A method for constructing a cell-level pathological feature and gene mutation data matching model, a cell-level pathological feature and gene mutation data matching model, and an application thereof

Abstract

The present application relates to the technical field of medical image analysis and gene detection, in particular to a cell-level pathological feature and gene mutation data matching model construction method, a cell-level pathological feature and gene mutation data matching model and application thereof. The present application realizes cell-level pathological feature extraction of pathological images and precise correlation analysis of gene mutations by combining high-resolution pathological image analysis and single-cell gene mutation sequencing technology. This innovation not only improves the accuracy and comprehensiveness of pathological diagnosis, but also provides more rich lesion information for clinicians, which helps to develop more personalized treatment plans. Through comprehensive analysis of pathological features such as cell morphology, nucleolus features and chromatin distribution and gene mutation data, a new way for precise medical treatment of renal cell carcinoma is further opened up, which has important clinical application value and social significance.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of medical image analysis and gene detection, in particular to a cell-level pathological feature and gene mutation data matching model construction method, a cell-level pathological feature and gene mutation data matching model and application thereof. BACKGROUND

[0002] With the rapid development of medical imaging technology and gene sequencing technology, pathological diagnosis plays an increasingly important role in tumor treatment. Renal cell carcinoma, as a common malignant tumor of the urinary system, its early diagnosis and personalized treatment plan are crucial for improving patient survival rate and quality of life.

[0003] The traditional pathological diagnosis method of renal cell carcinoma mainly relies on the experience of pathologists and subjective judgment of tissue morphology. This method has several limitations. First, macroscopic observation cannot accurately quantify the subtle differences in cell morphology, nucleolus features and chromatin distribution in pathological images, resulting in a certain subjectivity and error in the diagnosis results. Second, the traditional method cannot directly obtain the gene mutation information of tumor cells, while gene mutations play a key role in the occurrence, development and treatment response of tumors. Although gene sequencing technology can provide gene mutation data, how to effectively combine these data with pathological image features to guide more accurate diagnosis and treatment is a problem that traditional methods cannot solve. In addition, the traditional diagnosis method is not up to the task when facing tumor heterogeneity, and it is difficult to fully reflect the complexity and diversity of the tumor.

[0004] In view of the limitations of the traditional pathological diagnosis method of renal cell carcinoma, it is crucial to provide a technical solution that can solve the above technical problems. SUMMARY

[0005] To solve the above problems, the purpose of the present application is to provide a cell-level pathological feature and gene mutation data matching model construction method, a cell-level pathological feature and gene mutation data matching model and application thereof. The present application realizes the extraction of cell-level pathological features of pathological images and the accurate correlation analysis of gene mutations by combining high-resolution pathological image analysis and single-cell gene mutation sequencing technology. This innovation not only improves the accuracy and comprehensiveness of pathological diagnosis, but also provides clinicians with more rich lesion information, which helps to develop more personalized treatment plans. By comprehensively analyzing pathological features such as cell morphology, nucleolus features, chromatin distribution and gene mutation data, the present application further opens up a new way for the precision medicine of renal cell carcinoma, which has important clinical application value and social significance.

[0006] The purpose of the present application can be realized by the following technical solutions:

[0007] The first object of the present application is to provide a method for constructing a cell-level pathological feature and gene mutation data matching model, comprising the following steps:

[0008] S1, preparing continuous pathological sections of pathological tissue samples of renal cell carcinoma patients, and obtaining high-resolution pathological images by using a part of the pathological sections; performing single cell separation on another part of the pathological sections to obtain single cell gene mutation data (whole genome or targeted gene mutation sequencing is performed on the separated single cells), and establishing a location mapping relationship of the single cells in the pathological images;

[0009] S2, extracting cell contours after denoising the high-resolution pathological images obtained in step S1; extracting cell pathological features based on the extracted cell contours, and standardizing all the extracted cell pathological features to obtain a standardized cell-level pathological feature data set;

[0010] S3, performing quality control on the single cell gene mutation data obtained in step S1 to obtain high-quality mutation data; obtaining a single cell gene mutation data set based on the high-quality mutation data, and performing gene typing on the single cells;

[0011] S4, according to the location mapping relationship of the single cells in the pathological images established in step S1, associating the standardized cell-level pathological feature data set obtained in step S2 with the single cell gene mutation data set obtained in step S3 to obtain "pathological feature-gene subtype" paired data;

[0012] S5, performing cluster analysis on the standardized cell-level pathological feature data set obtained in step S2 to obtain cell-level feature clustering;

[0013] The "pathological feature-gene subtype" paired data obtained in step S4 is optimized by combining cell-level feature clustering with the single cell gene typing obtained in step S3 to obtain a joint optimization model;

[0014] S6, improving the joint optimization model obtained in step S5 to obtain a cell-level pathological feature and gene mutation data matching model.

[0015] In an embodiment of the present application, in step S1, high-resolution pathological images are obtained by digital pathology scanners after HE staining;

[0016] Laser capture microdissection or microfluidic technology is used for single cell separation;

[0017] The gene mutation data includes one or both of the whole genome mutation data of the single cells or the targeted gene mutation data of the single cells.

[0018] In one embodiment of the present application, in step S2, a multi-threshold fusion algorithm is used to extract the cell contour (separate the cytoplasm and the nucleus);

[0019] All the extracted cell pathological features are standardized by using a dynamic scaling algorithm;

[0020] The cell pathological features include cell morphological features, nucleolus features and chromatin distribution features.

[0021] In one embodiment of the present application, an adaptive denoising algorithm based on cell edge features is used to remove background noise;

[0022] Adaptive denoising algorithm:

[0023] (1) The cell edge contour in the pathological image is extracted by using a Canny edge detection operator, and the gray gradient distribution of all edge pixel points is counted to determine the gray fluctuation range of the edge region, i.e. the minimum and maximum values of the gray values of the edge pixel points;

[0024] (2) The absolute value of the difference between the gray value of a single pixel point and the average gray value of the edge region pixel points is divided by the gray fluctuation range of the edge region to obtain the gray difference coefficient of the pixel point and the edge region. Then, the gray difference coefficient is multiplied by the edge protection coefficient, and the product is subtracted from 1 to obtain the filtering weight of the pixel point;

[0025] Wherein, the edge protection coefficient is 0.65 (determined by testing 150 renal cell carcinoma pathological images containing different noise intensities), when the coefficient is 0.65, the peak signal-to-noise ratio of the denoised image reaches more than 38dB, and the cell edge retention rate is more than 92%;

[0026] When the pixel point is close to the edge region, the gray difference coefficient is small, and the corresponding filtering weight tends to 1, so the filtering strength is weakened, thereby avoiding the loss of cell edge details;

[0027] When the pixel point is in the background noise region, the gray difference coefficient is large, and the corresponding filtering weight is reduced, so the filtering strength is enhanced, thereby effectively removing the background noise.

[0028] Further, a multi-threshold fusion algorithm is used to extract the cell contour, specifically a multi-threshold fusion algorithm based on the gray difference between the nucleus and the cytoplasm (aiming to solve the problem of incomplete cell adhesion segmentation caused by traditional single threshold segmentation);

[0029] Multi-threshold fusion algorithm based on the gray difference between the nucleus and the cytoplasm:

[0030] (1) The gray mean value of the nucleus region in the pathological image is counted and the gray mean value of the cytoplasm region , calculate the gray value difference between the nucleus and the cytoplasm ;

[0031] wherein, the value range is 50-80, the value range is 120-150;

[0032] (2) build the first segmentation threshold and the calculation model of the second segmentation threshold , the formula is as follows:

[0033] , ,

[0034] wherein, is the segmentation threshold of the nucleus region, used for preliminary separation of the nucleus and the cytoplasm, is the segmentation threshold of the cytoplasm region, used for preliminary separation of the cytoplasm and the background;

[0035] Further, a dynamic scaling algorithm is used to standardize all extracted cell pathological features, which is a dynamic scaling algorithm based on the overall feature distribution of the sample, aiming to solve the problem of poor standardization effect of traditional Z-score standardization when the sample feature distribution is uneven.

[0036] Dynamic scaling algorithm based on overall feature distribution of sample:

[0037] (1) statistics of the maximum value , minimum value and median of a certain pathological feature of all cells, calculate the interquartile range , 25% quantile of the feature value, 75% quantile of the feature value, determined by sorting the feature values of all cells;

[0038] (2) build the calculation model of the feature standardization value, the formula is as follows:

[0039] ,

[0040] wherein, is the standardization value of the cell area, is the actual area of a single cell;

[0041] For other pathological features such as nucleolus number and chromatin distribution entropy value, the same dynamic scaling algorithm is used for standardization, only the corresponding feature maximum value, minimum value, median and interquartile range parameters need to be replaced;

[0042] The cell-level pathological feature data set standardized by the algorithm has a more uniform feature value distribution, and can eliminate feature deviations caused by differences in staining conditions and scanning parameters between different samples, ensuring that the feature data used in subsequent clustering analysis and model construction are comparable.

[0043] In one embodiment of the application, in step S3, the high-quality mutation data is annotated (to determine the clinical significance and gene function of the mutation) using a clinical relevance scoring algorithm to obtain annotation results, and a single-cell gene mutation data set is established based on the annotation results;

[0044] The single cells are preliminarily genotyped by the mutation weight score.

[0045] Further, the quality control specifically filters low-quality sequencing data using a multi-index fusion algorithm, which is specifically constructed based on sequencing depth, base quality, and gene coverage, aiming to more accurately filter low-quality sequencing data.

[0046] A multi-index fusion algorithm is constructed based on sequencing depth, base quality, and gene coverage:

[0047] (1) First, define three quality evaluation indexes: sequencing depth compliance rate , high-quality base proportion , and gene coverage .

[0048] (2) Construct a sequencing data quality score model, the formula is as follows:

[0049] ,

[0050] wherein, Q is the quality score of single-cell sequencing data, , , are the weight coefficients of the three indexes, respectively.

[0051] The weight coefficients are determined by the analytic hierarchy process (AHP): first, construct a judgment matrix, invite five experts in the field of gene sequencing to compare the importance of the three indexes pairwise (sequencing depth has the greatest impact on data quality, followed by high-quality base proportion, and finally gene coverage); then determine , , by calculating the eigenvectors of the judgment matrix, and satisfy .

[0052] When Q is greater than or equal to 0.88, the single-cell sequencing data is determined to be high-quality data and is retained; when Q is less than 0.88, the single-cell sequencing data is determined to be low-quality data and is filtered out.

[0053] Further, when annotating high-quality mutation data by a clinical relevance scoring algorithm, specifically a relevance scoring algorithm based on clinical evidence level, aiming to quantify the clinical relevance of gene mutations to renal cell carcinoma;

[0054] Relevance scoring algorithm based on clinical evidence level:

[0055] (1) The classification criteria for the clinical evidence level of gene mutations, the clinical evidence of renal cell carcinoma-related gene mutations is divided into 4 levels:

[0056] Level 1 has clear clinical guideline recommended targeted therapy targets, level 2 has clinical research confirmed to be associated with treatment response, level 3 has basic research confirmed to be associated with tumor occurrence and development, and level 4 has unknown clinical significance;

[0057] (2) Construct a gene mutation clinical relevance scoring model, the formula is as follows:

[0058] ,

[0059] Among them, is the clinical relevance score of gene mutation, the value range is 0-4; is the clinical evidence level score, level 1 is 4 points, level 2 is 3 points, level 3 is 2 points, and level 4 is 1 point; is the mutation frequency score, the mutation frequency in renal cell carcinoma patients ≥10% is 3 points, 5%≤ the mutation frequency in renal cell carcinoma patients <10% is 2 points, 1%≤ the mutation frequency in renal cell carcinoma patients <5% is 1 point, and the mutation frequency in renal cell carcinoma patients <1% is 0 point, the frequency data is from TCGA renal cell carcinoma database; is the prognosis correlation score, 3 points for research confirming that the mutation is associated with poor prognosis of patients, 1 point for no significant correlation with prognosis, and 0 point for no prognosis correlation research, based on the statistics of renal cell carcinoma prognosis research literature published in the past 5 years; The weight coefficients 0.4, 0.3 and 0.3 in the formula are determined by multiple linear regression analysis, aiming to maximize the correlation between the clinical relevance score of gene mutation and the results of expert clinical evaluation, and finally determine the weight of each index;

[0060] When , it is determined that the mutation is a mutation with important clinical significance, which is given priority in subsequent genotyping and treatment plan recommendation; when , it is determined that the mutation is a mutation with lower clinical significance, and its weight is reduced in the analysis;

[0061] The gene mutation data annotated by this algorithm can more accurately screen out key mutations related to the diagnosis and treatment of renal cell carcinoma, providing more clinically valuable evidence for subsequent single-cell genotyping and precision treatment recommendations.

[0062] In one embodiment of the present invention, in step S5, a density-distance fusion algorithm is used to perform cluster analysis on the standardized cell-level pathological feature dataset.

[0063] Specifically, this refers to a density- and distance-based fusion clustering algorithm;

[0064] A density- and distance-based fusion clustering algorithm:

[0065] (1) Calculate the Euclidean distance between each cell feature vector and all other cell feature vectors, and count the distance threshold for each cell. The number of neighboring cells within the range, The value is 1.2 times the median of the Euclidean distances between all cells;

[0066] (2) Construct an evaluation index model for cluster center selection. The specific formula is as follows:

[0067] ,

[0068] in, For the first The evaluation score of each cell as the cluster center For the first Density of individual cells For the first The cell and the first Euclidean distance of cells For the first Individual cells The formula includes neighboring cells within the range. This is the distance attenuation coefficient; the closer two cells are... The smaller the value, the closer the coefficient is to 1, and the greater the contribution of the density of neighboring cells to the evaluation score; the greater the distance, the smaller the coefficient, and the smaller the contribution.

[0069] By calculating all cells Value, select Top The cells serve as the initial cluster centers. To minimize the number of clusters, the initial cluster centers are ensured to have a high density and be far apart from each other. During the clustering iteration process, each cell is assigned to the corresponding cluster group based on the weighted distance to each initial cluster center, and the cluster centers are iteratively updated until the clustering results are stable.

[0070] The adjustment Lande index of the clustering result is improved by 20% than the traditional K-means clustering, and the sensitivity to the initial clustering center is reduced, so that the cells with similar pathological characteristics can be more stably and accurately classified into the same class, and a more reliable clustering result is provided for subsequent joint optimization with genotyping.

[0071] In an embodiment of the application, in step S6, the optimal parameters are determined by maximizing the consistency of clustering and genotyping, and an improved support vector machine algorithm based on feature-gene correlation degree weighting and rare subtype loss optimization is used to improve the joint optimization model.

[0072] Further, the improved support vector machine algorithm based on feature-gene correlation degree weighting and rare subtype loss optimization:

[0073] (1) Calculate the correlation degree of each cell-level pathological feature and gene mutation typing , is the pathological feature number, and the Spearman rank correlation coefficient is used for calculation, and the specific formula is as follows:

[0074] ,

[0075] wherein, is the total number of cell samples, is the difference between the ranking of the i th pathological feature of the j th cell and the ranking of the gene mutation typing, , the greater the absolute value, the stronger the distinguishing ability of the feature and the gene typing; (2) Integrate the correlation degree as a feature weight into the radial basis kernel function to construct an improved kernel function, and the specific formula is as follows:

[0076]

[0077] ,

[0078] wherein, , are the pathological feature vectors of two cells, is the total number of pathological features, , are the i th feature value of , respectively, , is the Euclidean distance norm, and the kernel function makes the high features contribute more in distance measurement; For the rare subtype loss optimization, the rare subtype loss function is added to the objective function of the support vector machine algorithm, and the specific formula is as follows:

[0079] ​​​​For rare gene mutation subtypes, a weighted loss function is used to train the model. The specific formula for the loss function is as follows:

[0080] ,

[0081] in, For the first Loss weights for each cell For cell genotyping tags, For SVM model pairs The predicted value;

[0082] The second objective of this invention is to provide a cell-level pathological feature matching model with gene mutation data, which is constructed using the method described above.

[0083] The third objective of this invention is to provide an application of a cell-level pathological feature and gene mutation data matching model in calculating the matching degree between drugs and gene mutation subtypes.

[0084] In one embodiment of the present invention, the method for calculating the matching degree between a drug and a gene mutation subtype includes the following steps:

[0085] Based on the above-mentioned cell-level pathological features and gene mutation data matching model, a multi-factor matching algorithm based on gene mutation subtype, cell number proportion and drug sensitivity is used (to improve the accuracy of treatment plan recommendation) to establish a correspondence library between renal cell carcinoma gene mutation subtypes and targeted drugs. At the same time, a sensitivity score is assigned to each drug, and the matching degree between the drug and gene mutation subtype is further calculated.

[0086] When the matching score is ≥3.5, the drug is considered a priority recommended drug for the corresponding gene mutation subtype; when 2.5 < matching score <3.5, it is considered an alternative recommended drug; when the matching score is ≤2.5, it is considered a non-recommended drug.

[0087] In one embodiment of the present invention, when scoring sensitivity, the score is determined based on the objective response rate of the drug clinical trial: 5 points for ORR≥50%, 4 points for ORR<50% for ORR<30% for ORR<2 ...

[0088] When calculating the matching degree between a drug and a gene mutation subtype, the drug sensitivity score, the proportion of cells of that subtype in the lesion, and the average frequency of occurrence of that subtype in patients with renal cell carcinoma are taken into account. Based on the drug sensitivity score to the corresponding subtype, the results are multiplied by the weight of the proportion of cells and the weight of the subtype frequency difference adjustment term, and then the two results are added together to obtain the final matching degree score.

[0089] Wherein, the subtype frequency difference adjustment term is calculated by subtracting the absolute value of the proportion of the number of the subtype in the lesion to the average occurrence frequency of the patient from 1, to correct the influence of the difference between the proportion of the lesion subtype and the general case on the treatment effect.

[0090] Further, the cell-level pathological feature and gene mutation data matching model are applied to predict the gene mutation state of all cells in the pathological image, a spatial distribution algorithm based on cell density is used to draw a heat map of cells of different gene mutation subtypes to realize spatial positioning, the quantitative analysis of the number proportion and morphological parameters of each subtype cell is performed, and a corresponding relationship library of the gene mutation subtype of renal cell carcinoma and the targeted drug is established through a multi-factor matching algorithm.

[0091] Further, the weight coefficients of 0.6 and 0.4 are determined through clinical efficacy backtracking analysis, the correlation between the matching degree score and the actual treatment response rate of the patient is maximized as the goal, and finally the weight of the cell number proportion is higher, because the proportion of the cells of the subtype in the lesion directly affects the overall efficacy of the drug.

[0092] Further, if there are multiple gene mutation subtypes in the lesion, the number proportion of each subtype cell is used as the weight, and the matching degree scores of the drugs corresponding to different subtypes are weighted and summed to obtain the comprehensive matching degree of the drug, and the drug or drug combination with the highest comprehensive matching degree is selected as the final treatment scheme.

[0093] The coincidence degree of the treatment scheme recommended by the algorithm and the actual treatment response rate of the patient is improved by 30% compared with the traditional recommendation method based on a single gene mutation, which can provide more precise targeted treatment suggestions for the clinician according to the cell-level features of the patient's lesion, and help to improve the treatment effect and reduce the risk of drug adverse reactions.

[0094] Compared with the prior art, the present application has the following beneficial effects:

[0095] (1) The present application realizes the extraction of cell-level pathological features of renal cell carcinoma pathological images and the accurate correlation analysis of gene mutations by combining high-resolution pathological image analysis and single-cell gene mutation sequencing technology. This method not only improves the accuracy of pathological diagnosis and can more accurately identify the type of lesion, but also provides more comprehensive lesion information for clinicians by comprehensively analyzing pathological features such as cell morphology, nucleolus features, and chromatin distribution, which helps to develop more personalized treatment plans.

[0096] (2) The target therapy scheme recommendation system based on the multi-factor matching algorithm comprehensively considers multiple factors such as gene mutation subtypes, cell quantity proportion and drug sensitivity, can more accurately match patient lesion characteristics and target drugs, and thus recommend the optimal treatment scheme. This method not only improves the accuracy of treatment scheme recommendation, but also continuously optimizes algorithm parameters through clinical efficacy backtracking analysis, ensures that the recommended drug combination can maximize the treatment effect, reduces the risk of drug adverse reactions, and brings better treatment experience and prognosis effect to patients. DETAILED DESCRIPTION

[0097] The technical solution provided by the present application is not a direct implementation object of a living human body, and is a technical solution for non-disease diagnosis purposes and non-treatment purposes.

[0098] The present application will be described in detail below with reference to specific embodiments.

[0099] In the following examples, unless otherwise specified, the reagents used are commercially available reagents, and the detection means and methods used are conventional detection means and methods in the art.

[0100] Example 1

[0101] The present embodiment provides a method for constructing a cell-level pathological feature and gene mutation data matching model, comprising the following steps:

[0102] S1, make continuous pathological sections of pathological tissue samples of renal cell carcinoma patients, use a part of the pathological sections to obtain high-resolution pathological images; perform single cell separation on another part of the pathological sections to obtain single cell gene mutation data (perform whole genome or targeted gene mutation sequencing on the separated single cells), and establish a position mapping relationship of the single cells in the pathological images;

[0103] Specifically, high-resolution pathological images are obtained by a digital pathology scanner after HE staining;

[0104] Laser capture microdissection or microfluidic technology is used for single cell separation;

[0105] The gene mutation data includes one or both of whole genome mutation data of single cells or targeted gene mutation data of single cells.

[0106] S2, extract cell contours after denoising processing of the high-resolution pathological images obtained in step S1; extract cell pathological features based on the extracted cell contours, and standardize all extracted cell pathological features to obtain a standardized cell-level pathological feature data set;

[0107] A multi-threshold fusion algorithm (to separate cytoplasm and nucleus) is used to extract cell contours;

[0108] All the extracted cell pathological features are standardized by using a dynamic scaling algorithm;

[0109] The cell pathological features include cell morphological features, nucleolus features and chromatin distribution features.

[0110] Specifically, a filtering algorithm based on cell edge features is used to remove background noise, specifically an adaptive denoising algorithm.

[0111] Adaptive denoising algorithm:

[0112] (1) The cell edge contour in the pathological image is extracted by a Canny edge detection operator, and the gray level gradient distribution of all edge pixel points is counted to determine the gray level fluctuation range of the edge region, i.e. the minimum and maximum values of the gray level of the edge pixel points;

[0113] (2) The absolute value of the difference between the gray level of a single pixel point and the average gray level of the edge region pixel points is divided by the gray level fluctuation range of the edge region to obtain the gray level difference coefficient of the pixel point and the edge region. Then, the gray level difference coefficient is multiplied by the edge protection coefficient, and the product is subtracted from 1 to obtain the filtering weight of the pixel point;

[0114] Wherein, the edge protection coefficient is 0.65 (determined by testing 150 renal cell carcinoma pathological images containing different noise intensities), when the coefficient is 0.65, the peak signal-to-noise ratio of the denoised image reaches more than 38dB, and the cell edge retention rate is more than 92%;

[0115] When the pixel point is close to the edge region, the gray level difference coefficient is small, and the corresponding filtering weight tends to 1, so the filtering strength is weakened, thereby avoiding the loss of cell edge details;

[0116] When the pixel point is in the background noise region, the gray level difference coefficient is large, the corresponding filtering weight is reduced, and the filtering strength is enhanced, thereby effectively removing the background noise.

[0117] In actual filtering operation, a 3x3 neighborhood window is used. For each pixel point, the product of the filtering weight and the corresponding gray level of all pixel points in the neighborhood window is calculated and summed, and then divided by the sum of the filtering weights of all pixel points in the neighborhood window to obtain the filtered gray level of the pixel point. The pathological image processed by this algorithm can effectively remove background noise and completely retain key detail features such as cell edges and nucleolus boundaries, providing higher quality image data support for subsequent cell-level pathological feature extraction.

[0118] Further, a multi-threshold fusion algorithm is used to extract the cell contour, specifically a multi-threshold fusion algorithm based on the gray level difference between the cell nucleus and the cytoplasm (aiming to solve the problem of incomplete cell adhesion segmentation caused by traditional single threshold segmentation).

[0119] Multi-threshold fusion algorithm based on gray difference between nucleus and cytoplasm:

[0120] (1) Statistics of the gray mean value of the nucleus region in the pathological image and the gray mean value of the cytoplasm region , calculate the gray difference value between the nucleus and the cytoplasm ;

[0121] Among them, the gray mean value of the corresponding region is calculated by manually marking the nucleus and cytoplasm region of 100 typical cells, the value range is 50-80, the value range is 120-150;

[0122] (2) Construct the calculation model of the first segmentation threshold and the second segmentation threshold , the formula is as follows:

[0123] , ,

[0124] Among them, is the segmentation threshold of the nucleus region, which is used to preliminarily separate the nucleus and the cytoplasm, is the segmentation threshold of the cytoplasm region, which is used to preliminarily separate the cytoplasm and the background; The coefficients of 0.3 and 0.4 in the formula are determined by testing 200 pathological images containing cell adhesion. When using this coefficient, the segmentation accuracy of the nucleus and the cytoplasm reaches 93%, and the segmentation accuracy of the cell adhesion region reaches 88%;

[0125] In the actual segmentation process, first use to segment the image into nucleus candidate region and non-nucleus region; Then use to segment the non-nucleus region into cytoplasm candidate region and background region; Finally, morphological closing operation is performed on the nucleus candidate region and the cytoplasm candidate region, and the holes in the region are filled, and the impurity region is filtered through the area threshold, and the complete cell contour is finally obtained. The cell contour extracted by the algorithm can effectively solve the problem of cell adhesion and accurately distinguish the boundary between the nucleus and the cytoplasm, providing accurate region division basis for subsequent nucleolus feature and chromatin distribution feature extraction.

[0126] Further, a dynamic scaling algorithm is used to standardize all extracted cell pathological features, which is a dynamic scaling algorithm based on the overall feature distribution of the sample, aiming to solve the problem of poor standardization effect of traditional Z-score standardization when the sample feature distribution is uneven;

[0127] Dynamic scaling algorithm based on the overall feature distribution of the samples:

[0128] (1) Calculate the maximum value of a certain pathological feature of all cells. Minimum value With median Calculate the interquartile range of the feature , The 25th percentile of the eigenvalues The 75th percentile of the eigenvalues ​​is determined by sorting the eigenvalues ​​of all cells.

[0129] (2) Construct a calculation model for the standardized feature values. The specific formula is as follows:

[0130] ,

[0131] in, This is a standardized value for cell area. This refers to the actual area of ​​a single cell;

[0132] The coefficient 0.8 in this formula was determined through testing with characteristic data from 1000 cells. When this coefficient is used, the distribution ratio of the standardized feature values ​​in the range of [-2,2] reaches 90%, effectively avoiding the impact of extreme outliers on the standardization results.

[0133] For other pathological features such as nucleolus number and chromatin distribution entropy, the same dynamic scaling algorithm is used for standardization, only requiring the replacement of the corresponding feature maximum, minimum, median and interquartile range parameters;

[0134] The standardized cellular pathological feature dataset obtained through this algorithm has a more uniform distribution of feature values, which can eliminate feature bias caused by differences in staining conditions and scanning parameters between different samples, ensuring that the feature data used in subsequent cluster analysis and model construction are comparable.

[0135] S3. Perform quality control on the single-cell gene mutation data obtained in step S1 to obtain high-quality mutation data; obtain a single-cell gene mutation dataset based on the high-quality mutation data, and perform genotyping on the single cells;

[0136] The clinical relevance scoring algorithm was used to annotate high-quality mutation data (clarifying the clinical significance and gene function of the mutations) to obtain annotation results, and a single-cell gene mutation dataset was established based on the annotation results.

[0137] Preliminary genotyping of single cells is performed using mutation weight scores.

[0138] Further, the quality control specifically filters low-quality sequencing data by using a multi-index fusion algorithm, specifically constructing a multi-index fusion algorithm based on sequencing depth, base quality and gene coverage, aiming to more accurately filter low-quality sequencing data.

[0139] Constructing a multi-index fusion algorithm based on sequencing depth, base quality and gene coverage:

[0140] (1) First, define three quality evaluation indexes: sequencing depth compliance rate , high-quality base proportion , and gene coverage . Through statistical analysis of 500 high-quality single-cell sequencing samples, determine the qualified threshold of the three indexes: , , ;

[0141] (2) Construct a sequencing data quality score model, the formula is as follows:

[0142] ,

[0143] wherein, is the quality score of single-cell sequencing data, , , are the weight coefficients of the three indexes, respectively;

[0144] The weight coefficients are determined by the analytic hierarchy process (AHP): first, construct a judgment matrix, invite five experts in the field of gene sequencing to compare the importance of the three indexes pairwise (sequencing depth has the greatest impact on data quality, followed by high-quality base proportion, and finally gene coverage); then determine , , by calculating the eigenvectors of the judgment matrix, and satisfy ;

[0145] When Q≥0.88, the single-cell sequencing data is determined to be high-quality data and is retained; when Q<0.88, it is determined to be low-quality data and is filtered out.

[0146] The threshold value of 0.88 is determined by testing 100 sequencing data of known quality, and the accuracy of the quality control reaches 95%. After quality control by this algorithm, the proportion of low-quality single-cell gene mutation data set is reduced from 15% by traditional methods to less than 5%, and potential effective data with slight sequencing depth deficiency but good base quality and gene coverage can be effectively retained, providing more reliable gene data for subsequent gene mutation annotation and typing.

[0147] Further, when annotating high-quality mutation data by a clinical relevance scoring algorithm, specifically a relevance scoring algorithm based on clinical evidence level, aiming to quantify the clinical relevance of gene mutations to renal cell carcinoma;

[0148] Relevance scoring algorithm based on clinical evidence level:

[0149] (1) The classification criteria for the clinical evidence level of gene mutations, the clinical evidence of renal cell carcinoma-related gene mutations is divided into 4 levels:

[0150] Level 1 has clear clinical guideline recommended targeted therapy targets, level 2 has clinical research confirmed to be associated with treatment response, level 3 has basic research confirmed to be associated with tumor occurrence and development, and level 4 has unknown clinical significance;

[0151] (2) Construct a gene mutation clinical relevance scoring model, the formula is as follows:

[0152] ,

[0153] Among them, is the clinical relevance score of gene mutation, the value range is 0-4; is the clinical evidence level score, level 1 is 4 points, level 2 is 3 points, level 3 is 2 points, and level 4 is 1 point; is the mutation frequency score, the mutation frequency in renal cell carcinoma patients ≥10% is 3 points, 5%≤ the mutation frequency in renal cell carcinoma patients <10% is 2 points, 1%≤ the mutation frequency in renal cell carcinoma patients <5% is 1 point, and the mutation frequency in renal cell carcinoma patients <1% is 0 point, the frequency data is from TCGA renal cell carcinoma database; is the prognosis correlation score, 3 points for research confirming that the mutation is associated with poor prognosis of patients, 1 point for no significant correlation with prognosis, and 0 point for no prognosis correlation research, based on the statistics of renal cell carcinoma prognosis research literature published in the past 5 years; The weight coefficients 0.4, 0.3 and 0.3 in the formula are determined by multiple linear regression analysis, aiming to maximize the correlation between the clinical relevance score of gene mutation and the results of expert clinical evaluation, and finally determine the weight of each index;

[0154] When , it is determined that the mutation is a mutation with important clinical significance, which is given priority in subsequent genotyping and treatment plan recommendation; when , it is determined that the mutation is a mutation with lower clinical significance, and its weight is reduced in the analysis;

[0155] The gene mutation data annotated by this algorithm can more accurately screen out key mutations related to the diagnosis and treatment of renal cell carcinoma, providing more clinically valuable evidence for subsequent single-cell genotyping and precision treatment recommendations.

[0156] S4. Based on the positional mapping relationship of single cells in pathological images established in step S1, associate the standardized cell-level pathological feature dataset obtained in step S2 with the single-cell gene mutation dataset obtained in step S3 to obtain “pathological feature-gene subtype” paired data.

[0157] S5. Perform cluster analysis on the standardized cell-level pathological feature dataset obtained in step S2 to obtain cell-level feature clusters;

[0158] Specifically, a density-distance fusion algorithm was used to perform cluster analysis on the standardized cellular pathological feature dataset;

[0159] Specifically, this refers to a density- and distance-based fusion clustering algorithm;

[0160] A density- and distance-based fusion clustering algorithm:

[0161] (1) Calculate the Euclidean distance between each cell feature vector and all other cell feature vectors, and count the distance threshold for each cell. The number of neighboring cells within the range, The value is set to 1.2 times the median Euclidean distance between all cells, and was determined through testing on a dataset of 100 cell features. This value ensures that most cells have a reasonable number of neighbors.

[0162] (2) Construct an evaluation index model for cluster center selection. The specific formula is as follows:

[0163] ,

[0164] in, For the first The evaluation score of each cell as the cluster center For the first Density of individual cells For the first The cell and the first Euclidean distance of cells For the first Individual cells The formula includes neighboring cells within the range. This is the distance attenuation coefficient; the closer two cells are... The smaller the value, the closer the coefficient is to 1, and the greater the contribution of the density of neighboring cells to the evaluation score; the greater the distance, the smaller the coefficient, and the smaller the contribution.

[0165] By calculating all cells Value, select Top The cells serve as the initial cluster centers. To minimize the number of clusters, the initial cluster centers are ensured to have a high density and be far apart from each other. During the clustering iteration process, each cell is assigned to the corresponding cluster group based on the weighted distance to each initial cluster center, and the cluster centers are iteratively updated until the clustering results are stable.

[0166] Cluster analysis using this algorithm improves the adjusted Rand index of the clustering results by 20% compared to traditional K-means clustering, and reduces the sensitivity to the initial cluster centers. It can more stably and accurately classify cells with similar pathological features into the same group, providing more reliable clustering results for subsequent joint optimization with genotyping.

[0167] The paired data of "pathological features-genotype" obtained in step S4 are jointly optimized by cell-level feature clustering and single-cell genotyping obtained in step S3 to obtain a joint optimization model;

[0168] S6. Improve the joint optimization model obtained in step S5 to obtain a cell-level pathological feature and gene mutation data matching model.

[0169] Specifically, the optimal parameters are determined with the goal of maximizing clustering-genotyping consistency, and an improved support vector machine algorithm based on feature-gene association weighting and rare subtype loss optimization is used to improve the joint optimization model.

[0170] Furthermore, an improved support vector machine algorithm based on feature-gene association weighting and rare subtype loss optimization is proposed:

[0171] (1) Calculate the correlation between each cell-level pathological feature and gene mutation type. , The pathological feature number is used, and the Spearman rank correlation coefficient is calculated using the following formula:

[0172] ,

[0173] in, The total number of cells in the sample. For the first The first cell The difference between the ranking of pathological features and the ranking of gene mutation types. The larger the absolute value, the stronger the ability of the feature to distinguish from the genotype;

[0174] (2) The feature weights are incorporated into the radial basis kernel function to construct an improved kernel function, as shown in the following formula:

[0175] ,

[0176] wherein, , is the pathological feature vector of two cells, is the total number of pathological features, , is the , th eigenvalue of , , is the Euclidean distance norm, and the kernel function makes the high features contribute more in the distance metric;

[0177] For the rare gene mutation subtype accounting for , the model is trained using a weighted loss function, and the loss function formula is as follows:

[0178] ,

[0179] wherein, is the loss weight of the th cell, is the cell genotyping label, is the predicted value of the SVM model for ;

[0180] The SVM penalty coefficient C is determined by 5-fold cross-validation and grid search method, the search interval is [1, 10, 100], and finally C = 10 is selected. After training, the overall prediction accuracy of the model is > 92%, and the prediction accuracy of the rare subtype is improved by 45% compared with the traditional SVM.

[0181] Embodiment 2

[0182] The embodiment provides a method for calculating the matching degree of a drug and a gene mutation subtype, comprising the following steps:

[0183] A cell-level pathological feature and gene mutation data matching model is applied to predict the gene mutation state of all cells in the pathological image, a spatial distribution algorithm based on cell density is used to draw a heat map of cells of different gene mutation subtypes to realize spatial positioning, the number proportion and morphological parameters of cells of each subtype are counted for quantitative analysis, and a corresponding relationship database of renal cell carcinoma gene mutation subtypes and targeted drugs is established through a multi-factor matching algorithm.

[0184] That is, on the basis of the cell-level pathological features provided in Embodiment 1 and the matching model of gene mutation data, a multi-factor matching algorithm based on gene mutation subtypes, cell quantity proportion, and drug sensitivity (to improve the accuracy of treatment recommendation) is used to establish a correspondence relationship database between renal cell carcinoma gene mutation subtypes and targeted drugs, and a sensitivity score is assigned to each drug, and the matching degree of the drug and the gene mutation subtype is further calculated.

[0185] In the sensitivity score, the score is determined according to the objective response rate of the drug clinical trial. When ORR≥50%, score 5 points; when 30%≤ORR<50%, score 4 points; when 20%≤ORR<30%, score 3 points; when ORR<20%, score 2 points.

[0186] In the calculation of the matching degree of the drug and the gene mutation subtype, the sensitivity score of the drug, the proportion of the number of cells in the lesion, and the average frequency of the subtype in renal cell carcinoma patients are considered. Based on the sensitivity score of the drug to the corresponding subtype, the weight of the cell quantity proportion and the weight of the subtype frequency difference adjustment term are multiplied respectively, and the two results are added to obtain the final matching degree score.

[0187] Among them, the subtype frequency difference adjustment term is calculated by subtracting the absolute value of the proportion of the number of cells in the lesion and the average frequency of the subtype in the patient from 1, which is used to correct the influence of the difference between the lesion subtype proportion and the general situation on the treatment effect.

[0188] The weight coefficients of 0.6 and 0.4 are determined by clinical efficacy backtracking analysis, with the goal of maximizing the correlation between the matching degree score and the actual treatment response rate of the patient, and finally determining that the weight of the cell quantity proportion is higher, because the proportion of the subtype cells in the lesion directly affects the overall efficacy of the drug.

[0189] Further, when the matching degree score≥3.5, the drug is determined as the preferred recommended drug for the corresponding gene mutation subtype; when 2.5<matching degree score<3.5, it is determined as the alternative recommended drug; when the matching degree score≤2.5, it is determined as the non-recommended drug. If there are multiple gene mutation subtypes in the lesion, the cell quantity proportion of each subtype is used as the weight, and the matching degree scores of the drugs corresponding to different subtypes are weighted and summed to obtain the comprehensive matching degree of the drug, and the drug or drug combination with the highest comprehensive matching degree is selected as the final treatment scheme. The treatment scheme recommended by the algorithm has a 30% improvement in the degree of coincidence with the actual treatment response rate of the patient compared with the traditional recommendation method based on a single gene mutation, which can provide more accurate targeted treatment suggestions for clinicians that are more in line with the cell-level characteristics of the patient's lesion, and help to improve the treatment effect and reduce the risk of drug adverse reactions.

[0190] Embodiment 3

[0191] The embodiment provides a method for constructing a cell-level pathological feature and gene mutation data matching model, and comprises the following steps.

[0192] S1: sample preparation and data acquisition

[0193] A patient suspected of having renal clear cell carcinoma through imaging examination and preliminary pathological screening is selected, and after obtaining the informed consent of the patient and family members, a pathological tissue sample of a renal lesion of the patient is obtained through surgical resection. The tissue sample is made into continuous thin sections according to pathological sectioning specifications, which can ensure that image analysis and gene detection can be simultaneously carried out from the same tissue area in the subsequent step, and analysis deviation caused by sample difference is avoided. Part of the sections are subjected to HE staining technology, and the specific staining effect of hematoxylin and eosin on cell nuclei and cytoplasm clearly presents cell morphology and tissue structure characteristics. The sections after staining are scanned by a high-resolution digital pathology scanner to generate pathological images that can be digitally analyzed. The images can provide a clear visual basis for subsequent cell feature extraction. The other part of the sections is subjected to laser capture microdissection technology, which can accurately locate and separate single complete cancer cells in the sections, effectively avoiding the interference of normal cells or other impurity cells on the gene detection result. The single cells separated are subjected to whole genome gene mutation sequencing to comprehensively capture the gene variation information of the cells. Meanwhile, the position mapping relationship of the single cells in the pathological images is established through the coordinate marking of the pathological sections. The establishment of the relationship is a key prerequisite for accurately correlating cell pathological features and gene mutation data in the subsequent step, which ensures that each gene detection result can correspond to a specific cell in the image.

[0194] S2: cell-level pathological feature extraction

[0195] The obtained high-resolution pathological image is first subjected to adaptive denoising treatment. A filtering algorithm based on cell edge features is adopted, which can filter out noise points in the image background by identifying the gray gradient features of the cell edges, and at the same time, the details of the cell edges are retained to the greatest extent to avoid the blurring of the cell contour caused by denoising, thereby laying a clear image foundation for subsequent cell contour extraction. After denoising, a multi-threshold fusion algorithm is used for cell contour extraction. First, the gray mean value of the cell nucleus region and the gray mean value of the cytoplasm region are respectively calculated, and the gray difference value of the cell nucleus and the cytoplasm is calculated . Then, a calculation model of a first segmentation threshold and a second segmentation threshold is constructed, and the specific formula is as follows:

[0196] 、 ,

[0197] wherein, The threshold for segmenting the cell nucleus region. The algorithm uses a segmentation threshold to define the cytoplasmic region. Based on the grayscale difference between the nucleus and cytoplasm, a reasonable segmentation threshold is set to achieve precise separation of the nucleus and cytoplasm, thereby completely extracting the outline of a single cell. This effectively avoids outline extraction errors caused by cell overlap or unclear boundaries, ensuring the accuracy of subsequent feature extraction. Based on the extracted cell outline, cell morphology features, nucleolar features, and chromatin distribution features are extracted. These features can reflect the pathological phenotype of the cell from different dimensions, providing a basis for judging whether the cell has malignant characteristics. Finally, a dynamic scaling algorithm is used to standardize all extracted pathological features. First, the maximum value of a certain pathological feature of all cells is counted. Minimum value With median Calculate the interquartile range of the feature , The 25th percentile of the eigenvalues The 75th percentile of the eigenvalues ​​is used as the basis for calculating the standardized eigenvalues. The specific formula is as follows:

[0198] ,

[0199] in, This is a standardized value for cell area. The algorithm eliminates the feature value deviation caused by the difference in measurement scale between different cells (such as the difference in area value caused by the difference in cell size), so that the feature data of all cells are in a unified analysis dimension. This ensures that each feature has equal weight and comparability in subsequent cluster analysis and model construction, and finally forms a standardized cell-level pathological feature dataset.

[0200] S3: Single-cell gene mutation data processing

[0201] The raw data obtained from single-cell whole-genome sequencing first undergoes quality control, employing a multi-index fusion algorithm, starting from sequencing depth. Base quality Gene coverage The three core indicators are used to comprehensively evaluate the data, and the specific formula is as follows:

[0202] ,

[0203] in, Quality scoring for single-cell sequencing data. , , The weighting coefficients for the three indicators are used to filter out low-quality data that do not meet quality standards, such as sequencing results with insufficient sequencing depth, high base error rate, or incomplete gene coverage. This effectively reduces the risk of false positives and false negatives in subsequent analyses, ensuring that the retained high-quality data can truly reflect the gene mutation status of cells. Subsequently, the high-quality mutation data is annotated using a clinical correlation scoring algorithm. This algorithm combines published clinical research evidence of renal cell carcinoma, gene function databases, and prognostic statistics to analyze each detected gene mutation, clarifying whether the mutation is related to the occurrence and development of clear cell renal cell carcinoma, whether it affects the normal function of genes, and the potential impact of the mutation on patient prognosis. Based on the annotation results, a single-cell gene mutation dataset containing the clinical significance of mutations is established, and preliminary genotyping of single cells is performed based on mutation weight scores. Cells with similar mutation profiles are grouped into the same subtype. This can preliminarily distinguish different gene mutation clonal populations within lesions, providing a basis for subsequent analysis of tumor heterogeneity.

[0204] S4: Construction of a model for matching cellular pathological features with gene mutation data

[0205] Based on the positional mapping relationship established in S1, each cell feature vector in the cell-level pathological feature dataset is associated with the gene mutation information of the corresponding cell in the single-cell gene mutation dataset, achieving a one-to-one correspondence between "cell image and gene information." This provides complete paired data for subsequent construction of the association model. Subsequently, a density-distance fusion algorithm is used to perform cluster analysis on the cell-level pathological features. First, the Euclidean distance between each cell feature vector and all other cell feature vectors is calculated, and the distance threshold for each cell is statistically analyzed. The number of neighboring cells within the range, The value is taken as 1.2 times the median Euclidean distance between all cells. Then, an evaluation index model for cluster center selection is constructed, and the specific formula is as follows:

[0206] ,

[0207] in, For the first The evaluation score of each cell as the cluster center For the first Density of individual cells For the first The cell and the first Euclidean distance of cells For the first Individual cells The formula includes neighboring cells within the range. This is the distance attenuation coefficient; the closer two cells are... The smaller,

[0208] The closer the coefficient is to 1, the greater the contribution of the density of neighboring cells to the evaluation score; the farther the distance is, the smaller the coefficient is, and the smaller the contribution is, the algorithm aggregates cells with similar pathological phenotypes by calculating the similarity of pathological features between cells and the degree of aggregation of cells in the feature space, such as classifying cells with high morphological heterogeneity and obvious nucleoli into the same cluster, which can classify the cell population from the perspective of pathological phenotype, and provide a basis for subsequent comparison with the genotyping result. Combined with the single-cell genotyping result obtained in S3, a joint optimization model is constructed, which aims to maximize the consistency of the clustering results of pathological features and the genotyping results. By continuously adjusting the clustering parameters and typing standards, cells of the same genetic subtype are classified into the same pathological cluster group as much as possible, and vice versa, thereby reducing the classification error caused by a single analysis dimension;

[0209] Finally, an improved support vector machine algorithm based on feature-gene correlation degree weighting and rare subtype loss optimization is used to construct a precise matching model, as follows: First, calculate the correlation degree of each cell-level pathological feature and gene mutation typing , is the pathological feature number, calculated using the Spearman rank correlation coefficient, and the formula is as follows:

[0210] ,

[0211] wherein, is the total number of cell samples, is the difference between the ranking of the th pathological feature of the th cell and the ranking of the gene mutation typing, The larger the absolute value is, the stronger the distinguishing ability of the feature and the gene typing is;

[0212] is taken as the feature weight and integrated into the radial basis kernel function to construct an improved kernel function, and the formula is as follows:

[0213] ,

[0214] wherein, , are the pathological feature vectors of two cells, is the total number of pathological features, , are the , th feature values of , , is the Euclidean distance norm, and this kernel function makes high features contribute more in distance measurement; ​

[0215] Regarding the proportion For rare gene mutation subtypes, a weighted loss function is used to train the model. The specific formula for the loss function is as follows:

[0216] ,

[0217] in, For the first Loss weights for each cell For cell genotyping tags, For SVM model pairs The algorithm can learn the correlation between pathological features and gene subtypes based on optimized pairing data, and ultimately accurately predict the corresponding gene mutation status by inputting the pathological features of cells, providing an efficient prediction tool for subsequent large-scale cell analysis.

[0218] Example 4

[0219] This embodiment provides a model for further analysis and result output of heterogeneous lesions in renal cell carcinoma based on the cell-level pathological features and gene mutation data matching model provided in Embodiment 3, as detailed below:

[0220] The pre-constructed precise matching model is applied to all cells in pathological images. The model rapidly predicts the gene mutation status of each cell, significantly improving analysis efficiency compared to traditional cell-by-cell gene detection. It also achieves full coverage analysis of all cells within the lesion, avoiding heterogeneity omissions caused by sampling. A spatial distribution algorithm based on cell density is employed to generate heatmaps of different cell subtypes based on their gene mutation subtypes and locations. These heatmaps visually display the distribution areas of each gene mutation subtype within the lesion, helping doctors clearly identify clonal distribution characteristics and spatial heterogeneity. The model also statistically analyzes the proportion of each gene mutation subtype and its morphological parameters, clarifying the characteristics of different clones within the lesion through quantitative analysis. The degree of advantage and differences in pathological characteristics provide data support for judging the malignancy and development stage of the tumor. Finally, a multi-factor matching algorithm is used, combined with the targeted therapy regimens for clear cell renal cell carcinoma recommended in current clinical guidelines, to comprehensively consider the sensitivity of each gene mutation subtype to different targeted drugs, the proportion of subtype cells, and the general response of subtypes in clinical patients, to screen the optimal targeted therapy drugs or drug combinations for the patient. The final output is a comprehensive analysis report including "the correlation between cytopathological characteristics and gene mutations", "thermograms of spatial distribution of cells of each subtype", "quantitative analysis results", and "personalized treatment suggestions". This report can provide clinicians with comprehensive and accurate reference for developing precise treatment plans, evaluating treatment effects, and judging prognosis.

[0221] Example 5

[0222] The embodiment provides a method for constructing a cell-level pathological feature and gene mutation data matching model, and comprises the following steps:

[0223] S1, sample preparation and data acquisition:

[0224] A patient diagnosed as metastatic renal cell carcinoma and having a recurrent lesion is selected, pathological tissue samples of the recurrent lesion are obtained through puncture biopsy, the samples obtained through puncture are made into continuous pathological sections, a part of the sections is subjected to HE staining, and after staining, a digital pathology scanner is used for scanning to obtain a high-resolution pathological image of the recurrent lesion, which clearly presents the cell morphology and structure of the recurrent lesion, and another part of the sections adopts microfluidic technology, which can efficiently and quickly separate single cells in the sections, and targeted gene mutation sequencing is performed on the separated single cells, and the mutation conditions of gene sites related to renal cell carcinoma recurrence are detected, in the operation process, the specific positions of the single cells in the recurrent lesion tissue are recorded in detail, the mapping relationship between the single cells and the positions of the corresponding cells in the pathological image is established, and accurate correspondence between the gene data and the cell image is ensured.

[0225] S2, cell-level pathological feature extraction:

[0226] For the obtained high-resolution pathological image of the recurrent lesion, first, adaptive denoising is performed, a filtering algorithm based on cell edge features is adopted, background noise in the image is effectively removed, cell edge details are highlighted, and interference of noise on subsequent analysis is avoided, after denoising, cell contour extraction is performed, a multi-threshold fusion algorithm is used, according to the gray characteristic difference between the cell nucleus and the cytoplasm, effective separation of the cell nucleus and the cytoplasm is realized, and then complete cell contours are extracted, based on the extracted cell contours, morphological features, nucleolus features and chromatin distribution features of the cells are extracted, then, a dynamic scaling algorithm is used to standardize the extracted pathological features, and the dimension difference between different cell features is eliminated, thereby forming a cell-level pathological feature data set, and providing uniform standard feature data for subsequent data analysis.

[0227] S3, single cell gene mutation data processing:

[0228] Quality control is carried out on single-cell targeted gene mutation sequencing data, and a multi-index fusion algorithm is used to comprehensively evaluate the sequencing data from the aspects of sequencing depth, base quality, and gene coverage, etc. Low-quality sequencing data that does not meet the quality requirements is filtered out, and high-quality mutation data is retained. Then, the high-quality mutation data is annotated, and a clinical relevance scoring algorithm is used to determine the clinical significance of each mutation by referring to existing clinical research evidence on metastatic renal cell carcinoma gene mutations, analyze the impact of mutations on gene function, and construct an annotated single-cell gene mutation dataset according to the annotation results. According to the mutation weight score, the single cells are preliminarily genotyped to distinguish different gene mutation subtypes, and prepare for subsequent matching analysis.

[0229] S4, cell-level pathological feature and gene mutation data matching model construction

[0230] With the help of the previously established position mapping relationship, the cell-level pathological feature dataset and the annotated single-cell gene mutation dataset are associated. The density-distance fusion algorithm is used to cluster the cell-level pathological features, and cells with similar pathological features are clustered together to form different cell clusters. Combined with the single-cell gene typing results, a joint optimization model is constructed, and the consistency of the clustering results and the gene typing results is the highest. The optimal settings of the model are determined, and finally, an improved support vector machine algorithm is used to construct a precise matching model. This model can accurately determine the corresponding gene mutation of the cell according to its pathological features, and realize the effective matching of cell pathological features and gene mutation data.

[0231] Embodiment 6

[0232] This embodiment provides a cell-level pathological feature and gene mutation data matching model based on the cell-level pathological feature and gene mutation data matching model provided in Embodiment 5, which is further used to analyze and output the results of renal cell carcinoma heterogeneity lesions. The specific steps are as follows:

[0233] The precision matching model is applied to all cells in the recurrent lesion pathological image, the gene mutation state of each cell is predicted by the model, a spatial distribution algorithm based on cell density is used, a heat map of cells of different gene mutation subtypes is drawn according to the gene mutation state and position information of the cells, the spatial distribution condition of cells of each subtype in the recurrent lesion is clearly displayed, the heterogeneity characteristics of the recurrent lesion are helped to be understood by the doctor, the number proportion of cells of each gene mutation subtype is counted, the advantage degree of cells of each subtype in the recurrent lesion is analyzed, meanwhile, the morphological parameters of cells of each subtype are counted, quantitative analysis is carried out, the change of cell characteristics of the recurrent lesion is deeply mined, finally, through a multi-factor matching algorithm, combining with the targeted treatment scheme recommendation in the clinical guideline for recurrent patients with metastatic renal cell carcinoma, considering multiple factors such as gene mutation subtype, cell number proportion and drug sensitivity, a comprehensive analysis report containing cell pathological characteristics and gene mutation correlation results, spatial distribution of cells of each subtype, quantitative analysis data and personalized targeted treatment suggestion is output for the recurrent patient, which assists the clinical doctor to make more suitable treatment strategy for the patient.

[0234] The above description of the embodiments is to facilitate those skilled in the art to understand and use the application. Those skilled in the art can easily make various modifications to the embodiments, and apply the general principles described herein to other embodiments without creative labor. Therefore, the present application is not limited to the above embodiments, and the improvements and modifications made by those skilled in the art according to the explanation of the present application without departing from the scope of the present application should be within the protection scope of the present application.

Claims

1. A method for constructing a model for matching cell-level pathological features with gene mutation data, characterized in that, Includes the following steps: S1. Prepare serial pathological sections from pathological tissue samples of patients with renal cell carcinoma, and use a portion of the pathological sections to obtain high-resolution pathological images. Single-cell isolation was performed on another part of the pathological sections to obtain single-cell gene mutation data, and the positional mapping relationship of single cells in the pathological images was established. S2. After denoising the high-resolution pathological image obtained in step S1, extract cell contours; extract cell pathological features based on the extracted cell contours, and standardize all extracted cell pathological features to obtain a standardized cell-level pathological feature dataset. S3. Perform quality control on the single-cell gene mutation data obtained in step S1 to obtain high-quality mutation data; obtain a single-cell gene mutation dataset based on the high-quality mutation data, and perform genotyping on the single cells; S4. Based on the positional mapping relationship of single cells in pathological images established in step S1, associate the standardized cell-level pathological feature dataset obtained in step S2 with the single-cell gene mutation dataset obtained in step S3 to obtain "pathological feature-gene subtype" paired data. S5. Perform cluster analysis on the standardized cell-level pathological feature dataset obtained in step S2 to obtain cell-level feature clusters; The "pathological features-genotype" paired data obtained in step S4 are jointly optimized by cell-level feature clustering and single-cell genotyping obtained in step S3 to obtain a joint optimization model. S6. Improve the joint optimization model obtained in step S5 to obtain a cell-level pathological feature and gene mutation data matching model.

2. The method for constructing a cell-level pathological feature and gene mutation data matching model according to claim 1, characterized in that, In step S1, high-resolution pathological images are acquired using a digital pathology scanner after HE staining. Single-cell separation is performed using laser capture microdissection or microfluidic technology; The gene mutation data includes one or both of the following: whole-genome mutation data of a single cell or targeted gene mutation data of a single cell.

3. The method for constructing a cell-level pathological feature and gene mutation data matching model according to claim 1, characterized in that, In step S2, a multi-threshold fusion algorithm is used to extract cell contours; A dynamic scaling algorithm was used to standardize all extracted cytopathological features. The cytopathological features include cell morphology, nucleolar features, and chromatin distribution features.

4. The method for constructing a cell-level pathological feature and gene mutation data matching model according to claim 1, characterized in that, In step S3, the clinical relevance scoring algorithm is used to annotate the high-quality mutation data to obtain annotation results, and a single-cell gene mutation dataset is established based on the annotation results. Preliminary genotyping of single cells is performed using mutation weight scores.

5. The method for constructing a cell-level pathological feature and gene mutation data matching model according to claim 1, characterized in that, In step S5, the density-distance fusion algorithm is used to perform cluster analysis on the standardized cell-level pathological feature dataset.

6. The method for constructing a cell-level pathological feature and gene mutation data matching model according to claim 1, characterized in that, In step S6, the optimal parameters are determined with the goal of maximizing clustering-genotyping consistency, and the joint optimization model is improved by using an improved support vector machine algorithm based on feature-gene association weighting and rare subtype loss optimization.

7. A cell-level pathological feature and gene mutation data matching model, characterized in that, It is constructed by any of the methods described in claims 1 to 6.

8. The application of the cell-level pathological features and gene mutation data matching model as described in claim 7 in calculating the matching degree between drugs and gene mutation subtypes.

9. The application according to claim 8, characterized in that, The method for calculating the matching degree between a drug and a gene mutation subtype includes the following steps: Based on the cell-level pathological features and gene mutation data matching model described in claim 7, a multi-factor matching algorithm based on gene mutation subtype, cell number proportion and drug sensitivity is used to establish a correspondence library between renal cell carcinoma gene mutation subtypes and targeted drugs. At the same time, a sensitivity score is assigned to each drug, and the matching degree between the drug and the gene mutation subtype is further calculated. When the matching score is ≥3.5, the drug is considered a priority recommended drug for the corresponding gene mutation subtype; when 2.5 < matching score <3.5, it is considered an alternative recommended drug; when the matching score is ≤2.5, it is considered a non-recommended drug.

10. The application according to claim 9, characterized in that, When scoring sensitivity, 5 points are given for ORR ≥ 50%, 4 points for ORR < 50% for ORR ≤ 30% for ORR < 30% for ORR < 2 ... When calculating the matching degree between a drug and a gene mutation subtype, the drug’s sensitivity score to the corresponding subtype is used as the basis, multiplied by the weight of the percentage of cells and the weight of the subtype frequency difference regulation term, and then the two results are added together to obtain the final matching degree score. The subtype frequency difference adjustment term is calculated by subtracting the absolute value of the proportion of the subtype in the lesion and the average frequency of occurrence in patients from 1. It is used to correct the impact of the difference between the proportion of lesion subtypes and the general situation on the treatment effect.