Bone age detection model construction method, prediction method and device

Abstract

The invention discloses a method for constructing a bone age detection model, which comprises the following steps of: inputting an X-ray film image in a training set into a bone age detection model to obtain a corresponding predicted bone age, and training the bone age detection model by minimizing the difference loss between the predicted bone age and a corresponding bone age label, wherein the bone age detection model comprises a global model, local models in one-to-one correspondence with k bone age segments and a fusion module, and when an X-ray film image is input into the bone age detection model, an X-ray film obtains first probability distribution and second probability distribution of the bone age through the global model and the corresponding local models, the first probability distribution and the second probability distribution are input into a fusion module, third probability distribution is obtained, the corresponding predicted bone age is obtained according to the third probability distribution, the X-ray image in the training set is an image containing bone age information, and the corresponding label is the bone age. According to the method, the bone age prediction precision is improved by considering the global features of the bone age and the specific features of the bone age.

Description

Technical Field

[0001] The present invention belongs to the intersection of medical image analysis and artificial intelligence, and specifically relates to a method for constructing a bone age detection model, a prediction method, and a device. Background Art

[0002] Traditional bone age assessment methods rely primarily on manual analysis of X-rays by doctors, a process that is not only time-consuming and labor-intensive but also highly subjective. Doctors must score each joint based on epiphyseal development and then calculate the corresponding age from a table. This method requires a high level of experience and expertise and is susceptible to human error and bias, leading to inconsistent assessment results.

[0003] With the development of artificial intelligence (AI), machine learning methods have gradually been introduced into the field of bone age assessment. For example, traditional machine learning algorithms such as support vector machines (SVM) and multilayer perceptrons (MLP) are used for feature extraction and classification. However, these methods have limitations in feature extraction and struggle to capture subtle changes in bone development, resulting in insufficient accuracy.

[0004] The emergence of deep learning models has brought new hope to bone age assessment. For example, models such as Faster R-CNN automatically extract features through convolutional neural networks, improving the automation level of bone age assessment.

[0005] However, these models still face the following challenges in practical applications:

[0006] Imbalance between global and local features: Skeletal development is a complex process that requires consideration of both the overall structure of the hand bones and the detailed features of key local regions (such as the epiphysis and metaphysis). A single deep learning model often struggles to extract both global and local features, leading to inaccurate evaluation results.

[0007] Second, existing research also lacks dataset diversity and model generalization capabilities. For example, some studies use only a single dataset for training and validation, lacking comprehensive consideration of bone age assessment for children of different genders and geographic regions. This can make the models unsuitable for diverse clinical needs in practical applications.

[0008] In summary, although deep learning models have made some progress in bone age assessment, many challenges remain. These issues not only limit the accuracy and reliability of the models but also hinder their widespread application in clinical practice. Therefore, developing a bone age assessment model that can balance global and local features and possess strong generalization capabilities has important clinical significance and application value. Summary of the Invention

[0009] In response to the defects and improvement needs of the existing technology, the present invention provides a method for constructing a bone age detection model, comprising: inputting X-ray images in a training set into the bone age detection model to obtain the corresponding predicted bone age, and training the bone age detection model by minimizing the difference loss between the predicted bone age and the corresponding bone age label, wherein the bone age detection model includes a global model and k local models corresponding to bone age segments and a fusion module. When the X-ray image is input into the bone age detection model, the X-ray film is respectively passed through the global model and the corresponding local model to obtain a first probability distribution and a second probability distribution of the bone age, and the first probability distribution and the second probability distribution are input into the fusion module to obtain a third probability distribution and obtain the corresponding predicted bone age according to the third probability distribution, wherein the X-ray image in the training set is an image containing bone age information, and the corresponding label is bone age.

[0010] The present invention also discloses a method for predicting bone age based on a bone age detection model, comprising the following steps:

[0011] Input the X-ray image containing bone age information into the trained bone age detection model;

[0012] The global model outputs the first probability distribution of bone age based on X-ray images;

[0013] The local model inputs the X-ray into the corresponding local model according to the bone age segment corresponding to the first probability distribution, and the corresponding local model outputs a second probability distribution of bone age;

[0014] Inputting the first probability distribution and the second probability distribution into a fusion model, and outputting a third probability distribution of bone age;

[0015] The third probability distribution is normalized and the predicted bone age is output.

[0016] The present invention discloses a device for predicting bone age, comprising: an acquisition module for acquiring a bone X-ray of a target person;

[0017] a first extraction module, configured to extract features from the bone X-ray film using a global model to obtain general features of the bone X-ray film, and obtain a first probability distribution of bone age based on the general features;

[0018] a second extraction module, configured to perform feature extraction on the bone X-ray film using a local model to obtain specific features of the bone X-ray film, and obtain a second probability distribution of bone age based on the specific features;

[0019] A fusion module, configured to obtain a third probability distribution based on the first probability distribution and the second probability distribution;

[0020] The evaluation module is used to obtain the predicted bone age based on the third probability distribution.

[0021] In general, the above technical solutions conceived by the present invention can achieve the following beneficial effects:

[0022] By training global and local models, and through the division of labor and cooperation between the global and local models, both the global and specific characteristics of bone age are taken into account; secondly, the fusion module dynamically integrates and learns the weighted fusion module, integrates the prediction results of multiple models, and reduces the deviation of a single model. The bone age detection model disclosed by the present invention predicts bone age, thereby improving the accuracy of bone age prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] Figure 1 Shown is a flow chart of a method for constructing a bone age detection model according to an embodiment of the present invention;

[0024] Figure 2 Shown is a flowchart of a method for predicting bone age based on a bone age detection model according to an embodiment of the present invention.

[0025] Figure 3 Shown is a comparison diagram of different models based on the RSNA training set according to an embodiment of the present invention. DETAILED DESCRIPTION

[0026] In order to make the objectives, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely for the purpose of explaining the present invention and are not intended to limit the present invention. In addition, the technical features involved in the various embodiments of the present invention described below may be combined with each other as long as they do not conflict with each other.

[0027] In the present invention, the terms "first", "second", etc. (if any) in the present invention and the drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence.

[0028] Example 1:

[0029] Figure 1 FIG. 1 is a flow chart of a method for constructing a bone age detection model according to an embodiment of the present invention. Figure 1As shown, the bone age detection model includes a global model A and k local models, where the number of local models corresponds to the number of bone age segments. For example, in the embodiment of the present invention, the number of bone age segments is divided into 3, and the corresponding local models are B1-B3. Those skilled in the art should understand that the number of segments here is not limited to 3, and can be a number greater than or equal to 2. In order to better explain the present invention, the present invention is described with three bone age segments as an example. For example, the first bone age segment is 3-12 years old, which is childhood, and the bones grow rapidly in this stage; 12-17 years old is adolescence, and the bones mature rapidly; 17-20 years old is the early adulthood, and the bones tend to be stable. Different bone age segments correspond to different bone characteristics, and the local models focus on learning different bone characteristics.

[0030] Specifically, the method for constructing the bone age detection model is as follows:

[0031] like Figure 1 As shown, three bone age segments are used as an example, that is, k is equal to 3, and the local model is 3, namely B1-B3. First, the X-ray images in the training set are input into the bone age detection model to obtain the corresponding predicted bone age, and the bone age detection model is trained by minimizing the difference loss between the predicted bone age and the corresponding bone age label. The bone age detection model includes a global model A and local models B1-B3 corresponding to the three bone age segments and a fusion module. When the X-ray image is input into the bone age detection model, the X-ray film passes through the global model A and the corresponding local model to obtain the first probability distribution A and the second probability distribution B of the bone age. The first probability distribution A and the second probability distribution B are input into the fusion module to obtain the third probability distribution C and obtain the corresponding predicted bone age according to the third probability distribution C. The X-ray images in the training set are images containing bone age information, and the corresponding labels are bone age. It should be noted that the X-ray images here can be images of the hand or leg bones.

[0032] In this embodiment of the present invention, the training set used is the RSNA-2017 (Radiological Society of North America) Pediatric Bone Age Challenge dataset, which contains approximately 12,000 hand X-ray images, each with a precise bone age annotation (in months). The training set was rigorously annotated by multiple medical experts to ensure data accuracy and reliability.

[0033] Based on hospital research, bone development is divided into three characteristics: childhood (3-12 years old, rapid bone growth), adolescence (12-17 years old, accelerated bone maturation), and early adulthood (17-20 years old, bones tend to be stable). Therefore, the embodiment of the present invention chooses to divide the bone age segment into three stages, and the three bone age segments are partitioned using a clustering algorithm, specifically including initialization, iterative optimization, and classification. The first is initialization, randomly selecting three cluster centers corresponding to the three bone age segments; then iterative optimization, by calculating the Euclidean distance of each bone age data point to the cluster center, redistributing the clusters and updating the center value until convergence, such as when the change threshold is <0.01, convergence is determined; finally, the bone age data is divided into three bone age intervals of 3-12 years old, 12-17 years old, and 17-20 years old, and the data in each interval has similar bone development characteristics. The skeletal characteristics of those aged 3-12 years are: characteristics corresponding to rapid bone growth; the skeletal characteristics of those aged 12-17 years are: characteristics corresponding to accelerated bone maturation; the skeletal characteristics of those aged 17-20 years are: characteristics corresponding to stable bones. The above three bone age intervals correspond to the training local models B1-B3 respectively.

[0034] It should be noted that the global model A and local models B1-B3 in the present invention are based on EfficientNet-v2, and a composite scaling strategy is used to balance the network depth, width and resolution. For example, the configuration of the global model A is: the depth coefficient is 1.2; the width coefficient is 1.1 and the resolution coefficient is 1.15. Secondly, the loss function of the global model A is the mean square error (MSE) loss, and the optimization goal is to minimize the gap between the predicted bone age and the actual bone age. The local model B1 focuses on learning the characteristics of rapid bone growth in childhood, such as epiphyseal width, the number of ossification centers, etc.; the local model B2 focuses on capturing the details of accelerated bone maturation in adolescence, such as the degree of epiphyseal closure, bone density changes, etc.; the local model B3 focuses on learning the characteristics of stable bones in early adulthood, such as the final shape of bone morphology. The local models all adopt the EfficientNet-v2 architecture, are trained independently, the loss function is MSE, and the training parameters are consistent with the global model A.

[0035] The global model A and the local models B1-B3 can be trained separately or simultaneously. After the training set is input into the global model A, the global model A outputs a preliminary bone age value and probability distribution, such as the first probability distribution A. Then, the training set is input into the local model B corresponding to the bone age segment, that is, one of B1-B3, to output a refined bone age value and probability distribution, such as the second probability distribution B. The fusion model outputs a fused third probability distribution C based on the first probability distribution A and the second probability distribution B. In one embodiment of the present invention, the fusion model adopts the EfficientNetV2 model, which dynamically adjusts the depth, width and resolution of the network through a composite scaling method to achieve efficient performance under limited computing resources. The fusion module includes an input layer, a feature extraction module, a regression branch and a classification branch. The training of the fusion module depends on a combined loss function and a loss weight update mechanism to optimize the performance of regression and classification tasks. The fusion module adopts a combined loss function, which is expressed as follows:

[0036] L=λ1·MSE+λ2·CE

[0037] Mean square error (MSE): used for regression tasks to measure the difference between predicted bone age and actual bone age. The MSE expression is as follows:

[0038]

[0039] Among them, y i is the actual bone age, To predict bone age, n is the number of samples.

[0040] Cross-Entropy Loss (CE): used for classification tasks to measure the difference between the predicted age range probability distribution and the true label. The expression of cross-entropy loss (CE) is as follows:

[0041]

[0042] Among them, y ij is the true label of sample i in interval j, p ij is the predicted probability, and k is the number of bone age intervals.

[0043] The loss weights λ1 and λ2 are hyperparameters used to balance the contributions of MSE and CE to ensure that the two tasks are optimized evenly during training.

[0044] Furthermore, the fusion module weights, i.e., convolution kernel and fully connected layer parameters, are updated by the stochastic gradient descent algorithm, driving the momentum mechanism. The update formula is:

[0045]

[0046] Δwt Weight update at time step t.

[0047] μ: Momentum factor, which accelerates the gradient direction and suppresses oscillations.

[0048] α: learning rate, controls the weight update step size.

[0049] The gradient of the loss function with respect to the weights calculated by backpropagation.

[0050] In an embodiment of the present invention, the training process runs for 200 cycles, and the data is shuffled and trained in batches in each cycle, converging to a state that minimizes the loss through iterative updates.

[0051] Furthermore, the loss weight parameters λ1 and λ2 are determined. In the embodiment of the present invention, λ1 and λ2 are fixed hyperparameters and are determined manually through experiments. The specific process is as follows:

[0052] Experimental setup: Test different combinations of λ1 and λ2 by trial and error or grid search.

[0053] Validation evaluation: The validation set is used to evaluate the performance of each group of λ1 and λ2. The indicators include MSE, MAE (Mean Absolute Error) and classification accuracy.

[0054] Selection criteria: Select the combination that makes the model perform best on both regression and classification tasks, usually with the goal of minimizing the validation set MAE, such as a target of 9 months.

[0055] Fixed value: Once selected, λ1 and λ2 remain unchanged during training to ensure that the loss balance is controllable.

[0056] This manual tuning method is widely used in multi-task learning, avoiding the complexity of dynamic weight adjustment while ensuring the stability and repeatability of the model.

[0057] When the training set is input into the bone age detection model, the global model A outputs a preliminary prediction value and the first probability distribution A, the local model B outputs a refined prediction value and the second probability distribution B, and the fully connected layer in the fusion module learns the weights α, β, and γ, and outputs the weighted bone age value and the corresponding third probability distribution C.

[0058] It should be noted that the X-ray images in the training set Z are obtained by segmenting the hand bone area using U-Net on the original image, generating a mask image, and performing contrast enhancement, brightness adjustment, and standardization on the image.

[0059] Example 2:

[0060] Figure 2As shown, the present invention discloses a method for predicting bone age based on the above-mentioned bone age detection model, comprising the following steps:

[0061] Input the X-ray image containing bone age information into the trained bone age detection model;

[0062] The global model outputs a preliminary prediction of bone age and the first probability distribution A based on the X-ray image;

[0063] The local model inputs the X-ray into the corresponding local model according to the bone age segment corresponding to the first probability distribution A, and the corresponding local model outputs the refined bone age and the second probability distribution B;

[0064] The first probability distribution A and the second probability distribution B are input into the fusion module, and the third probability distribution C is output; the fusion module here weightedly fuses the results of the first probability distribution A and the second probability distribution B to obtain the third probability distribution C.

[0065] The third probability distribution C is normalized and then the predicted bone age is output. After the probability of the third probability distribution C is normalized, the maximum probability corresponding to the predicted bone age is used as the final predicted bone age.

[0066] It should be noted that the global model extracts general bone features in X-ray images and establishes a mapping relationship between X-ray images and bone age. The local model extracts specific bone features in X-ray images, including rapid bone growth features, accelerated bone maturation features, or bone stabilization features.

[0067] The global model, local model and fusion module in Example 2 have the same structure and function as those in Example 1, and will not be described in detail to avoid redundancy.

[0068] Example 3:

[0069] The present invention also discloses a device for predicting bone age, comprising: an acquisition module for acquiring a skeletal X-ray of a target person; the X-rays herein include but are not limited to hand X-rays, and the target person may be a person in the three bone age zones in Example 1. A first extraction module for extracting features from the skeletal X-ray using a global model to obtain general features of the skeletal X-ray, and obtaining a first probability distribution of bone age based on the general features; the structure of the global model herein refers to the global model A in Example 1, and is used to extract general features of the bones and predict a preliminary bone age and a corresponding first probability distribution. A second extraction module for extracting features from the skeletal X-ray using a local model to obtain specific features of the skeletal X-ray, and obtaining a second probability distribution of bone age based on the specific features; the structure of the local model herein refers to one of the local models B1-B3 in Example 1, and is used to extract bone age features corresponding to the bone age segment, with specific reference to the description in Example 1, which will not be described in detail to avoid redundancy. A fusion module for obtaining a third probability distribution based on the first probability distribution and the second probability distribution; the fusion module herein weightedly fuses the first probability distribution and the second probability distribution to obtain a final predicted bone age and the corresponding third probability distribution. The evaluation module is used to obtain a predicted bone age based on the third probability distribution. Here, the evaluation module selects the maximum probability distribution value corresponding to the bone age as the final predicted bone age.

[0070] It should be noted that the acquisition module in Example 3 is used to acquire X-ray images of the target person. The X-ray images here can be images of the hands or legs. In the embodiment of the present invention, the target persons are children, adolescents, and early adults, that is, the three bone age groups in Example 1. The global model, local model, and fusion module in Example 3 have the same structure and function as those in Example 1 and are not described in detail to avoid redundancy.

[0071] To better validate the effectiveness of the bone age detection model disclosed in this invention, the RSNA dataset was used as a training set. The bone age prediction model based on EfficientNetV2 achieved a mean absolute error (MAE) of 0.752 years, significantly outperforming other common models. For example, the MAE of the ResNet model was 0.80 years, while the MAE of a single EfficientNet model was less convergent and stable than that of EfficientNetV2. This demonstrates that EfficientNetV2 has higher accuracy in bone age prediction tasks.

[0072] EfficientNetV2 demonstrates significant advantages in training efficiency. Under the same experimental conditions, EfficientNetV2 requires only 10 TPU (Tensor Processing Unit) days to complete training, while EfficientNet requires 25 TPU days and ResNet-RS requires 40 TPU days. This demonstrates that EfficientNetV2 can complete model training in a shorter time, significantly improving development efficiency.

[0073] In terms of accuracy, EfficientNetV2 also performed well. Under the same training time, the accuracy of EfficientNetV2 reached 86.0%, which is much higher than EfficientNet's 83.0% and ResNet-RS's 84.0%. This shows that EfficientNetV2 not only has a fast training speed, but also has a clear advantage in prediction accuracy. Figure 3 shown.

[0074] In summary, by training the global model and the local model, through the division of labor and cooperation between the global model and the local model, the global characteristics of bone age and the specific characteristics of bone age are taken into account; secondly, the fusion module dynamically fuses, and the fusion module with learnable weights integrates the prediction results of multiple models to reduce the deviation of a single model. The bone age detection model disclosed by the present invention predicts bone age, improving the accuracy of bone age prediction.

[0075] It will be easily understood by those skilled in the art that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A method for constructing a bone age detection model, characterized in that: include: The X-ray images in the training set are input into the bone age detection model to obtain the corresponding predicted bone age, and the bone age detection model is trained by minimizing the difference loss between the predicted bone age and the corresponding bone age label, wherein the bone age detection model includes a global model and k local models corresponding to bone age segments and a fusion module. When the X-ray images are input into the bone age detection model, the X-ray films are respectively passed through the global model and the corresponding local model to obtain the first probability distribution and the second probability distribution of the bone age, and the first probability distribution and the second probability distribution are input into the fusion module to obtain the third probability distribution and obtain the corresponding predicted bone age according to the third probability distribution, wherein the X-ray images in the training set are images containing bone age information, and the corresponding label is bone age.

2. The method for constructing a bone age detection model according to claim 1, wherein: The fusion module is used to weight the results of the fusion of the global model and the local model, and the training of the fusion module is based on the combined loss function and the loss weight parameter.

3. The method for constructing a bone age detection model according to claim 2, wherein: The combined loss function of the fusion module is as follows: L=λ1·MSE+λ2·CE Among them, MSE is the mean square error, which is used to measure the difference between the predicted bone age and the actual bone age, y i is the actual bone age, is the predicted bone age, n is the number of samples; CE is the cross entropy loss, which is used to measure the difference between the predicted age interval probability distribution and the true label, y ij is the true label of sample i in interval j, p ij is the prediction probability, k is the number of bone age intervals, λ1 and λ2 are loss weight parameters, and are fixed hyperparameters used to balance MSE and CE.

4. The method for constructing a bone age detection model according to claim 1, wherein: The classification method of the k bone age segments is as follows: Randomly select k cluster centers; Calculate the Euclidean distance from each bone age data point to the cluster center, redistribute the clusters and update the center value until convergence; The bone age data are divided into k intervals, where the data sets in each interval have similar bone development characteristics, and k is greater than or equal to 2, where k bone age segments correspond to k local models.

5. The method for constructing a bone age detection model according to claim 1, wherein: The method also includes using U-Net to segment the hand bone area of the X-ray images in the training set, generating mask images, and performing contrast enhancement, brightness adjustment, and normalization on the images.

6. The method for constructing a bone age detection model according to claim 1, wherein: The global model and local model adopt the EfficientNet-v2 network.

7. The method for constructing a bone age detection model according to claim 6, wherein: The global model extracts general features of bones in X-ray images and establishes a mapping relationship between X-ray images and bone age. When k is equal to 3, the k local models extract specific features of bones in X-ray images, which are rapid bone growth features, accelerated bone maturation features, or bone stabilization features.

8. A method for predicting bone age based on the bone age detection model according to claims 1-7, characterized in that: The following steps are involved: Input the X-ray image containing bone age information into the trained bone age detection model; The global model outputs the first probability distribution of bone age based on X-ray images; The local model inputs the X-ray into the corresponding local model according to the bone age segment corresponding to the first probability distribution, and the corresponding local model outputs a second probability distribution of bone age; Inputting the first probability distribution and the second probability distribution into a fusion model, and outputting a third probability distribution of bone age; The third probability distribution is normalized and the predicted bone age is output.

9. The method for predicting bone age using a bone age detection model according to claim 8, wherein: The global model extracts general features of bones in X-ray images and establishes a mapping relationship between X-ray images and bone age. The local model extracts specific features of bones in X-ray images, including rapid bone growth features, accelerated bone maturation features, or bone stabilization features.

10. A device for predicting bone age, characterized in that: include: An acquisition module, used to obtain a bone X-ray of the target person; a first extraction module, configured to extract features from the bone X-ray film using a global model to obtain general features of the bone X-ray film, and obtain a first probability distribution of bone age based on the general features; a second extraction module, configured to perform feature extraction on the bone X-ray film using a local model to obtain specific features of the bone X-ray film, and obtain a second probability distribution of bone age based on the specific features; A fusion module, configured to obtain a third probability distribution based on the first probability distribution and the second probability distribution; The evaluation module is used to obtain the predicted bone age based on the third probability distribution.

Images