A method of bone age prediction

Abstract

The application discloses a bone age prediction method, which comprises the following steps: pre-training a bone age prediction model through a preset training set, and using the bone age prediction model to predict the bone age. The technical scheme adds an attention mechanism according to the coupling degree of different convolution layer information in the bone age prediction model, improves the problem that the traditional attention mechanism is not easy to converge, and directly adds the gender information into the neural network model for end-to-end training, thereby optimizing the training process and further improving the accuracy of the bone age prediction. The bone age prediction model used in the bone age prediction method is verified through the attention mechanism, and the model pays more attention to the metacarpal part of the patient, thereby having a good prediction effect.

Description

Technical Field

[0001] This invention relates to the field of bone age prediction, and more particularly to a method for bone age prediction. Background Technology

[0002] Bone age is a physiological age characterizing the developmental stage of adolescents. It has wide applications in height prediction, athlete selection, and medical and health fields. The method of quantitatively measuring bone age is called bone age prediction. Bone age prediction is divided into traditional artificial bone age prediction methods, machine learning-based bone age prediction methods, and recently emerging deep learning-based bone age prediction methods.

[0003] Traditional bone age prediction methods generally include counting methods, atlas methods (such as the GP atlas method), and scoring methods (such as the TW, CNH, and Chinese O5 methods). The counting method involves counting the ossification centers and epiphyseal formation areas of the metacarpal bones; different values ​​correspond to different bone age values. This method is simple but has poor effectiveness and limited practicality. The GP atlas method is the most typical atlas method. This type of method mainly determines the specific bone age value by comparing X-ray images with standard bone age images spaced one year apart. If the comparison result is in the middle of two standard images, the one that is most similar is selected. This method is also easy to implement and is more accurate than the counting method. The TW series bone age scoring method assigns a weight to the developmental level of each bone and then sums them to obtain the score value. This score value ranges from 0 to 1000, while the corresponding bone age ranges from 0 to 15 years, with the smallest accuracy to 0.1 years. Therefore, this type of method is much more accurate than the previous two methods. However, this type of method is relatively slow in calculation, and inexperienced doctors are prone to confuse the weights, leading to misjudgments.

[0004] Traditional bone age prediction methods are time-consuming and require a high level of clinical experience from physicians. Doctors need a fast and relatively accurate auxiliary method for bone age prediction, leading to the development of automated bone age prediction methods that combine machine learning with traditional image processing. The most successful example of this type is BoneXpert, which has been successfully integrated into some medical devices and commercially deployed. BoneXpert uses an Active Appearance Model (AAM) to automatically segment the 15 bones of the hand and wrist, then determines the GP or TW2 bone age based on their shape, strength, and texture features, ultimately converting the bone age of each bone into the overall bone age. While BoneXpert has achieved a certain level of accuracy, it has several key limitations. BoneXpert cannot directly predict bone age; its prediction is primarily determined by the correlation between age and bone age. Furthermore, the system is quite fragile and will reject X-rays with excessive image noise.

[0005] In recent years, with the rapid development of artificial intelligence, especially deep learning, some methods for bone age prediction using deep learning have emerged. (Tajmir et al., 2019) proposed an automatic bone age prediction process. First, the input image is normalized to remove contrast differences. Then, a supervised convolutional neural network is trained to segment a coarse mask covering the region of interest (ROI) of the image. The correct mask is then obtained using the maximum connected component algorithm. Finally, the ROI is obtained by finding the intersection and performing classification. (C. Spampinato et al., 2017) considered the problem of the non-fixed position of the palm in X-ray images and added a spatial transformer network (STN) to the model for unsupervised registration of the palm position. However, this made network training extremely difficult and even affected the accuracy of the final prediction. (Iglovikov et al., 2017) also proposed a preprocessing procedure. First, a segmentation network is trained to segment the region of the hand bones in the image. Then, a registration network is trained to register the ROI to a fixed position. Considering the potential impact of gender on bone age prediction, the authors differentiated the model by gender. They also considered the varying impacts of different skeletal regions (the entire metacarpal, carpal, metacarpal, and phalangeal bones) on prediction accuracy, further differentiating the model by region. Furthermore, they addressed the definition of classification versus regression in this problem. Finally, they combined these factors to train a total of 15 networks. The results showed that gender differentiation was more effective than no differentiation, classification networks performed slightly better than regression networks, models based on the entire metacarpal region generally achieved the highest accuracy, while in some female models, carpal-based models performed slightly better than those based on the entire metacarpal region. Ensemble analysis of all models yielded the best results. While these methods utilize deep learning techniques to improve the speed and accuracy of bone age prediction, they do not focus on key skeletal regions crucial for prediction, resulting in poor robustness and interpretability. There is still room for improvement in accuracy. Summary of the Invention

[0006] To address the aforementioned problems in existing technologies, a bone age prediction method is provided, the specific technical solution of which is shown below:

[0007] A bone age prediction method, applied to bone age prediction from X-ray images, is characterized by pre-training a bone age prediction model using a preset training set;

[0008] The pre-set training set includes multiple X-ray images for bone age determination, the gender information of the test subject corresponding to each X-ray image for bone age determination, and the determined bone age value of the test subject corresponding to each X-ray image for bone age determination.

[0009] The bone age prediction model is trained using a baseline convolutional neural network. The training process for the bone age prediction model specifically includes:

[0010] Step A1: Extract X-ray images for bone age measurement and corresponding gender information from the preset training set in sequence and use them as a training data set.

[0011] Step A2: Preprocess the dataset to be trained and output the corresponding set of parameters to be trained.

[0012] Step A3: Input the set of parameters to be trained into the benchmark convolutional neural network for forward propagation to obtain a bone age prediction value;

[0013] Step A4: Calculate the error based on the predicted bone age value and the corresponding determined bone age value to obtain a loss function;

[0014] Step A5: Obtain a loss function gradient based on the loss function, and perform backpropagation on the baseline convolutional neural network based on the loss function gradient to complete one training cycle.

[0015] Step A6: Determine whether the fluctuation range of the loss function falls within a preset range.

[0016] If not, return to step A1;

[0017] If so, the trained baseline convolutional neural network will be output as the bone age prediction model.

[0018] After training to form a bone age prediction model, the bone age prediction model is applied to bone age prediction methods. The bone age prediction process specifically includes:

[0019] Step S1: Obtain the X-ray image of the test subject for bone age measurement and the test subject's gender information as a set of data to be predicted;

[0020] Step S2: Preprocess the dataset to be predicted and output the corresponding set of parameters to be predicted;

[0021] Step S3: Input the set of parameters to be predicted into the bone age prediction model to obtain a corresponding bone age prediction value and output it.

[0022] Preferably, in this bone age prediction method, the baseline convolutional neural network includes a feature extraction network and a bone age calculation network;

[0023] Feature extraction networks are used to extract features from a set of parameters to be trained and output a feature vector.

[0024] The input of the bone age calculation network is connected to the output of the feature extraction network, which is used to process the feature vector and output the predicted bone age value.

[0025] Preferably, in this bone age prediction method, the baseline convolutional neural network is implemented using the Inception-V3 model;

[0026] The feature extraction network consists of ten sequentially connected convolutional layers and one fully pooled layer;

[0027] A spatial domain attention module is set in the feature extraction network. The input of the spatial domain attention module is connected to the output of the seventh convolutional layer, and the output of the spatial domain attention module is connected to the input of the fully pooling layer. The spatial domain attention module is used to add spatial attention weights to the first feature vector output by the seventh convolutional layer and output a first weighted feature vector to the fully pooling layer.

[0028] A channel domain attention module is set in the feature extraction network. The input of the channel domain attention module is connected to the output of the tenth convolutional layer, and the output of the channel domain attention module is connected to the input of the full pooling layer. The channel domain attention module is used to derive attention weights and add them to the second feature vector output by the tenth convolutional layer, and output a second weighted feature vector to the full pooling layer.

[0029] The full pooling layer is used to perform global average pooling based on the first weighted feature vector and the second weighted feature vector, and outputs a feature vector.

[0030] Preferably, in this bone age prediction method, step A2 further includes:

[0031] Step A21: Extract the X-ray images used for bone age determination from the dataset to be trained, and perform metacarpal image segmentation on the X-ray images used for bone age determination to obtain a metacarpal image;

[0032] Step A22: Perform random data augmentation on the metacarpal images to obtain the corresponding enhanced metacarpal images;

[0033] Step A23: Normalize each enhanced image of a metacarpal bone;

[0034] Step A24: Obtain the grayscale values ​​of the three channels of the normalized metacarpal image. The set of grayscale values ​​of the three channels is included in the set of parameters to be trained.

[0035] Preferably, in this bone age prediction method, step S2 further includes:

[0036] Step S21: Extract the X-ray images used for bone age determination from the dataset to be predicted, and perform metacarpal image segmentation on the X-ray images used for bone age determination to obtain a metacarpal image.

[0037] Step S22: Perform random data augmentation on the metacarpal images three times to obtain three corresponding enhanced metacarpal images;

[0038] Step S23: Normalize each enhanced image of a metacarpal bone to obtain three corresponding normalized metacarpal bone images;

[0039] Step S24: Obtain the three-channel image grayscale values ​​of each normalized metacarpal image. The set of three-channel image grayscale values ​​is included in the set of parameters to be predicted.

[0040] Preferably, in this bone age prediction method, random data augmentation includes random center rotation of the metacarpal image, random horizontal translation of the metacarpal image, random magnification of the metacarpal image, and random horizontal flipping of the metacarpal image, wherein:

[0041] The rotation angle range of the random center rotation is [-20°, 20°];

[0042] The range of random horizontal translation is 20% of the horizontal width of the metacarpal image;

[0043] The magnification range for random magnification is [0, 0.2].

[0044] Preferably, in this bone age prediction method, step A3 further includes:

[0045] Step A31: Based on the grayscale values ​​of the three-channel image corresponding to the normalized metacarpal image, obtain a corresponding initial bone age prediction value.

[0046] Step A32: All initial bone age prediction values ​​are denormalized to obtain a corresponding denormalized bone age prediction value, which is the final bone age prediction value.

[0047] Preferably, in this bone age prediction method, step S3 further includes:

[0048] Step S31: Based on the grayscale values ​​of the three-channel images corresponding to each normalized metacarpal image, obtain a corresponding initial bone age prediction value.

[0049] Step S32: Perform inverse normalization on each initial bone age prediction value to obtain a corresponding inverse normalized bone age prediction value.

[0050] Step S33: Take the average value of each denormalized bone age prediction value, and record the average value as the final bone age prediction value.

[0051] Preferably, in this bone age prediction method, the loss function is:

[0052]

[0053] in:

[0054] yi is the predicted value for bone age;

[0055] Determine the value for bone age;

[0056] m represents the number of X-ray images used for bone age determination in the preset training set that have undergone the training process.

[0057] This technical solution has the following advantages and beneficial effects:

[0058] This technical solution employs a novel attention mechanism addition method, selectively adding an attention mechanism based on the coupling degree of different convolutional layer information in the bone age prediction model, thus improving the convergence problem of traditional attention mechanisms. Gender information is directly incorporated into the neural network model for end-to-end training, further enhancing the accuracy of bone age prediction while optimizing the training process. The bone age prediction model used in this method is visualized and validated through the attention mechanism, demonstrating that the model focuses more on presenting the developmental stage of the patient's hand bones and exhibits excellent predictive performance. Attached Figure Description

[0059] Figure 1 This is a flowchart illustrating the bone age prediction model training process in a bone age prediction method according to the present invention.

[0060] Figure 2 This is a schematic flowchart of a bone age prediction method according to the present invention.

[0061] Figure 3 This is a training loss diagram of a bone age prediction model based on CBAM in a bone age prediction method of the present invention.

[0062] Figure 4 This is a training loss diagram of a cross-domain attention-based bone age prediction model in a bone age prediction method of the present invention. Detailed Implementation

[0063] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0064] It should be noted that, unless otherwise specified, the embodiments and features described in the present invention can be combined with each other.

[0065] The present invention will be further described below with reference to the accompanying drawings and specific embodiments, but this is not intended to limit the scope of the invention.

[0066] To address the aforementioned problems in existing technologies, a bone age prediction method is provided, the specific technical solution of which is shown below:

[0067] A bone age prediction method, applied to bone age prediction from X-ray images, is characterized by pre-training a bone age prediction model using a preset training set;

[0068] The pre-set training set includes multiple X-ray images for bone age determination, the gender information of the test subject corresponding to each X-ray image for bone age determination, and the determined bone age value of the test subject corresponding to each X-ray image for bone age determination.

[0069] The bone age prediction model is trained using a baseline convolutional neural network, such as... Figure 1 As shown, the training process for the bone age prediction model specifically includes:

[0070] Step A1: Extract X-ray images for bone age measurement and corresponding gender information from the preset training set in sequence and use them as a training data set.

[0071] Step A2: Preprocess the dataset to be trained and output the corresponding set of parameters to be trained.

[0072] Step A3: Input the set of parameters to be trained into the benchmark convolutional neural network for forward propagation to obtain a bone age prediction value;

[0073] Step A4: Calculate the error based on the predicted bone age value and the corresponding determined bone age value to obtain a loss function;

[0074] Step A5: Obtain a loss function gradient based on the loss function, and perform backpropagation on the baseline convolutional neural network based on the loss function gradient to complete one training cycle.

[0075] Step A6: Determine whether the fluctuation range of the loss function falls within a preset range.

[0076] If not, return to step A1;

[0077] If so, the trained baseline convolutional neural network will be output as the bone age prediction model.

[0078] After training to form a bone age prediction model, the bone age prediction model is applied to bone age prediction methods, such as... Figure 2 As shown, the bone age prediction process specifically includes:

[0079] Step S1: Obtain the X-ray image of the test subject for bone age measurement and the test subject's gender information as a set of data to be predicted;

[0080] Step S2: Preprocess the dataset to be predicted and output the corresponding set of parameters to be predicted;

[0081] Step S3: Input the set of parameters to be predicted into the bone age prediction model to obtain a corresponding bone age prediction value and output it.

[0082] As a preferred embodiment, the bone age prediction method includes a baseline convolutional neural network comprising a feature extraction network and a bone age calculation network.

[0083] Feature extraction networks are used to extract features from a set of parameters to be trained and output a feature vector.

[0084] The input of the bone age calculation network is connected to the output of the feature calculation network, which is used to process the feature vector and output the predicted bone age value.

[0085] In a preferred embodiment of the present invention, the benchmark convolutional neural network includes a feature extraction network and a bone age calculation network. The feature extraction network is used to extract features from the image based on the preprocessed set of training parameters and output a feature vector. The bone age calculation network outputs a bone age prediction value based on the feature vector obtained by the feature extraction network through two fully connected layers.

[0086] As a preferred embodiment, this bone age prediction method uses the Inception-V3 model to implement the baseline convolutional neural network.

[0087] The feature extraction network consists of ten sequentially connected convolutional layers and one fully pooled layer;

[0088] A spatial domain attention module is set in the feature extraction network. The input of the spatial domain attention module is connected to the output of the seventh convolutional layer, and the output of the spatial domain attention module is connected to the input of the fully pooling layer. The spatial domain attention module is used to add spatial attention weights to the first feature vector output by the seventh convolutional layer and output a first weighted feature vector to the fully pooling layer.

[0089] A channel domain attention module is set in the feature extraction network. The input of the channel domain attention module is connected to the output of the tenth convolutional layer, and the output of the channel domain attention module is connected to the input of the full pooling layer. The channel domain attention module is used to derive attention weights and add them to the second feature vector output by the tenth convolutional layer, and output a second weighted feature vector to the full pooling layer.

[0090] The full pooling layer is used to perform global average pooling based on the first weighted feature vector and the second weighted feature vector, and outputs a feature vector.

[0091] In another preferred embodiment of the present invention, the introduction of the cross-domain attention model is further explained and described in detail:

[0092] In existing technologies, the mainstream attention models are SENet (Squeeze-and-Excitation Networks) and CBAM (Conversion Block Attention Model). SENet models the output of each layer of the baseline model on the channels to obtain channel attention weights, and multiplies these channel attention weights by the original features to obtain features with the attention mechanism. CBAM, in addition to generating channel attention weights, also generates spatial attention weights. Each layer's output feature needs to be multiplied by both the channel attention weights and the spatial attention weights, as shown in the following formula:

[0093]

[0094]

[0095] Here, F represents the output feature map of each layer of the model. F is passed through AvgPool and MaxPool to model the statistical information of the spatial domain and compress the feature map into a one-dimensional vector. Then, these two one-dimensional features are fully connected (Mlp) and activated by the Sigmoid activation function to obtain their respective attention to the channel domain (where σ represents the Sigmoid activation function). Finally, they are added together to obtain the final attention weight M. c (F). The spatial attention weight calculation method is similar, but the pooling method is changed to channel-domain pooling, and the fully connected layer is changed to f. 7×7 That is, a 7×7 convolution operation.

[0096] Although the CBAM model incorporates both spatial and channel attention mechanisms and achieves better results than SENet, in practice, simply adding the CBAM module to every layer of the model leads to premature convergence. This is because the attention mechanism essentially multiplies each feature by a mask with an absolute value less than 1 for feature selection. CBAM incorporates both channel and spatial attention, thus multiplying the features by the mask twice. This causes the model's gradient to decrease during backpropagation, as expressed in the following formula:

[0097]

[0098]

[0099] Where k represents the k-th layer of the model, Represents the channel domain attention weight. f(x) represents the attention weights in the spatial domain. k ) represents the convolution operation in the k-th layer. In the previous formula... Represents input features, This represents the output characteristics of the last CBAM block. This represents the final output value of the model. The local gradient of each layer needs to be multiplied by the weights of the two masks, which causes the gradient to become smaller and smaller, resulting in the gradient vanishing problem and making it difficult for the model to converge.

[0100] To address this issue, this invention proposes a novel attention-addition method. Considering the high coupling degree of channel domain information in shallow features, adding a channel domain attention mechanism to the shallow layers cannot effectively distinguish the details of correlation between features, thus failing to effectively focus on the regions of interest in the original image. Similarly, adding a spatial domain attention mechanism to deeper layers also fails to effectively focus on the regions of interest in the original image due to the larger receptive field and higher coupling degree between feature pixels. Therefore, this invention adds a spatial domain attention mechanism to the shallower layers and a channel domain attention mechanism to the deeper layers. Compared to the CBAM model, this invention only adds one attention module after the baseline network, which can alleviate the gradient vanishing problem and effectively help the entire bone age prediction model converge.

[0101] In the preferred embodiment described above, this bone age prediction model relies on the existing InceptionV3 as the baseline convolutional network. This convolutional neural network includes ten convolutional layers, referred to as Mix1 to Mix10. When Mix7 is reached, the feature map resolution is reduced to its minimum; therefore, a spatial domain attention module is added. The output of Mix7 is connected to the input of the spatial domain attention module and the input of Mix8. The spatial domain attention module adds spatial attention weights to the first feature vector output by Mix7 and outputs a first weighted feature vector to the full pooling layer. Simultaneously, Mix10 is used as a deeper convolutional layer, and a channel domain attention module is added here to derive attention weights for the second feature vector output by the tenth convolutional layer and outputs a second weighted feature vector to the full pooling layer. The input of the full pooling layer is connected to the output of the spatial domain attention module and the output of the channel domain attention module. The addition of these attention modules facilitates better model fitting.

[0102] Specifically, in the preferred embodiments described above, to make the model more sensitive to fine-grained features, this invention adopts a feature pyramid structure, using features from two different layers to predict the final bone age value. This structure is called cross-domain cascade: since the patient's gender information is also an important factor affecting the prediction results in the bone age prediction task, the gender information F is incorporated into the feature extraction network. gAs the input of a binary number to a feature extraction network, a 24-dimensional feature vector is generated after passing through two fully connected layers. This feature vector is then concatenated with the output feature vector of the previous convolutional network to obtain the final feature vector. Therefore, the entire feature extraction process is represented by the following formula:

[0103]

[0104]

[0105] F g_out =Mlp(F g )=(W1(W0(F g ))),F g ∈{0,1}

[0106] F out =Concat(F c_out ;F s_out ;F g_out )

[0107] Where F c_out F represents the output of the channel domain attention mechanism. s_out F represents the spatial domain attention output. g_out F represents the output of gender features. out The final output feature vector, This is element-wise multiplication. Spatial domain attention weight M s (F) is obtained by multiplying the feature F of the layer element by element to obtain the spatial domain attention output. The channel domain attention output is similar.

[0108] Extensive experiments have demonstrated that incorporating gender information improves the overall bone age prediction model. Furthermore, compared to existing practices that train different prediction models for different patients based on their gender, the end-to-end training method employed in this invention significantly optimizes the training process without compromising prediction performance.

[0109] In a preferred embodiment, step A2 of this bone age prediction method further includes:

[0110] Step A21: Extract the X-ray images used for bone age determination from the dataset to be trained, and perform metacarpal image segmentation on the X-ray images used for bone age determination to obtain a metacarpal image;

[0111] Step A22: Perform random data augmentation on the metacarpal images respectively;

[0112] Step A23: Normalize each enhanced image of a metacarpal bone;

[0113] Step A24: Obtain the grayscale values ​​of the three channels of each normalized metacarpal image. The set of grayscale values ​​of the three channels is included in the set of parameters to be trained.

[0114] In a preferred embodiment, step S2 of this bone age prediction method further includes:

[0115] Step S21: Extract the X-ray images used for bone age determination from the dataset to be predicted, and perform metacarpal image segmentation on the X-ray images used for bone age determination to obtain a metacarpal image.

[0116] Step S22: Perform random data augmentation on the metacarpal images three times to obtain three corresponding enhanced metacarpal images;

[0117] Step S23: Normalize each enhanced image of a metacarpal bone to obtain three corresponding normalized metacarpal bone images;

[0118] Step S24: Obtain the three-channel image grayscale values ​​of each normalized metacarpal image. The set of three-channel image grayscale values ​​is included in the set of parameters to be predicted.

[0119] In a preferred embodiment, this bone age prediction method includes random data augmentation comprising random center rotation of the metacarpal image, random horizontal translation of the metacarpal image, random magnification of the metacarpal image, and random horizontal flipping of the metacarpal image, wherein:

[0120] The rotation angle range of the random center rotation is [-20°, 20°];

[0121] The range of random horizontal translation is 20% of the horizontal width of the metacarpal image;

[0122] The magnification range for random magnification is [0, 0.2].

[0123] In another preferred embodiment of the present invention, the preprocessing of the training dataset is further described and defined: for the obtained X-ray images used for bone age determination, image extraction is required first to extract the metacarpal images from the X-ray images; for the extracted metacarpal images, three random data augmentations are required to prevent the bone age prediction model from overfitting; similarly, when using this bone age prediction model to predict the bone age of new patients, the above image augmentation and normalization steps are also required.

[0124] In a preferred embodiment, step A3 of this bone age prediction method further includes:

[0125] Step A31: Based on the grayscale values ​​of the three-channel images corresponding to the normalized metacarpal images, obtain a corresponding initial bone age prediction value.

[0126] Step A32: All initial bone age prediction values ​​are denormalized to obtain a corresponding denormalized bone age prediction value, which is the final bone age prediction value.

[0127] In a preferred embodiment, step S3 of this bone age prediction method further includes:

[0128] Step S31: Based on the grayscale values ​​of the three-channel images corresponding to each normalized metacarpal image, obtain a corresponding initial bone age prediction value.

[0129] Step S32: Perform inverse normalization on each initial bone age prediction value to obtain a corresponding inverse normalized bone age prediction value.

[0130] Step S33: Take the average value of each denormalized bone age prediction value, and use the average value as the final bone age prediction value.

[0131] In a preferred embodiment, the bone age prediction method uses the following loss function:

[0132]

[0133] in:

[0134] y i This is a predicted value for bone age.

[0135] Determine the value for bone age;

[0136] m represents the number of X-ray images used for bone age determination in the preset training set that have undergone the training process.

[0137] In another preferred embodiment of the present invention, since bone age prediction is a typical regression problem, this bone age prediction method directly uses the mean absolute error as the loss function. The mean absolute error is chosen as the evaluation metric because, compared to the mean squared absolute error, it is less sensitive to outliers and can more intuitively reflect the training effect of the regression model, making it more suitable as an evaluation metric.

[0138] A specific embodiment is now provided to further demonstrate and illustrate this technical solution:

[0139] In a specific embodiment of the present invention, this bone age prediction method uses a specially designed bone age prediction model to predict bone age.

[0140] In the above specific embodiment, the preset training set used is the RSNA2017 bone age prediction competition dataset, which includes 12,000 PNG format X-ray images of the hand bones of European and American adolescents and corresponding bone age labels, with bone age in months. Since the data in this dataset varies greatly across different age groups, it may lead to class imbalance. In this specific embodiment, the dataset is divided into ten groups according to age groups. Groups with fewer than 700 images are oversampled to 700 images to make the amount of data in each age group as similar as possible. Finally, 2,000 images are selected as the validation set, 200 images as the test set, and the rest as the training set.

[0141] In the specific embodiments described above, the training and parameters during the training process are set as follows:

[0142] The model loss function is the mean squared error, and the index is the mean absolute error.

[0143] The optimization method uses the Adam optimizer with beta1 of 0.9 and beta2 of 0.999. The initial learning rate is 0.0001. The learning rate is decayed after the validation set loss has been stable for a long time. The observation range is 10 epochs, and the decay coefficient is 0.8.

[0144] The regularization method uses random deactivation with a coefficient of 0.5; it also terminates the algorithm early and does not use weight penalties; the normalization method uniformly uses batch normalization.

[0145] Training was conducted using a Mini-Batch approach, with a batch size of 15.

[0146] In the above specific embodiments, to demonstrate the beneficial effects of this technical solution through comparison, the winning model of the RSNA 2017 bone age prediction competition was selected for comparison: This model was implemented based on the content of its team's blog. This model also uses InceptionV3 as the baseline convolutional model, and the features before the bone age regression layer are not subjected to global mean pooling but are directly flattened:

[0147] like Figure 3-4 As shown, Figure 3 The training loss diagram of the CBAM-based bone age prediction model is given. Figure 4 The training loss diagram of the bone age prediction model in this technical solution is given. It can be seen that the bone age model directly based on CBAM fails to converge after 10 epochs; while the model used in this technical solution converges stably and completes training after 120 epochs.

[0148] In the above specific embodiments, it can be observed through attention visualization that the model mainly focuses on the metacarpal region and the ossification center region of the phalanges, which is consistent with the key focus areas of traditional bone age prediction methods, proving the predictive effectiveness of this technical solution.

[0149] In summary, this technical solution improves upon the traditional attention mechanism by adding a novel attention mechanism tailored to the coupling degree of different convolutional layers in the bone age prediction model, thus overcoming the convergence problem. Furthermore, by directly incorporating gender information into the neural network model for end-to-end training, the accuracy of bone age prediction is further enhanced while optimizing the training process. The bone age prediction model used in this method is validated through visualization of the attention mechanism, demonstrating that the model focuses more on the hand bones, which are centrally responsible for presenting the patient's developmental stage, and exhibits excellent predictive performance.

[0150] The above description is merely a preferred embodiment of the present invention and does not limit the implementation and protection scope of the present invention. Those skilled in the art should realize that any equivalent substitutions and obvious changes made based on the description and illustrations of the present invention should be included within the protection scope of the present invention.

Claims

1. A bone age prediction method, applied to bone age prediction from X-ray images, characterized in that, A bone age prediction model is pre-trained using a pre-defined training set. The preset training set includes multiple X-ray images for bone age determination, the gender information of the test subject corresponding to each X-ray image for bone age determination, and the bone age determination value of the test subject corresponding to each X-ray image for bone age determination. The bone age prediction model is trained using a baseline convolutional neural network. The training process for the bone age prediction model specifically includes: Step A1: Extract the X-ray image used for bone age measurement and the corresponding gender information from the preset training set in sequence and use them as a training data set. Step A2: Preprocess the dataset to be trained and output the corresponding set of parameters to be trained; Step A3: Input the set of parameters to be trained into the benchmark convolutional neural network for forward propagation to obtain a bone age prediction value; Step A4: Calculate the error based on the predicted bone age value and the corresponding determined bone age value to obtain a loss function; Step A5: Obtain a loss function gradient based on the loss function, and the baseline convolutional neural network performs backpropagation based on the loss function gradient to complete one training cycle; Step A6: Determine whether the fluctuation range of the loss function falls within a preset range. If not, return to step A1; If so, the trained baseline convolutional neural network will be output as the bone age prediction model. After training and forming the bone age prediction model, the bone age prediction model is applied to the bone age prediction method. The bone age prediction process specifically includes: Step S1: Obtain the X-ray image of the test subject for bone age measurement and the gender information of the test subject, and use them as a set of data to be predicted; Step S2: Perform the preprocessing on the dataset to be predicted and output the corresponding set of parameters to be predicted; Step S3: Input the set of parameters to be predicted into the bone age prediction model to obtain a corresponding bone age prediction value and output it. The baseline convolutional neural network includes a feature extraction network and a bone age calculation network; The feature extraction network is used to extract features from the set of parameters to be trained and output a feature vector. The input of the bone age calculation network is connected to the output of the feature extraction network, and is used to process the feature vector and output the bone age prediction value. The baseline convolutional neural network is implemented using the Inception-V3 model. The feature extraction network comprises ten sequentially connected convolutional layers and one fully pooled layer; A spatial domain attention module is set in the feature extraction network. The input of the spatial domain attention module is connected to the output of the seventh convolutional layer, and the output of the spatial domain attention module is connected to the input of the fully pooling layer. The spatial domain attention module is used to add spatial attention weights to the first feature vector output by the seventh convolutional layer and output a first weighted feature vector to the fully pooling layer. A channel domain attention module is set in the feature extraction network. The input of the channel domain attention module is connected to the output of the tenth convolutional layer, and the output of the channel domain attention module is connected to the input of the fully pooling layer. The channel domain attention module is used to derive attention weights and add them to the second feature vector output by the tenth convolutional layer, and output a second weighted feature vector to the fully pooling layer. The global pooling layer is used to perform global average pooling on the first weighted feature vector and the second weighted feature vector, and output the feature vector.

2. The bone age prediction method as described in claim 1, characterized in that, Step A2 further includes: Step A21: Extract the X-ray images used for bone age determination from the dataset to be trained, and perform metacarpal image segmentation on the X-ray images used for bone age determination to obtain a metacarpal image; Step A22: Perform random data augmentation on the metacarpal images to obtain corresponding enhanced metacarpal images; Step A23: Normalize each of the enhanced metacarpal images to obtain the corresponding normalized metacarpal images; Step A24: Obtain the three-channel image grayscale values ​​of each normalized metacarpal image, wherein the set of the three-channel image grayscale values ​​is included in the set of parameters to be trained.

3. The bone age prediction method as described in claim 1, characterized in that, Step S2 further includes: Step S21: Extract the X-ray image for bone age measurement from the dataset to be predicted, and perform metacarpal image segmentation on the X-ray image for bone age measurement to obtain a metacarpal image. Step S22: Perform random data augmentation on the metacarpal image three times to obtain three corresponding enhanced metacarpal images; Step S23: Normalize each of the enhanced metacarpal images to obtain three corresponding normalized metacarpal images; Step S24: Obtain the three-channel image grayscale values ​​of each normalized metacarpal image, wherein the set of the three-channel image grayscale values ​​is included in the set of parameters to be predicted.

4. The bone age prediction method according to any one of claims 2 and 3, characterized in that, The random data augmentation includes random center rotation of the metacarpal image, random horizontal translation of the metacarpal image, random magnification of the metacarpal image, and random horizontal flipping of the metacarpal image, wherein: The rotation angle range of the random center rotation is [-20°, 20°]; The range of the random horizontal translation is 20% of the horizontal width of the metacarpal image; The magnification factor of the random magnification ranges from [0, 0.2].

5. The bone age prediction method as described in claim 2, characterized in that, Step A3 further includes: Step A31: Based on the grayscale values ​​of the three-channel image corresponding to the normalized metacarpal image, obtain a corresponding initial bone age prediction value. Step A32: The initial bone age prediction values ​​are all denormalized to obtain a corresponding denormalized bone age value, which is the bone age prediction value.

6. The bone age prediction method as described in claim 3, characterized in that, Step S3 further includes: Step S31: Based on the grayscale values ​​of the three-channel images corresponding to each normalized metacarpal image, obtain a corresponding initial bone age prediction value. Step S32: Each of the initial bone age prediction values ​​is denormalized to obtain a corresponding denormalized bone age prediction value. Step S33: Take the average value of all the inverse normalized bone age prediction values, and record the average value as the bone age prediction value.

7. The bone age prediction method as described in claim 1, characterized in that, The loss function is: ； in: The bone age prediction value; Determine a value for the bone age; The number of X-ray images used for bone age determination in the preset training set that have undergone the training process.

Images