obtain a gray image of the sample image data; applying a contour detection algorithm, an erosion algorithm and the like to remove the connected domains that do not meet the requirements and extract the cell image to be analyzed, the cell image to be analyzed including one or more of the malignant tumor cell image, the benign cell image and the inflammatory cell image.

[0078] Since the tumor cells and normal cells have certain differences in protein, nucleic acid, water, structure and the like, the spectral curves thereof are different. Therefore, the spectral features in the cell image to be analyzed are extracted by means of the spectral features contained in the cell image to be analyzed to obtain the spectral feature vector corresponding to the cell image to be analyzed. In the embodiment, the spectral features are hyperspectral data, and the number of wavebands of the hyperspectral data is large and the correlation between the wavebands is strong. In order to eliminate the redundant information between the wavebands, reduce the data amount and calculation time, the dimensionality of the hyperspectral data of the sample needs to be reduced.

[0079] Since the sizes of the malignant tumor cells, the benign cells and the inflammatory cells and the nuclear-cytoplasmic ratios thereof are very different, the spectral feature vector extracted after the dimensionality reduction is used to extract the color matrix features of the cell image to be analyzed to obtain the color matrix feature vector corresponding to the cell image to be analyzed, and the spectral feature vector and the color feature vector are fused to obtain a fusion feature vector. At this time, the fusion feature vector integrates the data of the color matrix feature vector and the spectral feature vector, so that the differences between the malignant tumor cells, the benign cells and the inflammatory cells are more obvious and are easier to distinguish, and therefore, the fusion feature vector can significantly identify the three types of cells.

[0080] The SVM classifier is used for classifying the fusion feature vector to obtain a cell image classification result. In this embodiment, the cell image classification result represents specific proportion values of three types of cells. Probability calculation and probability calibration are performed based on the cell image classification result to output a cancer probability corresponding to the current tumor cell slice image. Through probability calculation and probability calibration of the cell image classification result, early tumor cell classification accuracy probability values and classification effect evaluation of the sample image data obtained by the hyperspectral imaging are realized, and algorithm evaluation of the cancer probability of the patient is performed. The cancer probability prediction accuracy is improved, and the next step of investigation by relevant medical staff is facilitated.

[0081] The tumor cell slice image is photographed by using the hyperspectral imaging system to obtain sample image data of tumor cells and corresponding spectral information. Since tumor cells and normal cells differ in protein, nucleic acid, water, structure, etc., the spectral curves thereof differ. By analyzing the size, shape, and spectral curve of the cells, it is determined whether tumor cells exist in the cell slice. This process does not require manual screening of pathological images, effectively improves the image screening efficiency, avoids the problems of manual omission and error detection, and effectively retains the characteristic information of tumor cells in the hyperspectral image data, which facilitates the identification of tumor cells using spectral information and improves the correct identification rate.

[0082] In some embodiments, the SVM-RFE model classifier is used for spectral feature extraction of the cell image to be analyzed. The spectral feature extraction includes the following steps:

[0083] The number of categories of the cell image to be analyzed is obtained, and the number of SVM-RFE model classifiers is determined according to the number of categories of the cell image to be analyzed. The number of SVM-RFE model classifiers is obtained according to formula (1), and formula (1) is as follows:

[0084] v = N(N-1) / 2,

[0085] wherein v is the number of SVM-RFE model classifiers, and v is a positive integer;

[0086] The cell image to be analyzed is input into multiple SVM-RFE models according to the categories in pairs to obtain multiple ranking criterion scores.

[0087] The multiple ranking criterion scores are summed to obtain a total ranking criterion score.

[0088] The spectral features are removed according to the total ranking criterion score, and the smallest spectral feature is iteratively removed until a preset condition is met to obtain a final spectral feature set.

[0089] In the embodiment, RFE is the abbreviation of Recursive Feature Elimination, which is a feature selection method with good performance and strong generalization ability. It is a backward search method based on model, starting from all features in the current feature set, searching for feature subsets, and deleting the least important features until the desired number of features is reached.

[0090] In the embodiment, the preset condition can be a threshold of the ranking criterion total score, and the iteration is stopped when the elimination reaches the threshold of the ranking criterion total score. Alternatively, the preset condition can also be a limit on the number of spectral features, and the iteration is stopped when the elimination reaches the preset number of spectral features.

[0091] The final SVM multi-classification algorithm is trained multiple times by using the Recursive Feature Elimination (RFE) to reduce the feature dimension and eliminate the redundancy between features. Since the cells to be analyzed have three categories, i.e., malignant tumor cells, benign cells, and inflammatory cells, the SVM needs to pair the categories two by two when processing the multi-classification problem, one of which is the positive class and the other is the negative class. Thus, N(N-1) / 2 classifiers need to be trained, which becomes multiple two-class problems. Each two-class problem is classified by using an SVM-RFE model classifier to select features, multiple ranking criterion scores are obtained by using multiple SVM-RFE model classifiers, and then the ranking criterion total score is obtained by adding the multiple ranking criterion scores, which is used as the basis for feature elimination. The smallest feature is eliminated each time, and all features are deleted until the iteration is stopped. The SVM classification accuracy is used to evaluate the advantages and disadvantages of the features, so as to obtain the optimal feature subset. In the embodiment, the preset condition is a limit on the number of spectral features, and the number of spectral features is k. Therefore, after the iteration is stopped, the spectral feature set contains k features.

[0092] In the embodiment, the spectral features are spectral features extracted from hyperspectral data. Alternatively, the number of wavebands of the hyperspectral data is large, and the correlation between the wavebands is strong. In order to eliminate the redundancy between the wavebands, reduce the data amount and calculation time, the dimensionality of the hyperspectral data of the sample needs to be reduced. In the embodiment, the spectral features after dimensionality reduction by using the Recursive Feature Elimination (RFE) are fused with the color features of the sample image and input into the SVM classifier, so as to achieve the purpose of dimensionality reduction of the hyperspectral data.

[0093] In some embodiments, the specific calculation process of the SVM-RFE model classifier is as follows:

[0094] Suppose the training samples are: where x i ∈R D , vi ∈ {1,2,3}, N is the sample size, x i is the i-th sample, D is the original sample feature number, that is, the original spectral band number;

[0095] Initialize the original feature set: S = {1,2,…,D}, and the feature ranking set: R = [];

[0096] Generate three training samples, and find the two-way combination of different categories in the training sample T to obtain the final training sample: T j (j = 1,2,3);

[0097] Use T j Train the SVM classifier, respectively, to obtain the weight ω j (j = 1,2,3);

[0098] Calculate the ranking criterion score: Find the feature with the minimum ranking criterion score: f = argmin(c k ), update the feature set R = [f,R];

[0099] Remove the minimum feature value from S to obtain a new feature set: S = S[1:f-1,f+1:lengh(S);

[0100] Iterate until S = [k], indicating that the number of features in S is k.

[0101] Obtain the feature ranking list R, that is, the order in which the features are eliminated.

[0102] In this embodiment, since R is empty at the beginning, there is no value in it. Based on the traditional SVM classification model, the spectral features after dimensionality reduction by the recursive feature elimination method (RFE) are fused with the sample image color features and input into the SVM classifier in this embodiment, so as to achieve the purpose of reducing the dimensionality of hyperspectral data.

[0103] In some embodiments, in the feature ranking list R, each to-be-analyzed cell image corresponds to k-dimensional spectral features. The hyperspectral data after RFE dimensionality reduction has a total of k bands, that is, each sample obtains k-dimensional spectral features.

[0104] Please refer to Figure 3 In some embodiments, the color matrix feature extraction of the to-be-analyzed cell image by using the spectral features comprises:

[0105] S21, determining the color components in the color matrix according to the spectral wavelengths corresponding to the spectral features;

[0106] S22, extracting the color information of the low-order moments in the color matrix corresponding to the to-be-analyzed cell image;

[0107] S23, obtaining a color matrix feature vector according to the color information.

[0108] In the embodiment, the k characteristic wavelengths obtained by the RFE method are selected as k color components of the color moment, and the corresponding color information is extracted. Since the color information is mainly distributed in the low-order moment, the first moment (representing the mean value of all pixels on the i-th color channel), the second moment (representing the standard deviation of all pixels on the i-th color channel), and the third moment (representing the cubic root of the slope of all pixels on the i-th channel) are sufficient to express the color distribution of the image.

[0109] In some embodiments, the low-order moment includes the first moment, the second moment, and the third moment, and the color information of the low-order moment in the color matrix corresponding to the cell image to be analyzed includes:

[0110] Let the k spectral wavelengths corresponding to the k-dimensional spectral features be the k color components of the color matrix, M represents the number of pixels in the cell image to be analyzed, the first moment μ i The definition is represented by formula (2), and formula (2) is as follows:

[0111]

[0112] The second moment σ i The definition is represented by formula (3), and formula (3) is as follows:

[0113]

[0114] The third moment s i The definition is represented by formula (4), and formula (4) is as follows:

[0115]

[0116] Wherein, the first moment represents the mean value of all pixels on the i-th color channel, the second moment represents the standard deviation of all pixels on the i-th color channel, and the third moment represents the cubic root of the slope of all pixels on the i-th channel, p i,j The pixel value of the jth pixel in the i-th color component (i=1, 2, …, k);

[0117] The color matrix feature vector is obtained by formula (5), and formula (5) is as follows:

[0118] F color =[μ1,σ1,s1,…,μ k ,σ k ,s k ].

[0119] In the embodiment, the color matrix feature vector is a 3k-dimensional color feature vector, the obtained spectral feature vector is fused with the color feature vector, and a 4k-dimensional fusion feature vector is obtained. At this time, the fusion feature vector integrates the data of the color matrix feature vector and the spectral feature vector, so that the differences between the malignant tumor cells, the benign cells and the inflammatory cells are more obvious and are more easily distinguished, and therefore, the fusion feature vector can significantly identify the three types of cells.

[0120] In some embodiments, the classification of the fusion feature vector by the SVM classifier includes:

[0121] Three SVM classifiers are trained, and the constraint condition of the hyperplane objective function of the SVM classifiers is represented by formula (6), and formula (6) is as follows:

[0122]

[0123] subject to y i [(wx i )+b]≥1-ζ i (i=1,2,…,N),

[0124] ζ i ≥0

[0125] wherein x i is a feature of each sample, y i is a label thereof, N is a sample number, w is a hyperplane normal vector, ζ i is a slack variable, C is a penalty factor, and the penalty factor represents the importance of the loss caused by an outlier.

[0126] In the embodiment, the cell classification result of the SVM classifier is only a category, and does not have a probability value.

[0127] In some embodiments, the probability calculation and the probability calibration are calculated by using an isotonic regression algorithm, and the calculation process of the isotonic regression algorithm is represented by formula (7), and formula (7) is as follows:

[0128]

[0129] wherein y i is a true label of the sample i, and y is an output probability of the sample i after calibration.

[0130] The probability calibration is performed by using the isotonic regression method, so that the predicted value of the sample feature x i and the true sample label y iThe mean square error is minimized. In this embodiment, the tumor cell classification result is calibrated by using the order-preserving regression probability calibration algorithm while the tumor cell classification result is given, which not only converts the output of the non-probabilistic classification model into probability, but also can further correct the result of the probabilistic classification model, improve the early tumor cell classification accuracy, and evaluate the classification effect. It is convenient for relevant medical staff to further investigate.

[0131] In a second aspect, the embodiment further provides a computer-readable storage medium, which stores computer program instructions, and the computer program instructions, when executed by a processor, implement the method in the first aspect.

[0132] Please refer to Figure 4 In a third aspect, the embodiment further provides an electronic device 1, comprising a memory 11 and a processor 12, the memory 11 is used to store one or more computer program instructions, wherein the one or more computer program instructions are executed by the processor 12 to implement the method in the first aspect.

[0133] The storage medium / memory 11 includes but is not limited to RAM, ROM, magnetic disk, magnetic tape, optical disk, flash memory, U disk, mobile hard disk, memory card, memory stick, network server storage, network cloud storage, etc. The processor 12 includes but is not limited to CPU (central processing unit 12), GPU (graphics processor 12), MCU (microprocessor 12), etc.

[0134] The above technical solution uses a hyperspectral imaging system to shoot tumor cell slice images, obtains sample image data of tumor cells and corresponding spectral information thereof, and since tumor cells and normal cells have certain differences in protein, nucleic acid, moisture, structure, etc., the spectral curves thereof are different. By analyzing the size, shape and spectral curve of the cells, it can be accurately judged whether tumor cells exist in the cell slice. This process does not need manual screening of pathological images, effectively improves the image screening efficiency, avoids the problem of manual omission and error detection, and at the same time, the hyperspectral image data contains spectral information, effectively retains the characteristic information of tumor cells, facilitates the identification of tumor cells using spectral information, and improves the correct identification rate.

[0135] Finally, it should be noted that although the above embodiments have been described in the specification and drawings of the present application, they should not limit the patent protection scope of the present application. Any equivalent structure or equivalent process substitution or modification based on the essential concept of the present application, using the content described in the specification and drawings of the present application, and directly or indirectly implementing the technical solutions of the above embodiments in other related technical fields, etc., are all included in the patent protection scope of the present application.

Claims

1. A method of recognizing a tumor cell section image, characterized by, The method comprises the following steps: obtaining sample image data of tumor cells and performing morphological processing on the sample image data to extract a to-be-analyzed cell image, wherein the sample image data is obtained by a hyperspectral imaging system, and the to-be-analyzed cell image comprises any one or more of a malignant tumor cell image, a benign cell image and an inflammatory cell image; extracting spectral features of the to-be-analyzed cell image to obtain a spectral feature vector corresponding to the to-be-analyzed cell image; extracting color matrix features of the to-be-analyzed cell image by using the extracted spectral feature vector to obtain a color matrix feature vector corresponding to the to-be-analyzed cell image, comprising: determining color components in a color matrix according to spectral wavelengths corresponding to the spectral features; extracting color information of low-order moments in the color matrix corresponding to the to-be-analyzed cell image; and obtaining a color matrix feature vector according to the color information; the low-order moments comprise first-order moments, second-order moments and third-order moments, and the extraction of the color information of the low-order moments in the color matrix corresponding to the to-be-analyzed cell image comprises: Let k spectral wavelengths corresponding to k spectral features in k-dimensional space be taken as k color components of a color matrix, M represents the number of pixels in the cell image to be analyzed, the first moment is defined by equation (1) as follows: ； Second moment The definition is represented by equation (2) as follows: ； Third moment The definition is represented by equation (3) as follows: ； where the first moment represents the mean of all pixels on the i-th color channel, the second moment represents the standard deviation of all pixels on the i-th color channel, and the third moment represents the cubic root of the skewness of all pixels on the i-th channel, represents the pixel value of the j-th pixel in the i-th color component (i = 1, 2,..., k); the color matrix feature vector is obtained by formula (4), and the formula (4) is as follows: ,…, ]； fusing the spectral feature vector and the color feature vector to obtain a fused feature vector; classifying the fused feature vector by using an SVM classifier to obtain a cell image classification result; performing probability calculation and probability calibration based on the cell image classification result to output a cancer probability corresponding to a current tumor cell slice image.

2. The method of claim 1, wherein the step of identifying the tumor cell section image is performed by using a machine learning algorithm. The spectral feature extraction of the to-be-analyzed cell image by using an SVM-RFE model classifier comprises the following steps: obtaining the number of categories of the to-be-analyzed cell image, and determining the number of SVM-RFE model classifiers according to the number of categories of the to-be-analyzed cell image, wherein the number of the SVM-RFE model classifiers is obtained according to formula (5), and the formula (5) is as follows: ， wherein is the number of SVM-RFE model classifiers, is a positive integer; inputting the to-be-analyzed cell image according to category combinations to multiple SVM-RFE models to obtain multiple ranking criterion scores; adding the multiple ranking criterion scores to obtain a ranking criterion total score; performing spectral feature elimination according to the ranking criterion total score, and iteratively eliminating minimum spectral features until a preset condition is met to obtain a final spectral feature set.

3. The method of claim 2, wherein the step of identifying the tumor cell section image is performed by using a neural network. The specific calculation process of the SVM-RFE model classifier is as follows: Let the training samples be: wherein N is the sample size, is the i-th sample, and D is the original sample feature number, i.e., the original spectral band number. Initialize the original feature set: , the feature ranking set: ; Three training samples are generated, and in the training sample T, two- two combinations of different categories are found to obtain the final training sample: ; Respectively use Train the SVM classifier, respectively, get weight ; computing the ranking criterion score: , finding the feature with the smallest ranking criterion score: , updating the feature set ; In S the smallest eigenvalue is removed to obtain a new feature set: ; The loop iterates until The number of features in the middle is k. obtaining a feature ranking list R, that is, the order in which the features are eliminated.

4. The method of claim 3, wherein the step of identifying the tumor cell section image is performed by using a neural network. In the feature ranking list R, each to-be-analyzed cell image corresponds to k-dimensional spectral features.

5. The method of claim 1, wherein the step of identifying the tumor cell section image comprises the steps of: The classification of the fused feature vector by using an SVM classifier comprises: ​ training three SVM classifiers, and the constraint condition of the hyperplane objective function of the three SVM classifiers is represented by formula (6), and the formula (6) is as follows: ， wherein, is the label of the sample; N is the number of samples; w is the hyperplane normal vector, is the label of the sample; N is the number of samples; w is the hyperplane normal vector, is the label of the sample; N is the number of samples; w is the hyperplane normal vector, 6. The method of claim 5, wherein the step of identifying the tumor cell section image is performed by using a neural network. the probability calculation and the probability calibration are calculated by using an isotonic regression algorithm, and the calculation process of the isotonic regression algorithm is represented by formula (7), and the formula (7) is as follows: ， wherein, is the true label of the sample i, is the output probability of the calibrated sample i.

7. A computer readable storage medium having stored thereon computer program instructions, wherein, the computer program instructions realize the method in any one of claims 1-6 when executed by the processor.

8. An electronic device comprising a memory and a processor, characterized in that The memory is configured to store one or more computer program instructions, wherein the one or more computer program instructions are executed by the processor to implement the method according to any one of claims 1-6.

Images