An AI model-based bone age assessment method and system, and a bone age tester
By fusing multimodal data based on the Vision Transformer model and deep learning regression model, the problems of low automation and insufficient accuracy in bone age assessment are solved, and efficient and accurate bone age prediction is achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2025-01-10
- Publication Date
- 2026-04-07
AI Technical Summary
Existing bone age assessment methods have low automation and insufficient accuracy, rely on manual interpretation and suffer from error accumulation. CNNs have limitations in capturing long-range dependencies and modeling global features, and multimodal fusion and integration are insufficient.
Feature extraction is performed using the Vision Transformer model, multimodal data fusion is combined with demographic information, and bone age is assessed using a deep learning regression model to output a bone age assessment report.
It has achieved automation and improved accuracy in bone age assessment, enhanced the precision and efficiency of bone age prediction, and reduced the subjectivity and error of manual interpretation.
Smart Images
Figure CN119862533B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the fields of medical image processing and artificial intelligence technology, specifically to a method and system for automated bone age assessment of wrist X-ray images using a Transformer model. Background Technology
[0002] Bone age is a crucial medical indicator for assessing the developmental maturity of children and adolescents. Accurate bone age assessment is essential for growth and development monitoring, clinical treatment planning, and athlete selection. Traditional bone age assessment methods primarily rely on experienced radiologists or pediatricians manually interpreting X-ray images and referring to the Greulich-Pyle (GP) or Tanner-Whitehouse (TW) methods to determine bone age. These methods depend heavily on the physician's experience and are significantly affected by the interpreter's experience and fatigue level. Furthermore, manual interpretation is inefficient for large-scale population screening and real-time monitoring, and manual feature extraction suffers from error accumulation. Although convolutional neural networks (CNNs) have been explored for bone age assessment, CNNs have limitations in capturing long-range dependencies and modeling global features, and still lack sufficient capacity to handle complex ossification points, age-group adaptability, and multimodal fusion integration.
[0003] In summary, existing bone age assessment methods suffer from low automation and insufficient accuracy. Summary of the Invention
[0004] To address the shortcomings of the existing technologies, this invention provides an AI-based bone age assessment method and bone age meter, enabling the bone age meter to automatically and accurately extract features and predict bone age, thereby improving the accuracy and efficiency of bone age prediction.
[0005] In a first aspect, the present invention provides a bone age assessment method based on an AI model, comprising:
[0006] Obtain X-ray images of the subject's left wrist as detected by a bone age analyzer and record the subject's demographic information;
[0007] The X-ray image is preprocessed to obtain a preprocessed image;
[0008] The pre-processed image is used to extract features by a trained Vision Transformer model to obtain the skeletal features of the subject.
[0009] The skeletal characteristics of the subjects are fused with their demographic information to obtain multi-dimensional bone age-related data.
[0010] The multi-dimensional bone age-related data are processed by a trained deep learning regression model to assess bone age.
[0011] Based on the bone age assessment results from the deep learning regression model, a bone age assessment report is output.
[0012] Secondly, the present invention provides a bone age assessment system based on an AI model, wherein the bone age assessment system based on an AI model uses the aforementioned bone age assessment method based on an AI model.
[0013] Thirdly, the present invention provides a bone age assessment device, including a main control module, wherein the main control module runs the above-mentioned bone age assessment method based on an AI model.
[0014] Compared with the prior art, the beneficial effects of this invention are as follows:
[0015] This invention provides a bone age assessment method, system, and bone age meter based on an AI model. It acquires an X-ray image of the subject's left wrist using the bone age meter and records the subject's demographic information. The X-ray image is preprocessed to obtain a preprocessed image. A trained Vision Transformer model is used to extract features from the preprocessed image to obtain the subject's skeletal features. The skeletal features are then fused with the subject's demographic information using multimodal data fusion to obtain multidimensional bone age-related data. A trained deep learning regression model is used to process this multidimensional bone age-related data to assess bone age. Based on the bone age assessment results from the deep learning regression model, a bone age assessment report is output. This allows the bone age meter to automatically and accurately extract features and predict bone age, thereby improving the accuracy and efficiency of bone age prediction. Attached Figure Description
[0016] The accompanying drawings, which are included to provide a further understanding of the invention and form part of this invention, illustrate exemplary embodiments of the invention and are used to explain the invention, but do not constitute an undue limitation of the invention. Some specific embodiments of the invention will be described in detail below with reference to the accompanying drawings in an exemplary and non-limiting manner. The same reference numerals in the drawings designate the same or similar parts or components. It should be understood by those skilled in the art that these drawings are not necessarily drawn to scale. In the drawings:
[0017] Figure 1 This is a schematic flowchart of a bone age assessment method using a bone age meter according to an embodiment of the present invention. Detailed Implementation
[0018] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments. Obviously, the described embodiments are merely 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 should fall within the scope of protection of the present invention.
[0019] Example 1
[0020] See Figure 1 This embodiment provides an AI-based bone age assessment method. This method can run in the main control module of a bone age analyzer. It acquires an X-ray image of the subject's left wrist detected in the bone age analyzer and records the subject's demographic information. The X-ray image is preprocessed to obtain a preprocessed image. A trained Vision Transformer model is used to extract features from the preprocessed image to obtain the subject's skeletal features. The skeletal features and the subject's demographic information are then fused using multimodal data to obtain multidimensional bone age-related data. A trained deep learning regression model processes this multidimensional bone age-related data to assess bone age. Based on the bone age assessment results from the deep learning regression model, a bone age assessment report is output. This allows the bone age analyzer to automatically perform accurate feature extraction and bone age prediction, thereby improving the accuracy and efficiency of bone age prediction.
[0021] See Figure 1 The X-ray monitoring method for bone age measurement provided in this embodiment includes the following steps:
[0022] S101. Obtain the X-ray image of the subject's left wrist detected in the bone age analyzer and record the subject's demographic information.
[0023] S102. The X-ray image is preprocessed to obtain a preprocessed image;
[0024] S103. The preprocessed image is used to extract features through the trained Vision Transformer model to obtain the skeletal features of the subject.
[0025] S104. The skeletal characteristics of the subject are fused with the demographic information of the subject to obtain multi-modal data, thereby obtaining multi-dimensional bone age-related data.
[0026] S105. The multi-dimensional bone age-related data are processed using a trained deep learning regression model to assess bone age.
[0027] S106. Based on the bone age assessment results of the deep learning regression model, output a bone age assessment report.
[0028] It should be noted that in step S101, an X-ray image of the subject's left wrist, obtained from a bone age analyzer, is acquired. The skeletal structure of the left wrist is a commonly used site for clinical bone age assessment. Simultaneously, the subject's demographic information (e.g., gender, height, weight, ethnicity, date of birth, etc.) is recorded. This information is closely related to bone age assessment, assisting in subsequent multimodal data fusion and regression analysis, and providing more reference dimensions for bone age assessment, thus improving the accuracy and reliability of the assessment.
[0029] In step S102, considering that raw X-ray images often contain noise, low contrast, or artifacts, directly using them for deep learning models can affect the accuracy of feature extraction. In this embodiment, by preprocessing the raw X-ray image of the left wrist (such as denoising, enhancement, distortion correction, cropping, etc.), the image quality can be significantly improved, reducing interference factors in the model training and inference process. This makes it easier for the subsequent Vision Transformer model to focus on skeletal features, thereby improving the accuracy and stability of bone age determination.
[0030] In step S103, the preprocessed image is used to extract features using a trained Vision Transformer model to obtain skeletal features. Vision Transformer is a visual feature extraction network based on the Transformer architecture. Compared to traditional convolutional neural networks (CNNs), it performs better in capturing long-range dependencies, global information, and detailed features. This step extracts higher-dimensional and finer-grained feature information of the skeletal structure, providing a more representative feature description for subsequent bone age calculation and high-quality feature input for the bone age assessment model, thereby effectively improving assessment accuracy.
[0031] In step S104, in medical applications, relying solely on imaging features may not fully reflect individual differences, while incorporating demographic information (such as age, sex, height, and weight) can further improve the accuracy of bone age prediction. Multimodal data fusion can integrate features from different sources, enabling the model to more comprehensively understand the growth and development of the examinee.
[0032] In step S105, the final prediction of bone age is a regression problem. The multimodal data obtained from the previous fusion is input into a deep learning regression model (such as one based on a fully connected network or other advanced network architecture) for training and inference. This model can be trained in advance and then labeled and corrected based on actual clinical assessment results to ensure the accuracy of the prediction. This allows for refined predictions that take into account individual differences, reducing the shortcomings of relying solely on traditional statistics or empirical formulas. In this step, the deep learning regression model enables automated, high-precision, and personalized bone age assessment, and it is easy to update and iterate (continuing to train based on more data and new algorithms).
[0033] In step S106, a bone age assessment report is output based on the bone age assessment results from the deep learning regression model. The final bone age assessment results need to be output in a clinically readable and user-friendly format, providing doctors or patients with intuitive bone age analysis. In addition to bone age values, bone age can be compared with actual age or a reference range can be provided to obtain a bone age assessment report, helping clinicians assess the growth and development of the examinee.
[0034] In some preferred embodiments, the X-ray image is preprocessed to obtain a preprocessed image, including:
[0035] The X-ray image is subjected to grayscale normalization and noise filtering.
[0036] The X-ray image after grayscale normalization and noise filtering is enhanced using a self-supervised learning model, a masking autoencoder, to obtain a preprocessed image.
[0037] It should be noted that grayscale normalization refers to the standardization of the grayscale values of the original X-ray image to give it a zero mean and a unit standard deviation. Grayscale normalization can reduce interference caused by imaging differences.
[0038] The algorithm for grayscale normalization of the X-ray image is as follows:
[0039]
[0040] Among them, I norm (x,y) represents the grayscale value at pixel coordinates (x,y) after grayscale normalization of the original X-ray image. (x,y) The grayscale value of the original X-ray image at pixel coordinates (x, y). μ and σ are the mean and standard deviation of the grayscale distribution of the original X-ray image, respectively.
[0041] It should be noted that after grayscale normalization of the X-ray image, Gaussian filtering is preferably used to smooth the image again to further reduce noise interference. A 7×7 Gaussian kernel can be used to perform convolution operations on the grayscale normalized image to smooth random noise while preserving as much edge information of key bone structures in the wrist as possible.
[0042] It should be noted that the self-supervised learning model, the Masking Autoencoder (MAE encoder), learns more comprehensive and in-depth skeletal features by randomly occluding some pixels (or image patches) in the input image, forcing the model to "reconstruct" the occluded parts based on contextual information from other regions. In bone age assessment, this higher-quality feature representation allows the model to more accurately capture the developmental status of various parts and enhances the model's robustness. The MAE encoder's feature enhancement of the X-ray image includes: dividing the X-ray image, after grayscale normalization and Gaussian filtering, into several image patches and randomly occluding a certain proportion (e.g., 30%–50%). The MAE encoder only encodes the unoccluded image patches, learning compact feature vectors; the MAE decoder reconstructs the occluded parts based on these feature vectors. By continuously optimizing the reconstruction error, the model gradually improves its ability to restore missing regions. After training, the MAE encoder is retained as a feature extractor to encode new input images and obtain the "feature-enhanced" image representation. These features are then input into the subsequent bone age assessment model (Vision Transformer) to improve the overall assessment effect.
[0043] In some preferred embodiments, the skeletal features include: the ratio of the width of the distal radial epiphysis to the metaphysis, the length of the third metaphysis, and the maturity grading of auxiliary ossification points.
[0044] It should be noted that the distal radial epiphysis and metaphysis are important indicators reflecting the degree of skeletal development. The distal radial epiphysis and metaphysis show regular changes with age and are easy to observe and measure. The ratio of the width of the distal radial epiphysis to the width of the metaphysis not only significantly reflects the regular changes in skeletal development and is easy to observe and measure, but its dimensionless ratio also eliminates the influence of imaging conditions, providing efficient and reliable indicators for bone age assessment. The length of the third metaphysis is a key indicator of metacarpal growth and development, highly correlated with bone age, and its measurement location is fixed and highly repeatable. Accessory ossification points are located at specific locations in the carpal bones or phalanges and are minor ossification centers (such as the pisiform and sesamoid bones), which will exhibit different morphologies in X-ray images during bone maturation. Based on the morphology, density, and fusion with surrounding bones, ossification points can be classified into different grades: no ossification point (grade 0), ossification point present (grade 1), ossification point developing (grade 2), and ossification point closed (grade 3), reflecting the developmental stage of accessory ossification points. Numerical indicators such as the ratio of distal radius epiphysis to metaphysis width and the length of the third metaphysis can quantitatively reflect morphological changes in bones during growth and development. The maturity grading of accessory ossification points, through the appearance and developmental degree of ossification centers, further supplements the assessment of bone development speed. By integrating these characteristics, bone age assessment becomes more accurate and comprehensive.
[0045] In some preferred embodiments, training the Vision Transformer model includes:
[0046] A labeled training set of wrist X-ray images is obtained, the labels including the ratio of distal radius epiphysis to metaphysis width, the length of the third metaphysis, and the maturity grading of auxiliary ossification points. The wrist X-ray images in the training set are divided into several image blocks. The image blocks are positionally encoded to obtain a sequence of image blocks with positional information. The image block sequence is input into the Vision Transformer model, and feature learning is performed through a multi-head self-attention mechanism. For the numerical features, the ratio of distal radius epiphysis to metaphysis width and the length of the third metaphysis, the mean squared error loss function is used. For the categorical feature, the maturity grading of auxiliary ossification points, the cross-entropy loss function is used. The error between the features output by the model and the manually labeled values is calculated. Based on the error, the parameters of the Vision Transformer model are optimized and updated until the model converges, resulting in the trained Vision Transformer model.
[0047] It should be noted that the specific training process of the Vision Transformer model includes: acquiring a labeled training set of wrist X-ray images; labeling each image with the following information: the ratio of the distal radius epiphysis to the metaphysis width; the length of the third metaphysis; and the maturity level of auxiliary ossification points. Each X-ray image is divided into 16×16 pixel blocks; each image block is flattened into a vector and mapped to a feature vector via linear projection; positional encoding is added to preserve the spatial location information of the image, resulting in an image block sequence, which is then input into a multi-layer Transformer encoder; in each layer, multi-head self-attention is used for computation. After processing by the multi-layer Transformer encoder, a feature vector is output, which contains a global understanding of structures such as the radius, metacarpals, and auxiliary ossification points. After the last layer of the Vision Transformer model, multiple fully connected layers are connected for regression and classification tasks, respectively. To output two continuous numerical indicators—the ratio of the distal radius epiphysis to the metaphysis width and the length of the third metacarpal metaphysis—two independent fully connected regression layers are used for each of these indicators. Each layer has one output node and is used to predict the corresponding continuous skeletal features. For the classification task of auxiliary ossification point maturity levels, a fully connected layer is used with four output nodes, corresponding to the four levels of ossification points: no ossification point (level 0), ossification point present (level 1), ossification point developing (level 2), and ossification point closed (level 3). Based on the extracted ossification point feature vector, the model selects the category with the highest probability value as the final maturity level of the auxiliary ossification point through the probability distribution of the classification output. For the numerical skeletal features of the distal radius epiphysis to the metaphysis width ratio and the third metacarpal metaphysis length, a mean squared error loss function can be used. For classifying skeletal features to assist in ossification point maturity levels, the cross-entropy loss function can be used. The total loss is obtained by weighting the mean squared error loss and the cross-entropy loss. Based on the total loss, the Vision Transformer model parameters are iteratively updated using a backpropagation algorithm combined with the Adam optimizer until the loss function converges, resulting in a trained Vision Transformer model. The Vision Transformer model automatically extracts skeletal features, achieving efficient feature extraction and providing a more objective and consistent evaluation, thus overcoming the subjectivity of traditional manual judgment.
[0048] In some preferred embodiments, the skeletal features are fused with the subject's demographic information in a multimodal manner, including: vectorizing the skeletal features to obtain a skeletal feature vector; standardizing and encoding the demographic information to obtain a demographic feature vector; and concatenating the skeletal feature vector with the subject's demographic feature vector to obtain the multidimensional bone age-related data.
[0049] It should be noted that skeletal structural features include: the ratio of the width of the distal radial epiphysis to the metaphysis, the length of the third metaphysis, and the maturity grading of accessory ossification points. Demographic features include: age, height, and sex. Among these, the numerical features—the ratio of the width of the distal radial epiphysis to the metaphysis, the length of the third metaphysis, age, and height—are standardized using Z-scores. The maturity grading of accessory ossification points and sex characteristics (male, female) are represented using one-hot coding. The one-hot coding for the maturity grading of accessory ossification points is as follows: Maturity level 0 [1,0,0,0], Maturity level 1 [0,1,0,0], Maturity level 2 [0,0,1,0], Maturity level 3 [0,0,0,1]. The one-hot coding for sex characteristics is: Male: [1,0], Female: [0,1]. Skeletal structural features are vectorized as [ratio of distal radius epiphysis to metaphysis width, third metaphysis, and maturity grading of auxiliary ossification points], while demographic features are vectorized as [age, height, male, female]. A concatenation operation is used to directly merge the skeletal structural feature vector and the demographic feature vector into a new vector along the feature dimension, forming a multimodal input feature vector. By fusing skeletal features with the subject's demographic information in a multimodal manner, the model can utilize more comprehensive feature information, improving the accuracy and reliability of bone age assessment.
[0050] In some preferred embodiments, training the deep learning regression model includes: using a multilayer perceptron model as the deep learning regression model; and training the multilayer perceptron model to obtain the trained deep learning regression model.
[0051] It should be noted that the multilayer perceptron model is a simple yet powerful feedforward neural network. Through its design of an input layer, multiple hidden layers, and an output layer, it can effectively handle multi-dimensional bone age-related features. The specific network structure of the multilayer perceptron model is as follows: Input layer: receives the fused multi-dimensional bone age-related feature vector. Hidden layers: contain 3 layers, each with 128 neurons, using the ReLU activation function. Output layer: outputs the predicted bone age value.
[0052] In some preferred embodiments, training the multilayer perceptron model includes: using the multidimensional bone age-related data as model input and the corresponding standard bone age value as the training target; calculating the error based on the difference between the predicted bone age value and the standard bone age value, and updating the model parameters using a backpropagation algorithm; repeating the steps of using the multidimensional bone age-related data as model input, using the corresponding standard bone age value as the training target, calculating the error based on the difference between the predicted bone age value and the standard bone age value, and updating the model parameters using a backpropagation algorithm, until the prediction error decreases to the expected range, thereby obtaining the trained deep learning regression model.
[0053] It should be noted that the training process of the multilayer perceptron model includes: inputting the multi-dimensional feature vector after multimodal fusion and the corresponding standard bone age value into the model; calculating the error between the predicted bone age value and the true bone age value, which can be achieved using the mean squared error loss function; using the backpropagation algorithm to update the network parameters using gradients; and repeating the above process iteratively. The training process ends when the mean squared error loss function loss is lower than a preset threshold, resulting in a trained multilayer perceptron model. During training, the Adam optimizer is used, with a learning rate of 0.001, a batch size of 32, and 100 training epochs. To prevent overfitting, an early stopping strategy is used to dynamically adjust the training process. By using the multilayer perceptron model to perform regression prediction on the fused feature vector, the nonlinear relationship between skeletal development indicators and demographic information can be effectively captured, accurately predicting bone age.
[0054] In some preferred embodiments, a bone age assessment report is output based on the bone age assessment results of the deep learning regression model. The bone age assessment report may include: predicted bone age value, skeletal feature interpretation, and demographic information, which are used to compare the prediction results and provide better diagnostic and consultation services for doctors and patients.
[0055] Example 2
[0056] See Figure 1This embodiment provides an AI-based bone age assessment system. The system utilizes the aforementioned AI-based bone age assessment method, acquiring an X-ray image of the subject's left wrist detected in a bone age analyzer and recording the subject's demographic information. The X-ray image is preprocessed to obtain a preprocessed image. A trained Vision Transformer model is used to extract features from the preprocessed image to obtain the subject's skeletal features. The skeletal features are then fused with the subject's demographic information using multimodal data fusion to obtain multidimensional bone age-related data. A trained deep learning regression model is used to process this multidimensional bone age-related data to assess bone age. Based on the bone age assessment results from the deep learning regression model, a bone age assessment report is output. This allows the bone age analyzer to automatically perform accurate feature extraction and bone age prediction, thereby improving the accuracy and efficiency of bone age prediction.
[0057] The above embodiments are merely preferred embodiments of the present invention, and the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention. The scope of protection of the present invention should be determined by the scope of the claims.
Claims
1. A bone age assessment method based on an AI model, characterized in that, Includes the following steps: Obtain an X-ray image of the subject's left wrist taken using a bone age analyzer and record the subject's demographic information, including gender, height, weight, ethnicity, and date of birth. The X-ray image is preprocessed to obtain a preprocessed image, specifically including: The X-ray image is subjected to grayscale normalization and noise filtering. The X-ray image after gray-scale normalization and noise filtering is enhanced using a self-supervised learning model masking autoencoder to obtain a preprocessed image. After gray-scale normalization of the X-ray image, Gaussian filtering is used to smooth the gray-scale normalized image again. A 7×7 Gaussian kernel is selected to perform convolution operation on the gray-scale normalized image to preserve the edge information of key bone structures in the wrist while smoothing random noise. The algorithm for grayscale normalization of the X-ray image is as follows: in, To normalize the original X-ray image in pixel coordinates grayscale value at that location Raw X-ray image in pixel coordinates The grayscale value at that location; and These are the mean and standard deviation of the grayscale distribution of the original X-ray image, respectively. The self-supervised learning model occlusion autoencoder performs feature enhancement on the X-ray image by: dividing the X-ray image after grayscale normalization and Gaussian filtering into several image blocks, randomly occluding 30% to 50% of them, encoding only the unoccluded image blocks, learning compact feature vectors, reconstructing the occluded parts based on the feature vectors, and continuously optimizing the reconstruction error. The preprocessed images were used to extract features using a trained Vision Transformer model to obtain the skeletal features of the subject. The skeletal features included: the ratio of the width of the distal radius epiphysis to the metaphysis, the length of the third metaphysis, and the maturity grading of auxiliary ossification points. The specific training process of the Vision Transformer model includes: acquiring a labeled training set of wrist X-ray images; labeling each image with the following information: the ratio of the distal radius epiphysis to the metaphysis width; the length of the third metaphysis; and the maturity level of auxiliary ossification points; dividing each X-ray image into 16×16 pixel blocks; flattening each image block into a vector and mapping it to a feature vector through linear projection; adding positional encoding to preserve the spatial location information of the image, obtaining an image block sequence, and inputting the image block sequence into a multi-layer Transformer encoder; calculating through a multi-head self-attention mechanism in each layer; and outputting a feature vector after processing by the multi-layer Transformer encoder, which contains a global understanding of the radius, metacarpals, and auxiliary ossification points; in Vision... Following the final layer of the Transformer model, multiple fully connected layers are connected for regression and classification tasks, respectively. To output the two continuous numerical indicators—the ratio of the distal radius epiphysis to the metaphysis and the length of the third metaphysis—two independent regression fully connected layers are connected for each of these two parameters. Each layer has one output node and is used to predict the corresponding continuous skeletal features. For the classification task that assists in determining the maturity level of ossification points, a fully connected layer is connected with four output nodes, corresponding to the four levels of ossification points: no ossification point, ossification point, etc. The model identifies the emergence, development, and closure of ossification points. Based on the extracted ossification point feature vectors, it selects the category with the highest probability value as the final maturity grade for the auxiliary ossification point by classifying the output probability distribution. For numerical skeletal features such as the ratio of distal radius epiphysis to metaphysis width and the length of the third metaphysis, a mean squared error loss function is used. For categorized skeletal features that assist in the maturity grade of ossification points, a cross-entropy loss function is used. The total loss is obtained by weighting the mean squared error loss and cross-entropy loss. Based on the total loss, the Vision Transformer model parameters are iteratively updated using a backpropagation algorithm combined with the Adam optimizer until the loss function converges, resulting in a trained Vision Transformer model. The skeletal features of the subject are fused with the subject's demographic information in a multimodal manner. The skeletal features are vectorized to obtain a skeletal feature vector. The demographic information is standardized and encoded to obtain a demographic feature vector. The skeletal feature vector is concatenated with the subject's demographic feature vector to obtain multidimensional bone age-related data. The multi-dimensional bone age-related data are processed by a trained deep learning regression model to assess bone age. Based on the bone age assessment results from the deep learning regression model, a bone age assessment report is output.
2. The bone age assessment method based on an AI model as described in claim 1, characterized in that, Training the deep learning regression model includes: using a multilayer perceptron model as the deep learning regression model; and training the multilayer perceptron model to obtain the trained deep learning regression model.
3. The bone age assessment method based on an AI model as described in claim 2, characterized in that, Training the multilayer perceptron model includes: using the multidimensional bone age-related data as model input and the corresponding standard bone age value as the training target; calculating the error based on the difference between the predicted bone age value and the standard bone age value, and updating the model parameters using the backpropagation algorithm; repeating the steps of using the multidimensional bone age-related data as model input, using the corresponding standard bone age value as the training target, calculating the error based on the difference between the predicted bone age value and the standard bone age value, and updating the model parameters using the backpropagation algorithm, until the prediction error decreases to the expected range, thereby obtaining the trained deep learning regression model.
4. The bone age assessment method based on an AI model as described in any one of claims 1-3, characterized in that, The bone age assessment report includes predicted bone age values and interpretations of skeletal characteristics.
5. A bone age assessment system based on an AI model, characterized in that, The AI-based bone age assessment system uses the AI-based bone age assessment method as described in any one of claims 1-4.
6. A bone age assessment device, comprising a main control module, characterized in that, The main control module runs the bone age assessment method based on the AI model as described in any one of claims 1-4.