A disease grade classification method, device, medium and product

Abstract

The application discloses a disease grade classification method, device, medium and product, and relates to the field of pattern recognition. The application processes acquired medical data by using a classification model with an automatic hierarchical decision process to obtain a probability prediction result, so that the classification model can obtain better performance in a medical grade classification task in the presence of label noise. Moreover, the application obtains a disease grading prediction result based on the probability prediction result, so that the accuracy of disease grading prediction can be further improved.

Description

Technical Field

[0001] This invention relates to the field of pattern recognition, and in particular to a method, device, medium, and product for classifying disease severity. Background Technology

[0002] Disease severity classification refers to determining the severity level of diseases in datasets related to the medical field. In the complex data of medical diagnosis, labeling errors are inevitable. Manually processing such complex data would be extremely time-consuming and resource-intensive; therefore, using artificial intelligence to build robust models can help alleviate resource pressure. Existing medical severity classification problems often assume completely correct labels, which does not reflect reality. Furthermore, there is a clear hierarchical relationship between categories in medical severity datasets; order information should also be considered.

[0003] Xiaofeng Liu et al. used Wasserstein loss with single-peak uniform constraints to train models for medical diagnosis and to mine the implicit order information in medical diagnosis classification problems; Bo Han et al. proposed a co-teaching method to address the label noise problem, which selects clean samples for model training based on network fitting effects; Feng Chen et al. proposed an SSR method that selects samples based on KNN and corrects labels based on predicted output probabilities.

[0004] In Chinese invention patent application CN201910357585.5, a method for classifying weakness in cardiovascular disease patients based on data mining technology uses principal component analysis information as training data items for a classification model. Chinese invention patent application CN202111346214.0 provides a brain disease classification method based on 3D attention convolution and self-supervised learning. This method uses a 3D CNN with an attention mechanism to extract spatial features of whole-brain voxels from fMRI data, while simultaneously using self-supervised learning to mine latent features of the data to improve classification performance. In invention patent application CN202311183569.1, a method for classifying labeled noisy data based on asynchronous co-teaching is proposed. This method replaces the two identical DNNs (Deep Neural Networks) of the classic co-teaching model with an early DNN and a mid-term DNN. After early training, the robustness of the mid-term DNN to labeled noise gradually decreases, while the early DNN retains its robustness to labeled noise.

[0005] However, the methods mentioned above that focus on disease classification do not take into account the problem of label noise, and the methods that focus on label noise do not take into account the implicit hierarchical information between disease classification categories, that is, the more adjacent the categories are, the stronger the correlation between the disease severity characteristics represented by the categories. Summary of the Invention

[0006] To address the aforementioned problems in the existing technology, the present invention provides a method, device, medium, and product for classifying disease severity.

[0007] To achieve the above objectives, the present invention provides the following solution:

[0008] A disease severity classification method, the method comprising:

[0009] Construct a classification model; the classification model is used to automate the hierarchical decision-making process.

[0010] Acquire medical data and input the medical data into the classification model to obtain probability prediction results;

[0011] Disease classification prediction results are obtained based on the probability prediction results.

[0012] Optionally, a classification model is constructed, specifically including:

[0013] Generate labeled noise to obtain samples with both real and noise labels, in order to generate a medical-grade dataset;

[0014] A hierarchical decision problem is constructed based on the aforementioned medical grade dataset;

[0015] The classification model is obtained by modifying the network structure of the initial classification model according to the hierarchical decision problem.

[0016] Optionally, the process of generating label noise includes:

[0017] Determine the noise level and determine the flip probability based on the noise level; the flip probability includes the probability that the real label flips into a noise label and the probability that the real label flips into itself;

[0018] The flipping probability is controlled by the probability density function of a normal distribution, so that the flipping probability decreases as the distance from the real label increases, in order to simulate the noise generation process in real-world situations.

[0019] The flip probability is normalized using the Softmax method to determine the noise transition matrix;

[0020] Based on the noise transfer matrix, the real labels are flipped into noise labels.

[0021] Optionally, the classification model is obtained by modifying the network structure of the initial classification model according to the hierarchical decision problem, specifically including:

[0022] Based on the hierarchical decision problem, the initial classification model outputs the conditional probabilities of the intermediate processes of each level of decision results, and performs recursive calculations to obtain the predicted probabilities of the next level of decision problems. This process is repeated so that the neurons output by the network structure correspond to the conditional probabilities of each decision layer, thus obtaining the classification model.

[0023] Optionally, acquiring medical data and inputting the medical data into the classification model to obtain probability prediction results specifically includes:

[0024] The medical data is input into the classification model to obtain the probability of the medical data.

[0025] In the classification model, the probabilities of the medical data are ordered and the probabilities of adjacent categories are merged to obtain multiple decision problems;

[0026] The sum of probabilities for each category in the decision problem is determined, and the sum of probabilities for each category is normalized using a Softmax normalization operation to obtain the probability prediction result.

[0027] Optionally, the disease classification prediction result is obtained based on the probability prediction result, specifically including:

[0028] The category containing the maximum value of the probability prediction result is taken as the disease classification prediction result.

[0029] A computer device includes: a memory, a processor, and a computer program stored in the memory and executable on the processor, the processor executing the computer program to implement the steps of the disease grading method described in any of the preceding claims.

[0030] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the disease grading method described in any of the preceding claims.

[0031] A computer program product includes a computer program that, when executed by a processor, implements the steps of the disease grading method described in any of the preceding claims.

[0032] According to specific embodiments provided by the present invention, the present invention discloses the following technical effects:

[0033] This invention processes acquired medical data using a classification model with an automated hierarchical decision-making process to obtain probabilistic prediction results. This enables the classification model to achieve better performance in medical grade classification tasks even in the presence of label noise. Furthermore, this invention obtains disease grade prediction results based on the probabilistic prediction results, which can further improve the accuracy of disease grade prediction. Attached Figure Description

[0034] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in 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 effort.

[0035] Figure 1 This is a flowchart of the disease severity classification method provided in Embodiment 1 of the present invention;

[0036] Figure 2 This is a flowchart illustrating the implementation of the disease severity classification method provided in Embodiment 1 of the present invention.

[0037] Figure 3 This is a schematic diagram of the hierarchical decision problem calculation process provided in Embodiment 1 of the present invention;

[0038] Figure 4 This is a schematic diagram of the calculation process for the final prediction result provided in Embodiment 1 of the present invention. Detailed Implementation

[0039] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0040] The purpose of this invention is to provide a disease classification method, device, medium, and product, which aims to achieve better performance in medical classification tasks under label noise conditions.

[0041] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0042] Example 1

[0043] This embodiment provides a disease severity classification method, such as Figure 1 and Figure 2 As shown, the method includes:

[0044] Step 100: Build a classification model. The classification model is used to automate the hierarchical decision-making process.

[0045] In practical applications, this step can be implemented as follows:

[0046] Step 100-1: Generate labeled noise to obtain samples with both real and noise labels to generate a medical grade dataset. For example, synthesize labeled noise for the medical grade dataset to simulate the real-world process of generating medical grade category label noise. From a realistic perspective, because the deterioration of a disease is a gradual process, categories with more adjacent labels are more likely to be mislabeled. The specific noise generation method can be mainly divided into the following steps:

[0047] Step 1: Control the overall noise level to ρ. Represent the samples in the dataset as x, their corresponding true labels as y, and the noise labels as... The number of medical grades is K, which represents the number of categories in the classification problem. This number is used for designing the classification method and judging the results. The probabilities of labels flipping each other form the noise transition matrix, denoted by T. The probability that a true label i flips to a noise label j is T. i,j The probability that the real label is flipped into itself is T. i,i ,in

[0048] Step 2: The flipping probability of each category is controlled using the probability density function of a normal distribution, causing the flipping probability to decrease as the distance from the true label increases, simulating the noise generation process in real-world scenarios. The difference between the noise label values ​​(e.g., 1, 2, and 3) and the true label values ​​(e.g., 1, 2, and 3) is used as the input to the probability density function to obtain the flipping probability for each category, i.e., each disease level. Where e represents the natural constant, and σ is a parameter used to control how much the flip probability changes with the difference in label values. The larger σ is, the less likely the true label is to be flipped into a label of a category with a significantly different label value.

[0049] Step 3: Normalize the probabilities of each category in Step 2 using the Softmax method so that the overall noise flip probability of each category is 1, and calculate the noise transition matrix.

[0050] The noise transfer matrix is ​​expressed as follows:

[0051] Step 4: Based on the noise transition matrix T, flip the true labels y in the dataset to noise labels. We obtain samples that simultaneously possess both true and noisy labels. Based on this, we design a noisy ordinal regression classification method, build a model, and validate the effectiveness of the classification method.

[0052] Step 100-2: Construct a hierarchical decision problem based on the medical grade dataset.

[0053] Step 100-3: Modify the network structure of the initial classification model according to the hierarchical decision problem to obtain the classification model.

[0054] Building a classification model primarily involves modifying its network structure to automate the hierarchical decision-making process. This process mainly consists of a theoretical decision-making process and a network structure implementation process. Specifically, the theoretical decision-making process involves dividing the network into hierarchical relationships based on the number of categories, thus obtaining the decision problem for each level. Network structure implementation refers to applying the hierarchical decision-making concept to the network structure, modifying the network structure according to the hierarchical decision problems, so that the output of the final neurons in the network obtains the conditional probabilities of the intermediate processes in the decision-making results at each level, thereby obtaining the outcome of each level's decision problem.

[0055] Specifically, in the hierarchical decision-making method, each level of the decision problem involves probability prediction of combined categories. The calculation process for the hierarchical decision problem is as follows: Figure 3 As shown. For example, for a problem with K categories, there are C = K-1 decision layers. Specifically, the c-th decision layer among the C decision layers performs the following probability predictions:

[0056] p(1≤y≤Kc|x).

[0057] p(2≤y≤K-c+1|x).

[0058] …

[0059] p(s≤y≤K-c+s-1|x).

[0060] …

[0061] p(c+1≤y≤K|x).

[0062] Here, p(s≤y≤K-c+s-1|x) represents the probability that needs to be predicted in the c-th decision layer. Similarly, the c-th decision layer needs to predict the probabilities of each disease level, i.e., each category, expressed as p(y=i|x), i∈{1,...,K}, where i represents a category among K categories. The probabilities of each category will be used to subsequently determine the category to which the input sample belongs.

[0063] To enable the network structure to automate the entire problem decision-making process and implement the recursive process for each level of the problem, the network structure was modified so that the output of the final neurons obtains the conditional probabilities of the intermediate processes in the decision results of each level. For example, the conditional probabilities related to the decision at level c (s≤y≤K-c+s-1|x) and level c+1 (s≤y≤K-c+s-2|s≤y≤K-c+s-1) are p(s+1≤y≤K-c+s-1|s≤y≤K-c+s-1). In other words, obtaining the decision result at level c+1 based on the decision result at level c requires considering the conditional probabilities between the decision problems. The neurons output by the network structure ultimately correspond to the conditional probabilities connecting each decision level.

[0064] The decision probability of each layer is represented as a point in the tree structure, and the conditional probability is calculated from the edges. Therefore, the calculation of the corresponding network structure is as follows:

[0065] o c,s =o c-1,s-1 ×e c,s +o c-1,s ×e c,s+1 .

[0066] Among them, o c,s This represents the probability value corresponding to the s-th node in the c-th layer of the tree structure of the modified network, e. c,s This represents the probability value corresponding to the s-th edge in the c-th layer of the edge layer.

[0067] After the network prediction is completed, the system automatically obtains the probability prediction results for the next level decision problem from the conditional probability of the network prediction, starting from the first-level decision problem. The s-th prediction probability p(s≤y≤K-c+s-1|x) of the (c+1)-th level is calculated from the prediction probabilities p(s≤y≤K-c+s|x), p(s-1≤y≤K-c+s|x), and the conditional probability of the c-th level:

[0068] p(s≤y≤K-c+s-1|x)

[0069] =p(s≤y≤K-c+s|x)×p(s≤y≤K-c+s-1|s≤y≤K-c+s)+.

[0070] p(s-1≤y≤K-c+s|x)×p(s≤y≤K-c+s-1|s-1≤y≤K-c+s)

[0071] Based on the above calculation process, the calculation is automatically performed up to the Cth layer, ultimately yielding the predicted probability p(y=i|x), i∈{1,...,K} for each category, where i represents one of the K categories. The probability of each category will be used to subsequently determine the category to which the input sample belongs.

[0072] The model is used to predict the test set data, and the prediction results are obtained in probabilistic form.

[0073] Step 101: Obtain medical data and input it into the classification model to obtain probability prediction results.

[0074] Step 102: Obtain disease classification prediction results based on probability prediction results.

[0075] Based on the obtained probability prediction results, category merging and normalization are performed to obtain the final disease classification prediction result. The implementation process can be as follows:

[0076] 1) Obtain the probability prediction results, then perform a sequential operation on these results, merging the probabilities of several adjacent categories to obtain a total of C = K-1 decision problems. Each decision problem is represented as follows:

[0077]

[0078] in, Let represent the label for the c-th binary classification problem, where category 1' is used to distinguish it from category 1 of the original problem. The formula for calculating the probability after merging categories is:

[0079]

[0080] Then, the sum of the probabilities of the decision problem included in the probability of each category is calculated, expressed as:

[0081]

[0082] in, This is an indicator function used to determine the class label of the c-th binary classification decision problem corresponding to the current class; it is 1 if the condition is met, and 0 otherwise. Finally, the sum of the probabilities corresponding to each class is subjected to a Softmax normalization operation to obtain the final probability prediction result, expressed as:

[0083]

[0084] The category containing the highest predicted probability is taken as the final classification result. The calculation process for the final prediction result is as follows: Figure 4 As shown.

[0085] Example 2

[0086] This embodiment uses the Knee, Aptos, and AD datasets as a medical grade classification problem to illustrate the method provided in Embodiment 1 above.

[0087] Among them, Knee represents the knee arthritis dataset, which was collected from different hospitals and diagnostic centers and contains X-ray images of the knee joint acquired using a PROTEC PRS 500E X-ray machine. The Aptos dataset is a dataset developed by the Asia-Pacific Teleophthalmology Association, containing fundus images of adults with diabetic retinopathy in rural India. The AD dataset contains MRI images of the four early stages of Alzheimer's disease development. To verify the performance of the disease grading classification method provided in Embodiment 1 of the present invention, this embodiment sets the label noise level ρ to 20%, 30%, and 40% when constructing artificially synthesized label noise, respectively. This verifies that the disease grading classification method provided in Embodiment 1 of the present invention achieves good prediction results on the three datasets. Therefore, the present invention has good application value.

[0088] Table 1 shows the results (mean ± standard deviation) of the classification model on the Knee, Aptos, and AD datasets under label noise conditions, using the commonly used ordinal regression evaluation criteria, Mean Absolute Error (MAE) and Root Mean Squared Error (RMSE). The lower the value (↓), the better the experimental results. The formulas for calculating MAE and RMSE are as follows:

[0089]

[0090] In the formula, N represents the number of samples in the test set, and y m This represents the true label of the m-th test sample. This indicates the predicted label.

[0091] Table 1 shows the experimental results of the disease grading method provided by this invention on the datasets Knee, Aptos, and AD under different noise rates.

[0092]

[0093] Based on the above description, the disease grading method provided by the present invention has the following advantages compared with the prior art:

[0094] (1) This invention focuses on the order information implied in the ordinal regression disease grading, making the classification model closer to the potential real classification model.

[0095] (2) This invention enables the prediction of disease grade classification results to theoretically conform to the order prediction relationship in a realistic sense.

[0096] (3) The present invention makes the noise level of the modified classification problem lower than that of the original classification problem, thereby improving the noise robustness of the classification process.

[0097] Example 3

[0098] A computer device includes: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the disease grading method in Embodiment 1.

[0099] Example 4

[0100] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the disease grading method in Embodiment 1.

[0101] Example 5

[0102] A computer program product includes a computer program that, when executed by a processor, implements the steps of the disease grading method in Embodiment 1.

[0103] Example 6

[0104] A computer device, which may be a database, includes a processor, memory, input / output (I / O) interfaces, and a communication interface. The processor, memory, and I / O interfaces are connected via a system bus, and the communication interface is also connected to the system bus via the I / O interfaces. The processor provides computational and control capabilities. The memory includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores an operating system, computer programs, and a database. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage medium. The database stores pending transactions. The I / O interfaces facilitate information exchange between the processor and external devices. The communication interface allows communication with external terminals via a network connection. When executed by the processor, the computer program implements the disease grading method described in Embodiment 1.

[0105] It should be noted that the object information (including but not limited to object device information, object personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this invention are all information and data authorized by the object or fully authorized by all parties, and the collection, use and processing of related data must comply with the relevant laws, regulations and standards of the relevant countries and regions.

[0106] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments described above. Any references to memory, databases, or other media used in the embodiments provided by this invention can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM). The databases involved in the embodiments provided by this invention may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided by this invention may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, etc., and are not limited to these.

[0107] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0108] This document uses specific examples to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of the present invention. Similar or identical parts between the various embodiments can be referred to mutually. Furthermore, those skilled in the art will recognize that, based on the ideas of the present invention, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A method for classifying disease severity, characterized in that, The method includes: The classification model is constructed by: generating labeled noise to obtain samples with both real and noise labels to generate a medical grade dataset; constructing a hierarchical decision problem based on the medical grade dataset; modifying the network structure of the initial classification model according to the hierarchical decision problem to obtain the classification model; the classification model is used to automate the hierarchical decision process. Acquire medical data and input the medical data into the classification model to obtain probability prediction results; the medical data is Knee; Knee represents the knee arthritis dataset, which is collected from different hospitals and diagnostic centers and contains X-ray images of the knee joint; Based on the probability prediction results, disease classification prediction results are obtained; The classification model is obtained by modifying the network structure of the initial classification model based on the hierarchical decision problem, specifically including: Based on the hierarchical decision problem, the initial classification model outputs the conditional probabilities of the intermediate processes of each level of decision results, and performs recursive calculations to obtain the predicted probabilities of the next level of decision problem. This process is repeated so that the neurons output by the network structure correspond to the conditional probabilities of each decision layer, thus obtaining the classification model. In this network, the decision probability of each layer is represented as a point in the tree structure, and the conditional probability is calculated from the edges. Therefore, the calculation of the corresponding network structure is as follows: ； In the formula, This indicates the node level in the modified network's tree structure. Layer Each node corresponds to a probability value. Represents the first in the edge layer Layer Each edge corresponds to a probability value.

2. The disease severity classification method according to claim 1, characterized in that, The process of generating labeled noise includes: Determine the noise level and determine the flip probability based on the noise level; the flip probability includes the probability that the real label flips into a noise label and the probability that the real label flips into itself; The flipping probability is controlled by the probability density function of a normal distribution, so that the flipping probability decreases as the distance from the real label increases, in order to simulate the noise generation process in real-world situations. The flip probability is normalized using the Softmax method to determine the noise transition matrix; Based on the noise transfer matrix, the real labels are flipped into noise labels.

3. The disease severity classification method according to claim 1, characterized in that, Acquiring medical data and inputting the medical data into the classification model to obtain probability prediction results specifically includes: The medical data is input into the classification model to obtain the probability of the medical data. In the classification model, the probabilities of the medical data are ordered and the probabilities of adjacent categories are merged to obtain multiple decision problems; The sum of probabilities for each category in the decision problem is determined, and the sum of probabilities for each category is normalized using a Softmax normalization operation to obtain the probability prediction result.

4. The disease severity classification method according to claim 1, characterized in that, Based on the probability prediction results, disease classification prediction results are obtained, specifically including: The category containing the maximum value of the probability prediction result is taken as the disease classification prediction result.

5. A computer device, comprising: A memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that the processor executes the computer program to implement the steps of the disease grading method according to any one of claims 1-4.

6. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the computer program implements the steps of the disease grading method according to any one of claims 1-4.

7. A computer program product, comprising a computer program, characterized in that, When executed by a processor, the computer program implements the steps of the disease grading method according to any one of claims 1-4.

Images