A fine-grained classification method for skin basal cell carcinoma based on cluster analysis algorithm

Abstract

The present invention proposes a fine-grained classification method for basal cell carcinoma of the skin based on a cluster analysis algorithm. The present invention combines the collected multi-source, multi-modal, and massive basal cell carcinoma data with the kmeans clustering method, and divides the data into K non-overlapping clusters through iterative calculations, providing a feasible solution for mining fine-grained classification of basal cell carcinoma of the skin. Through fine-grained classification, doctors can more accurately understand the disease status of patients with basal cell carcinoma of the skin, provide important support for precision medicine of basal cell carcinoma of the skin, help doctors choose targeted treatment plans that are more suitable for patients, improve treatment effects and reduce unnecessary side effects, and also help monitor the recurrence and metastasis of the disease, timely discover and take corresponding intervention measures, and provide new tools and methods for medical research and clinical practice.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and in particular to a fine-grained classification method for skin basal cell carcinoma based on a cluster analysis algorithm. Background Art

[0002] In the medical field, "precise classification" of diseases is the key to accurately determine the degree of disease progression, develop precise clinical treatment plans, and improve patient prognosis. The ideal classification method for a disease should be able to accurately reflect the etiology and biological characteristics, and be related to clinical behavior and treatment requirements. Basal cell carcinoma of the skin is one of the most common skin malignancies. Currently, the traditional clinical lesions of basal cell carcinoma of the skin are divided into 5 types. Limited by human subjectivity and limited knowledge, this classification method has not fully referred to all known clinical characteristics of the disease, and the clinical pathway boundaries overlap a lot. It is a coarse-grained classification method that cannot truly reflect the heterogeneity of tumors, cannot explain the correlation between the various factors affecting the classification, and is difficult to adapt to the needs of clinical individualized diagnosis and treatment.

[0003] With the recognition of accurate classification and the development of artificial intelligence technology, AI can combine multi-source, massive disease data with high-performance computing and data mining, providing a method for achieving fine-grained classification of basal cell carcinoma of the skin. Fine-grained classification has been a research hotspot in the fields of computer vision and pattern recognition in recent years. It focuses on distinguishing subtle differences between different categories. Combining visual information with text description information, AI achieves a more refined classification by capturing subtle differences. Fine-grained classification of basal cell carcinoma of the skin refers to the further subdivision of multiple specific subcategories within a broad category, especially in biological manifestations such as cell morphology, tissue structure, molecular markers and other aspects to identify and extract subtle differences between subcategories, thereby achieving guidance for disease treatment and prognosis.

[0004] The data-driven AI model based on cluster analysis is one of the most widely used methods. It can comprehensively hierarchically process multimodal disease information, produce finer-grained clustering than existing classifications, and provide undiscovered correlations between symptoms and disease evolution. At present, good sub-classification performance has been achieved in the research of congestive heart failure, new coronavirus pneumonia, sarcoidosis, eosinophilic gastrointestinal diseases, and dermatomyositis, which helps doctors make faster and more accurate clinical decisions. However, the above-mentioned fine-grained classification method has not yet considered the inclusion of all clinical characteristics of the disease (case attributes and dimensions) for classification, and there is no fine-grained classification method suitable for basal cell carcinoma of the skin.

[0005] As of the end of 2022, the number of new cases of basal cell carcinoma of the skin has reached 10 million cases per year worldwide, making it the skin malignancy with the highest incidence rate. From 1990 to 2019, the number of new cases of basal cell carcinoma in China has increased from 40,000 cases / year to 240,000 cases / year, showing a rapid growth trend. It has attracted much attention due to its huge harm and rapid development. Basal cell carcinoma has an insidious onset and diverse clinical manifestations. It is easy to misdiagnose, miss, and misdiagnose, resulting in complex disease evolution, and ultimately leading to locally advanced basal cell carcinoma and metastatic basal cell carcinoma, severe organ damage and dysfunction, and even death and other complications, which seriously endanger the patient's physical and mental health. A large number of clinical practices have shown that "accurate typing" is the key to improving the prognosis of basal cell carcinoma, accurately judging the degree of tumor progression, and formulating accurate clinical treatment plans. It is also an effective way to solve problems such as misdiagnosis and mistreatment.

[0006] Therefore, developing a fine-grained classification method for cutaneous basal cell carcinoma has important clinical value. Summary of the invention

[0007] In order to solve the above problems, the present invention proposes a fine-grained classification method for basal cell carcinoma of the skin based on a cluster analysis algorithm. The technical solution adopted by the present invention is:

[0008] The present invention provides a fine-grained classification method for basal cell carcinoma of the skin based on a cluster analysis algorithm, and the specific steps of the method are as follows:

[0009] S1. Construct a multimodal normative dataset of skin basal cell carcinoma and perform data preprocessing.

[0010] S1-1 collects and integrates skin basal cell carcinoma data from different medical institutions, retains only case data confirmed as skin basal cell carcinoma, and excludes case data with unclear or controversial diagnoses; the case data is multimodal data, including case text data and medical image data;

[0011] The case text data includes: ① Demographic information, including natural conditions: basic information of the patient, hobbies, etc.; Environmental factors: living environment, working environment, etc. of the patient. ② Clinical medical records, including chief complaint: symptoms, course of disease, etc.; Medical history: current medical history, past medical history, family history, etc.; Vital signs: blood pressure, heart rate, blood oxygen, respiration; Physical examination: comprehensive clinical physical examination; Laboratory examination: blood type, blood routine, tumor markers, etc.

[0012] Medical image data includes: ①Clinical images, including clinical manifestations and classification, etc. ②Pathological images, including pathological manifestations and classification, etc. ③Imaging images, including dermatoscope, skin CT, skin ultrasound, skin OCT, skin RCM, etc.

[0013] S1-2 performs data preprocessing and data enhancement to ensure data quality and consistency.

[0014] For case text data, mean filling and regression algorithms were used to handle missing values, and normalization methods were used for feature scaling to ensure that the features had the same scale; spatial transformation algorithms were used for image data enhancement, and images were transformed through rotation, horizontal flipping, and shearing transformations.

[0015] S2. Use k-means algorithm to mine fine-grained classification of skin basal cell carcinoma and verify its accuracy.

[0016] S2-1 relies on the multimodal normative data set of basal cell carcinoma of the skin. First, the data features in the data set are divided into three types: logical, categorical and quantitative. The logical type is converted into 0, 1, the categorical type is converted into 0, 1, 2, 3,..., and the quantitative type is standardized and converted into R+; then the above processed data is used to construct the clinical feature vector of basal cell carcinoma of the skin X = {k1x1, k2x2, k3x3,...}, where x1, x2, x3,... represent the three types of logical, categorical and quantitative data after standardized processing, k1, k2, k3,... represent the weight value corresponding to the feature, and the weight value is set by performing multivariate regression analysis on Y = k1x1+k2x2+k3x3...

[0017] S2-2 uses the k-means algorithm based on Euclidean distance to perform cluster analysis on the feature vector X, to mine and summarize the statistical distribution of the characteristics of basal cell carcinoma of the skin, that is, the fine-grained classification of basal cell carcinoma of the skin; the optimal number of clusters k in the k-means algorithm is determined using the Elbow method.

[0018] S2-3 Use confirmatory factor analysis, hybrid cluster factor analysis, and / or FP-growth algorithm to verify the accuracy of cluster analysis;

[0019] Among them, confirmatory factor analysis (CFA) is used to evaluate and verify the fitness of the measurement model to determine the relationship between the observed variables and the latent factors. The fit between the model and the data is evaluated by the fit index. The cluster analysis model with a high fit is also more accurate.

[0020] Mixture Clustering Factor Analysis (MCFA) was used to combine cluster analysis and factor analysis methods, while considering the similarity between samples and the correlation between variables. Fitting indicators were used to evaluate the fit between the model and the data, including chi-square fit index, CFI and RMSEA.

[0021] The FP-growth algorithm is used to analyze the frequently occurring patterns and item sets in the data set, and the mined frequent patterns and association rules are applied to data analysis, while verifying the accuracy of clustering analysis.

[0022] S3. Perform principal component analysis (PCA) or factor analysis (EFA) on the cluster analysis model data that does not meet the accuracy standards, remove outliers that affect the accuracy of the cluster analysis, then re-perform the cluster analysis until the standards are met, and save the cluster analysis model.

[0023] S4. Match the basal cell carcinoma patient data with the classification in the cluster analysis model to obtain the fine-grained classification corresponding to the patient and formulate a precise clinical treatment plan based on the classification

[0024] S5. Verify the reliability and adequacy of fine-grained classification based on the prognosis of patients with basal cell carcinoma of the skin, and adjust the collected factors and characteristics of clinical diagnostic classification according to the verification results.

[0025] The specific process of the k-means algorithm in step S2-2 is as follows:

[0026] ① Randomly select K samples as the initial cluster center, denoted as μ (1) ,μ (2) ,…μ (k) ; K is a hyperparameter, representing the number of clusters, that is, the number of categories;

[0027] ②For each sample x in the data set (i) , calculate its Euclidean distance to the K cluster centers, and assign it to the class corresponding to the cluster center with the smallest distance In the process, a clustering division is completed;

[0028]

[0029] ③ For each cluster, recalculate its cluster center position. If the geometric center does not coincide with the cluster center, use the geometric center as the new cluster center and re-divide the clusters.

[0030]

[0031] ④ Repeat the above steps ② and ③ until the center position of the cluster remains unchanged, stop and output the clustering results.

[0032] The steps of using the Elbow method to obtain the optimal number of clusters K in step S2-2 are:

[0033] ① Draw the clustering performance indicators under different numbers of clusters into a line graph.

[0034] ②Observe the line chart and find an inflection point, namely the Elbow point.

[0035] ③The cluster number K corresponding to the Elbow point is the best choice.

[0036] The specific steps of using the FP-growth algorithm to analyze frequently occurring patterns and item sets in the data set in step S2-3 are:

[0037] ①Traverse the data set to count the frequency of each item and build the FP tree and the nodes of the FP tree.

[0038] ② For each transaction, the items are sorted by frequency and inserted into the FP tree. Each transaction starts from the root node and is inserted or extended level by level according to the order of the items.

[0039] ③ For each frequent item, construct its conditional pattern base, that is, the set of paths that contain the item. For each frequent item's conditional pattern base, recursively construct a conditional FP tree.

[0040] ④Trace back from the leaf nodes of the FP tree to build frequent item sets. For each conditional FP tree of frequent items, repeat steps ①②③ until no more frequent item sets can be built.

[0041] ⑤ Based on the constructed FP tree and frequent item sets, frequent patterns and corresponding support can be mined. Based on the mined frequent patterns, association rules are generated and confidence is calculated.

[0042] ⑥ Prune the generated frequent item sets, remove items with low support, reduce the number of rules and improve the quality of rules.

[0043] Step S2 also includes step S2-4, which is a visualization display of fine-grained typing results. The visualization display diagram of the fine-grained typing results is a scatter plot, where each point represents a sample point in a data set, and points of different colors represent different clusters; and a large point of a specified color represents the cluster center point of each cluster; the x-axis and y-axis represent two features in the data set respectively; if the number of features is greater than 2, principal component analysis (PCA) is used to reduce the dimension to 2 features and then compose the diagram.

[0044] Step S3 performs principal component analysis (PCA) or factor analysis (EFA) on the cluster analysis model data whose accuracy does not meet the standard, and removes outliers that affect the accuracy of the cluster analysis. The specific steps are as follows:

[0045] The specific steps of S3-1 principal component analysis PCA are as follows:

[0046] (1) Standardized data:

[0047] For each feature of the original data, the mean is subtracted and then divided by the standard deviation to ensure that each feature has the same importance.

[0048] (2) Calculate the covariance matrix:

[0049] Calculate the covariance matrix of the standardized dataset to understand the correlation between different features.

[0050] (3) Calculate eigenvalues ​​and eigenvectors:

[0051] Perform eigenvalue decomposition on the covariance matrix to obtain eigenvalues ​​and corresponding eigenvectors. The eigenvectors represent the principal component directions in the data set, while the eigenvalues ​​represent the variance of the data in these principal component directions.

[0052] (4) Select the number of principal components:

[0053] The number of principal components to be retained is selected based on the size of the eigenvalues. Usually the number of principal components selected to be retained is the number of principal components that can explain a larger part of the data variance.

[0054] (5) Construct the projection matrix:

[0055] Select the number of principal components selected in the previous step and construct the projection matrix. The projection matrix is ​​a matrix composed of eigenvectors and is used to project the original data into the new principal component space.

[0056] (6) Dimensionality reduction transformation:

[0057] Multiplying the normalized dataset by the projection matrix can transform the dataset from the original high-dimensional space to the new low-dimensional principal component space.

[0058] The specific steps of S3-2 factor analysis EFA are as follows:

[0059] (1) Standardized data:

[0060] Data were standardized using z-score standardization or minimum and maximum scaling methods to ensure that all variables had the same scale.

[0061] (2) Extraction factors:

[0062] Statistical software is used to calculate and obtain factor loading matrices, commonality matrices and other results. These results can reveal the relationship between factors and variables and the internal structure of factors.

[0063] (3) Determine the number of factors:

[0064] The Kaiser criterion, Scree plot and explained cumulative variance were used to explain the variance contribution of each factor and the cumulative variance contribution to determine the number of factors to be retained.

[0065] (4) Factor rotation:

[0066] Varimax and Promax were used to rotate the extracted factors to make them easier to interpret.

[0067] (5) Factor explanation:

[0068] Analyze the rotated factor loading matrix, determine the relationship between each variable and factor according to the absolute value and sign of the loading, and give the factor a reasonable explanation.

[0069] (6) Factor score calculation:

[0070] The score of each sample on each factor is calculated to further analyze the position and distribution of the sample in the factor space.

[0071] The present invention combines the collected multi-source, multi-modal, massive skin basal cell carcinoma data with the kmeans clustering method. The k-means algorithm is an unsupervised clustering algorithm that can divide the data into k non-overlapping clusters through iterative calculations, so that the sum of the distances between each data point and the center of the cluster to which it belongs is minimized, providing a feasible solution for mining the fine-grained typing of skin basal cell carcinoma. Through fine-grained typing, doctors can more accurately understand the disease status of patients with skin basal cell carcinoma, thereby selecting a more targeted treatment plan that is more suitable for the patient, improving the treatment effect and reducing unnecessary side effects. For example, different surgical methods, drug treatments, or radiotherapy strategies may be required for different subtypes of skin basal cell carcinoma. In addition, fine-grained typing also helps to monitor the recurrence and metastasis of the disease, and promptly detect and take corresponding intervention measures.

[0072] This invention proposes a fine-grained classification method for basal cell carcinoma of the skin based on a cluster analysis algorithm, combined with medical data analysis, to achieve more accurate and objective classification of basal cell carcinoma of the skin. According to the characteristics and conditions of different patients, personalized fine-grained classification and prediction will provide important support for precision medicine of basal cell carcinoma of the skin, help doctors formulate individualized treatment plans, improve the accuracy of treatment effects and prognosis assessments, improve the quality of life of patients, promote the development of personalized treatment and precision medicine, and provide new tools and methods for medical research and clinical practice. BRIEF DESCRIPTION OF THE DRAWINGS

[0073] Figure 1 It is a schematic diagram of the process of the present invention;

[0074] Figure 2 Kmeans algorithm flow chart;

[0075] Figure 3 Visualization image of cluster analysis results. DETAILED DESCRIPTION

[0076] The technical solution of the present invention is further explained and illustrated in the form of specific embodiments below.

[0077] In this embodiment, a fine-grained classification method for basal cell carcinoma of the skin based on a cluster analysis algorithm is provided. The specific steps of the method are as follows:

[0078] S1. Construct a multimodal normative dataset of skin basal cell carcinoma and perform data preprocessing.

[0079] S1-1 relies on a multi-center basal cell carcinoma data resource integration platform to collect and integrate basal cell carcinoma data from different medical institutions.

[0080] A multi-center skin basal cell carcinoma data resource integration platform was established, and the platform and client were deployed in 1+N mode. A data collection and aggregation network was established to collect multi-source, massive case text data and image data of skin basal cell carcinoma patients, and obtain multi-modal original data, layered data, and thematic data of skin basal cell carcinoma.

[0081] [Inclusion Criteria] All cases were diagnosed with basal cell carcinoma of the skin by histopathology; clinical images of skin lesions were routinely taken for each patient; detailed information on the patients and skin lesions was complete.

[0082] [Exclusion criteria] The diagnosis in the pathology report is unclear or controversial; the clinical information is missing, wrong, or incomplete; the imaging data or clinical images do not meet the requirements, such as low image clarity, out-of-focus or blurred images, etc.

[0083] Clinical data resources were combined with web crawler tools to obtain skin basal cell carcinoma literature, guidelines and consensus, forming an original database of skin basal cell carcinoma.

[0084] S1-2 trains the entity extraction model, relationship extraction model and preprocessing model of the knowledge graph of basal cell carcinoma of the skin, mines the knowledge graph of basal cell carcinoma of the skin, and uses the pattern layer of the knowledge graph to guide the construction of basal cell carcinoma of the skin data resources.

[0085] S1-3 Construction of a multimodal normative dataset for basal cell carcinoma of the skin.

[0086] In the skin basal cell carcinoma data resources, case text data and medical image data exist in a one-to-many (1:m) relationship. This relationship is mapped to the skin basal cell carcinoma standard dataset to ensure entity alignment and knowledge fusion, integrate case text data and medical image data to form a multimodal dataset that meets the requirements of balance, uniformity, and variance homogeneity.

[0087] Multimodal data includes:

[0088] 1) Case text data:

[0089] ① Demographic information, including natural conditions: basic information of the patient, hobbies, etc.; environmental factors: living environment, working environment, etc. of the patient. ② Clinical medical records, including chief complaint: symptoms, course of disease, etc.; medical history: current medical history, past medical history, family history, etc.; vital signs: blood pressure, heart rate, blood oxygen, respiration; physical examination: comprehensive clinical physical examination; laboratory examination: blood type, blood routine, tumor markers, etc.

[0090] 2) Medical image data: ① Clinical images, including clinical manifestations and classification, etc. ② Pathological images, including pathological manifestations and classification, etc. ③ Imaging images, including dermatoscope, skin CT, skin ultrasound, skin OCT, skin RCM, etc.

[0091] S1-4 performs data preprocessing and data enhancement to ensure data quality and consistency.

[0092] For case text data, mean value filling and regression algorithms were used to handle missing values, and normalization methods were used for feature scaling to ensure that the features had the same scale. A spatial transformation algorithm was used for image data enhancement, and the image was transformed by 180° rotation, horizontal flipping, and shearing transformation, with a transformation coefficient of 0.2.

[0093] S2. Use k-means algorithm to mine fine-grained classification of skin basal cell carcinoma, and verify its accuracy and visualize it (e.g. Figure 2 shown).

[0094] S2-1 relies on the multimodal normative dataset of basal cell carcinoma of the skin, refers to the current clinical classification, and combines the clinical characteristics of basal cell carcinoma of the skin obtained in previous studies (Table 1) to construct the clinical feature vector of basal cell carcinoma of the skin;

[0095] First, the clinical characteristics in Table 1 were divided into three types: logical, categorical, and quantitative. The logical type was converted into 0, 1, the categorical type was converted into 0, 1, 2, 3, …, and the quantitative type was standardized and converted into R+.

[0096] For example, "Do you smoke?" is described by logical type, yes: 1, no: 0

[0097] "Blood type" is described by classification type, A: 0, B: 1, AB: 2, O: 3

[0098] "Age" is described quantitatively, and the data is mapped to [0, 1] using the following formula:

[0099]

[0100] Then, the processed data are used to construct the clinical feature vector X={k1x1, k2x2, k3x3, ...} of basal cell carcinoma of the skin, where x1, x2, x3, ... represent the three types of logical, classification and quantitative data after standardization processing, and k1, k2, k3, ... represent the weight value corresponding to the feature. The weight value is set by performing multiple regression analysis on Y=k1x1+k2x2+k3x3...

[0101] Table 1: Detailed list of factors and characteristics affecting the clinical diagnosis and classification of basal cell carcinoma of the skin

[0102]

[0103]

[0104] S2-2 is based on Euclidean distance and uses the k-means algorithm, which has good interpretability and can refer to subjective clustering inference intentions, to cluster X, mine and summarize the statistical distribution of basal cell carcinoma characteristics, which is the fine-grained classification of basal cell carcinoma.

[0105] (1) k-means algorithm process

[0106] ① Randomly select K samples as the initial cluster center, denoted as μ (1) ,μ (2) ,…μ (k) ; K is a hyperparameter, representing the number of clusters, that is, the number of categories;

[0107] ②For each sample x in the data set (i) , calculate its Euclidean distance to the K cluster centers, and assign it to the class corresponding to the cluster center with the smallest distance In the process, a clustering division is completed;

[0108]

[0109] ③ For each cluster, recalculate its cluster center position. If the geometric center does not coincide with the cluster center, use the geometric center as the new cluster center and re-divide the clusters.

[0110]

[0111] ④ Repeat the above steps ② and ③ until the center position of the cluster remains unchanged, stop and output the clustering results.

[0112] (2) Use the Elbow method to determine the optimal number of clusters k. Elbow is the elbow method, which is usually used to help determine the optimal number of clusters or hyperparameters. The basic idea is that as the number of clusters gradually increases, the performance of the clusters (such as the compactness within the clusters) will gradually improve. However, when the number of clusters increases to a certain extent, the performance improvement will slow down, forming an inflection point, just like the elbow of an arm. The number of clusters corresponding to this inflection point is the best choice.

[0113] ① Draw the clustering performance indicators under different numbers of clusters into a line graph.

[0114] ②Observe the line chart and find an inflection point, namely the Elbow point.

[0115] ③The number of clusters corresponding to the Elbow point is the best choice.

[0116] S2-3 Use confirmatory factor analysis, mixed cluster factor analysis or FP-growth algorithm to verify the accuracy of cluster analysis.

[0117] (1) Use confirmatory factor analysis (CFA) to evaluate and verify the goodness of fit of the measurement model to determine the relationship between observed variables and latent factors. CFA is often used to test whether a theoretical model can explain the variability of observed data and whether the observed variables can measure the latent factors well.

[0118] ① Identify the potential factors involved in the study and the hypotheses about their relationship with the observed variables. These hypotheses are usually based on theory or previous research.

[0119] ②Construct measurement model: According to the research hypothesis, establish the relationship model between latent factors and observed variables. In CFA, it is usually assumed that each latent factor is associated with a set of observed variables, and the measurement model includes the correlation between latent factors.

[0120] ③ Collect data: Collect samples containing observed variable data.

[0121] ④ Evaluate the degree of fit between the model and the data. Use the silhouette coefficient, DBI, CH and other fitting indicators to measure the matching degree between categories, the similarity of samples within a class, the difference of samples between classes, the distance between classes and the ratio of distance within a class.

[0122] ⑤Evaluate model fit: Evaluate the fit between the model and the data based on the fit index. If the model does not fit well, you may need to modify the model or consider other factors.

[0123] ⑥Interpret the results: Based on the model fitting results, explain the relationship between potential factors and observed variables and verify the research hypothesis.

[0124] (2) Mixture Clustering Factor Analysis (MCFA) was used to combine cluster analysis and factor analysis methods, while taking into account the similarity between samples and the correlation between variables.

[0125] ① Use fitting indicators (chi-square fit index, CFI, RMSEA, etc.) to evaluate the fit between the model and the data.

[0126] ② Based on the model fitting results, explain the differences between groups and the relationship between potential factors. Through mixed cluster factor analysis, explore the group structure and potential factor structure at the same time.

[0127] (3) Analyze frequently occurring patterns and item sets in the dataset based on the FP-growth algorithm.

[0128] ①Traverse the data set to count the frequency of each item and build the FP tree and the nodes of the FP tree.

[0129] ② For each transaction, the items are sorted by frequency and inserted into the FP tree. Each transaction starts from the root node and is inserted or extended level by level according to the order of the items.

[0130] ③ For each frequent item, construct its conditional pattern base, that is, the set of paths that contain the item. For each frequent item's conditional pattern base, recursively construct a conditional FP tree.

[0131] ④Trace back from the leaf nodes of the FP tree to build frequent item sets. For each conditional FP tree of frequent items, repeat steps ①②③ until no more frequent item sets can be built.

[0132] ⑤ Based on the constructed FP tree and frequent item sets, frequent patterns and corresponding support can be mined. Based on the mined frequent patterns, association rules are generated and confidence is calculated.

[0133] ⑥ Prune the generated frequent item sets, remove items with low support, reduce the number of rules and improve the quality of rules.

[0134] ⑦ Apply the mined frequent patterns and association rules to data analysis, and verify the accuracy of cluster analysis.

[0135] S2-4 Visual display and interpretation of fine-grained typing results.

[0136] In the above k-means clustering visualization, there are several important pieces of information that need to be explained:

[0137] (1) Scatter plot:

[0138] The scatter plot shows the sample points in the data set, and each point represents a data sample;

[0139] Points of different colors represent different clusters, and points of the same color belong to the same cluster.

[0140] By distinguishing colors, you can intuitively see the clustering effect of the k-means algorithm on the data set.

[0141] (2) Cluster center:

[0142] The large dots of the specified color represent the cluster centers of each cluster.

[0143] The cluster center point is the average point of each cluster and represents the center position of the cluster.

[0144] The location of the cluster center depends on the results calculated by the k-means algorithm when clustering the data.

[0145] (3) Coordinate axis:

[0146] The x-axis and y-axis represent two features in the dataset (Feature 1 and Feature 2). If the number of features is greater than 2, PCA is used to reduce the dimension to 2 feature structures.

[0147] The scales on the coordinate axes represent different numerical ranges, which helps to understand the distribution of data points in the feature space.

[0148] Through this visualization, we can clearly see how the sample points in the data set are divided into different clusters by the k-means algorithm. The location of the cluster center and the distribution of the clusters can also be intuitively displayed, helping us understand the effect of the clustering algorithm and the structure of the data set.

[0149] S3. Perform principal component analysis (PCA) or factor analysis (EFA) on the cluster analysis model data that does not meet the accuracy standards, remove outliers that affect the accuracy of the cluster analysis, then re-perform the cluster analysis until the standards are met, and save the cluster analysis model.

[0150] S3-1 Principal Component Analysis PCA is a data dimensionality reduction technique and feature extraction method. In the process of converting high-dimensional data into low-dimensional data, some highly correlated original variables may be merged into the same principal component, thereby simplifying the data structure. This helps us see more clearly which features are highly correlated and the impact of these correlated features on the overall data.

[0151] (1) Standardized data:

[0152] For each feature of the original data, the mean is subtracted and then divided by the standard deviation to ensure that each feature has the same importance.

[0153] (2) Calculate the covariance matrix:

[0154] Calculate the covariance matrix of the standardized dataset to understand the correlation between different features.

[0155] (3) Calculate eigenvalues ​​and eigenvectors:

[0156] Perform eigenvalue decomposition on the covariance matrix to obtain eigenvalues ​​and corresponding eigenvectors. The eigenvectors represent the principal component directions in the data set, while the eigenvalues ​​represent the variance of the data in these principal component directions.

[0157] (4) Select the number of principal components:

[0158] The number of principal components to be retained is selected based on the size of the eigenvalues. Usually the number of principal components selected to be retained is the number of principal components that can explain a larger part of the data variance.

[0159] (5) Construct the projection matrix:

[0160] Select the number of principal components selected in the previous step and construct the projection matrix. The projection matrix is ​​a matrix composed of eigenvectors and is used to project the original data into the new principal component space.

[0161] ⑥ Dimensionality reduction transformation:

[0162] Multiplying the normalized dataset by the projection matrix can transform the dataset from the original high-dimensional space to the new low-dimensional principal component space.

[0163] S3-2 Factor Analysis EFA is a statistical technique used to explore the potential factor structure in a data set, which can reveal the potential structure and relationships in the data set, thereby better understanding the patterns and laws behind the data.

[0164] (1) Standardized data:

[0165] Data were standardized using z-score standardization or minimum and maximum scaling methods to ensure that all variables had the same scale.

[0166] (2) Extraction factors:

[0167] Statistical software is used to calculate and obtain factor loading matrices, commonality matrices and other results. These results can reveal the relationship between factors and variables and the internal structure of factors.

[0168] (3) Determine the number of factors:

[0169] The Kaiser criterion, Scree plot and explained cumulative variance were used to explain the variance contribution of each factor and the cumulative variance contribution to determine the number of factors to be retained.

[0170] (4) Factor rotation:

[0171] Varimax and Promax were used to rotate the extracted factors to make them easier to interpret.

[0172] (5) Factor explanation:

[0173] Analyze the rotated factor loading matrix, determine the relationship between each variable and factor according to the absolute value and sign of the loading, and give the factor a reasonable explanation.

[0174] (6) Factor score calculation:

[0175] Calculate the score of each sample on each factor to further analyze the position and distribution of the sample in the factor space.

[0176] S4. Match the basal cell carcinoma patient data with the classification in the cluster analysis model to obtain the fine-grained classification corresponding to the patient and formulate a precise clinical treatment plan based on the classification

[0177] S5. Verify the reliability and adequacy of fine-grained classification based on the prognosis of patients with basal cell carcinoma of the skin, and adjust the collected factors and characteristics of clinical diagnostic classification according to the verification results.

Claims

1. A fine-grained classification method for basal cell carcinoma of the skin based on a cluster analysis algorithm, characterized in that: The specific steps of this method are as follows: S1. Construct a multimodal normative dataset of skin basal cell carcinoma and perform data preprocessing; S1-1 collects and integrates skin basal cell carcinoma data from different medical institutions, retains only case data confirmed as skin basal cell carcinoma, and excludes case data with unclear or controversial diagnoses; the case data is multimodal data, including case text data and medical image data; The case text data includes: ① demographic information, including natural conditions: basic information and hobbies of the patient; environmental factors: the patient's living environment and work environment; ② clinical medical records, including chief complaints: symptoms and course of disease; medical history: current medical history, past medical history and family history; Vital signs: blood pressure, heart rate, blood oxygen and respiration; physical examination: comprehensive clinical examination; laboratory tests: blood type, blood routine and tumor markers; Medical image data includes: ① clinical images, including clinical manifestations and classification; ② pathological images, including pathological manifestations and classification; ③ imaging images, including dermatoscope, skin CT, skin ultrasound, skin OCT and skin RCM; S1-2 performs data preprocessing and data enhancement to ensure data quality and consistency; For case text data, mean value filling and regression algorithms were used to handle missing values, and normalization methods were used for feature scaling to ensure that the features had the same scale. A spatial transformation algorithm was used for image data enhancement, and the images were transformed by rotation, horizontal flipping, and shearing transformations. S2. Use k-means algorithm to mine fine-grained classification of skin basal cell carcinoma and verify its accuracy; S2-1 relies on the multimodal normative data set of basal cell carcinoma of the skin. First, the data features in the data set are divided into three types: logic, classification and quantitative. The logical type is converted to 0, 1, the classification type is converted to 0, 1, 2, 3, ..., and the quantitative type is standardized and converted to R+; then the clinical feature vector of basal cell carcinoma of the skin is constructed using the above processed data X = {k1x1, k2x2, k3x3, ...}, where x1, x2, x3, ... represent the three types of logic, classification and quantitative data after standardization, k1, k2, k3, ... represent the weight value corresponding to the feature, and the weight value is set by performing multivariate regression analysis on Y = k1x1+k2x2+k3x3...; S2-2 uses the k-means algorithm based on Euclidean distance to perform cluster analysis on the feature vector X, and mines and summarizes the statistical distribution of the characteristics of basal cell carcinoma of the skin, which is the fine-grained classification of basal cell carcinoma of the skin; the optimal number of clusters k in the k-means algorithm is determined using the Elbow method; S2-3 Use confirmatory factor analysis, hybrid cluster factor analysis, and / or FP-growth algorithm to verify the accuracy of cluster analysis; Among them, confirmatory factor analysis is used to evaluate and verify the fitness of the measurement model to determine the relationship between the observed variables and the latent factors. The fit between the model and the data is evaluated by the fit index. The cluster analysis model with a high fit is also highly accurate. A hybrid cluster factor analysis was used to combine cluster analysis and factor analysis, taking into account the similarity between samples and the correlation between variables. Fit indices were used to evaluate the fit between the model and the data, including chi-square fit index, CFI, and RMSEA. Use the FP-growth algorithm to analyze the frequently occurring patterns and item sets in the data set, apply the mined frequent patterns and association rules to data analysis, and verify the accuracy of cluster analysis; S3, performing principal component analysis or factor analysis on the cluster analysis model data whose accuracy does not meet the standard, eliminating outliers that affect the accuracy of the cluster analysis, and then re-performing the cluster analysis until the standard is met, and saving the cluster analysis model; S4. Match the basal cell carcinoma patient data with the classification in the cluster analysis model to obtain the fine-grained classification corresponding to the patient and formulate a precise clinical treatment plan based on the classification S5. Verify the reliability and adequacy of fine-grained classification based on the prognosis of patients with basal cell carcinoma of the skin, and adjust the collected factors and characteristics of clinical diagnostic classification according to the verification results.

2. The fine-grained classification method for basal cell carcinoma of the skin based on cluster analysis algorithm according to claim 1, characterized in that: The specific process of the k-means algorithm in step S2-2 is: ① Randomly select K samples as the initial cluster center, denoted as μ (1) ,μ (2) ,…μ (k) ; K is a hyperparameter, representing the number of clusters, that is, the number of categories; ②For each sample x in the data set (i) , calculate its Euclidean distance to the K cluster centers, and assign it to the class corresponding to the cluster center with the smallest distance In the process, a clustering division is completed; ③ For each cluster, recalculate its cluster center position. If the geometric center does not coincide with the cluster center, use the geometric center as the new cluster center and re-divide the clusters. ④ Repeat the above steps ② and ③ until the center position of the cluster remains unchanged, stop and output the clustering results.

3. The fine-grained classification method for basal cell carcinoma of the skin based on cluster analysis algorithm according to claim 1, characterized in that: The steps of using the Elbow method to obtain the optimal number of clusters K in step S2-2 are: ① Draw the clustering performance indicators under different cluster numbers into a line graph; ②Observe the line chart and find an inflection point, namely the Elbow point; ③The cluster number K corresponding to the Elbow point is the best choice; The specific steps of using the FP-growth algorithm to analyze frequently occurring patterns and item sets in the data set in step S2-3 are: ①Traverse the data set to count the frequency of each item and build the FP tree and the nodes of the FP tree; ② For each transaction, sort the items by frequency and insert them into the FP tree. Each transaction starts from the root node and is inserted or extended level by level according to the order of the items. ③For each frequent item, construct its conditional pattern base, that is, the path set containing the item; for the conditional pattern base of each frequent item, recursively construct the conditional FP tree; ④Trace back from the leaf nodes of the FP tree to build frequent item sets; for each conditional FP tree of frequent items, repeat steps ①②③ until no more frequent item sets can be built; ⑤ Based on the constructed FP tree and frequent item sets, frequent patterns and corresponding support can be mined; based on the mined frequent patterns, association rules are generated and confidence is calculated; ⑥ Prune the generated frequent item sets, remove items with low support, reduce the number of rules and improve the quality of rules.

4. The fine-grained classification method for basal cell carcinoma of the skin based on cluster analysis algorithm according to claim 1, characterized in that: Step S2 also includes step S2-4 of visually displaying the fine-grained typing results; the visual display diagram of the fine-grained typing results is a scatter plot, each point represents a sample point in a data set, and points of different colors represent different clusters; The large dots of specified colors represent the cluster center points of each cluster; the x-axis and y-axis represent the two features in the data set respectively; If the number of features is greater than 2, principal component analysis is used to reduce the dimension to 2 features and then construct the graph.

5. The fine-grained classification method for basal cell carcinoma of the skin based on cluster analysis algorithm according to claim 1, characterized in that: Step S3 performs principal component analysis or factor analysis on the cluster analysis model data whose accuracy does not meet the standard, and removes outliers that affect the accuracy of the cluster analysis. The specific steps are as follows: The specific steps of S3-1 principal component analysis PCA are as follows: (1) Standardized data: Subtract the mean from each feature of the original data and divide by the standard deviation to ensure that each feature has the same importance; (2) Calculate the covariance matrix: Calculate the covariance matrix of the standardized dataset to understand the correlation between different features; (3) Calculate eigenvalues ​​and eigenvectors: Perform eigenvalue decomposition on the covariance matrix to obtain eigenvalues ​​and corresponding eigenvectors; the eigenvectors represent the principal component directions in the data set, and the eigenvalues ​​represent the variance of the data in these principal component directions; (4) Select the number of principal components: The number of principal components to be retained is selected based on the size of the eigenvalues; usually the number of principal components selected to be retained is the number of principal components that can explain a larger part of the data variance; (5) Construct the projection matrix: Select the number of principal components selected in the previous step and construct a projection matrix; the projection matrix is ​​a matrix composed of eigenvectors used to project the original data into the new principal component space; (6) Dimensionality reduction transformation: Multiplying the standardized data set by the projection matrix can transform the data set from the original high-dimensional space to a new low-dimensional principal component space; The specific steps of S3-2 factor analysis EFA are as follows: (1) Standardized data: Standardize the data using z-score standardization or minimum and maximum scaling methods to ensure that all variables have the same scale; (2) Extraction factors: Statistical software is used to calculate and obtain factor loading matrices and commonality matrix results; these results can reveal the relationship between factors and variables and the internal structure of factors; (3) Determine the number of factors: Kaiser criterion, Scree plot and explained cumulative variance were used to explain the variance contribution of each factor and the cumulative variance contribution to determine the number of factors to be retained; (4) Factor rotation: The extracted factors were rotated using maximization of Varimax and maximum likelihood estimation Promax to make the factors easier to interpret; (5) Factor explanation: Analyze the rotated factor loading matrix, determine the relationship between each variable and factor according to the absolute value and sign of the loading, and give the factor a reasonable explanation; (6) Factor score calculation: Calculate the score of each sample on each factor to further analyze the position and distribution of the sample in the factor space.

Images