A method and system for assessing children's bone age with autonomous learning ability

Abstract

The present invention discloses a method and system for evaluating children's bone age with autonomous learning ability. First, the YOLOv5 model is used for coarse-grained classification to locate the positions and categories of 13 bone regions from the original image, and the 13 bone regions are cropped out from the original image. Then, the CAP model is used for fine-grained classification to predict the bone age grade of each bone region. Finally, the bone age is calculated according to the Chinese 05 bone age scoring standard. The system includes two deep neural networks, namely the object detection model YOLOv5 for coarse-grained classification and the context-aware attention pooling model CAP for fine-grained classification, which perform data preprocessing, feature extraction, bone age grade classification and calculation on the input image. The model update module is used to automatically update the CAP model and perform incremental training on the CAP model by using the incremental learning method based on knowledge distillation. It has the ability of autonomous learning. It has high prediction accuracy and low prediction cost.

Description

Technical Field

[0001] The present invention relates to a method and system for evaluating children's bone age with autonomous learning ability, which is applicable to the field of children's bone age evaluation under the target detection task in computer vision. Background Art

[0002] The age of human growth and development can be divided into calendar age and biological age. In medicine, the biological age is generally judged by bone age. Different bones in the human body show continuity and stages during the development process, and there are obvious differences in bone morphology at different ages. Bone age is determined by observing the development status of the left hand bones, such as their shape, size, position, and degree of closure. Bone age assessment was first applied in medical research to help understand and promote children's growth retardation, prevent precocious puberty, predict children's adult height, and assist in diagnosing endocrine diseases. Later, the application scenarios of bone age assessment have been widely expanded, such as determining the actual age of athletes and using it as a non-conviction standard in forensic medicine.

[0003] The initial bone age assessment method was based on left hand X-ray images, and doctors performed manual assessments through the GP atlas method or the TW method. The GP atlas method compares the image with a reference atlas and selects the most similar profile for the estimated bone age. The TW method uses the cumulative scores of different bone development morphologies to evaluate bone age. The latest version of the TW3 method is more objective than the GP mapping method, so it is considered to have higher accuracy and reproducibility than the GP mapping method. However, manual bone age assessment is cumbersome and time-consuming, costly, and easily affected by subjective factors, making bone age assessment not popular in township hospitals. Therefore, it is of great significance to research and develop an automatic, accurate, stable, simple, and inexpensive end-to-end automated bone age assessment method. Summary of the Invention

[0004] Aiming at the deficiencies of the prior art, a method and system for scoring children's bone age with autonomous learning ability are provided, which have low evaluation cost, are easy to popularize, are accurate in evaluation, and are convenient to use.

[0005] To achieve the above technical objectives, the method for evaluating children's bone age with autonomous learning ability of the present invention comprises the following steps:

[0006] S101: Preprocess the desensitized hand bone images, scale the long side of the X-ray image of the children's left hand bones to a unified number of pixels, and then scale the short side proportionally. The X-ray images of the children's left hand bones all contain 13 bone regions;

[0007] S102: Send the preprocessed X-ray image of the children's left hand bones into the target detection model YOLOv5 for position calibration and class detection of 13 hand bone regions;

[0008] S103: The YOLOv5 model outputs a tensor containing the detection box coordinates and detection categories of 13 bone regions. Since a large number of detection boxes will be generated for the same hand bone region during the object detection process of YOLOv5, the non-maximum suppression (NMS) method is used to eliminate redundant detection boxes and only keep the detection box with the highest confidence;

[0009] S104: Segment the X-ray image of the child's left hand bone according to the output result of the YOLOv5 model. Using the detection box coordinates of the 13 hand bone regions output by YOLOv5 in S103, 13 images corresponding to the 13 hand bone regions are segmented from the original X-ray image of the child's left hand bone;

[0010] S105: Preprocess the 13 segmented hand bone region images. First, scale the 13 bone region images to a unified preset size, and then crop each hand bone region image into multiple sub-images of different sizes for feature extraction at different scales. At the same time, normalize the pixel values in the cropped sub-images to between 0 and 1 to complete the preprocessing;

[0011] S106: Input the preprocessed hand bone region sub-images into the corresponding dedicated CAP model for feature extraction and prediction. Each CAP model receives sub-images of different scales from the same hand bone region, uses the convolutional neural network Xception to extract the feature representations of each sub-image at different scales. After obtaining the feature maps at each scale, the context information of each sub-image is integrated through the context-aware attention pooling module;

[0012] This module first uses the Bilinear pooling method to represent the feature maps at different scales as feature matrices of the same size, and then calculates the similarity values between these feature matrices. The similarity values between the feature map at each scale and the feature maps of other scales are used as weights to perform weighted summation on the feature maps at different scales to obtain the context vector of each scale feature map. The context vector contains the correlation between this scale and the features of other scales and has a richer feature representation than the original feature map. The number of generated context vectors is the same as the number of sub-images;

[0013] S107: Input the context vectors of the sub-images at different scales generated in S106 into the long short-term memory network (LSTM) after global average pooling to obtain the final feature map. The feature map is normalized to a probability distribution through a learnable pooling layer and the Softmax function. This probability distribution represents the confidence that the input hand bone region image belongs to which growth and development level. Sort the probabilities in the probability distribution, and output the bone age level with the highest probability as the predicted bone age level;

[0014] S108: According to the China05 skeletal age scoring standard, based on the growth status of 13 hand bone regions, the ulna includes 7 skeletal age grades, and each of the remaining 12 bone regions includes 8 skeletal age grades. Assign scores to each hand bone region according to the skeletal age grade output by the CAP model in S107, then sum up the scores of each region, and query the China05 skeletal age scoring table to calculate the final skeletal age.

[0015] Furthermore, desensitization is implemented using the Template Matching method. By pre - defining a template and matching it with medical images, it is determined whether sensitive information is included, thus achieving the purpose of desensitization.

[0016] Furthermore, pre - process the input hand bone image. Use relative scaling to scale the long side of the image to 640 pixels, then scale the short side proportionally, and then divide each pixel by 255 to normalize it to the range [0, 1].

[0017] Furthermore, the YOLOv5 model includes a feature extraction network, an SPP module, a Feature Pyramid Network, a detection head, and a post - processing module connected in sequence. Specifically, the feature extraction network adopted by YOLOv5 is CSPDarknet53, which can improve the performance and efficiency of the model; the SPP module is a variable - scale pooling module that can handle targets of different sizes in the image while reducing the computational amount; the Feature Pyramid Network (FPN) is used for multi - scale feature fusion to fuse feature maps of different scales, improving the accuracy and robustness of detection and finally outputting; the detection head consists of three convolutional layers, which are used to process the feature maps of three scales from the FPN respectively. Use convolution to unify the number of channels of the three feature maps and output the abscissa, ordinate, height, width, and confidence of the center point of the detection box; the YOLOv5 model uses the Non - Maximum Suppression (NMS) algorithm in the post - processing module to eliminate redundant prediction boxes and uses the Generalized Intersection over Union (GIOU) to measure the gap between the prediction box and the ground truth box.

[0018] Furthermore, the YOLOv5 model uses the CSPDarknet53 lightweight convolutional neural network as the feature extraction network. The tensor size output by the CSPDarknet53 network is:

[0019] (batch_size, num_anchore*(num_classes + 5), grid_size, grid_size);

[0020] Among them, batch_size represents the batch size, num_anchors represents the number of prediction boxes generated by the center points of the prediction boxes, num_classes represents the number of target classes, 5 represents the number of parameters predicted for each anchor box, including the center coordinates x, y, width w, height h and confidence of each anchor box, and grid_size represents the size of the feature map.

[0021] Furthermore, the 13 bone regions included in each left hand bone X-ray image of children are: ulna, radius, first proximal phalanx, first distal phalanx, third proximal phalanx, third middle phalanx, third distal phalanx, fifth proximal phalanx, fifth middle phalanx, fifth distal phalanx, first metacarpal bone, third metacarpal bone, and fifth metacarpal bone. The YOLOv5 model outputs the classification position coordinates of the 13 bone regions, and the original left hand bone X-ray image of children is cropped into 13 bone region images according to the coordinates.

[0022] Furthermore, since the bones in different regions have different morphological characteristics and also have differences in the bone age assessment levels, 13 specialized CAP models are respectively set for the 13 bone regions for classification. The 13 bone region images are respectively sent into the corresponding CAP models for fine-grained classification. According to the Chinese 05 bone age scoring standard, the ulna is divided into 7 levels according to different growth stages, and the bones in other regions are divided into 8 levels;

[0023] Each CAP model uses a context-aware attention module. First, the bilinear pooling method is used to represent the feature maps of different scales as feature matrices of the same size, and then the similarity values between these feature matrices are calculated. The similarity values of each scale of feature map with the feature maps of other scales are used as weights to perform weighted summation on the feature maps of different scales, obtaining the context vector of each scale of feature map. The context vector contains the correlation between this scale and the features of other scales, has a richer feature representation than the original feature map, and the number of generated context vectors is the same as the number of sub-images.

[0024] Furthermore, when the number of left hand bone X-ray images of children collected from the user side exceeds 5000, the incremental learning method based on knowledge distillation is used to perform incremental training on the CAP model. Among them, the left hand bone X-ray image data of children used to train the initial YOLOv5 model and CAP model is the original data, and the hand bone images collected from the user side are the incremental data;

[0025] The incremental learning method based on knowledge distillation is used to perform incremental training on the CAP model, adopting a two-stage training method:

[0026] First, use the YOLOv5 network to perform object detection on the original data to obtain the 13 segmented bone region images;

[0027] Then, using the original CAP model as the teacher model, predict the original data, and use the predicted probability distribution of the bone age grade of a bone region as the soft label. The soft label contains the prediction information of the teacher model. Using the soft label to supervise the training of the student model can alleviate the forgetting of the student model about the original data;

[0028] Subsequently, train a new CAP model. Use the incremental data to train CAP, and use the weighted L of KL divergence and cross-entropy loss YO-CAP as the loss function for model training. Specifically:

[0029] L YO-CAP (p t , p s , p′ s , y) = αL KL (p t , p s ) + (1 - α)L CE (p s , y)

[0030] L KL (p t , p s ) = ∑ i -p ti log(p si / p ti )

[0031]

[0032] Among them, p t represents the soft label generated by the original CAP model predicting the incremental data, p s represents the prediction result of the new CAP model for the incremental data, p′ s represents the prediction result of the new CAP model for the original data, y represents the label of the original data, and α represents the weight factor used to balance the weight ratio between the two losses; L CE represents the cross-entropy loss;

[0033] Finally, replace the original CAP model with the newly trained CAP model to achieve autonomous learning.

[0034] Furthermore, before adding the hand bone image to the incremental dataset, the user needs to be asked whether they agree to upload it; if they agree, add the hand bone image to the incremental dataset; if they do not agree, destroy the hand bone image after detection.

[0035] A children's bone age assessment system with autonomous learning ability, comprising: a data collection module, a bone age assessment module and a model update module; wherein the data collection module is connected to the bone age assessment module, and sends the collected data to the bone age assessment module for evaluation. At the same time, after updating the data, the data collection module continuously provides new training samples for the bone age assessment module;

[0036] The data collection module is used to collect the left hand bone X-ray images of children under 18 years old, desensitize the X-ray images, and remove sensitive information in the image content to protect user privacy; at the same time, collect the uploaded hand bone images and store them with the predicted bone age as the annotation;

[0037] The bone age assessment module is used to predict the bone age of the left hand bone X-ray images of children, and is composed of two deep neural networks, YOLOv5 and CAP, connected in sequence, and performs data preprocessing, feature extraction, bone age grade classification and calculation on the input left hand bone X-ray images of children;

[0038] The model update module is used to automatically update the model, and perform incremental training on the CAP model by using an incremental learning method based on knowledge distillation, so that the evaluation system has autonomous learning ability.

[0039] Beneficial effects:

[0040] Compared with the existing bone age assessment methods, the present invention has the following characteristics: adopting a two-stage bone age assessment method of coarse-grained + fine-grained based on deep learning, which can better mine the feature information of hand bone images, make full use of the context information of the images, and has higher accuracy and stronger generalization ability. Compared with the single-stage deep learning bone age assessment, the method proposed by the present invention can not only output the bone age prediction value, but also provide the detection results of 13 bone regions, which is convenient for doctors to further observe and helps in the next stage of diagnosis. Brief description of the drawings

[0041] Figure 1 It is the structural block diagram of the children's bone age assessment system with autonomous learning ability of the present invention.

[0042] Figure 2 It is the schematic diagram of the bone age assessment process in the embodiment of the present invention. Detailed implementation manners

[0043] The following will describe in detail the embodiments of the present invention with reference to the drawings:

[0044] The present invention discloses a method for scoring children's bone age with autonomous learning ability. First, the YOLOv5 model is used for coarse-grained classification to locate the positions and categories of 13 bone regions from the original image, and the 13 bone regions are cropped out from the original image. Then, the CAP model is used for fine-grained classification to predict the bone age grade of each bone region. Finally, the bone age is calculated according to the Chinese 05 bone age scoring standard. Using coarse-grained + fine-grained prediction can achieve higher prediction accuracy and significantly reduce the predicted categories (12*8 + 7 categories in a single stage, only 13 categories need to be detected by the two-stage YOLOv5, and only 7(8) categories need to be detected by a single CAP model).

[0045] The specific steps are as follows:

[0046] S101: Preprocess the desensitized hand bone image. Use relative scaling to scale the long side of the image to 640 pixels, and then scale the short side proportionally. At the same time, normalize the pixel values to between 0 and 1.

[0047] S102: Feed the preprocessed image into the neural network of YOLOv5 for prediction. YOLOv5 uses the CSPDarknet53 lightweight model as the backbone network, and the model outputs a tensor with a shape of (batch_size, num_anchors * (num_classes + 5), grid_size, grid_size). Among them, batch_size is the batch size, num_anchors is the number of anchor boxes predicted at each position, num_classes is the number of target categories, 5 represents the number of parameters predicted for each anchor box (center coordinate x, v, width w, height h, and confidence), and grid_size represents the size of the feature map.

[0048] S103: Decode the output information of YOLOv5. First, convert the prediction results to real coordinates, that is, convert the anchor box coordinates and sizes to actual coordinates and add them to the grid offset. Then, use non-maximum suppression (NMS) to remove overlapping prediction boxes and only keep the box with the highest score.

[0049] S104: Cut the image according to the output results. YOLOv5 outputs the positions and categories of 13 bone regions, and each bone region is cut out independently into a picture using the position coordinates.

[0050] S105: Preprocess the 13 cut bone region pictures. First, scale the image to the specified size, and then crop it into multiple sub-images for feature extraction at different scales. At the same time, normalize the pixel values to between 0 and 1.

[0051] S106: Feed the preprocessed bone region image into the corresponding CAP model for feature extraction and prediction. First, use a feature pyramid to extract feature representations of the image at different scales. After obtaining the feature maps at each scale, the CAP model uses a context-aware module to better utilize the context information of the image. The context-aware module includes two sub-modules: a global context module and a local context module. The global context module divides the feature map into multiple regions of different sizes through spatial pyramid pooling and performs average pooling on each region to obtain the feature representation of each region. The local context module processes the local receptive field of the feature map through convolutional operations to obtain a more detailed local feature representation. The feature representations obtained by the two sub-modules are respectively fed into different attention pooling modules for weighted fusion to obtain a richer context-aware feature representation.

[0052] S107: Feed the feature representation into a classifier to obtain the classification prediction result of the bone age grade. Use a multi-layer perceptron (MLP) classifier, and by sorting the probability distribution output by the classifier, obtain the bone age grade of the bone region.

[0053] S108: Calculate the bone age score according to the bone age grade, look up the China 05 bone age score standard table, and output the predicted bone age.

[0054] Specifically, the model update module is responsible for updating the CAP model using incremental data to implement the self-learning function. When the number of hand bone images collected by the data collection module exceeds 5000, use the model update module to update the CAP model. The model update mainly includes the following steps:

[0055] S201: Extract the knowledge in the original CAP model. Use the original CAP model to predict the incremental data set, and use the bone age grade probability distribution output by the model as a soft label, denoted as p t .

[0056] S202: Retrain the CAP model using the original data and incremental data. Use the weighted L of KL divergence and cross-entropy loss YO-CAP as the loss function for model training. Specifically:

[0057] L YO-CAP (p t , p s , p′ s , y) = αL KL (p t , p s ) + (1 - α)L CE (p s , y)

[0058] LKL (p t ,p s ) = ∑ i -P ti log(p si / p ti )

[0059]

[0060] Where ps is the prediction result of the new CAP model for the incremental data, p' s is the prediction result of the new CAP model for the original data, y is the label of the original data, and α is the weight factor used to balance the weight ratio between the two losses. Iterate repeatedly until the model converges. Combine the original data set with 5000 incremental data sets as the original data set for the next update.

[0061] S203: Replace the original CAP model with the new CAP model to achieve autonomous learning.

[0062] In this embodiment, the trained YOLOv5 model and CAP model are used, and the original data set is from the clinical data of Xuzhou Central Hospital. The data set contains the left hand X-ray images of 2518 children under 18 years old. These images are labeled with bone categories and bone age grades by multiple radiologists.

[0063] See Figure 1 , this embodiment provides a child bone age assessment system with autonomous learning ability. This system includes: a data collection module, a bone age assessment module, and a model update module. Refer to Figure 1 , this embodiment includes the following steps:

[0064] S301: Desensitize the left hand bone image uploaded by the user. Considering the certain standardization of the hand bone image, the template matching method (TemplateMatching) is used for desensitization: this method is to pre-define a template, match the medical image to determine whether it contains sensitive information, and achieve the purpose of desensitization.

[0065] S302: Ask the user whether to agree to include the uploaded hand bone image in the incremental data set. If agreed, the image is added to the incremental data set; if not, the image is destroyed after detection.

[0066] S303: Preprocess the input hand bone image. Use relative scaling to scale the long side of the image to 640 pixels, then scale the short side proportionally, and then divide each pixel by 255 to normalize it to the [0, 1] interval.

[0067] S304: Classify the preprocessed images using the trained YOLOv5 model.

[0068] The YOLOv5 network mainly consists of five parts: a feature extraction network, an SPP module, a feature pyramid network, a detection head, and post-processing. Specifically, the feature extraction network adopted by YOLOv5 is CSPDarknet53, which is an improved version of Darknet53. It uses the CrossStagePartialNetwork (CSP) structure, which can improve the performance and efficiency of the model. The SPP (SpatialPyramidPooling) module is a variable-scale pooling module that can handle objects of different sizes while reducing the computational complexity of the model. In YOLOv5, the SPP module is added to the feature extraction network to increase the receptive field and detection accuracy of the model. The feature pyramid network (FPN) is a method of multi-scale feature fusion that can fuse feature maps of different levels to improve the accuracy and robustness of detection. The detection head is a key part of YOLOv5. It consists of multiple convolutional layers and pooling layers and is responsible for predicting the class and location of the target. In YOLOv5, the detection head uses the multi-scale prediction method in YOLOv3 to predict targets of different scales. YOLOv5 adopts some new technologies in the post-processing stage, such as NMS (Non-Maximum Suppression) and GIou (Generalized Intersection over Union), etc., to further improve the accuracy and robustness of detection.

[0069] Refer to Figure 2 , each hand bone image contains 13 bone regions, namely the ulna, radius, first proximal and distal phalanges, third proximal, middle and distal phalanges, fifth proximal, middle and distal phalanges, first, third, and fifth metacarpals, that is, 13 categories. YOLOv5 outputs the classification position coordinates of the 13 bone regions, and the original image is cropped into 13 bone region images according to the coordinates.

[0070] S305: Send the 13 images into the corresponding CAP models for fine-grained classification respectively. Since the bones in different regions have different morphological characteristics and also have differences in the bone age assessment levels, in this embodiment, 13 CAP models are trained for classification for the 13 bone regions. Among them, the ulna is divided into 7 grades according to different growth stages, and the bones in other regions are divided into 8 grades.

[0071] The CAP model uses a context-aware module to better utilize the context information of images. The context-aware module includes two sub-modules: a global context module and a local context module. The global context module divides the feature map into multiple regions of different sizes through spatial pyramid pooling and performs average pooling on each region to obtain the feature representation of each region. The local context module processes the local receptive fields of the feature map through convolution operations to obtain a more detailed local feature representation. The feature representations obtained by the two sub-modules are respectively fed into different attention pooling modules for weighted fusion to obtain a richer context-aware feature representation.

[0072] S306: Feed the feature representation into a classifier to obtain the classification prediction result of the bone age grade. An MLP (Multi-Layer Perceptron) classifier is used, and by sorting the probability distribution output by the classifier, the bone age grade of the bone region is obtained.

[0073] S307: Calculate the bone age score based on the bone age grade, look up the China 05 bone age scoring standard table, and output the predicted bone age.

[0074] When the number of hand bone images collected by the data collection module exceeds 5000, the model update module is used to update the CAP model. First, extract the knowledge in the original CAP model. Use the original CAP model to predict the incremental dataset, and use the probability distribution of the bone age grade output by the model as the soft label, denoted as p t ; then, retrain the CAP model using the original data and the incremental data. Use KL divergence and cross-entropy loss weighted by L YO-CAP as the loss function for model training. Specifically:

[0075] L YO-CAP (p t , p s , p' s , y) = αL KL (p t , p s )+(1 - α)L CE (p s , y)

[0076] L KL (p t , p s ) = ∑ i -p ti log(p si / p ti )

[0077]

[0078] Among them, p sIt is the prediction result p' of the new CAP model for incremental data. s It is the prediction result of the new CAP model for the original data. y is the label of the original data, and α is the weight factor used to balance the weight ratio between the two losses. Iterate repeatedly until the model converges. Merge the original dataset with 5000 incremental datasets as the original dataset for the next update. Finally, replace the original CAP model with the new CAP model to achieve autonomous learning.

[0079] The hardware environment used in this embodiment is NVIDIA Tesla K80 GPU, 12GB video memory, and 8GB running memory. The CPU model is Intel(R) Xeon(R) CPU E5-2678 v3 @ 2.50GHz. The operating system is Ubuntu 18.04, CUDA 10.0, and CUDNN 7.6. The software versions are Python 3.5, TensorFlow 1.13.1, and Keras 2.2.

[0080] Evaluation metrics

[0081] In this embodiment, accuracy is used to evaluate the classification performance of the CAP model, and mean absolute error (MAE) is used to evaluate the accuracy of bone age assessment. The MAE expression is as follows:

[0082]

[0083] Among them, N is the number of pictures in the test set, R i represents the true bone age of the i-th picture, and P i represents the predicted bone age of the i-th picture.

[0084] Table 1 Comparison of accuracies of different classification models on the clinical dataset

[0085]

[0086]

[0087] CAP was compared with other classification models, and the results of bone maturity levels were output. Table 1 shows the classification accuracies of bone age grade classification using different classifiers on the clinical dataset. The results show that compared with other classification methods, the CAP model used in this embodiment achieved the highest classification accuracy (an average of 86.93%) in bone age grading, verifying the effectiveness of using CAP as a fine-grained classifier.

[0088] In this embodiment, the results of bone age assessment algorithms published in existing literature were also compared, with MAE used as the evaluation criterion. As shown in Table 3, the YOLOv5+CAP two-stage hierarchical bone age assessment method used in this embodiment is superior to the comparative methods, and the accuracy is much higher than that of the traditional manual bone age assessment method.

[0089] Table 2 Bone Age Assessment Accuracy

[0090] Bone age assessment method Mean absolute error (MAE: month) RAN 7.38 Agumented + GPR 6.80 CNN 6.89 BoneXpert 8.16 YOLOv5 + CAP 6.53

[0091] Although the embodiments of the present invention have been shown and described, those of ordinary skill in the art can understand that various changes, modifications, substitutions, and variations can be made to these embodiments without departing from the principles and spirit of the present invention. The scope of the present invention is defined by the claims and their equivalents.

Claims

1. A method for evaluating children's bone age with autonomous learning ability, characterized in that The steps are as follows: S101: Preprocess the desensitized hand bone images. Scale the long side of the X-ray images of children's left hand bones to a unified number of pixels, and then scale the short side proportionally. The X-ray images of children's left hand bones all contain 13 bone regions; S102: Send the preprocessed X-ray images of children's left hand bones into the target detection model YOLOv5 for position calibration and class detection of 13 hand bone regions; S103: The output of the YOLOv5 model is a tensor containing the detection box coordinates and detection classes of 13 bone regions. Use the NMS (Non-Maximum Suppression) method to eliminate redundant detection boxes and only retain the detection box with the highest confidence; S104: Segment the X-ray images of children's left hand bones according to the output results of the YOLOv5 model. Use the detection box coordinates of the 13 hand bone regions output by YOLOv5 in S103 to segment 13 images corresponding to the 13 hand bone regions from the original X-ray images of children's left hand bones; S105: Preprocess the 13 segmented hand bone region images. First, scale the 13 bone region images to a unified preset size, and then crop each hand bone region image into multiple sub-images of different sizes for feature extraction at different scales. At the same time, normalize the pixel values in the cropped sub-images to between 0 and 1 to complete the preprocessing; S106: Input the preprocessed sub-images of hand bone regions into the corresponding dedicated CAP model for feature extraction and prediction. Each CAP model receives sub-images of different scales from the same hand bone region, uses the convolutional neural network Xception to extract the feature representations of each sub-image at different scales. After obtaining the feature maps at each scale, use the context-aware attention pooling module to integrate the context information of each sub-image; This module first uses the Bilinear pooling method to represent the feature maps at different scales as feature matrices of the same size, and then calculates the similarity values between these feature matrices. Use the similarity values between the feature map at each scale and the feature maps of other scales as weights to perform weighted summation on the feature maps at different scales to obtain the context vector of the feature map at each scale. The context vector contains the correlation between this scale and the features of other scales and has a richer feature representation than the original feature map. The number of generated context vectors is the same as the number of sub-images; S107: After global average pooling, input the context vectors of sub-images at different scales generated in S106 into the long short-term memory network LSTM to obtain the final feature map. Normalize the feature map into a probability distribution through a learnable pooling layer and the Softmax function. This probability distribution represents the confidence that the input hand bone region image is in a certain growth and development level. Sort the probabilities in the probability distribution, and output the bone age level with the highest probability as the predicted bone age level; S108: According to the Chinese 05 skeletal age scoring standard, based on the growth status of 13 hand bone regions, the ulna includes 7 skeletal age grades, and each of the remaining 12 bone regions includes 8 skeletal age grades. Assign scores to each hand bone region according to the skeletal age grade output by the CAP model in S107, then sum up the scores of each region, and query the Chinese 05 skeletal age scoring table to calculate the final skeletal age; Use the incremental learning method based on knowledge distillation to incrementally train the CAP model, adopting a two-stage training method: Weighting of KL Divergence and Cross-Entropy Loss As the loss function for model training, specifically: ； ； ； Among them, represents the soft label generated by the original CAP model for predicting incremental data, represents the prediction result of the new CAP model for incremental data, represents the prediction result of the new CAP model for the original data, represents the label of the original data, represents the weight factor, which is used to balance the weight ratio between the two losses; L CE represents the cross-entropy loss; Finally, replace the original CAP model with the newly trained CAP model to achieve autonomous learning.

2. The method for evaluating children's bone age with autonomous learning ability according to claim 1, characterized in that: Desensitization is implemented using the Template Matching method. By pre-defining a template, match the medical image to determine whether it contains sensitive information, so as to achieve the purpose of desensitization.

3. The method for evaluating children's bone age with autonomous learning ability according to claim 1, characterized in that: Preprocess the input hand bone image. Use relative scaling to scale the long side of the image to 640 pixels, then scale the short side proportionally, and then divide each pixel by 255 to normalize it to interval.

4. The method for evaluating children's bone age with autonomous learning ability according to claim 1, wherein: The YOLOv5 model includes a feature extraction network, an SPP module, a feature pyramid network, a detection head, and a post-processing module connected in sequence. Specifically, the feature extraction network adopted by YOLOv5 is CSPDarknet53; the SPP module is a variable-scale pooling module that processes targets of different sizes in the image; the Feature Pyramid Network FPN is used for multi-scale feature fusion to fuse feature maps of different scales and finally output; The detection head consists of three convolutional layers, which are used to process the feature maps of three scales from the FPN respectively. Use convolution to unify the number of channels of the three feature maps, and output the abscissa, ordinate, height, width, and confidence of the center point of the detection box; The YOLOv5 model uses the Non-Maximum Suppression algorithm NMS in the post-processing module to eliminate redundant prediction boxes and uses GIOU to measure the gap between the prediction box and the ground truth box.

5. The method for evaluating children's bone age with autonomous learning ability according to claim 4, characterized in that: Among them, the YOLOv5 model uses the CSPDarknet53 lightweight convolutional neural network as the feature extraction network, and the tensor size output by the CSPDarknet53 network is: ； Among them, batch_size represents the batch size, num_anchors represents the number of prediction boxes generated by the center point of the prediction box, num_classes represents the number of target categories, 5 represents the number of parameters predicted by each anchor box, including the center coordinates x, y, width w, height h, and confidence of each anchor box, and grid_size represents the size of the feature map.

6. The method for evaluating children's bone age with autonomous learning ability according to claim 1, wherein The 13 bone regions included in each X-ray image of a child's left hand bone are: ulna, radius, first proximal phalanx, first distal phalanx, third proximal phalanx, third middle phalanx, third distal phalanx, fifth proximal phalanx, fifth middle phalanx, fifth distal phalanx, first metacarpal bone, third metacarpal bone, and fifth metacarpal bone. The YOLOv5 model outputs the classification position coordinates of the 13 bone regions, and crops the original X-ray image of the child's left hand bone into 13 bone region images according to the coordinates.

7. A method for evaluating children's bone age with autonomous learning ability according to claim 1, characterized in that Thirteen dedicated CAP models are set up for classification for 13 skeletal regions respectively. The 13 skeletal region images are sent into the corresponding CAP models for fine-grained classification. According to the Chinese 05 bone age scoring standard, the ulna is divided into 7 grades according to different growth stages, and the bones in other regions are divided into 8 grades.

8. A method for evaluating children's bone age with autonomous learning ability according to claim 1, characterized in that After more than 5000 left hand bone X-ray images of children are collected from the user side, an incremental learning method based on knowledge distillation is used to incrementally train the CAP model. Among them, the left hand bone X-ray image data of children used to train the initial YOLOv5 model and the CAP model are the original data, and the hand bone images collected from the user side are the incremental data. First, use the YOLOv5 network to perform object detection on the original data to obtain 13 skeletal region images after segmentation. Then, taking the original CAP model as the teacher model, predict the original data, and use its predicted probability distribution of the bone age grade of a skeletal region as the soft label, and the soft label contains the prediction information of the teacher model. Subsequently, train a new CAP model and use the incremental data to train the CAP.

9. The method for evaluating children's bone age with autonomous learning ability according to claim 8, characterized in that: Before adding the hand bone image to the incremental data set, the user needs to be asked whether they agree to upload it; if they agree, the hand bone image will be added to the incremental data set; if they do not agree, the hand bone image will be destroyed after detection.

10. A children's bone age assessment system with autonomous learning ability, characterized in that, To implement the above-mentioned child bone age assessment method with autonomous learning ability, the system includes: a data collection module, a bone age assessment module and a model update module; among them, the data collection module is connected to the bone age assessment module, and sends the collected data to the bone age assessment module for evaluation. At the same time, the data collection module continuously provides new training samples for the bone age assessment module after updating the data. The data collection module is used to collect left hand bone X-ray images of children under 18 years old, desensitize the X-ray images, remove sensitive information in the image content to protect user privacy; at the same time, collect the uploaded hand bone images and store them with the predicted bone age as the annotation. The bone age assessment module is used to predict the bone age of the left hand bone X-ray image of children. It is composed of two deep neural networks, YOLOv5 and CAP, connected in sequence, and performs data preprocessing, feature extraction, bone age grade classification and calculation on the input left hand bone X-ray image of children. The model update module is used to automatically update the model, and uses an incremental learning method based on knowledge distillation to incrementally train the CAP model, so that the evaluation system has autonomous learning ability.

Images