Bone mineral density prediction method and device, electronic equipment and storage medium

Abstract

The invention relates to the technical field of data processing, in particular to a bone mineral density prediction method and device, electronic equipment and a storage medium, the method is applied to a bone mineral density prediction model based on TabNet and self-paced learning, and pre-processed clinical scale data is processed by a batch standardization layer and a feature converter module to obtain initial extraction features; obtaining a target decision feature from the first front column extraction feature and the first rear column extraction feature obtained through segmentation through G decision step modules; and inputting the target decision feature into the full connection layer to obtain a bone mineral density prediction value of the to-be-predicted object. Wherein each decision step module comprises an attention converter module, a mask layer, a feature converter module, a cutting layer and an activation function layer. Therefore, feature learning is carried out by using the strong ability of deep learning, the dependence on professional feature engineering is reduced, meanwhile, the capture ability of the model for key features is enhanced through an attention mechanism, and the accuracy and reliability of model prediction can be improved.

Description

Technical Field

[0001] The present application relates to the field of data processing technology, and in particular to a bone density prediction method, device, electronic device and storage medium. Background Art

[0002] In existing prediction models, traditional machine learning methods often rely on complex feature engineering and expertise, which not only increases the difficulty of model development, but also limits the generalization ability and automation of the model. In addition, these methods are usually unable to effectively handle missing values ​​and outliers in the data, which are particularly common in clinical scale data, thus affecting the accuracy and reliability of the prediction model. Summary of the invention

[0003] The embodiments of the present application provide a bone density prediction method, device, electronic device and storage medium, aiming to solve the accuracy and reliability problems of model prediction.

[0004] In a first aspect, an embodiment of the present application provides a bone density prediction method, which is applied to a trained bone density prediction model, wherein the bone density prediction model is a deep learning model based on TabNet and self-paced learning, and the bone density prediction model includes a batch normalization layer, a feature converter module, a slicing layer, G decision step modules and a fully connected layer, where G is a positive integer, and the method includes: Acquiring clinical scale data of the subject to be predicted, and preprocessing the clinical scale data; Extracting features from the preprocessed clinical scale data through the batch normalization layer and the feature converter module to obtain initial extracted features; Inputting the initial extracted features into the segmentation layer for segmentation operation to obtain the first front column extracted features and the first back column extracted features after segmentation; The first front column extracted features and the first rear column extracted features are subjected to feature decision making by G decision step modules to obtain target decision features; each decision step module includes an attention converter module, a mask layer, a feature converter module, a slicing layer, and an activation function layer; The target decision feature is input into the fully connected layer to obtain the bone density prediction value of the object to be predicted.

[0005] In a second aspect, an embodiment of the present application further provides a bone density prediction device, which includes a unit for executing the above method.

[0006] In a third aspect, an embodiment of the present application further provides an electronic device, which includes a memory and a processor, wherein a computer program is stored in the memory, and the processor implements the above method when executing the computer program.

[0007] In a fourth aspect, an embodiment of the present application further provides a computer-readable storage medium, wherein the storage medium stores a computer program, and the computer program can implement the above method when executed by a processor.

[0008] The embodiment of the present application provides a bone density prediction method, device, electronic device and storage medium. The method includes: obtaining initial extracted features from the preprocessed clinical scale data through a batch normalization layer and a feature converter module; obtaining target decision features from the first front column extracted features and the first rear column extracted features obtained through segmentation through G decision step modules; inputting the target decision features into the fully connected layer to obtain the bone density prediction value of the object to be predicted. Each decision step module includes an attention converter module, a mask layer, a feature converter module, a segmentation layer and an activation function layer. The embodiment of the present application reduces the reliance on professional feature engineering by utilizing the powerful ability of deep learning for feature learning, and at the same time enhances the model's ability to capture key features through the attention mechanism, which can improve the accuracy and reliability of model prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0009] The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with the present application and, together with the description, serve to explain the principles of the present application.

[0010] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, for ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative labor.

[0011] One or more embodiments are exemplarily described by pictures in the corresponding drawings, and these exemplified descriptions do not constitute limitations on the embodiments. Elements with the same reference numerals in the drawings represent similar elements, and unless otherwise stated, the figures in the drawings do not constitute proportional limitations.

[0012] Figure 1 A flowchart of a bone density prediction method provided in an embodiment of the present application; Figure 2 A framework diagram of a feature converter module provided in an embodiment of the present application; Figure 3 An overall framework diagram of a bone density prediction model provided in an embodiment of the present application; Figure 4 A framework diagram of an attention converter module provided in an embodiment of the present application; Figure 5 A flowchart of a training process of a bone density prediction model provided in an embodiment of the present application; Figure 6 A schematic diagram of the structure of a bone density prediction device provided in an embodiment of the present application; Figure 7 A schematic diagram of the structure of an electronic device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0013] In order to make the purpose, technical solution and advantages of the embodiments of the present application clearer, the technical solution in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of this application.

[0014] The disclosure below provides many different embodiments or examples to realize the different structures of the present application. In order to simplify the disclosure of the present application, the parts and settings of specific examples are described below. Of course, they are only examples, and the purpose is not to limit the present application. In addition, the present application can repeat reference numbers and / or letters in different examples. This repetition is for the purpose of simplification and clarity, and does not itself indicate the relationship between the various embodiments and / or settings discussed.

[0015] It should be understood that when used in this specification, the terms "include" and "comprising" indicate the presence of described features, integers, steps, operations, elements and / or components, but do not exclude the presence or addition of one or more other features, integers, steps, operations, elements, components and / or collections thereof.

[0016] It should also be understood that the terms used in this application specification are only for the purpose of describing specific embodiments and are not intended to limit the application. As used in this application specification, unless the context clearly indicates otherwise, the singular forms "a", "an" and "the" are intended to include plural forms.

[0017] It should be further understood that the term “and / or” used in the specification of the present application refers to any combination and all possible combinations of one or more of the associated listed items, and includes these combinations.

[0018] As used in the present specification, the term "if" can be interpreted as "when" or "uponce" or "in response to determining" or "in response to detecting" depending on the context. Similarly, the phrase "if it is determined" or "if [the described condition or event] is detected" can be interpreted as meaning "uponce it is determined" or "in response to determining" or "uponce [the described condition or event] is detected" or "in response to detecting [the described condition or event]" depending on the context.

[0019] It should be noted that the feasibility of predicting bone density with artificial intelligence algorithms and the idea of ​​productization is a very promising field. Artificial intelligence models can also have an important impact on bone density prediction by training clinical data of the population (such as age, gender, weight, medical history, etc.). The feasibility of bone density prediction depends on a large amount of labeled and reliable clinical data. By collecting a large amount of data, effectively labeling and cleaning it, the labeled data set is used to train and verify the model to ensure that the model can generalize to unseen data. Existing machine learning models often require complex data preprocessing, data cleaning and other steps, and it is difficult to mine the high-order information contained in clinical data. In addition, with the development of deep learning, models such as convolutional neural networks and self-attention mechanisms have gradually been used in this field. Although these models have advantages in deep information mining, problems such as poor interpretability and high training costs have caused difficulties in practical applications.

[0020] However, in existing prediction models, traditional machine learning methods often rely on complex feature engineering and expertise, which not only increases the difficulty of model development, but also limits the generalization ability and automation of the model. In addition, these methods are usually unable to effectively handle missing values ​​and outliers in the data, which are particularly common in clinical scale data, thus affecting the accuracy and reliability of the prediction model.

[0021] In order to solve the technical problem of low accuracy and reliability of prediction models in the prior art, the present application provides a bone density prediction method, device, electronic device and storage medium. By utilizing the powerful ability of deep learning for feature learning, the dependence on professional feature engineering is reduced. At the same time, the model's ability to capture key features is enhanced through the attention mechanism, which can improve the accuracy and reliability of model prediction.

[0022] Figure 1A flowchart of a bone density prediction method provided in an embodiment of the present application. The bone density prediction method is applied to an electronic device, and the electronic device is installed or deployed with a trained bone density prediction model, and the bone density prediction model is a deep learning model based on TabNet and self-step learning, and the bone density prediction model includes a batch normalization layer, a feature converter module, a slicing layer, G decision step modules and a fully connected layer, where G is a positive integer. Here, TabNet is the abbreviation of Tabular Network, which is mainly used to process tabular data. Therefore, the bone density prediction model provided in the embodiment of the present application can process tabular data based on deep learning, and the bone density prediction model processes feature selection and conversion in tabular data through a series of steps. It is designed to automatically learn feature representations in tabular data, and can effectively select the most relevant features when processing high-dimensional tabular data.

[0023] like Figure 1 As shown, the method comprises the following steps: S101, obtaining clinical scale data of a subject to be predicted, and preprocessing the clinical scale data.

[0024] In the specific implementation, the collected clinical scale data of the object to be predicted is obtained. The clinical scale data refers to the data collected in the medical clinical environment by using various standardized measurement tools (scales). Clinical scale data includes information such as age, occupation, gender, height, weight, smoking status, drinking status, medical history, etc. Then, the clinical scale data is preprocessed to obtain a feature set, which is obtained by preprocessing the clinical scale data.

[0025] Among them, during model training, clinical scale data of a large number of sample objects and the true value of bone density of each sample object can be obtained. The true value of bone density is used as the label, and the remaining features are used as the sample feature set. The sample feature set may include the following features: age, occupation, gender, height, weight, smoking status, drinking status, parents with hip fracture / osteoporosis, history of falls, diabetes, dementia, cancer, heart disease / angina pectoris / stroke or acute cerebral ischemia, asthma or chronic obstructive pulmonary disease, chronic liver disease, chronic kidney disease, Parkinson's disease, rheumatoid arthritis or systemic lupus erythematosus, gastrointestinal dyspepsia, epilepsy or taking anticonvulsant drugs, antidepressants, regular use of steroids, hormone replacement therapy using only estrogen, endocrine diseases, calcium supplementation, use of anti-osteoporosis drugs and other medical history information.

[0026] Specifically, clinical scale data can be preprocessed with the help of Pandas toolkit, Sklearn toolkit, feature_engine toolkit, etc. Among them, first, the Pandas toolkit can be used to read the scale data as a pandas object. Then, the .astype() method can be used to convert numerical features to float64 type and categorical features to object type. Furthermore, .astype() can be used to convert binary features in categorical features to int64 type. Finally, the LabelEncoder method in the Sklearn toolkit and the MeanEncoder method in the feature_engine toolkit are used to encode categorical features other than binary features.

[0027] For example, the data preprocessing process in a bone density prediction method provided in an embodiment of the present application may include the following steps: Step 1011, reading the clinical scale data of the object to be predicted; Step 1012: define data types; Step 1013, category coding; Step 1014: construct a feature set.

[0028] S102, extracting features from the preprocessed clinical scale data through the batch normalization layer and the feature converter module to obtain initial extracted features.

[0029] In the specific implementation, batch normalization layers and feature transformer modules are used to extract features from the preprocessed clinical scale data.

[0030] Among them, the batch normalization layer is used to normalize the input data so that the distribution of the data is more in line with the requirements of model training. The feature transformer module is a concept in the field of machine learning and data processing. It mainly refers to a series of methods or modules used to transform the original data features, with the purpose of converting the original features into new feature forms that are more suitable for model processing, better reflect the internal structure of the data, or are more conducive to improving model performance.

[0031] S103: Input the initial extracted features into a segmentation layer for segmentation to obtain a first front column extracted feature and a first back column extracted feature after segmentation.

[0032] In a specific implementation, a segmentation operation is performed on the extracted initial extracted features to obtain the first front column extracted features and the first back column extracted features.

[0033] Among them, in TabNet, the splitting operation (Feature Splitting) is a key step. Its main purpose is to process the features after feature selection. This operation is somewhat similar to grouping or dividing features, but it is more skillful than simple grouping. The splitting operation divides the features selected in each decision step into two groups. One group is used to generate an evaluation of the importance of the feature (by calculating and updating masks, etc.), and the other group is used for actual feature conversion and subsequent calculations to produce the output or intermediate results of the model.

[0034] For example, during the model training process, after passing through the feature converter module, the extracted initial extraction features are obtained. ,in, It is the size of Batch Size. Batch Size is an important parameter in the deep learning and machine learning training process. It refers to the number of samples input into the model at the same time in one training iteration. Represents the size of the hidden unit. For the extracted initial features, first perform feature segmentation to obtain the first front-row extracted features And the first and last columns extract features ,in Before the Representative Column features, After the representative Column features. As the input of the attention transformer model in the decision step module for subsequent feature selection.

[0035] S104. Perform feature decision on the first front column extracted features and the first rear column extracted features through G decision step modules to obtain target decision features; each decision step module includes an attention converter module, a mask layer, a feature converter module, a slicing layer and an activation function layer.

[0036] In the specific implementation, a split operation is performed on the initial extracted features, and the features are input into the attention converter module and the mask layer to realize feature selection. In addition, the selected features are passed through the feature converter module, the split operation and the activation function to obtain the output of the decision step. The above operation is repeated G times to obtain the output of G decision steps as the target decision features.

[0037] Here, feature selection is implemented through the mask layer of each decision step module, and the selection of features is implemented by the attention converter module of the decision step. The attention converter module realizes feature selection for the current decision step by learning the mask.

[0038] It can be understood that the embodiment of the present application provides a bone density prediction method, which can give an accurate bone density prediction result based on clinical scale data and comprehensive consideration of multi-dimensional clinical factors.

[0039] S105: Input the target decision feature into the fully connected layer to obtain the bone density prediction value of the object to be predicted. The bone density prediction value of the object to be predicted can also be sent to a preset terminal.

[0040] It should be noted that the bone density prediction method proposed in the embodiment of the present application is applied to a deep learning model based on TabNet and self-paced learning. By leveraging TabNet's powerful deep-level feature mining capabilities and effectively integrating the self-paced learning mechanism, the robustness of the prediction model is improved. Specifically, a bone density prediction backbone network is constructed based on TabNet, which can extract high-dimensional features from complex tabular data and enhance the model's ability to capture key features through an attention mechanism. This prediction method can not only provide people with preliminary self-testing, but also improve the efficiency of bone density prediction.

[0041] In a possible implementation, the feature converter module includes a cross-decision step sharing unit and a single-step independent decision unit, and the cross-decision step sharing unit and the single-step independent decision unit each include a plurality of decision step networks, and each decision step network includes a fully connected layer, a batch normalization layer, and a gated linear unit layer; S102 extracts features from the preprocessed clinical scale data through the batch normalization layer and the feature converter module to obtain initial extracted features, including the following steps: Step 1021: input the preprocessed clinical scale data into the batch normalization layer for normalization operation to obtain normalized normalized features.

[0042] During the model training process, the clinical scale data of a large number of samples after preprocessing can be defined as ,in is the batch size, is the number of dimensions of the feature. Before entering the feature converter module, the batch normalization layer is used to normalize the features to obtain the normalized features. , and then input into the feature converter for feature extraction.

[0043] Step 1022: Input the normalized feature into the cross-decision step sharing unit to extract common features to obtain initial intermediate features.

[0044] Step 1023: input the initial intermediate features into the single-step independent decision unit to extract specific features to obtain the initial extracted features.

[0045] Here, the shared units across decision steps are common throughout the entire TabNet, while the single-step independent decision units are passed through the gated linear unit layer to prevent gradient vanishing. For example, Figure 2 A framework diagram of a feature converter module provided in an embodiment of the present application; Figure 2 As shown, here, the feature converter module consists of two parts, a shared unit across decision steps and a single-step independent decision unit. Each decision step network includes a fully connected layer, a batch normalization layer, and a gated linear unit layer. Among them, the purpose of the shared unit across decision steps is to extract the common information of the features while reducing the number of parameters of the model. The function of the single-step independent decision unit is to extract the features in each decision step. In both parts of the feature converter module, the fully connected layer + batch normalization layer + gated linear unit are stacked. At the same time, in order to ensure the stability of network training, each unit is multiplied by, for example Residual connections are used to ensure that the variance of the entire network does not change significantly.

[0046] In a possible implementation, the cross-decision step sharing unit includes Q decision step networks, where Q is a positive integer; step 1022, inputting the normalized feature into the cross-decision step sharing unit to extract common features to obtain initial intermediate features, includes the following steps: Step (11), inputting the normalized feature into the first decision step network in the cross-decision step sharing unit to obtain a first shared intermediate feature; Step (12), inputting the first shared intermediate feature into the second decision step network in the cross-decision step sharing unit to obtain a second shared intermediate feature; Step (13), repeating the above process in other decision step networks until the Q-1th shared intermediate feature is input into the Qth decision step network in the cross-decision step sharing unit to obtain the Qth shared intermediate feature; Step (14): perform residual connection on the Q shared intermediate features to obtain the initial intermediate features.

[0047] In a possible implementation, the single-step independent decision unit includes K decision-step networks, where K is a positive integer; step 1023, inputting the initial intermediate feature into the single-step independent decision unit to extract specific features to obtain the initial extracted features, includes the following steps: Step (21), inputting the initial intermediate feature into the first decision step network in the single-step independent decision unit to obtain a first independent intermediate feature, and performing a residual connection between the first independent intermediate feature and the initial intermediate feature to obtain the processed first independent intermediate feature; Step (22), inputting the processed first independent intermediate feature into the second decision step network in the single-step independent decision unit to obtain a second independent intermediate feature, and performing a residual connection between the second independent intermediate feature and the first independent intermediate feature to obtain the processed second independent intermediate feature; Step (23), repeatedly performing the above processing in other decision step networks until the processed K-1th independent intermediate feature is input into the Kth decision step network in the single-step independent decision unit to obtain the Kth independent intermediate feature, and the Kth independent intermediate feature and the K-1th independent intermediate feature are residually connected to obtain the processed Kth independent intermediate feature; Step (24), determining the processed Kth independent intermediate feature as the initial extracted feature.

[0048] In a possible implementation manner, the step of performing feature decision on the first front-column extracted features and the first rear-column extracted features through G decision-step modules in S104 to obtain target decision features includes the following steps: Step 1041, input the first front column extracted features into the attention converter module in the first decision step module to obtain the first selected features; Step 1042: input the first selected feature and the normalized feature into the attention converter module in the first decision step module to obtain a first mask feature; the normalized feature is obtained by inputting the preprocessed clinical scale data into the batch normalization layer; Step 1043, the first mask feature is passed through the feature converter module and the slice layer in the first decision step module to obtain a second front column extraction feature and a second back column extraction feature; Step 1044, input the second rear column extracted features into the activation function layer in the first decision step module to obtain the first decision feature; Step 1045: Use the second front-row extracted features as the attention converter module in the first decision step module, start a new decision step cycle, and obtain the decision features corresponding to each decision step module; Step 1046: Aggregate the decision features corresponding to the G decision-step modules to obtain the target decision features.

[0049] Here, for any decision step module, its role is to select the most representative features through the attention mechanism to improve the model effect. The attention converter module in the decision step module realizes the feature selection of the current decision step by learning the mask: ,in, is the feature mask of the current decision step, Represents the current decision step. Sparsemax (sparse maximum activation function) is a variant of the Softmax activation function, which can obtain sparser output results and promote feature selection. Represents the feature information obtained by the previous decision step, and the dimension is , represents the fully connected layer and the batch normalization layer, The term "prior scales" is a specific term in the decision-making process. "Prior" usually means "prior". In fields such as Bayesian statistics, prior information is a kind of cognition or assumption about parameters before observing data. "Scales" means "scale, proportion". In the current decision step, the "prior scales" item may refer to a scale or proportion-related factor based on prior knowledge or determined in the previous step. Calculated in the previous decision step, the calculation formula is: ; in, The prior scales item represents the current decision step, indicating how many specific features were used before the decision step. is the sparse regularization term weight, which is used to add constraints to the feature sparsity in the feature selection stage. The smaller the feature selection, the sparser it is. Then, Multiply the feature elements obtained by the feature selection to achieve the feature selection of the current decision step. Input the selected features into the feature converter of the current decision step to start a new decision step cycle. After feature selection, the processed features are divided into two parts by the split operation, one part is used as the output of the current decision step, and the other part is used as the input information of the next decision step. All the above steps are repeated G times. Finally, borrowing the idea of ​​tree model aggregation, the features of all decision steps are aggregated: ; in, The output after feature aggregation for all decision steps, i.e., the target decision feature, represents the linear rectification function, For the The output of the decision step. Input into the fully connected layer to obtain the final bone density prediction value.

[0050] In addition, the bone density prediction model also outputs the importance of the features. Specifically, the input features are At the decision step, the feature dimension is obtained through the feature converter module , and then through the segmentation operation, one of the features is recorded as , and then go through After the activation function, the sum is obtained Contribution within a decision step: ; in, for One dimension in , it has no contribution to the final output. In the The contribution value of the decision step With sample In the The feature importance matrix within the decision step Multiply, and then take the sample The results of each decision step are summed. Intuitively, when The larger the value, the more obvious the impact on the model output. Can be used as the The feature importance matrix within the decision step The weighting of To define the importance value of each feature of the sample; ; in is the number of features. In the embodiments of this application .

[0051] Figure 3 An overall framework diagram of a bone density prediction model provided for an embodiment of the present application, wherein the bone density prediction model includes a batch normalization layer, a feature converter module, a slicing layer, G decision step modules and a fully connected layer, each decision step module includes an attention converter module, a mask layer, a feature converter module, a slicing layer and an activation function layer.

[0052] Figure 4 A framework diagram of the attention converter module provided in the embodiment of the present application; Figure 4 As shown in the figure, the attention converter module includes a fully connected layer, a batch normalization layer, and a Sparsemax activation function layer. In the attention converter module architecture, the scaling factor mainly appears in the scaled dot-product attention mechanism. It is used to scale the calculated attention score (obtained by the dot product of the query vector and the key vector).

[0053] In a possible implementation, the bone density prediction model is trained according to the following steps: obtaining clinical scale data of multiple sample objects, as well as the true bone density value and sample difficulty weight of each sample object; inputting the clinical scale data of each sample object into a deep learning model to obtain a predicted bone density value for each sample object; optimizing the network parameters of the deep learning model based on the mean square error loss value between the true bone density value and the predicted bone density value of each sample object, the sample difficulty weight, and a self-paced learning mechanism to obtain a trained bone density prediction model.

[0054] In addition, the self-paced learning mechanism is used to optimize the robustness of the bone density prediction model. Since the application embodiment is intended to predict the bone density of the object to be predicted, the training goal of the model is essentially a regression task. Therefore, for the final output obtained by TabNet, we use the mean square error loss to supervise the training of the model, and the loss is defined as follows: ; in, Represents the true value of bone density. Inspired by the human learning process, the concept of self-paced learning (SPL) was proposed, which can adjust the learning process of the model, allowing the model to learn simple samples first, and gradually introduce difficult samples for training as the training progresses, thereby improving the robustness of the model. Therefore, the system intends to adopt self-paced learning to constrain the training process of TabNet in bone density and fracture prediction tasks.

[0055] In one possible implementation, the following objective function is minimized to optimize the network parameters of the deep learning model: ; in, is the number of sample objects, is the sample training difficulty threshold, are the network parameters of the deep learning model, is the sample difficulty weight of each sample object, , is the mean square error loss value, is the self-paced learning regularization term, is a positive integer, .

[0056] Specifically, the sample difficulty weight is introduced , and use the mean squared error loss value for training As a measure of sample difficulty, here, we use represents the training data in each training subprocess, then the above equation can be rewritten as: ; ; Among them, the self-paced learning regularization term The constrained training process selects low training loss, that is, low-difficulty samples to enter the training process. represents the number of samples in this training round, Indicates conditional restrictions, limiting This application adopts a hard self-paced regularization term, which is expressed as follows: ; Then, an approximate solution to the above objective can be obtained: ; in, Represents the mean square error loss value of training TabNet. In the initial stage of training, a small subset of training data is randomly selected to train the model. During the training process, more difficult samples are added according to the training loss of the samples, which enables the model to learn more robust information related to the patient's bone density and further improve the performance of the bone density prediction model.

[0057] Here, the robustness of TabNet is improved by introducing a self-paced learning mechanism, and the generalization ability of the model is enhanced by imitating the human learning mechanism from easy to difficult. By using the self-paced learning mechanism, the model can gradually add samples to the model training process in the order of low training difficulty to high training difficulty, so as to better learn the deep information in the data and improve the robustness and generalization ability of the model.

[0058] Figure 5 A flowchart of a training process of a bone density prediction model provided in an embodiment of the present application is shown in FIG. Figure 5 As shown in FIG. 1 , a possible specific training implementation process of a bone density prediction model based on TabNet and self-paced learning includes the following steps: S1. Acquire clinical scale data collected from multiple sample subjects; S2, preprocessing the clinical scale data of multiple sample subjects to obtain a preprocessed sample matrix; S3, after using the batch normalization layer to normalize the samples, input them into the feature converter module for feature extraction; S4, after the extracted features are segmented, they are input into the attention converter module and the mask layer to achieve feature selection; S5, the selected features are passed through the feature converter module, the slicing layer and the ReLU activation function layer to obtain the output of the decision step; S6. Input the final features into the fully connected layer to obtain the bone density prediction value of each sample object; S7. Based on self-paced learning, samples with low learning difficulty are gradually input into the model to improve the robustness of the model.

[0059] See also Figure 6 , Figure 6 is a schematic block diagram of a bone density prediction device 600 provided in an embodiment of the present application. Corresponding to the above bone density prediction method, the present application also provides a bone density prediction device 600. The bone density prediction device 600 includes a unit for executing the above bone density prediction method, and the bone density prediction device 600 can be configured in a desktop computer, a tablet computer, a laptop computer, and other terminals. Specifically, the bone density prediction device 600 includes an acquisition unit 610, a first determination unit 620, a second determination unit 630, a third determination unit 640, and a prediction unit 650.

[0060] An acquisition unit 610 is used to acquire clinical scale data of the object to be predicted and preprocess the clinical scale data; A first determining unit 620 is used to extract features from the preprocessed clinical scale data through the batch normalization layer and the feature converter module to obtain initial extracted features; The second determining unit 630 is used to input the initial extracted features into the segmentation layer for segmentation operation to obtain the first front column extracted features and the first back column extracted features after segmentation; A third determination unit 640 is used to perform feature decision on the first front column extracted features and the first rear column extracted features through G decision step modules to obtain target decision features; each decision step module includes an attention converter module, a mask layer, a feature converter module, a slicing layer, and an activation function layer; The prediction unit 650 is used to input the target decision feature into the fully connected layer to obtain the bone density prediction value of the object to be predicted.

[0061] In an embodiment of the present application, the preprocessed clinical scale data is subjected to a batch normalization layer and a feature converter module to obtain initial extracted features; the first front column extracted features and the first rear column extracted features obtained by segmentation are subjected to G decision step modules to obtain target decision features; the target decision features are input into the fully connected layer to obtain the bone density prediction value of the object to be predicted. Each decision step module includes an attention converter module, a mask layer, a feature converter module, a segmentation layer, and an activation function layer. In this way, the present application reduces the reliance on professional feature engineering by utilizing the powerful capabilities of deep learning for feature learning, and at the same time enhances the model's ability to capture key features through the attention mechanism, which can improve the accuracy and reliability of model predictions.

[0062] like Figure 7As shown, an embodiment of the present application provides an electronic device, including a processor 111, a communication interface 112, a memory 113 and a communication bus 114, wherein the processor 111, the communication interface 112, and the memory 113 communicate with each other through the communication bus 114. Memory 113, used for storing computer programs; In one embodiment of the present application, the processor 111 is used to execute the program stored in the memory 113 to implement the bone density prediction provided by any of the above method embodiments, including: Acquiring clinical scale data of the subject to be predicted, and preprocessing the clinical scale data; Extracting features from the preprocessed clinical scale data through the batch normalization layer and the feature converter module to obtain initial extracted features; Inputting the initial extracted features into the segmentation layer for segmentation operation to obtain the first front column extracted features and the first back column extracted features after segmentation; The first front column extracted features and the first rear column extracted features are subjected to feature decision making by G decision step modules to obtain target decision features; each decision step module includes an attention converter module, a mask layer, a feature converter module, a slicing layer, and an activation function layer; The target decision feature is input into the fully connected layer to obtain the bone density prediction value of the object to be predicted.

[0063] It is understood by those skilled in the art that all or part of the processes in the method for implementing the above embodiment can be completed by instructing the relevant hardware through a computer program. The computer program can be stored in a storage medium, which is a computer-readable storage medium. The computer program is executed by at least one processor in the computer system to implement the process steps of the embodiment of the above method.

[0064] Therefore, an embodiment of the present application further provides a computer-readable storage medium having a computer program stored thereon, and when the computer program is executed by a processor, the steps of the bone density prediction method provided in any of the aforementioned method embodiments are implemented.

[0065] The storage medium is a physical, non-transient storage medium, such as a USB flash drive, a mobile hard disk, a read-only memory (ROK), a magnetic disk or an optical disk, etc., which can store program codes. The computer-readable storage medium can be non-volatile or volatile.

[0066] Those of ordinary skill in the art will appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of the two. In order to clearly illustrate the interchangeability of hardware and software, the composition and steps of each example have been generally described in terms of function in the above description. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of this application.

[0067] In the several embodiments provided in the present application, it should be understood that the disclosed devices and methods can be implemented in other ways. For example, the device embodiments described above are only schematic. For example, the division of each unit is only a logical function division, and there may be other division methods in actual implementation. For example, multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed.

[0068] The steps in the method of the embodiment of the present application can be adjusted in order, combined and deleted according to actual needs. The units in the device of the embodiment of the present application can be combined, divided and deleted according to actual needs. In addition, the functional units in the various embodiments of the present application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0069] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a storage medium. Based on this understanding, the technical solution of the present application is essentially or the part that contributes to the prior art, or all or part of the technical solution can be embodied in the form of a software product, and the computer software product is stored in a storage medium, including a number of instructions for an electronic device (which can be a personal computer, terminal, or network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present application.

[0070] In the above embodiments, the description of each embodiment has its own emphasis. For parts that are not described in detail in a certain embodiment, reference can be made to the relevant descriptions of other embodiments.

[0071] Obviously, those skilled in the art can make various changes and modifications to the present application without departing from the spirit and scope of the present application. Thus, if these modifications and variations of the present application are within the scope of the present application and its equivalents, the present application is also intended to include these modifications and variations.

[0072] The above description is only a specific implementation mode of the present application, but the protection scope of the present application is not limited thereto. Any technician familiar with the technical field can easily think of various equivalent modifications or substitutions within the technical scope disclosed in the present application, and these modifications or substitutions should be included in the protection scope of the present application.

Claims

1. A method for predicting bone density, characterized in that: Applied to an electronic device, the electronic device is deployed with a trained bone density prediction model, the bone density prediction model is a deep learning model based on TabNet and self-paced learning, the bone density prediction model includes a batch normalization layer, a feature converter module, a slicing layer, G decision step modules and a fully connected layer, G is a positive integer, and the method includes: Acquiring clinical scale data of the subject to be predicted, and preprocessing the clinical scale data; Extracting features from the preprocessed clinical scale data through the batch normalization layer and the feature converter module to obtain initial extracted features; Inputting the initial extracted features into the segmentation layer for segmentation operation to obtain the first front column extracted features and the first back column extracted features after segmentation; The first front column extracted features and the first rear column extracted features are subjected to feature decision making by G decision step modules to obtain target decision features; each decision step module includes an attention converter module, a mask layer, a feature converter module, a slicing layer, and an activation function layer; The target decision feature is input into the fully connected layer to obtain the bone density prediction value of the object to be predicted.

2. The method according to claim 1, characterized in that The feature converter module includes a cross-decision step sharing unit and a single-step independent decision unit, each of which includes a plurality of decision step networks, each of which includes a fully connected layer, a batch normalization layer, and a gated linear unit layer; the pre-processed clinical scale data is subjected to feature extraction through the batch normalization layer and the feature converter module to obtain initial extracted features, including: Inputting the preprocessed clinical scale data into the batch normalization layer for normalization operation to obtain normalized normalized features; Inputting the normalized features into the cross-decision step sharing unit to extract common features to obtain initial intermediate features; The initial intermediate features are input into the single-step independent decision unit to extract specific features, thereby obtaining the initial extracted features.

3. The method according to claim 2, characterized in that The cross-decision step sharing unit includes Q decision step networks, where Q is a positive integer; the step of inputting the normalized feature into the cross-decision step sharing unit to extract common features to obtain initial intermediate features includes: Inputting the normalized feature into the first decision step network in the cross-decision step sharing unit to obtain a first shared intermediate feature; Inputting the first shared intermediate feature into the second decision step network in the cross-decision step sharing unit to obtain a second shared intermediate feature; Repeat the above process in other decision-step networks until the Q-1th shared intermediate feature is input into the Qth decision-step network in the cross-decision-step sharing unit to obtain the Qth shared intermediate feature; The Q shared intermediate features are residually connected to obtain the initial intermediate features.

4. The method according to claim 2, characterized in that: The single-step independent decision unit includes K decision-step networks, where K is a positive integer; the initial intermediate feature is input into the single-step independent decision unit to extract specific features to obtain the initial extracted features, including: Inputting the initial intermediate feature into the first decision step network in the single-step independent decision unit to obtain a first independent intermediate feature, and performing a residual connection between the first independent intermediate feature and the initial intermediate feature to obtain the processed first independent intermediate feature; Inputting the processed first independent intermediate feature into the second decision step network in the single-step independent decision unit to obtain a second independent intermediate feature, and performing a residual connection between the second independent intermediate feature and the first independent intermediate feature to obtain the processed second independent intermediate feature; Repeat the above process in other decision-step networks until the processed K-1th independent intermediate feature is input into the Kth decision-step network in the single-step independent decision unit to obtain the Kth independent intermediate feature, and the Kth independent intermediate feature and the K-1th independent intermediate feature are residually connected to obtain the processed Kth independent intermediate feature; The processed Kth independent intermediate feature is determined as the initial extracted feature.

5. The method according to claim 1, characterized in that The step of performing feature decision on the first front column extracted features and the first rear column extracted features through G decision step modules to obtain target decision features includes: Inputting the first front column extracted features into the attention converter module in the first decision step module to obtain a first selected feature; Inputting the first selected feature and the normalized feature into the attention converter module in the first decision step module to obtain a first mask feature; the normalized feature is obtained by inputting the preprocessed clinical scale data into the batch normalization layer; Passing the first mask feature through the feature converter module and the slice layer in the first decision step module to obtain a second front column extracted feature and a second back column extracted feature; Inputting the second rear column extracted features into the activation function layer in the first decision step module to obtain the first decision feature; The second front-row extracted feature is used as the attention converter module in the first decision step module, a new decision step cycle is started, and the decision feature corresponding to each decision step module is obtained; The decision features corresponding to the G decision-step modules are aggregated to obtain the target decision features.

6. The method according to claim 1, characterized in that The bone density prediction model is obtained by training according to the following steps: Obtain clinical scale data of multiple sample subjects, as well as the true value of bone density and sample difficulty weight of each sample subject; The clinical scale data of each sample subject is input into the deep learning model to obtain the predicted bone density value of each sample subject; Based on the mean square error loss value between the true bone density value and the predicted bone density value of each sample object, the sample difficulty weight, and the self-paced learning mechanism, the network parameters of the deep learning model are optimized to obtain a trained bone density prediction model.

7. The method according to claim 6, characterized in that Minimize the following objective function to optimize the network parameters of the deep learning model: ； in, is the number of sample objects, is the sample training difficulty threshold, are the network parameters of the deep learning model, is the sample difficulty weight of each sample object, , is the mean square error loss value, is the self-paced learning regularization term, is a positive integer, .

8. A bone density prediction device, characterized in that: The method comprises a unit for executing the method according to any one of claims 1 to 7.

9. An electronic device, characterized in that: The electronic device includes a memory and a processor, the memory stores a computer program, and the processor implements the method according to any one of claims 1 to 7 when executing the computer program.

10. A computer-readable storage medium, characterized in that: The storage medium stores a computer program, and when the computer program is executed by a processor, the method according to any one of claims 1 to 7 can be implemented.

Images