Bone age prediction method and device, equipment, storage medium and computer program product

Abstract

The invention relates to a bone age prediction method and device, equipment, a storage medium and a computer program product. The method comprises the following steps: acquiring skeleton input data, and preprocessing the skeleton input data to obtain preprocessed skeleton input data; detecting all joints in the preprocessed skeleton input data through a target detection algorithm to obtain a key skeleton region detection result; target joint screening is carried out according to the key skeleton area detection result, and key skeleton joint screening features are obtained; and according to the key bone joint screening features, performing bone age grade prediction through a pre-trained deep neural network model, and outputting a bone age prediction result. By adopting the method, the bone age evaluation efficiency and the bone age evaluation accuracy can be improved.

Description

Technical Field

[0001] The present application relates to the technical field of bone age prediction, and in particular to a bone age prediction method, apparatus, device, storage medium and computer program product. Background Art

[0002] Although there are some computer-assisted assessment tools and machine learning methods in the existing technology that can accelerate the assessment process to a certain extent, these methods still have certain limitations, such as small data sets, limited model accuracy and generalization capabilities. In addition, most existing automatic assessment systems rely on a single assessment standard or technology, lack integrated and intelligent functions, and cannot fully adapt to the complex needs of clinical practice. Therefore, how to improve the efficiency and accuracy of bone age assessment is a technical problem that needs to be solved urgently. Summary of the Invention

[0003] Based on this, it is necessary to provide a bone age prediction method, device, equipment, storage medium and computer program product to address the above technical problems, which can improve the efficiency and accuracy of bone age assessment.

[0004] In a first aspect, the present application provides a method for predicting bone age, the method comprising:

[0005] Acquiring skeleton input data, and preprocessing the skeleton input data to obtain preprocessed skeleton input data;

[0006] Detecting all joints in the pre-processed skeletal input data using a target detection algorithm to obtain key skeletal region detection results;

[0007] Perform target joint screening based on the key bone region detection results to obtain key bone joint screening features;

[0008] Based on the key skeletal joint screening features, a pre-trained deep neural network model is used to predict the bone age grade and output the bone age prediction result;

[0009] The target detection algorithm includes a lightweight multi-scale feature extraction module, which is configured to capture multi-scale features through parallel convolution kernels of different sizes and enhance key feature channels in combination with a channel attention mechanism.

[0010] In one embodiment, the acquiring of the skeleton input data and preprocessing the skeleton input data to obtain preprocessed skeleton input data includes:

[0011] Performing data enhancement processing on the skeleton input data to obtain preprocessed skeleton input data, wherein the data enhancement processing includes at least one of image rotation, image translation, image scaling, image flipping, adding Gaussian noise, and histogram equalization;

[0012] The preprocessed skeleton input data is divided into a training set, a validation set and a test set according to a preset ratio.

[0013] In one embodiment, the target detection algorithm includes an input layer, a backbone network layer and a head network layer. The input layer is used to obtain preprocessed bone input data, the backbone network layer is used to extract bone key image features from the preprocessed bone input data, and the head network layer is used to fuse the feature information of the bone key image features, predict the category and position of the target bone, and obtain the key bone area detection result.

[0014] In one embodiment, the backbone network layer includes a first feature extraction module, a second feature extraction module, and a third feature extraction module;

[0015] The first feature extraction module includes a convolution layer, a batch normalization layer and an activation function layer;

[0016] The second feature extraction module includes a first sampling module and a second sampling module;

[0017] The third feature extraction module includes an efficient layer aggregation network module.

[0018] In one embodiment, the target joint screening is performed based on the key bone region detection result to obtain key bone joint screening features, including:

[0019] Target joints are screened according to the key bone region detection results and preset bone age standards to obtain key bone joint screening features, wherein the key bone joint screening features include a key bone region, and the key bone region includes at least two of the following target joints:

[0020] Radius, ulna, metacarpal I, metacarpal III, metacarpal V, proximal phalanx I, proximal phalanx III, proximal phalanx V, middle phalanx III, middle phalanx V, distal phalanx I, distal phalanx III and distal phalanx V.

[0021] In one embodiment, the bone age grade prediction is performed based on the key skeletal joint screening features using a pre-trained deep neural network model, and the bone age prediction result is output, including:

[0022] The key skeletal joint screening features are divided into multiple groups along the channel dimension through a pre-trained deep neural network model and cross-group feature aggregation calculation is performed to predict the level of each target joint and obtain the prediction result of each target joint;

[0023] The bone age prediction result is calculated and output according to the prediction result of each target joint and the preset bone age standard.

[0024] In a second aspect, the present application also provides a bone age prediction device. The device comprises:

[0025] A preprocessing module, configured to obtain skeleton input data and preprocess the skeleton input data to obtain preprocessed skeleton input data;

[0026] A target detection module is used to detect all joints in the pre-processed skeleton input data using a target detection algorithm to obtain key skeleton area detection results;

[0027] A joint screening module is used to screen target joints based on the key bone region detection results to obtain key bone joint screening features;

[0028] A bone age prediction module is used to predict bone age grade based on the key skeletal joint screening features using a pre-trained deep neural network model and output a bone age prediction result;

[0029] The target detection algorithm includes a lightweight multi-scale feature extraction module, which is configured to capture multi-scale features through parallel convolution kernels of different sizes and enhance key feature channels in combination with a channel attention mechanism.

[0030] In a third aspect, the present application further provides a computer device. The computer device includes a memory and a processor, wherein the memory stores a computer program, and when the processor executes the computer program, the following steps are performed:

[0031] Acquiring skeleton input data, and preprocessing the skeleton input data to obtain preprocessed skeleton input data;

[0032] Detecting all joints in the pre-processed skeletal input data using a target detection algorithm to obtain key skeletal region detection results;

[0033] Perform target joint screening based on the key bone region detection results to obtain key bone joint screening features;

[0034] Based on the key skeletal joint screening features, a pre-trained deep neural network model is used to predict the bone age grade and output the bone age prediction result;

[0035] The target detection algorithm includes a lightweight multi-scale feature extraction module, which is configured to capture multi-scale features through parallel convolution kernels of different sizes and enhance key feature channels in combination with a channel attention mechanism.

[0036] In a fourth aspect, the present application further provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the following steps:

[0037] Acquiring skeleton input data, and preprocessing the skeleton input data to obtain preprocessed skeleton input data;

[0038] Detecting all joints in the pre-processed skeletal input data using a target detection algorithm to obtain key skeletal region detection results;

[0039] Perform target joint screening based on the key bone region detection results to obtain key bone joint screening features;

[0040] Based on the key skeletal joint screening features, a pre-trained deep neural network model is used to predict the bone age grade and output the bone age prediction result;

[0041] The target detection algorithm includes a lightweight multi-scale feature extraction module, which is configured to capture multi-scale features through parallel convolution kernels of different sizes and enhance key feature channels in combination with a channel attention mechanism.

[0042] In a fifth aspect, the present application further provides a computer program product. The computer program product includes a computer program that, when executed by a processor, implements the following steps:

[0043] Acquiring skeleton input data, and preprocessing the skeleton input data to obtain preprocessed skeleton input data;

[0044] Detecting all joints in the pre-processed skeletal input data using a target detection algorithm to obtain key skeletal region detection results;

[0045] Perform target joint screening based on the key bone region detection results to obtain key bone joint screening features;

[0046] Based on the key skeletal joint screening features, a pre-trained deep neural network model is used to predict the bone age grade and output the bone age prediction result;

[0047] The target detection algorithm includes a lightweight multi-scale feature extraction module, which is configured to capture multi-scale features through parallel convolution kernels of different sizes and enhance key feature channels in combination with a channel attention mechanism.

[0048] The embodiments of the present application have the following beneficial effects:

[0049] The bone age prediction method, device, equipment, storage medium and computer program product provided in the embodiments of the present application can perform bone age prediction in multiple stages through the corresponding implementation methods of the data preprocessing unit, the target detection unit, the target joint screening unit and the grade prediction unit, reducing manual intervention and making the bone age assessment process more automated and standardized. The use of multiple stages for bone age prediction can significantly improve the assessment efficiency while ensuring high accuracy, and has high clinical application value. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] Figure 1 Schematic diagram of a flow chart of a bone age prediction method in one embodiment;

[0051] Figure 2 is a structural block diagram of an object detection algorithm in one embodiment;

[0052] Figure 3 is a structural block diagram of a backbone network layer in a target detection algorithm in another embodiment;

[0053] Figure 4 A schematic flow chart of a bone age prediction method according to another embodiment;

[0054] Figure 5 2 is a structural block diagram of a bone age prediction device in one embodiment. DETAILED DESCRIPTION

[0055] In order to make the purpose, technical solutions and advantages of this application more clear, the following further describes this application in detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain this application and are not intended to limit this application.

[0056] The bone age prediction method provided in the embodiment of the present application can be applied to a terminal or a server. The data storage system can store the data that the server needs to process. The data storage system can be integrated on the server, or it can be placed on the cloud or other network servers. Among them, the terminal can be, but is not limited to, various personal computers, laptops, smart phones, tablets, Internet of Things devices and portable wearable devices. The Internet of Things devices can be smart speakers, smart TVs, smart air conditioners, smart car-mounted devices, etc. Portable wearable devices can be smart watches, smart bracelets, head-mounted devices, etc. The server can be implemented as an independent server or a server cluster consisting of multiple servers.

[0057] Example 1

[0058] In one embodiment, a bone age prediction method is provided, referring to Figure 1 , methods include:

[0059] S1. Obtain skeleton input data, and preprocess the skeleton input data to obtain preprocessed skeleton input data;

[0060] S2. Detect all joints in the preprocessed skeletal input data using a target detection algorithm to obtain key skeletal region detection results;

[0061] S3. Screen target joints based on key bone region detection results to obtain key bone joint screening features;

[0062] S4. Based on the key bone joint screening features, the pre-trained deep neural network model is used to predict the bone age grade and output the bone age prediction results;

[0063] The target detection algorithm includes a lightweight multi-scale feature extraction module, which is configured to capture multi-scale features through parallel convolution kernels of different sizes and enhance key feature channels in combination with a channel attention mechanism.

[0064] Specifically, this embodiment performs bone age prediction in multiple stages, including a data preprocessing unit, a target detection unit, a target joint screening unit, and a grade prediction unit. First, the data preprocessing unit acquires skeletal input data, which can include medical imaging data (e.g., X-rays) and relevant user data (e.g., user age, gender, and other basic user data). The data preprocessing unit preprocesses the acquired skeletal input data to ensure its quality and consistency. The data preprocessing unit effectively removes noise from the images, improving image quality and providing reliable input for subsequent feature extraction and bone age prediction. The target detection unit detects all joints in the preprocessed skeletal input data, obtaining key skeletal region detection results and identifying key skeletal areas, such as the wrist bones, bone growth plates, and joint regions. The target detection unit employs convolutional neural networks (CNNs) for feature extraction, leveraging deep learning techniques to detect and locate the input images. This ensures that the system accurately locates the skeletal regions required for evaluation, reducing manual intervention and improving automation. The target joint screening unit then further analyzes the key skeletal region detection results extracted by the target detection unit, performing joint screening to obtain key skeletal joint screening features. By combining the target region coordinate information, the model can adaptively select the most informative areas. By carefully screening the joint development in the key skeletal region detection results, the joint screening unit can more accurately identify key information related to bone age and obtain key skeletal joint screening features. The grade prediction unit then applies the key skeletal joint screening features to a pre-trained deep neural network model to predict bone age grade, generating a specific bone age grade score based on the skeletal features identified by the joint screening. The grade prediction unit can learn the complex relationship between skeletal development and age from training data, enabling efficient prediction of bone age at different stages. Finally, the prediction results are weighted and aggregated to calculate the final bone age score. The score is based on various skeletal developmental features, such as growth lines, bone morphology, and joint development, ultimately yielding a comprehensive assessment result. By adopting such a technical solution, bone age prediction can be performed in multiple stages through data preprocessing unit, target detection unit, target joint screening unit and grade prediction unit, which reduces manual intervention and makes the bone age assessment process more automated and standardized. The use of multiple stages for bone age prediction can significantly improve the assessment efficiency while ensuring high accuracy, and has high clinical application value.

[0065] In one embodiment, S1 includes:

[0066] S11, performing data enhancement processing on the skeleton input data to obtain preprocessed skeleton input data, where the data enhancement processing includes at least one of image rotation, image translation, image scaling, image flipping, adding Gaussian noise, and histogram equalization;

[0067] S12. Divide the preprocessed skeletal input data into a training set, a validation set, and a test set according to a preset ratio.

[0068] Specifically, the skeletal input data can include medical imaging data (e.g., X-rays) and relevant user data (e.g., user age, gender, and other basic user data). Data augmentation can be performed on the skeletal input data to convert the raw image data into the standard format required by the system and perform denoising to remove interfering noise. Data augmentation techniques are used to improve the robustness of the model, resulting in preprocessed skeletal input data. For example, the data preprocessing unit can apply data augmentation techniques such as image rotation, image translation, image scaling, image flipping, and the addition of Gaussian noise to public datasets to expand the diversity of the training dataset and reduce the model's reliance on data bias, thereby improving the model's prediction accuracy and robustness across different patients, imaging conditions, and image quality. Furthermore, the data preprocessing unit can utilize histogram adaptive equalization and histogram matching techniques to enhance image quality. The data preprocessing unit can also perform normalization on the input skeletal images to ensure the quality of subsequent analysis. Image enhancement enhances image detail contrast, while normalization uniformly scales the image data for ease of processing by deep learning models. After obtaining the preprocessed skeletal input data, it can be divided into a training set, a validation set, and a test set according to a preset ratio, and the model training, model validation, and model testing are performed respectively to improve the accuracy of the model. The preset ratio can be 6:2:2, 7:1:2, or 98:1:1, etc., which can be determined according to different data levels or different training division scenarios. It should be noted that the above preset ratios are only exemplary and not practical limitations.

[0069] For example, assume that the original image is , the transformed image is , these transformations can be expressed as:

[0070] Rotation:

[0071] Pan:

[0072] Zoom:

[0073] Flip:

[0074] Gaussian noise:

[0075] Histogram equalization:

[0076]

[0077] in, Indicates the angle of rotation, represents the translation value, represents the scaling factor, Indicates the image width, The mean is 0 and the variance is Gaussian noise, is the grayscale of the image The probability of occurrence, is the grayscale after equalization.

[0078] By adopting such a technical solution, the data preprocessing unit can perform data enhancement processing on the skeleton input data to obtain preprocessed skeleton input data. Through these preprocessing steps, the system can better cope with images under different shooting conditions and improve the accuracy of the evaluation.

[0079] In one embodiment, referring to Figure 2 The target detection algorithm includes an input layer, a backbone network layer, and a head network layer. The input layer is used to obtain the preprocessed bone input data. The backbone network layer is used to extract the key bone image features from the preprocessed bone input data. The head network layer is used to fuse the feature information of the key bone image features, predict the category and position of the target bone, and obtain the key bone area detection results.

[0080] Specifically, the object detection algorithm includes an input layer, a backbone network layer, and a head network layer. For example, the object detection algorithm can use the YOLOv7 (You Only Look Once version 7) object detection model to detect all phalangeal joints. The object detection unit inputs preprocessed skeletal input data into the YOLOv7 object detection model input layer. The detection categories, from top to bottom (radius to distal phalanges), can include the following seven: radius, ulna, metacarpal I (MCPFirst), metacarpal (MCP), proximal phalanx (ProximalPhalanx), middle phalanx (MiddlePhalanx), and distal phalanx (DistalPhalanx). The backbone network layer is responsible for extracting image features, employing a multi-branch stacking module for feature extraction. The head network layer is used to predict the object category and location. This layer utilizes a multi-scale feature fusion architecture to effectively handle objects of varying scales and integrate feature information from different layers throughout the network, thereby improving model performance and robustness. This section predicts the various features output by the feature fusion network. The multi-scale feature fusion module uses a pyramid pooling structure to increase perception and uses four branches with different expansion rates to process features separately to accommodate inputs of different sizes.

[0081] For example, when an object detection algorithm uses an algorithm structure similar to YOLOv7, YOLOv7 introduces E-ELAN (Extended Efficient Layer Aggregation Network), which enhances the model's ability to extract multi-scale features by optimizing the feature fusion path. Compared to YOLOv5, E-ELAN improves computational efficiency through grouped convolution and dynamic channel allocation. Furthermore, a multi-branch structure is used during training to enhance feature learning, while merging the branches into a single branch during inference to reduce computational overhead. For example, RepConv (Reparameterized Convolution) combines multiple convolution operations during training and simplifies them to a single convolution during inference, achieving a balance between accuracy and speed. Furthermore, a dynamic label assignment strategy can be adopted, replacing the static assignment method used in YOLOv5. By dynamically adjusting the matching strategy between positive and negative samples, the model's ability to learn difficult samples is enhanced. By adopting a target detection algorithm with an input layer, a backbone network layer, and a head network layer, it can simultaneously complete the target positioning and classification tasks in a single forward propagation process, making it more suitable for real-time detection; through multi-scale feature extraction technology, it can effectively process targets of different sizes, whether small joints or large hand areas; background interference is suppressed through global context information, improving the accuracy and robustness of detection; due to its end-to-end training method, the model can not only simplify the training process and deployment, but also reduce the complexity of data preprocessing and post-processing.

[0082] In some examples, the object detection algorithm may also include a lightweight multi-scale feature extraction module. This module can be configured to capture multi-scale features using parallel convolution kernels of different sizes and enhance key feature channels using a channel-attention mechanism. Features can be extracted in parallel using 1×1, 3×3, and 5×5 convolution kernels. 1×1 convolutions are used to capture inter-channel relationships, achieving feature compression and dimensionality adjustment; 3×3 convolutions balance local feature extraction with computational efficiency, capturing mid-scale information; and 5×5 convolutions capture wider contextual information, enhancing global feature perception. Furthermore, a channel-attention mechanism is introduced during the feature fusion stage. Global average pooling is used to compress spatial information into channel descriptors. Two layers of convolution learn inter-channel dependencies, generating dynamic weights for features at different scales and highlighting important feature channels. In traditional object detection models, multi-scale feature extraction in the backbone network primarily relies on operations such as feature pyramids and dilated convolutions. However, the multi-branch structure significantly increases the number of parameters and computational complexity, and lacks the ability to select important features for features at different scales. By adopting a model structure such as a lightweight multi-scale feature extraction module, compared to the ELAN module in the above example, the innovative combination of parallel convolution structure and channel attention mechanism can significantly improve the multi-scale feature extraction capability while maintaining lightweight and reducing computational complexity.

[0083] In one embodiment, referring to Figure 3 , the backbone network layer includes a first feature extraction module, a second feature extraction module and a third feature extraction module;

[0084] The first feature extraction module includes a convolution layer, a batch normalization layer, and an activation function layer;

[0085] The second feature extraction module includes a first sampling module and a second sampling module;

[0086] The third feature extraction module includes an efficient layer aggregation network module.

[0087] Specifically, the first feature extraction module includes a convolution layer, a batch normalization layer, and an activation function layer, namely a CBS module. The CBS module mainly includes Conv (convolution layer), BN (Batch Normalization, batch normalization layer) and SiLU (activation function layer). For example, the first feature extraction module first uses Convolution is used to change the channel, and then the step size is 1. Convolution is used for feature extraction, and finally a step size of 2 is used Convolution is used for downsampling. The second feature extraction module includes a first sampling module and a second sampling module. The second feature extraction module can be a maximum pooling module (MP module), which includes two branches. The first sampling module first performs downsampling through a maximum pooling operation, and then performs a Convolution changes the number of channels; the second sampling module first passes Convolution changes the number of channels, and then passes through a step size of 2 Convolution is used for downsampling; finally, the output results of the two branches are combined through channel splicing to achieve the effect of super downsampling, forming the feature map output by the second feature extraction module. The third feature extraction module includes an efficient layer aggregation network module. The efficient layer aggregation network module is an efficient network architecture that enables the network to learn more features and enhance the robustness of the network by controlling the shortest and longest gradient paths. It consists of two branches. The first branch passes through a Convolution makes channel changes; the second branch first passes through a Convolution changes the channel and then passes through four Convolution is used for feature extraction; finally, the four extracted features are superimposed together to obtain the feature extraction result output by the third feature extraction module. By adopting this backbone network layer structure, multi-branch stacking modules can be used for feature extraction. Through the first feature extraction module, the second feature extraction module, and the third feature extraction module, more emphasis is placed on dynamic feature aggregation. Compared with other target detection network structures, it can achieve multi-branch training and single-branch reasoning, balancing precision and speed, and improving detection accuracy and robustness. This model not only simplifies the training process and deployment, but also reduces the complexity of data pre-processing and post-processing.

[0088] In one embodiment, S3 includes:

[0089] S31. Screen target joints according to the key bone region detection results and preset bone age standards to obtain key bone joint screening features. The key bone joint screening features include key bone regions, and the key bone regions include at least two of the following target joints:

[0090] Radius, ulna, metacarpal I, metacarpal III, metacarpal V, proximal phalanx I, proximal phalanx III, proximal phalanx V, middle phalanx III, middle phalanx V, distal phalanx I, distal phalanx III and distal phalanx V.

[0091] Specifically, in the target joint screening unit, the system performs joint screening on the bone areas in the key bone area detection results based on the coordinate frame information output by the target detection unit, and can focus on the bone parts that are closely related to bone age assessment, that is, the above-mentioned key bone joint screening features, with growth lines, joint development status and other information. The target joint screening unit directly screens based on the detected bone areas to ensure that each joint and bone part can be effectively analyzed. By adopting such a technical solution, the key areas required for bone age assessment can be accurately extracted through the target joint screening unit, avoiding the tediousness of traditional manual labeling, and improving the evaluation efficiency while ensuring the accuracy of the evaluation.

[0092] In one embodiment, S4 includes:

[0093] S41. Using a pre-trained deep neural network model, the key skeletal joint screening features are divided into multiple groups along the channel dimension and cross-group feature aggregation calculation is performed to predict the level of each target joint, thereby obtaining a prediction result for each target joint.

[0094] S42. Calculate and output the bone age prediction result based on the prediction result of each target joint according to the preset bone age standard.

[0095] Specifically, the key skeletal joint screening features can be divided into multiple groups along the channel dimension through a pre-trained deep neural network model and cross-group feature aggregation calculation can be performed to predict the level of each target joint. For example, Res2Net (Residual-like connections within a single residual block) can be used as the backbone network, and an alternative architecture with stronger multi-scale feature extraction capabilities can be adopted while maintaining a similar ResNet (ResidualNetwork) computational load, that is, a smaller filter group can be used to replace a group of filters, while connecting different filter groups in a hierarchical residual style. Specifically, after After convolution, the feature map is evenly divided into feature map subsets, denoted as ,in Each feature subset It has the same spatial size as the input feature map, but has the same number of channels. .Apart from In addition, each There is a corresponding Convolution, denoted as . use express The output of and The outputs are summed and then fed into In order to increase While reducing parameters, the network omits The 3×3 convolution, therefore, the output can be expressed as:

[0096]

[0097] This process predicts the grade of each target joint, facilitating subsequent score calculation. Based on the grade prediction unit's results, the corresponding grade scores are then added together, using pre-set bone age standards (e.g., the "Chinese Wrist Bone Development Standard - Zhonghua 05"). The resulting total score is then compared to the standard to produce the final bone age result. The grade prediction unit utilizes multi-layer perceptron technology, leveraging input skeletal features to optimize network weights layer by layer. Non-linear activation functions, such as linear rectification functions, map the complex relationships between bone age and imaging features. Multi-layer perceptrons effectively capture these complex non-linear relationships, thereby improving the accuracy of bone age prediction. Finally, the evaluation results from different stages are weighted and aggregated, and the final prediction is compared with the standard, such as the Zhonghua 05 standard, to ensure scientific and accurate results. This technical solution allows the key skeletal joint screening features extracted from the target joint screening unit to be grouped and aggregated across these groups. The final bone age prediction is then combined with the pre-set bone age standard to ensure scientific and accurate results.

[0098] For example, refer to Figure 4 The target detection algorithm can use the YOLOv7 target detection model to detect all joints of the phalanges. Then, based on the coordinate frame information output by the target detection unit, the bone areas in the key bone area detection results are screened for joints, which can focus on the bone parts closely related to bone age assessment, namely the key bone joint screening features. Then, through the pre-trained deep neural network model, the key bone joint screening features are divided into multiple groups along the channel dimension and cross-group feature aggregation calculations are performed to predict the grade of each target joint. Then, based on the results of the grade prediction unit, the scores of the corresponding grades are found and added according to the preset bone age standards. After obtaining the total score, the final bone age result is obtained by comparing it with the standards.

[0099] In this embodiment, bone age prediction can be performed in multiple stages through the corresponding implementation methods of the data preprocessing unit, the target detection unit, the target joint screening unit, and the grade prediction unit, which reduces manual intervention and makes the bone age assessment process more automated and standardized. The multi-stage bone age prediction can significantly improve the assessment efficiency while ensuring high accuracy, and has high clinical application value.

[0100] It should be understood that, although the various steps in the flowcharts involved in the various embodiments described above are displayed in sequence according to the instructions of the arrows, these steps are not necessarily executed in sequence in the order indicated by the arrows. Unless otherwise specified herein, there is no strict order restriction on the execution of these steps, and these steps can be executed in other orders. Moreover, at least a portion of the steps in the flowcharts involved in the various embodiments described above can include multiple steps or multiple stages, and these steps or stages are not necessarily executed and completed at the same time, but can be executed at different times, and the execution order of these steps or stages is not necessarily to be carried out in sequence, but can be executed in turn or alternately with other steps or at least a portion of steps or stages in other steps.

[0101] Example 2

[0102] Based on the same inventive concept, the present application also provides a bone age prediction device for implementing the bone age prediction method described above. The solution provided by the device is similar to the solution described in the method described above. Therefore, the specific limitations of one or more bone age prediction device embodiments provided below can be found in the above-mentioned limitations of the bone age prediction method and will not be repeated here.

[0103] In one embodiment, a bone age prediction device is provided, referring to Figure 5 ,include:

[0104] A preprocessing module, configured to obtain skeleton input data and preprocess the skeleton input data to obtain preprocessed skeleton input data;

[0105] A target detection module is used to detect all joints in the pre-processed skeleton input data using a target detection algorithm to obtain key skeleton area detection results;

[0106] A joint screening module is used to screen target joints based on the key bone region detection results to obtain key bone joint screening features;

[0107] A bone age prediction module is used to predict bone age grade based on the key skeletal joint screening features using a pre-trained deep neural network model and output a bone age prediction result;

[0108] The target detection algorithm includes a lightweight multi-scale feature extraction module, which is configured to capture multi-scale features through parallel convolution kernels of different sizes and enhance key feature channels in combination with a channel attention mechanism.

[0109] Furthermore, the preprocessing module is also used to perform data enhancement processing on the skeleton input data to obtain preprocessed skeleton input data, wherein the data enhancement processing includes at least one of image rotation, image translation, image scaling, image flipping, adding Gaussian noise and histogram equalization; and is used to divide the preprocessed skeleton input data into a training set, a verification set and a test set according to a preset ratio.

[0110] Furthermore, the target detection algorithm includes an input layer, a backbone network layer and a head network layer. The input layer is used to obtain preprocessed bone input data, the backbone network layer is used to extract bone key image features for the preprocessed bone input data, and the head network layer is used to fuse the feature information of the bone key image features, predict the category and position of the target bone, and obtain the key bone area detection result.

[0111] Furthermore, the backbone network layer includes a first feature extraction module, a second feature extraction module and a third feature extraction module;

[0112] The first feature extraction module includes a convolution layer, a batch normalization layer and an activation function layer;

[0113] The second feature extraction module includes a first sampling module and a second sampling module;

[0114] The third feature extraction module includes an efficient layer aggregation network module.

[0115] Furthermore, the joint screening module is also used to screen target joints according to the preset bone age standards based on the detection results of the key bone area to obtain key bone joint screening features, and the key bone joint screening features include bone key areas, and the bone key areas include at least two of the following target joints: radius, ulna, metacarpal I, metacarpal III, metacarpal V, proximal phalanx I, proximal phalanx III, proximal phalanx V, middle phalanx III, middle phalanx V, distal phalanx I, distal phalanx III and distal phalanx V.

[0116] Furthermore, the bone age prediction module is also used to divide the key skeletal joint screening features into multiple groups along the channel dimension through a pre-trained deep neural network model and perform cross-group feature aggregation calculations, predict the level of each target joint, and obtain the prediction results of each target joint; and is used to calculate and output the bone age prediction results according to the preset bone age standards based on the prediction results of each target joint.

[0117] Each module in the bone age prediction device can be implemented in whole or in part through software, hardware, or a combination thereof. Each module can be embedded in or independent of a processor in a computer device in hardware form, or can be stored in a computer device memory in software form, so that the processor can call and execute the corresponding operations of each module.

[0118] Example 3

[0119] In one embodiment, a computer device is provided, including a memory and a processor, wherein a computer program is stored in the memory, and when the processor executes the computer program, the following steps are implemented:

[0120] Acquiring skeleton input data, and preprocessing the skeleton input data to obtain preprocessed skeleton input data;

[0121] Detecting all joints in the pre-processed skeletal input data using a target detection algorithm to obtain key skeletal region detection results;

[0122] Perform target joint screening based on the key bone region detection results to obtain key bone joint screening features;

[0123] Based on the key skeletal joint screening features, a pre-trained deep neural network model is used to predict the bone age grade and output the bone age prediction result;

[0124] The target detection algorithm includes a lightweight multi-scale feature extraction module, which is configured to capture multi-scale features through parallel convolution kernels of different sizes and enhance key feature channels in combination with a channel attention mechanism.

[0125] In one embodiment, when the processor executes the computer program, the processor further implements the following steps:

[0126] Performing data enhancement processing on the skeleton input data to obtain preprocessed skeleton input data, wherein the data enhancement processing includes at least one of image rotation, image translation, image scaling, image flipping, adding Gaussian noise, and histogram equalization;

[0127] The preprocessed skeleton input data is divided into a training set, a validation set and a test set according to a preset ratio.

[0128] In one embodiment, the target detection algorithm includes an input layer, a backbone network layer, and a head network layer. The input layer is used to obtain preprocessed bone input data, the backbone network layer is used to extract bone key image features from the preprocessed bone input data, and the head network layer is used to fuse the feature information of the bone key image features, predict the category and position of the target bone, and obtain the key bone area detection results.

[0129] In one embodiment, the backbone network layer includes a first feature extraction module, a second feature extraction module, and a third feature extraction module;

[0130] The first feature extraction module includes a convolution layer, a batch normalization layer and an activation function layer;

[0131] The second feature extraction module includes a first sampling module and a second sampling module;

[0132] The third feature extraction module includes an efficient layer aggregation network module.

[0133] In one embodiment, when the processor executes the computer program, the processor further implements the following steps:

[0134] Target joints are screened according to the key bone region detection results and preset bone age standards to obtain key bone joint screening features, wherein the key bone joint screening features include a key bone region, and the key bone region includes at least two of the following target joints:

[0135] Radius, ulna, metacarpal I, metacarpal III, metacarpal V, proximal phalanx I, proximal phalanx III, proximal phalanx V, middle phalanx III, middle phalanx V, distal phalanx I, distal phalanx III and distal phalanx V.

[0136] In one embodiment, the key skeletal joint screening features are divided into multiple groups along the channel dimension through a pre-trained deep neural network model and cross-group feature aggregation calculation is performed to predict the level of each target joint to obtain a prediction result for each target joint;

[0137] The bone age prediction result is calculated and output according to the prediction result of each target joint and the preset bone age standard.

[0138] Example 4

[0139] In one embodiment, a computer-readable storage medium is provided, on which a computer program is stored. When the computer program is executed by a processor, the following steps are implemented:

[0140] Acquiring skeleton input data, and preprocessing the skeleton input data to obtain preprocessed skeleton input data;

[0141] Detecting all joints in the pre-processed skeletal input data using a target detection algorithm to obtain key skeletal region detection results;

[0142] Perform target joint screening based on the key bone region detection results to obtain key bone joint screening features;

[0143] Based on the key skeletal joint screening features, a pre-trained deep neural network model is used to predict the bone age grade and output the bone age prediction result;

[0144] The target detection algorithm includes a lightweight multi-scale feature extraction module, which is configured to capture multi-scale features through parallel convolution kernels of different sizes and enhance key feature channels in combination with a channel attention mechanism.

[0145] In one embodiment, when the computer program is executed by a processor, the following steps are further implemented:

[0146] Performing data enhancement processing on the skeleton input data to obtain preprocessed skeleton input data, wherein the data enhancement processing includes at least one of image rotation, image translation, image scaling, image flipping, adding Gaussian noise, and histogram equalization;

[0147] The preprocessed skeleton input data is divided into a training set, a validation set and a test set according to a preset ratio.

[0148] In one embodiment, the target detection algorithm includes an input layer, a backbone network layer, and a head network layer. The input layer is used to obtain preprocessed bone input data, the backbone network layer is used to extract bone key image features from the preprocessed bone input data, and the head network layer is used to fuse the feature information of the bone key image features, predict the category and position of the target bone, and obtain the key bone area detection results.

[0149] In one embodiment, the backbone network layer includes a first feature extraction module, a second feature extraction module, and a third feature extraction module;

[0150] The first feature extraction module includes a convolution layer, a batch normalization layer and an activation function layer;

[0151] The second feature extraction module includes a first sampling module and a second sampling module;

[0152] The third feature extraction module includes an efficient layer aggregation network module.

[0153] In one embodiment, when the computer program is executed by a processor, the following steps are further implemented:

[0154] Target joints are screened according to the key bone region detection results and preset bone age standards to obtain key bone joint screening features, wherein the key bone joint screening features include a key bone region, and the key bone region includes at least two of the following target joints:

[0155] Radius, ulna, metacarpal I, metacarpal III, metacarpal V, proximal phalanx I, proximal phalanx III, proximal phalanx V, middle phalanx III, middle phalanx V, distal phalanx I, distal phalanx III and distal phalanx V.

[0156] In one embodiment, when the computer program is executed by a processor, the following steps are further implemented:

[0157] The key skeletal joint screening features are divided into multiple groups along the channel dimension through a pre-trained deep neural network model and cross-group feature aggregation calculation is performed to predict the level of each target joint and obtain the prediction result of each target joint;

[0158] The bone age prediction result is calculated and output according to the prediction result of each target joint and the preset bone age standard.

[0159] Example 5

[0160] In one embodiment, a computer program product is provided, comprising a computer program, which, when executed by a processor, implements the following steps:

[0161] Acquiring skeleton input data, and preprocessing the skeleton input data to obtain preprocessed skeleton input data;

[0162] Detecting all joints in the pre-processed skeletal input data using a target detection algorithm to obtain key skeletal region detection results;

[0163] Perform target joint screening based on the key bone region detection results to obtain key bone joint screening features;

[0164] Based on the key skeletal joint screening features, a pre-trained deep neural network model is used to predict the bone age grade and output the bone age prediction result;

[0165] The target detection algorithm includes a lightweight multi-scale feature extraction module, which is configured to capture multi-scale features through parallel convolution kernels of different sizes and enhance key feature channels in combination with a channel attention mechanism.

[0166] In one embodiment, when the computer program is executed by a processor, the following steps are further implemented:

[0167] Performing data enhancement processing on the skeleton input data to obtain preprocessed skeleton input data, wherein the data enhancement processing includes at least one of image rotation, image translation, image scaling, image flipping, adding Gaussian noise, and histogram equalization;

[0168] The preprocessed skeleton input data is divided into a training set, a validation set and a test set according to a preset ratio.

[0169] In one embodiment, the target detection algorithm includes an input layer, a backbone network layer, and a head network layer. The input layer is used to obtain preprocessed bone input data, the backbone network layer is used to extract bone key image features from the preprocessed bone input data, and the head network layer is used to fuse the feature information of the bone key image features, predict the category and position of the target bone, and obtain the key bone area detection results.

[0170] In one embodiment, the backbone network layer includes a first feature extraction module, a second feature extraction module, and a third feature extraction module;

[0171] The first feature extraction module includes a convolution layer, a batch normalization layer and an activation function layer;

[0172] The second feature extraction module includes a first sampling module and a second sampling module;

[0173] The third feature extraction module includes an efficient layer aggregation network module.

[0174] In one embodiment, when the computer program is executed by a processor, the following steps are further implemented:

[0175] Target joints are screened according to the key bone region detection results and preset bone age standards to obtain key bone joint screening features, wherein the key bone joint screening features include a key bone region, and the key bone region includes at least two of the following target joints:

[0176] Radius, ulna, metacarpal I, metacarpal III, metacarpal V, proximal phalanx I, proximal phalanx III, proximal phalanx V, middle phalanx III, middle phalanx V, distal phalanx I, distal phalanx III and distal phalanx V.

[0177] In one embodiment, when the computer program is executed by a processor, the following steps are further implemented:

[0178] The key skeletal joint screening features are divided into multiple groups along the channel dimension through a pre-trained deep neural network model and cross-group feature aggregation calculation is performed to predict the level of each target joint and obtain the prediction result of each target joint;

[0179] The bone age prediction result is calculated and output according to the prediction result of each target joint and the preset bone age standard.

[0180] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, stored data, displayed data, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties.

[0181] Those skilled in the art will appreciate that all or part of the processes in the above-mentioned embodiments can be implemented by instructing the relevant hardware through a computer program. The computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the above-mentioned embodiments. In particular, any reference to memory, database, or other media used in the embodiments provided in this application 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 various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM). The databases involved in the various embodiments provided herein may include at least one of a relational database and a non-relational database. Non-relational databases may include, but are not limited to, distributed databases based on blockchains. The processors involved in the various embodiments provided herein may be, but are not limited to, general-purpose processors, central processing units (CPUs), graphics processing units (GPUs), digital signal processors (DSPs), programmable logic devices (PLDs), data processing logic devices based on quantum computing, and the like.

[0182] The technical features of the above embodiments can be combined arbitrarily. To make the description concise, 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.

[0183] The above-described embodiments merely represent several implementation methods of the present application. While the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the present application. It should be noted that a person of ordinary skill in the art may make various modifications and improvements without departing from the spirit of the present application, and these modifications and improvements fall within the scope of protection of the present application. Therefore, the scope of protection of the present application shall be determined by the appended claims.

Claims

1. A bone age prediction method, characterized in that: The method comprises: Acquiring skeleton input data, and preprocessing the skeleton input data to obtain preprocessed skeleton input data; Detecting all joints in the pre-processed skeletal input data using a target detection algorithm to obtain key skeletal region detection results; Perform target joint screening based on the key bone region detection results to obtain key bone joint screening features; Based on the key skeletal joint screening features, a pre-trained deep neural network model is used to predict the bone age grade and output the bone age prediction result; The target detection algorithm includes a lightweight multi-scale feature extraction module, which is configured to capture multi-scale features through parallel convolution kernels of different sizes and enhance key feature channels in combination with a channel attention mechanism.

2. The method according to claim 1, characterized in that The step of obtaining the skeleton input data and preprocessing the skeleton input data to obtain preprocessed skeleton input data includes: Performing data enhancement processing on the skeleton input data to obtain preprocessed skeleton input data, wherein the data enhancement processing includes at least one of image rotation, image translation, image scaling, image flipping, adding Gaussian noise, and histogram equalization; The preprocessed skeleton input data is divided into a training set, a validation set and a test set according to a preset ratio.

3. The method according to claim 1, characterized in that The target detection algorithm includes an input layer, a backbone network layer and a head network layer. The input layer is used to obtain preprocessed bone input data, the backbone network layer is used to extract bone key image features from the preprocessed bone input data, and the head network layer is used to fuse the feature information of the bone key image features, predict the category and position of the target bone, and obtain the key bone area detection results.

4. The method according to claim 3, characterized in that The backbone network layer includes a first feature extraction module, a second feature extraction module and a third feature extraction module; The first feature extraction module includes a convolution layer, a batch normalization layer and an activation function layer; The second feature extraction module includes a first sampling module and a second sampling module; The third feature extraction module includes an efficient layer aggregation network module.

5. The method according to claim 1, wherein The target joint screening is performed based on the key bone region detection result to obtain key bone joint screening features, including: Target joints are screened according to the key bone region detection results and preset bone age standards to obtain key bone joint screening features, wherein the key bone joint screening features include a key bone region, and the key bone region includes at least two of the following target joints: Radius, ulna, metacarpal I, metacarpal III, metacarpal V, proximal phalanx I, proximal phalanx III, proximal phalanx V, middle phalanx III, middle phalanx V, distal phalanx I, distal phalanx III and distal phalanx V.

6. The method according to claim 5, characterized in that The method of predicting bone age grade based on the key bone joint screening features using a pre-trained deep neural network model and outputting a bone age prediction result includes: The key skeletal joint screening features are divided into multiple groups along the channel dimension through a pre-trained deep neural network model and cross-group feature aggregation calculation is performed to predict the level of each target joint and obtain the prediction result of each target joint; The bone age prediction result is calculated and output according to the prediction result of each target joint and the preset bone age standard.

7. A bone age prediction device, characterized in that: The device comprises: A preprocessing module, configured to obtain skeleton input data and preprocess the skeleton input data to obtain preprocessed skeleton input data; A target detection module is used to detect all joints in the pre-processed skeleton input data using a target detection algorithm to obtain key skeleton area detection results; A joint screening module is used to screen target joints based on the key bone region detection results to obtain key bone joint screening features; A bone age prediction module is used to predict bone age grade based on the key skeletal joint screening features using a pre-trained deep neural network model and output a bone age prediction result; The target detection algorithm includes a lightweight multi-scale feature extraction module, which is configured to capture multi-scale features through parallel convolution kernels of different sizes and enhance key feature channels in combination with a channel attention mechanism.

8. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the method according to any one of claims 1 to 6 are implemented.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 6 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 according to any one of claims 1 to 6 are implemented.

Images