Artificial intelligence feature extraction method and system based on lymphoma MICM phenotypic typing

Abstract

The invention belongs to the field of intelligent medical treatment, and particularly relates to an artificial intelligence feature extraction method and system based on lymphoma MICM phenotypic typing. The method comprises the steps of S1, acquiring a diagnosis report of a to-be-tested person; s2, obtaining a structured diagnosis report based on the diagnosis report; s3, vectorizing the structured diagnosis report to obtain a vectorized diagnosis report; s4, the vectorized diagnosis report is input into a Transform based on quantum coding for feature extraction, and MICM phenotype typing features are obtained; and the Transformer based on quantum coding converts input data into a quantum state, then extracts quantum state feature information, converts the quantum state feature information into non-quantum state data, and then outputs the MICM phenotype typing features. According to the method, the required MICM features are extracted through the quantum coding-based Transform, and the capture capability of the model on the data features is enhanced by adopting the quantum coding-based Transform algorithm.

Description

Technical Field

[0001] The present invention relates to the field of intelligent medical treatment, and more specifically, to an artificial intelligence feature extraction method, system, device, medium and program product based on MICM phenotyping of lymphoma. Background Art

[0002] Structured reporting standards are imperative in the era of medical big data. Not only can they establish a good data foundation in the clinical diagnosis and treatment of hematological tumors and scientific research, but they can also serve as a better bridge in the communication process with clinicians and patients, promote the gradual standardization, dataization and intelligence of hematological tumor treatment centers, and carry out in-depth clinical research and innovative treatment methods in the field of hematological tumor treatment. Summary of the invention

[0003] In view of the above problems, the present invention provides an artificial intelligence feature extraction method based on MICM phenotyping of lymphoma, which uses data processing and feature extraction to capture information in time series data, and proposes improvements to the parameters of the disease prediction model and performs effective optimization, thereby constructing a disease warning model suitable for clinical use.

[0004] The present application (first aspect) discloses an artificial intelligence feature extraction method based on MICM phenotyping of lymphoma, comprising:

[0005] S101: Obtaining a diagnosis report of the subject to be tested;

[0006] S102: Obtaining a structured diagnosis report based on the diagnosis report;

[0007] S103: vectorizing the structured diagnostic report to obtain a vectorized diagnostic report;

[0008] S104: The vectorized diagnostic report is input into a quantum coding-based Transformer for feature extraction to obtain MICM phenotypic typing features; the quantum coding-based Transformer converts the input data into a quantum state and then extracts quantum state feature information, converts the quantum state feature information back into non-quantum state data and then outputs the MICM phenotypic typing features.

[0009] Furthermore, the Transformer based on quantum coding includes an input layer, a quantum coding layer, a quantum gate layer and an output layer; the processing steps of the Transformer based on quantum coding include:

[0010] Step 1: The input layer configures an initial quantum state for each vector of the vectorized diagnostic report to obtain a vector of initial quantum states;

[0011] Step 2: The quantum coding layer performs quantum coding on the vector of the initial quantum state to obtain a quantum coded vector;

[0012] Step 3: The quantum-encoded vector is forward-propagated through the quantum gate layer to obtain quantum state characteristics extracted by forward propagation;

[0013] Step 4: the output layer maps the extracted quantum state features back to the non-quantum state to obtain the MICM phenotyping features;

[0014] Optionally, the steps further include a step 3' between step 2 and step 3: optimizing the correlation distance between quantum bits of the quantum-encoded vector by a dynamic quantum topology optimization strategy.

[0015] Furthermore, the manner in which each vector configures the initial quantum state is expressed as:

[0016]

[0017] In the formula, represents the initial quantum state, α i is the complex amplitude, |i> is the ground state of the qubit, and n is the number of qubits used to represent the vector;

[0018] Optionally, the conversion of the quantum coding layer can be expressed as:

[0019] ψ enc =U(θ)|ψ>

[0020] In the formula, ψ enc represents the encoded quantum state, U(θ) is the quantum gate adjusted according to the characteristics of the input data, θ represents the parameters of the quantum gate, and |ψ> is the initial quantum state;

[0021] Optionally, the quantum gate layer is represented as:

[0022] ψ pro =U pro (γ)ψ enc

[0023] In the formula, ψ pro represents the quantum state after forward propagation, U pro (γ) is the quantum gate used in the forward propagation process, and γ is the quantum gate parameter;

[0024] Optionally, the output layer is expressed as:

[0025]

[0026] In the formula, ψ dec represents the MICM phenotyping characteristics obtained after decoding, ψ proIt represents the quantum state characteristic information extracted after forward propagation. is the conjugate transpose of the quantum coding operation, used to map the quantum state characteristic information into the MICM phenotyping characteristics;

[0027] Optionally, the quantum gate U(θ) is expressed as:

[0028] U(θ)=e -iθH

[0029] Where H is the Hamiltonian;

[0030] Optionally, a dynamic quantum topology optimization strategy is used to adjust and obtain an adjusted Hamiltonian, and the quantum gate U(θ) is obtained using the adjusted Hamiltonian;

[0031] Optionally, the dynamic adjustment method of the dynamic quantum topology optimization strategy can be expressed as:

[0032]

[0033] λ jk =softmax(-d jk / τ te )

[0034] In the formula, softmax() is the preset Softmax classification function, H DQTO is the Hamiltonian after dynamic adjustment; λ jk is the coupling strength between qubits j and k, and is the Pauli-Z operation acting on qubits j and k, d jk represents the characteristic correlation distance between qubits j and k, τ te is the temperature parameter.

[0035] Furthermore, the steps of constructing the quantum-coded Transformer include:

[0036] The initial quantum coded Transformer includes an initial input layer, an initial quantum coding layer, an initial quantum gate layer and an initial output layer, wherein the parameters of the initial quantum coding layer and the initial quantum gate layer are randomly initialized parameters;

[0037] After the vectorized diagnostic report of the training set is passed through the initial quantum-coded Transformer to obtain the initial MICM phenotyping features, the model prediction output is obtained based on the initial MICM phenotyping features, the difference between the actual value of the training set and the model prediction output is compared to the loss function, and the loss function is iteratively trained until the stop condition is reached to obtain the quantum-coded Transformer;

[0038] The stopping condition includes reaching a preset maximum number of iterations;

[0039] Optionally, the loss function calculated based on the initial MICM phenotyping features is expressed as:

[0040]

[0041] Where L represents the loss function, i∈[1,m], i represents the i-th sample in the training set, m represents the total number of samples in the training set, and y i is the target output of the i-th sample, ψ dec,i represents the MICM phenotyping characteristics of the i-th sample, f(ψ dec,i ) is the output of the model’s prediction for the i-th sample;

[0042] Optionally, the updating method of the parameters of the initial quantum coding layer is expressed as:

[0043]

[0044] In the formula, θ new and θ old represent the initial quantum coding layer parameters before and after the update, η rate is the learning rate, is the gradient of the loss function with respect to the initial quantum coding layer parameter θ;

[0045] Optionally, the updating method of the parameters of the initial quantum gate layer is expressed as:

[0046]

[0047] In the formula, γ new and γ old denote the parameters of the initial quantum gate layer before and after the update, η2 is the learning rate, is the gradient of the loss function with respect to the initial quantum gate layer parameter γ.

[0048] Optionally, the stopping condition includes that the quantum entanglement degree is less than a preset threshold: after forward propagation through the quantum gate layer to obtain the quantum state characteristics extracted by forward propagation, the entanglement degree between different quantum bits is calculated. When, it indicates that the model converges, then the training is stopped;

[0049] Optionally, the entanglement degree is expressed as:

[0050] E=Tr(ρ A logρ A )

[0051] Where E represents the quantum entanglement degree, ρ Ais the reduced density matrix of system A, which represents the quantum state characteristic information extracted from the previous iterations, and Tr() represents the trace operation.

[0052] Furthermore, the reduced density matrix ρ A It is obtained by trace operation from the total density matrix ρ, which is the characteristic matrix of the quantum state characteristic information extracted in this iteration, and the reduced density matrix ρ A It can be expressed as:

[0053] ρ A =Tr B (ρ)

[0054] In the formula, Tr B () indicates taking a trace of part B of the system, leaving the state of part A of the system. System B is the quantum state characteristic information extracted in this iteration.

[0055] further,

[0056] The diagnostic report includes morphological information, immunological information, cytogenetic information, and molecular information;

[0057] Optionally, the method for obtaining a structured diagnostic report based on the diagnostic report includes using machine vision technology to extract the structured diagnostic report according to a standard report model;

[0058] Optionally, the method of vectorizing the structured diagnostic report to obtain a vectorized diagnostic report includes one or more of the following methods: Word2Vec, Doc2Vec, and BERT.

[0059] Furthermore, the quantum coding-based artificial intelligence feature extraction method based on lymphoma MICM phenotyping also includes S105: obtaining the patient's diagnosis and treatment information based on the MICM phenotyping features;

[0060] Optionally, the method further comprises S105': inputting the MICM phenotyping features into a classifier to obtain a classification result of whether the subject suffers from a blood tumor;

[0061] Optionally, the MICM diagnostic report includes the patient's treatment information and the classification result.

[0062] Optionally, the classifier includes one or more of the following: conditional random field algorithm, support vector machine, decision tree, random forest, logistic regression;

[0063] Furthermore, the classifier is a trained conditional random field, and the method for constructing the trained conditional random field includes:

[0064] Obtaining a diagnostic report training set, wherein the training set also includes a label of whether the patient has a blood tumor;

[0065] Obtaining a MICM phenotyping feature training set based on the diagnostic report training set;

[0066] Initialize the parameters of the conditional random field to obtain an initialized conditional random field;

[0067] Performing Monte Carlo random sampling on the parameters of the initialized conditional random field to obtain sample parameters;

[0068] Obtaining predicted labels of the MICM phenotyping feature training set based on the sample parameters, calculating a loss function of a conditional random field based on the predicted labels of the MICM phenotyping feature training set and the actual labels of the diagnosis report, and gradually minimizing the loss function of the conditional random field through iteration, and obtaining the trained conditional random field after reaching a stop condition;

[0069] Optionally, the loss function of the conditional random field is expressed as:

[0070]

[0071] In the formula, L(q θ ) represents the parameter q based on the conditional random field θ The loss function is is the label data of the i-th sample, q θ (t) is the parameter of the conditional random field obtained by Monte Carlo sampling at the tth iteration in the iterative process, In a given feature and parameter q θ (t) The label prediction probability under L1 is the L1 regularization coefficient, ||q θ ||1 is parameter q θ (t) The L1 regularization term of ;

[0072] Optionally, the sample parameters obtained by the Monte Carlo random sampling are expressed as:

[0073] q θ (t) =q θ (t-1) +∈ (t)

[0074] In the formula, q θ (t) represents the sample parameter of the tth iteration, q θ (t-1) represents the sample parameter of the t-1th iteration, ∈ (t)is the sampling step length;

[0075] Optional, sampling step ∈ (t) For dynamic adaptive settings, the setting method is:

[0076] ∈ (t) =κ cy ·exp(-ρ cy t)

[0077] In the formula, κ cy is the initial step size, ρ cy is the attenuation factor;

[0078] Optionally, after optimizing the sample parameters by adopting a sparse regularization strategy to obtain optimized sample parameters, the optimized sample parameters are used to replace the sample parameters to calculate the loss function of the conditional random field;

[0079] Optionally, the sparse regularization strategy is expressed as:

[0080]

[0081] In the formula, represents the sample parameter optimized at the tth iteration, Δq θ (t) is the sample parameter q in the tth iteration θ (t) The increment of α gz is the learning rate, is the gradient of the loss function of the conditional random field with respect to the parameters, δ gz is an adjustment factor calculated based on the difference between the model output and the actual classification result, sign(q θ (t) ) indicates q θ (t) The symbolic function of

[0082] Optionally, the stopping condition includes that the information gain of the iteration is less than a preset threshold: after each iteration, the information gain is calculated based on the sample parameters of the current conditional random field, and the iteration is stopped when the information gain is less than the preset threshold;

[0083] Optional, Indicates that given a feature and parameter q θ (t) The label prediction probability under the conditional random field is determined, and all the Where c∈[1,C], c represents the classification label, C is all possible categories, and the information gain is expressed as follows:

[0084]

[0085] In the formula, Es i represents the information gain of the i-th sample, is a given feature The probability of category c under the condition of , p(c) is the prior probability of category c.

[0086] The second aspect of the present application discloses an artificial intelligence feature extraction system based on MICM phenotyping of lymphoma, comprising:

[0087] Acquisition module 201: used to obtain a diagnosis report of a subject;

[0088] Structured processing module 202: used to obtain a structured diagnosis report based on the diagnosis report;

[0089] Vectorization processing module 203: used for vectorizing the structured diagnosis report to obtain a vectorized diagnosis report;

[0090] Feature extraction module 204: used for inputting the vectorized diagnostic report into a quantum-coded based Transformer to extract features and obtain MICM phenotypic typing features; the quantum-coded based Transformer converts the input data into a quantum state and then extracts quantum state feature information, converts the quantum state feature information back into non-quantum state data and then outputs the MICM phenotypic typing features.

[0091] The third aspect of the present application discloses a computer device, which includes: a memory and a processor; the memory is used to store program instructions; the processor is used to call the program instructions, and when the program instructions are executed, it is used to execute the steps of the above method.

[0092] A fourth aspect of the present application discloses a computer-readable storage medium having a computer program stored thereon, wherein the computer program implements the steps of the above method when executed by a processor.

[0093] A fifth aspect of the present application discloses a computer program product, including a computer program, which implements the steps of the above method when executed by a processor.

[0094] This application has the following beneficial effects:

[0095] 1. This application extracts the required MICM phenotyping features through a quantum-encoded Transformer, and uses a quantum-encoded Transformer algorithm to enhance the model's ability to capture data features;

[0096] 2. Use dynamic quantum topology optimization strategy to optimize the connection topology between quantum bits to improve the efficiency of quantum coding in the feature extraction process and the information processing capability of the model;

[0097] 3. The present invention also uses a conditional random field algorithm based on Monte Carlo optimization to train the model, and improves processing efficiency and accuracy through a sparse representation strategy;

[0098] 4. Based on the innovative technology of the present invention, the data processing capability can be effectively improved in MICM diagnostic tasks, the model's adaptability to different data characteristics is enhanced, and the classification accuracy and prediction reliability of the model in medical diagnostic tasks are improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0099] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative work.

[0100] Figure 1 is a schematic diagram of a method flow chart provided by the first aspect of an embodiment of the present invention;

[0101] Figure 2 is a schematic diagram of a program product provided by the second aspect of an embodiment of the present invention;

[0102] Figure 3 is a schematic diagram of a computer device provided by an embodiment of the present invention;

[0103] Figure 4 is a schematic diagram of the architecture of an exemplary computing device provided by an embodiment of the present invention;

[0104] Figure 5 is a schematic diagram of a storage medium provided by an embodiment of the present invention;

[0105] Figure 6 It is a schematic diagram of a process of generating and publishing a MICM diagnostic report provided by an embodiment of the present invention;

[0106] Figure 7 It is a schematic diagram of a process of extracting MICM phenotyping features based on a quantum-coded Transformer algorithm provided in an embodiment of the present invention;

[0107] Figure 8 It is a schematic diagram of a training process of a Transformer algorithm based on quantum coding provided by an embodiment of the present invention;

[0108] Fig. 9 This is a diagram of the training process of the conditional random field algorithm based on Monte Carlo optimization. DETAILED DESCRIPTION

[0109] In order to enable those skilled in the art to better understand the solutions of the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention.

[0110] In some of the processes described in the specification and claims of the present invention and the above-mentioned figures, multiple operations that appear in a specific order are included, but it should be clearly understood that these operations may not be executed in the order in which they appear in this article or executed in parallel. The sequence numbers of the operations, such as S101, S102, etc., are only used to distinguish between different operations, and the sequence numbers themselves do not represent any execution order. In addition, these processes may include more or fewer operations, and these operations may be executed in sequence or in parallel. It should be noted that the descriptions of "first", "second", etc. in this article are used to distinguish different messages, devices, modules, etc., do not represent the order of precedence, and do not limit the "first" and "second" to be different types.

[0111] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative work are within the scope of protection of the present invention.

[0112] Figure 1 The present invention provides a flowchart of an artificial intelligence feature extraction method based on evaluating MICM phenotype classification of lymphomas. Specifically, the method comprises the following steps:

[0113] S101: Obtaining a diagnosis report of the subject to be tested;

[0114] In some embodiments, the obtained report is a medical report, including a cell morphology examination report, a flow cytometer analysis report, a cytogenetic examination report, and a fluorescence in situ hybridization examination report.

[0115] S102: Obtaining a structured diagnosis report based on the diagnosis report;

[0116] In some embodiments, the obtained diagnostic report is a structured report, and the report content is extracted according to the standard report model to obtain a structured diagnostic report ( Figure 6 ).

[0117] In some embodiments, the acquired diagnostic report is an unstructured report, and the report content is extracted using computer vision technology according to a standard report model to obtain a structured report ( Figure 6 )

[0118] S103: vectorizing the structured diagnostic report to obtain a vectorized diagnostic report;

[0119] The method of vectorizing the structured diagnostic report to obtain a vectorized diagnostic report includes one or more of the following methods: Word2Vec, Doc2Vec, and BERT.

[0120] In some embodiments, BERT is used to vectorize the structured diagnostic report to obtain a vectorized diagnostic report.

[0121] In some embodiments, Word2Vec is used to vectorize the structured diagnosis report to obtain a vectorized diagnosis report.

[0122] S104: The quantitative diagnostic report in step 1 is input into a quantum coding-based Transformer for feature extraction to obtain MICM phenotypic typing features; the quantum coding-based Transformer converts the input data into a quantum state and then extracts quantum state feature information, converts the quantum state feature information back into non-quantum state data and then outputs the features of the MICM.

[0123] In one embodiment, a blood tumor MICM comprehensive diagnosis method and system mainly consists of the following modules:

[0124] (1) Machine vision (OCR) enhanced recognition model module: recognize reports in image format and extract the content in the file.

[0125] (2) Report content model library: collects structured reports, such as PDF and Word file format reports.

[0126] (3) Database operation module: Use oledb to connect to the database.

[0127] (4) SFTP file upload module: secure file transfer protocol, ensuring the security of files during the upload process.

[0128] (5) Clinical publishing module: The final report is published to the clinical department and accessed via a browser.

[0129] (6) HIS interface module: automatically extract clinical medical order information into the system.

[0130] (7)CA signature module: tamper-proof and traceable.

[0131] (8) Report automatic fusion module: Identify the information in the report file and merge the file with the medical order information.

[0132] (9) File conversion module: convert jpg, word and other format files into PDF format.

[0133] (10) File encryption module: PDF file password encryption.

[0134] Data model involved in this method and system:

[0135] 1. Cell morphology examination report, flow cytometry analysis report, cytogenetic examination report, fluorescence in situ hybridization examination report. 1000 reports of each type were randomly selected, and a total of 4000 reports were used as training data sets.

[0136] 2Convert semi-structured data into structured data and store it in the database.

[0137] 3 Unstructured data, mainly in the form of pictures. First, train the machine vision text recognition model library for the samples, and then convert all sample data into the database through the machine vision text recognition model library. Analyze the sample data collected in the database.

[0138] 4. Form a report content model library and a machine vision text recognition model library.

[0139] This method and system generate itemized reports:

[0140] Monitor report folders, upload reports to the server, identify report content, and merge with patient order information ( Figure 6 ).in,

[0141] ① Determine whether the collected reports in the monitoring folder are complete.

[0142] The system generates an independent "monitoring folder" based on each patient's examination application, and determines whether the "monitoring folder" contains the automatic cell morphology examination report, flow cytometer analysis report, cytogenetic examination report, fluorescence in situ hybridization examination report and other reports listed in the patient's application form based on the examination contents listed in the patient's application form.

[0143] ② Determine the format of the reports collected in the monitoring folder.

[0144] The system automatically determines whether the report format collected in the monitoring folder is a structured report or an unstructured report, and whether the report format is PDF or Jpg, etc.

[0145] ③ Use machine vision technology to extract unstructured report content according to the standard report model.

[0146] ④The structured standard diagnostic report and clinical bereaved information are edited through artificial intelligence to generate a MICM diagnostic report.

[0147] Among them, the process of generating MICM diagnostic reports through artificial intelligence editing of structured standard diagnostic reports and clinical medical order information (i.e. Figure 6④) in the above is generated by using a natural language model, where the natural language model is a generative model obtained through training. The specific training process is as follows: Figure 8 shown.

[0148] In a specific embodiment, the training of the natural language model includes the following steps:

[0149] Step 1: Data collection and annotation

[0150] The present invention is used to train a natural language model as a generative model, and the training data is collected from various medical reports, including cell morphology examination reports, flow cytometer analysis reports, cytogenetic examination reports, and fluorescent in situ hybridization examination reports. The report content is extracted according to the standard report model, or a structured report is obtained from the report content extracted according to the standard report model using machine vision technology. The collected data is annotated, and the annotation method of the present invention is manual annotation. Professional medical experts analyze the report according to medical standards and obtain a structured diagnosis report. The structured diagnosis report constitutes training data in the form of "question and answer pairs" (as shown in Table 1), and supervised training of the natural language model Transformer algorithm model is performed; further, the diagnosis results are annotated according to the collected data, that is, the training data is constituted in the form of "question + type".

[0151] Table 1 Example of question-answer pairs

[0152]

[0153]

[0154] Step 2: Transformer algorithm model training based on quantum coding

[0155] The training data collected in step 1 is subjected to feature extraction to obtain MICM phenotypic typing features. The present invention adopts a Transformer algorithm based on quantum coding to perform feature extraction, converts input data into a quantum state through a quantum coding layer, and utilizes the entanglement and superposition characteristics of quantum information to enhance the model's ability to capture data features.

[0156] The training process of the Transformer algorithm based on quantum coding is as follows Fig. 9 As shown:

[0157] Step 2-1: Encode the training data in text format into vector data.

[0158] The Word2Vec algorithm is used to encode the training data in text format into vector data, that is, a shallow neural network is used to map words into a vector space so that semantically similar words are close to each other in the vector space.

[0159] Furthermore, the initial quantum state is configured for each input vector data encoded as vector data, and the configuration of the quantum state is expressed as:

[0160]

[0161] In the formula, ψ represents the quantum state, α i is the complex amplitude, |i> is the ground state of the qubit, and n is the number of qubits.

[0162] The initial value of the parameter θ for initializing the quantum gate is set to π / 4.

[0163] Step 2-2: Encode and map the initial quantum state of the input data into quantum state data through the quantum coding layer;

[0164] Different from the traditional Transformer algorithm model, the encoding method of the Transformer algorithm based on quantum coding is encoded and mapped through the quantum coding layer: the input data is converted into a quantum state through the quantum coding layer, that is, each feature of the data is encoded on the quantum bit, which can be expressed as:

[0165] ψ enc =U(θ)|ψ>

[0166] In the formula, ψ enc represents the encoded quantum state, U(θ) is the quantum gate operation adjusted according to the characteristics of the input data, θ represents the parameters of the quantum gate, and |ψ> is the original quantum state.

[0167] The calculation method of |ψ> can be expressed as:

[0168]

[0169] In the formula, |ψ> represents the quantum state, ns is the total number of characteristics, and p i is the probability that the i-th feature is selected; output features based on the probability that the feature is selected;

[0170] Furthermore, the probability p of a feature being selected i The calculation method can be expressed as:

[0171]

[0172] In the formula, β cy is the adjustment parameter; Es i is the information gain score based on the i-th feature, calculated by the conditional random field algorithm based on Monte Carlo optimization; μ cy is the average value of information gain.

[0173] Furthermore, the implementation of the quantum gate operation U(θ) can be expressed as:

[0174] U(θ)=e -iθH

[0175] Where H is the Hamiltonian.

[0176] In one embodiment, the Hamiltonian H is characterized by a Pauli matrix, which can be expressed as:

[0177]

[0178] In the formula, is the Pauli matrix corresponding to the X, Y, and Z operations on the qubit.

[0179] Step 2-3: Use dynamic quantum topology optimization strategy to optimize the connection topology between quantum bits to improve the efficiency of quantum coding and the information processing capability of the model.

[0180] Specifically, at the quantum coding layer, the initial connection of each qubit is automatically configured based on the preliminary analysis of the data. As the training progresses, the connection is dynamically adjusted according to the changes in the data flow. The adjustment method can be expressed as:

[0181]

[0182] λ jk =softmax(-d jk / τ te )

[0183] In the formula, softmax() is the preset Softmax classification function, H DQTO is the Hamiltonian after dynamic adjustment; λ jk is the coupling strength between qubits j and k, and is the Pauli-Z operation acting on qubits j and k. jk represents the characteristic correlation distance between qubits j and k, τ te is a temperature parameter. Preferably, τ te Set to 3.

[0184] Step 2-4: The quantum states are entangled and disentangled through quantum logic gates, which are controlled by the parameters learned during training to maximize the information exchange between features, which can be expressed as:

[0185] ψ pro =U pro (γ)ψ enc

[0186] In the formula, ψ pro represents the quantum state after forward propagation, U pro (γ) is the quantum gate operation used in the forward propagation process, and γ is the quantum gate parameter. Preferably, γ is set to π / 2.

[0187] Furthermore, the quantum gate operation U pro The calculation method of (γ) can be expressed as:

[0188]

[0189] In the formula, each H k represents the Hamiltonian operating at different quantum levels, γ k is the corresponding parameter, and m represents the total number of operation levels. Preferably, γ k Set to π / 4.

[0190] Steps 2-5 calculate the degree of entanglement between different quantum bits after the quantum gate operation.

[0191] The calculation method can be expressed as:

[0192] E=Tr(ρ A logρ A )

[0193] Where E represents the quantum entanglement degree, ρ A is the reduced density matrix of system A, Tr() represents the trace operation. Furthermore, the reduced density matrix ρ A Obtained from the total density matrix ρ through trace operation, it can be expressed as:

[0194] ρ A =Tr B (ρ)

[0195] In the formula, Tr B () means taking a trace of part B of the system, leaving the state of part A. If system B is the remaining n-1 qubits, it means that only the state of a specific qubit is considered from the state of the total system.

[0196] In this embodiment, system B and system A represent the transformation features extracted in this iteration and the transformation features extracted in the previous iteration, respectively, and the transformation features are the quantum state features output by the quantum gate layer.

[0197] Step 2-6: Different from the traditional Transformer algorithm model, the decoding method of the Transformer algorithm based on quantum coding maps the entangled quantum state back to classical information through the inverse quantum coding operation.

[0198] It can be expressed as:

[0199]

[0200] In the formula, ψ dec represents the decoded quantum state, It is the conjugate transpose of the quantum coding operation and is used for the inverse operation.

[0201] Step 2-7: Calculate the loss function and adjust the parameters of the quantum coding layer and quantum gate layer through the back-propagation algorithm, which can be expressed as:

[0202]

[0203] In the formula, L represents the loss function, y i is the target output, f(ψ dec,i ) is the predicted output of the model, is the gradient of the loss function with respect to the quantum gate parameter θ.

[0204] Furthermore, by bringing the loss function into the partial derivative calculation formula, we can get The calculation method is:

[0205]

[0206] Further, Calculated by the chain rule, it can be expressed as:

[0207]

[0208] Further, It involves the differentiation of the inverse quantum coding operation, and the calculation method can be expressed as:

[0209]

[0210] In the formula, H DQTO is the dynamically adjusted Hamiltonian.

[0211] Furthermore, jk The update is based on the back propagation algorithm and can be expressed as:

[0212]

[0213] Where η cg is the learning rate, is the gradient of the loss function with respect to the coupling strength, calculated by the chain rule. Preferably, η cg Set to 0.03.

[0214] Furthermore, the update method of the parameters of the quantum gate can be expressed as:

[0215]

[0216] In the formula, θ new and θ old represent the quantum gate parameters before and after the update, η rate is the learning rate. Preferably, η rate Set to 0.01.

[0217] Steps 2-8: Repeat the above steps until the preset stop iteration condition is met, indicating that the model training is completed.

[0218] The preset condition for stopping iteration is reaching a preset maximum number of iterations. In one embodiment, preferably, the preset maximum number of iterations is set to 1000 times.

[0219] In some embodiments, the preset condition for stopping iteration is that when the entanglement degree is less than a preset threshold, it means that the model has converged, and the training is stopped.

[0220] Step 3: Conditional Random Field Algorithm Model Training Based on Monte Carlo Optimization

[0221] The output features of the encoding part of the Transformer algorithm based on quantum coding are input into the conditional random field algorithm to train the auxiliary diagnosis model. The conditional random field is a statistical modeling method. The present invention adopts the conditional random field algorithm based on Monte Carlo optimization and adopts a sparse representation strategy to improve the efficiency and accuracy of the model when processing data. Specifically, the training process of the conditional random field algorithm based on Monte Carlo optimization is as follows: Fig. 9 shown.

[0222] Step 3-1: Let the output feature of the encoding part of the Transformer algorithm based on quantum coding be q x , initialize the parameters q of the conditional random field model θ , the initialization method is random initialization, which can be expressed as:

[0223] q θ =N(0,σ 2 I)

[0224] In the formula, q θ represents the parameter vector of the conditional random field model, N(0,σ 2 I) indicates that the parameters are initialized to a mean of 0 and a variance of σ 2 Normal distribution, I is the unit matrix. Preferably, σ 2 Set to 0.01.

[0225] Step 3-2: In each iteration, the model parameters q are adjusted using the Monte Carlo method. θ Random sampling can be expressed as:

[0226] q θ (t) =q θ (t-1) +∈ (t)

[0227] In the formula, q θ (t) and q θ (t-1) are the parameters after the tth and t-1th iterations, ∈ (t) is the sampling step length;

[0228] Furthermore, the sampling step ∈ (t) For dynamic adaptive settings, the setting method is:

[0229] ∈ (t) =κ cy ·exp(-ρ cy t)

[0230] In the formula, κ cy is the initial step size, ρ cy is the attenuation factor, and the step size is adjusted with the number of iterations t to ensure the stability and convergence of the sampling process. Preferably, κ cy Set to 1, ρ cy Set to 0.95.

[0231] Step 3-3: Use the parameters obtained from Monte Carlo sampling to train the conditional random field model by optimizing the objective function L(q θ ), the model will learn how to effectively classify based on the input feature vector, and the calculation method of the objective function can be expressed as:

[0232]

[0233] In the formula, is the label data, In a given feature and parameter q θ The label prediction probability under L1 is the L1 regularization coefficient, ||q θ ||1 is the L1 regularization term. Preferably, λ L1 Set to 0.3.

[0234] Furthermore, the probability The calculation method can be expressed as:

[0235]

[0236] In the formula, represents the feature function, which converts the input features and labels into numerical values ​​for model training; y′ represents all possible label configurations.

[0237] Step 3-4: Use sparse regularization strategy to optimize parameter q during model training θ , by calculating the increment Δq θ Update parameter q θ , to ensure that the model focuses on the features that are most critical to classification. The calculation method can be expressed as:

[0238]

[0239] In the formula, α gz is the learning rate, is the gradient of the loss function with respect to the parameter, δ gz is an adjustment factor calculated based on the difference between the model output and the actual classification result, sign(q θ ) indicates q θ Preferably, α gz Set to 0.01.

[0240] Step 3-5: Further, Gradient The calculation method can be expressed as:

[0241]

[0242] In the formula, It represents the expectation of the predicted distribution of all possible label configurations y′ under the current parameters.

[0243] In one embodiment, the adjustment factor δ gz The calculation method can be expressed as:

[0244]

[0245] Where η gz is the adjustment factor learning rate, y i is the actual class label, is the category predicted by the model, and tanh() is the hyperbolic tangent function. Preferably, η gz Set to 0.05.

[0246] Step 3-5: After each iteration, calculate the information gain Es i , the calculation method can be expressed as:

[0247]

[0248] In the formula, c represents the classification label, C is all possible categories, and p(c|x i ) is a given feature xi The probability of category c under the condition of , p(c) is the prior probability of category c.

[0249] Furthermore, the performance of the model on the training set is evaluated. In one embodiment, the performance of the model on the training set is measured by the training accuracy, and the prediction label method for calculating the training accuracy can be expressed as:

[0250]

[0251] In the formula, is the indicator function, In a given feature and model parameters q θ The predicted label under In a given feature and parameter q θ The predicted probability of the label under .

[0252] Repeat the above steps until the preset stop iteration condition is met, indicating that the model training is completed. In one embodiment, the preset stop iteration condition is that the training accuracy of the model on the training set reaches more than 95%, or reaches the preset maximum number of iterations. Preferably, the preset maximum number of iterations is set to 1000 times.

[0253] Step 4: Generate MICM diagnostic report

[0254] Generate a MICM diagnosis report using the trained model.

[0255] In one embodiment, a MICM diagnostic report is generated for the acquired report content. The MICM diagnostic report is in a standardized output format and includes two parts, one part is the diagnostic process and treatment suggestions output by the Transformer algorithm model based on quantum coding, and the other part is the auxiliary diagnosis disease category output by the conditional random field algorithm model optimized based on the Monte Carlo method.

[0256] In one embodiment, for a report whose content is: "The patient has symptoms of irregular heart rhythm and chest pain", the standardized output format is: "The patient may have non-ST-segment elevation myocardial infarction (NSTEMI), and it is recommended to undergo immediate blood biochemical tests and electrocardiogram examinations." and "Auxiliary diagnosis category of the disease: acute coronary syndrome".

[0257] In one embodiment, the MICM diagnostic report includes the patient's treatment information and the classification result.

[0258] The MICM report generation of this method and system: the system automatically completes the standardized sub-item diagnosis report after extracting the content of unstructured reports such as cell morphology examination report, flow cytometer analysis report, cytogenetic examination report, fluorescence in situ hybridization examination report, etc., collects the patient's medical order information and automatically generates the MICM standard report. After the doctor's review, the report is released to the clinic.

[0259] Figure 3 is a schematic diagram of a computer device provided by an embodiment of the present invention, such as Figure 3 As shown, the device may include: one or more processors, and one or more memories; wherein the memories store computer-readable codes, and when the computer-readable codes are run by the one or more processors, the method described above may be executed.

[0260] The processor in this embodiment can be an integrated circuit chip with signal processing capabilities. The above processor can be a general-purpose processor, a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA) or other programmable logic devices, discrete gates or transistor logic devices, discrete hardware components. The disclosed methods, operations and logic block diagrams in the embodiments of the present disclosure can be implemented or executed. The general-purpose processor can be a microprocessor or the processor can also be any conventional processor, etc., which can be an X86 architecture or an ARM architecture.

[0261] In general, various example embodiments of the present disclosure may be implemented in hardware or dedicated circuits, software, firmware, logic, or any combination thereof. Certain aspects may be implemented in hardware, while other aspects may be implemented in firmware or software that may be executed by a controller, microprocessor, or other computing device. When various aspects of the disclosed embodiments are illustrated or described as block diagrams, flow charts, or using some other graphical representation, it will be understood that the blocks, devices, systems, techniques, or methods described herein may be implemented in hardware, software, firmware, dedicated circuits or logic, general purpose hardware or controllers or other computing devices, or some combination thereof as non-limiting examples.

[0262] For example, the method or device according to the embodiment of the present disclosure may also be implemented by Figure 4 The architecture of the computing device 3000 shown in FIG. Figure 4As shown, the computing device 3000 may include a bus 3010, one or more CPUs 3020, a read-only memory (ROM) 3030, a random access memory (RAM) 3040, a communication port 3050 connected to a network, an input / output component 3060, a hard disk 3070, etc. The storage device in the computing device 3000, such as ROM 3030 or hard disk 3070, may store various data or files used for processing and / or communication of the method provided by the present disclosure and program instructions executed by the CPU. The computing device 3000 may also include a user interface 3080. Of course, Figure 4 The architecture shown is only exemplary and can be omitted according to actual needs when implementing different devices. Figure 4 One or more components of a computing device are shown.

[0263] The embodiment of the present invention also provides a computer-readable storage medium, such as Figure 5 As shown, it is a schematic diagram of a storage medium provided by an embodiment of the present invention, and a computer readable instruction 4010 is stored on the computer storage medium 4020. When the computer readable instruction 4010 is executed by the processor, the method according to the embodiment of the present disclosure described with reference to the above figures can be executed. The computer readable storage medium in the embodiment of the present disclosure can be a volatile memory or a non-volatile memory, or can include both volatile and non-volatile memories. The non-volatile memory can be a read-only memory (ROM), a programmable read-only memory (PROM), an erasable programmable read-only memory (EPROM), an electrically erasable programmable read-only memory (EEPROM) or a flash memory. The volatile memory can be a random access memory (RAM), which is used as an external cache. By way of example and not limitation, many forms of RAM are available, such as static random access memory (SRAM), dynamic random access memory (DRAM), synchronous dynamic random access memory (SDRAM), double data rate synchronous dynamic random access memory (DDRSDRAM), enhanced synchronous dynamic random access memory (ESDRAM), synchronous link dynamic random access memory (SLDRAM), and direct memory bus random access memory (DR RAM). It should be noted that the memory of the methods described herein is intended to include, but is not limited to, these and any other suitable types of memory. It should be noted that the memory of the methods described herein is intended to include, but is not limited to, these and any other suitable types of memory.

[0264] The present disclosure also provides a computer program product or a computer program, which implements the steps of the above method when executed by a processor. Figure 2 As shown, the computer program product or computer program comprises:

[0265] Acquisition module 201: used to obtain a diagnosis report of a subject;

[0266] Structured processing module 202: used to obtain a structured diagnosis report based on the diagnosis report;

[0267] Vectorization processing module 203: used for vectorizing the structured diagnosis report to obtain a vectorized diagnosis report;

[0268] Feature extraction module 204: used for inputting the vectorized diagnostic report into a quantum-coded based Transformer to extract features and obtain MICM phenotypic typing features; the quantum-coded based Transformer converts the input data into a quantum state and then extracts quantum state feature information, converts the quantum state feature information back into non-quantum state data and then outputs the MICM phenotypic typing features.

[0269] It should be noted that the flowcharts and block diagrams in the accompanying drawings illustrate the possible architecture, functions and operations of the systems, methods and computer program products according to various embodiments of the present disclosure. In this regard, each box in the flowchart or block diagram can represent a module, a program segment, or a part of a code, and the module, program segment, or a part of the code contains one or more executable instructions for implementing the specified logical function. It should also be noted that in some alternative implementations, the functions marked in the box can also occur in a different order from the order marked in the accompanying drawings. For example, two boxes represented in succession can actually be executed substantially in parallel, and they can sometimes be executed in the opposite order, depending on the functions involved. It should also be noted that each box in the block diagram and / or flowchart, and the combination of boxes in the block diagram and / or flowchart can be implemented with a dedicated hardware-based system that performs a specified function or operation, or can be implemented with a combination of dedicated hardware and computer instructions.

[0270] In general, various example embodiments of the present disclosure may be implemented in hardware or dedicated circuits, software, firmware, logic, or any combination thereof. Certain aspects may be implemented in hardware, while other aspects may be implemented in firmware or software that may be executed by a controller, microprocessor, or other computing device. When various aspects of the disclosed embodiments are illustrated or described as block diagrams, flow charts, or using some other graphical representation, it will be understood that the blocks, devices, systems, techniques, or methods described herein may be implemented in hardware, software, firmware, dedicated circuits or logic, general purpose hardware or controllers or other computing devices, or some combination thereof as non-limiting examples.

[0271] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working processes of the systems, devices and units described above can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0272] In the several embodiments provided in the present application, it should be understood that the disclosed systems, devices and methods can be implemented in other ways. For example, the device embodiments described above are only schematic. For example, the division of the units is only a logical function division. There may be other division methods in actual implementation, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be an indirect coupling or communication connection through some interfaces, devices or units, which can be electrical, mechanical or other forms.

[0273] The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed on multiple network units. Some or all of the units may be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0274] In addition, each functional unit in each embodiment of the present invention may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit. The above-mentioned integrated unit may be implemented in the form of hardware or in the form of software functional units.

[0275] The exemplary embodiments of the present disclosure described in detail above are merely illustrative and not restrictive. It should be understood by those skilled in the art that various modifications and combinations may be made to these embodiments or their features without departing from the principles and spirit of the present disclosure, and such modifications should fall within the scope of the present disclosure.

Claims

1. An artificial intelligence feature extraction method based on MICM phenotyping of lymphoma, characterized in that: The method comprises: S1: Obtain the diagnostic report of the subject; S2: Obtaining a structured diagnostic report based on the diagnostic report; S3: vectorizing the structured diagnostic report to obtain a vectorized diagnostic report; S4: The vectorized diagnostic report is input into a quantum coding-based Transformer for feature extraction to obtain MICM phenotypic typing features; the quantum coding-based Transformer converts the input data into a quantum state and then extracts the quantum state feature information, converts the quantum state feature information back into non-quantum state data and then outputs the MICM phenotypic typing features.

2. The artificial intelligence feature extraction method based on MICM phenotyping of lymphoma according to claim 1, characterized in that: The Transformer based on quantum coding includes an input layer, a quantum coding layer, a quantum gate layer and an output layer; the processing steps of the Transformer based on quantum coding include: Step 1: The input layer configures an initial quantum state for each vector of the vectorized diagnostic report to obtain a vector of initial quantum states; Step 2: The quantum coding layer performs quantum coding on the vector of the initial quantum state to obtain a quantum coded vector; Step 3: The quantum-encoded vector is forward-propagated through the quantum gate layer to obtain quantum state characteristics extracted by forward propagation; Step 4: the output layer maps the extracted quantum state features back to the non-quantum state to obtain the MICM phenotyping features; Optionally, the steps further include a step 3' between step 2 and step 3: optimizing the correlation distance between quantum bits of the quantum-encoded vector by a dynamic quantum topology optimization strategy.

3. The artificial intelligence feature extraction method based on lymphoma MICM phenotyping according to claim 2 is characterized in that: The way in which each vector configures the initial quantum state is expressed as: In the formula, represents the initial quantum state, α i is the complex amplitude, |i> is the ground state of the qubit, and n is the number of qubits used to represent the vector; Optionally, the conversion of the quantum coding layer can be expressed as: ψ enc =U(θ)|ψ> In the formula, ψ enc represents the encoded quantum state, U(θ) is the quantum gate adjusted according to the characteristics of the input data, θ represents the parameters of the quantum gate, and |ψ< is the initial quantum state; Optionally, the quantum gate layer is represented as: ψ pro =U pro (c)ψ enc In the formula, ψ pro represents the quantum state after forward propagation, U pro (γ) is the quantum gate used in the forward propagation process, and γ is the quantum gate parameter; Optionally, the output layer is expressed as: In the formula, ψ dec represents the MICM phenotyping characteristics obtained after decoding, ψ pro It represents the quantum state characteristic information extracted after forward propagation. is the conjugate transpose of the quantum coding operation, used to map the quantum state characteristic information into the MICM phenotyping characteristics; Alternatively, the quantum gate U(θ) is expressed as: U(θ)=e -iθH Where H is the Hamiltonian; Optionally, a dynamic quantum topology optimization strategy is used to adjust and obtain an adjusted Hamiltonian, and the quantum gate U(θ) is obtained using the adjusted Hamiltonian; Optionally, the dynamic adjustment method of the dynamic quantum topology optimization strategy can be expressed as: l jk =softmax(-d jk / t te ) In the formula, softmax() is the preset Softmax classification function, H DQTO is the Hamiltonian after dynamic adjustment; λ jk is the coupling strength between qubits j and k, and is the Pauli-Z operation acting on qubits j and k, d jk represents the characteristic correlation distance between qubits j and k, τ te is the temperature parameter.

4. The artificial intelligence feature extraction method based on lymphoma MICM phenotyping according to claim 1, characterized in that: The steps of constructing the quantum-coded Transformer include: The initial quantum coded Transformer includes an initial input layer, an initial quantum coding layer, an initial quantum gate layer and an initial output layer, wherein the parameters of the initial quantum coding layer and the initial quantum gate layer are randomly initialized parameters; After the vectorized diagnostic report of the training set is passed through the initial quantum-coded Transformer to obtain the output MICM phenotyping features, the model prediction output is obtained based on the output MICM phenotyping features, the difference between the actual value of the training set and the model prediction output is compared to the loss function, and the loss function is iteratively trained until the stop condition is reached to obtain the quantum-coded Transformer; The stopping condition includes reaching a preset maximum number of iterations; Optionally, a loss function is calculated based on the output MICM phenotyping features, expressed as: Where L represents the loss function, i∈[1,m], i represents the i-th sample in the training set, m represents the total number of samples in the training set, and y i is the target output of the i-th sample, ψ dec,i represents the MICM phenotyping characteristics of the output of the i-th sample, f(ψ dec,i ) is the output of the model’s prediction for the i-th sample; Optionally, the updating method of the parameters of the initial quantum coding layer is expressed as: In the formula, θ new and θ old represent the initial quantum coding layer parameters before and after the update, η rate is the learning rate, is the gradient of the loss function with respect to the initial quantum coding layer parameter θ; Optionally, the updating method of the parameters of the initial quantum gate layer is expressed as: In the formula, γ new and γ old denote the parameters of the initial quantum gate layer before and after the update, η2 is the learning rate, is the gradient of the loss function with respect to the initial quantum gate layer parameter γ; Optionally, the stopping condition includes that the quantum entanglement degree is less than a preset threshold: after forward propagation through the quantum gate layer to obtain the quantum state characteristics extracted by forward propagation, the entanglement degree between different quantum bits is calculated. When, it indicates that the model converges, then the training is stopped; Optionally, the entanglement degree is expressed as: E=Tr(ρ A logρ A ) Where E represents the quantum entanglement degree, ρ A is the reduced density matrix of system A. System A represents the quantum state characteristic information extracted from the last iteration. Tr() represents the trace operation. Furthermore, the reduced density matrix ρ A It is obtained by trace operation from the total density matrix ρ, which is the characteristic matrix of the quantum state characteristic information extracted in this iteration, and the reduced density matrix ρ A It can be expressed as: r A =Tr B (p) In the formula, Tr B () indicates taking a trace of part B of the system, leaving the state of part A of the system. System B is the quantum state characteristic information extracted in this iteration.

5. The artificial intelligence feature extraction method based on MICM phenotyping of lymphoma according to claim 1, characterized in that: The diagnostic report includes morphological information, immunological information, cytogenetic information, and molecular information; Optionally, the method for obtaining a structured diagnostic report based on the diagnostic report includes using machine vision technology to extract the structured diagnostic report according to a standard report model; Optionally, the method of vectorizing the structured diagnostic report to obtain a vectorized diagnostic report includes one or more of the following methods: Word2Vec, Doc2Vec, and BERT.

6. The artificial intelligence feature extraction method based on lymphoma MICM phenotyping according to claim 1, characterized in that: The quantum coding-based artificial intelligence feature extraction method based on lymphoma MICM phenotyping also includes S5: obtaining the patient's diagnosis and treatment information based on the MICM phenotyping features; optionally, it also includes S5': inputting the MICM phenotyping features into a classifier to obtain a classification result of whether the subject suffers from a blood tumor; Optionally, the classifier includes one or more of the following: conditional random field algorithm, support vector machine, decision tree, random forest, logistic regression; Optionally, a MICM diagnosis report is output based on the patient's diagnosis and treatment information and the classification result.

7. The artificial intelligence feature extraction method based on MICM phenotyping of lymphoma according to claim 6, characterized in that: The classifier is a trained conditional random field, and the method for constructing the trained conditional random field includes: Obtaining a diagnostic report training set, wherein the training set also includes a label of whether the patient has a blood tumor; Obtaining a MICM phenotyping feature training set based on the diagnostic report training set; Initialize the parameters of the conditional random field to obtain an initialized conditional random field; Performing Monte Carlo random sampling on the parameters of the initialized conditional random field to obtain sample parameters; Obtaining predicted labels of the MICM phenotyping feature training set based on the sample parameters, calculating a loss function of a conditional random field based on the predicted labels of the MICM phenotyping feature training set and the actual labels of the diagnosis report, and gradually minimizing the loss function of the conditional random field through iteration, and obtaining the trained conditional random field after reaching a stop condition; Optionally, the loss function of the conditional random field is expressed as: In the formula, L(q θ ) represents the parameter q based on the conditional random field θ The loss function is is the label data of the i-th sample, q θ (t) is the parameter of the conditional random field obtained by Monte Carlo sampling at the tth iteration in the iterative process, In a given feature and parameter q θ (t) The label prediction probability under L1 is the L1 regularization coefficient, ||q θ ||1 is parameter q θ (t) The L1 regularization term of ; Optionally, the sample parameters obtained by the Monte Carlo random sampling are expressed as: q θ (t) =q θ (t-1) +∈ (t) In the formula, q θ (t) represents the sample parameter of the tth iteration, q θ (t-1) represents the sample parameter of the t-1th iteration, ∈ (t) is the sampling step length; Optional, sampling step ∈ (t) For dynamic adaptive settings, the setting method is: ∈ (t) =k cy ·exp(-ρ cy t) In the formula, κ cy is the initial step size, ρ cy is the attenuation factor; Optionally, after optimizing the sample parameters by adopting a sparse regularization strategy to obtain optimized sample parameters, the optimized sample parameters are used to replace the sample parameters to calculate the loss function of the conditional random field; Optionally, the sparse regularization strategy is expressed as: In the formula, represents the sample parameter optimized at the tth iteration, Δq θ (t) is the sample parameter q in the tth iteration θ (t) The increment of α gz is the learning rate, is the gradient of the loss function of the conditional random field with respect to the parameters, δ gz is an adjustment factor calculated based on the difference between the model output and the actual classification result, sign(q θ (t) ) indicates q θ (t) The symbolic function of Optionally, the stopping condition includes that the information gain of the iteration is less than a preset threshold: after each iteration, the information gain is calculated based on the sample parameters of the current conditional random field, and the iteration is stopped when the information gain is less than the preset threshold; Optional, Indicates that given a feature and parameter q θ (t) The label prediction probability under the conditional random field is determined, and all the Where c∈[1,C], c represents the classification label, C is all possible categories, and the information gain is expressed as follows: In the formula, Es i represents the information gain of the i-th sample, p(c|q xi ) is a given feature q xi The probability of category c under the condition of , p(c) is the prior probability of category c.

8. A computer device, characterized in that: The device comprises: a memory and a processor; the memory is used to store a computer program; the processor executes the computer program to implement the steps of the method according to any one of claims 1 to 7.

9. A computer-readable storage medium, characterized in that: A computer program is stored thereon, and when the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 7 are implemented.

10. A computer program product, comprising a computer program, characterized in that When the computer program is executed by a processor, the steps of the method described in any one of claims 1 to 7 are implemented.

Images