Children bone age image acquisition and positioning device and detection method thereof

Abstract

The invention discloses a child bone age image collecting and positioning device and a detection method thereof, and belongs to the technical field of bone age detection.The detection method specifically comprises the steps that a standard examination period is set, a child bone age image is collected every other period, and an examination record written by medical staff is obtained; the records are then input into a preset BERT model to extract bone age feature keywords; constructing an auxiliary detection model by using a convolutional neural network and a recurrent neural network, and training through the labeled historical data set; converting the age and the keyword of the child into vectors, inputting the vectors into the trained auxiliary examination model, and outputting a bone age difference value between the actual age and the bone age of the child; the system counts a bone age difference value sequence of the children every n examination periods, calculates an examination period adjustment coefficient according to indexes and through a specific formula, and then adjusts the number of days of the next n examination periods; the popularization rate and effectiveness of bone age detection are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of bone age detection, and in particular to a children's bone age image acquisition and positioning device and a detection method thereof. Background Art

[0002] In the medical field, especially in the field of pediatrics, the assessment of bone age is of great significance for judging the growth and development of children. Traditional bone age assessment methods rely on doctors' observation and empirical judgment of wrist X-rays. This method is not only time-consuming and labor-intensive, but also greatly affected by doctors' subjective factors, and accuracy is difficult to guarantee. With the development of deep learning and artificial intelligence technology, bone age assessment methods based on image data have gradually emerged. These methods automatically assess bone age by analyzing X-ray images of children's wrists. Although this improves the efficiency and objectivity of the assessment, it also brings new problems.

[0003] First, the existing bone age examinations are either planned by parents themselves, or followed by doctors’ instructions. The examination time cannot be scientifically planned according to the actual bone age characteristics of the children. In addition, the bone age assessment method based on image data requires a large amount of labeled data for training, and the acquisition of these data often requires professional medical equipment and technicians, which is difficult. Secondly, image data processing and analysis require high computing resources, especially when processing high-resolution X-ray images, the computing power cost increases significantly. In addition, the quality of image data also directly affects the accuracy of the assessment, and high-quality image acquisition further increases the cost.

[0004] Therefore, how to reduce the cost of bone age assessment while ensuring the accuracy of the assessment has become an important technical issue. The present invention proposes a method for judging through the text of the bone age examination report, aiming to use the existing medical record text information, combined with a deep learning model, to achieve efficient and low-cost bone age assessment. This method can not only reduce the dependence on expensive medical equipment, but also reduce the requirements for professional medical personnel, thereby increasing the popularity and convenience of bone age assessment. Summary of the invention

[0005] The purpose of the present invention is to provide a children's bone age image acquisition and positioning device and a detection method thereof to solve the following technical problems:

[0006] The existing bone age examination time intervals cannot be scientifically planned according to the actual situation of children, and the bone age assessment method based on image data requires a large amount of labeled data for training, which is difficult and requires high computing resources.

[0007] The purpose of the present invention can be achieved through the following technical solutions:

[0008] A method for collecting and detecting children's bone age images comprises the following steps:

[0009] A standard inspection cycle d is set, where d represents the number of days, and a bone age image of the child is collected every inspection cycle d, and an inspection record written by a medical staff based on the bone age image is obtained;

[0010] Inputting the inspection record into a preset BERT model to extract keywords of bone age characteristics;

[0011] An auxiliary detection model is constructed based on a convolutional neural network and a recurrent neural network, the model is trained through annotated historical data sets, the child's age and keywords in the examination record are extracted, the child's age and keywords are converted into vectors and input into the trained auxiliary inspection model, and the bone age difference between the child's actual age and bone age is output;

[0012] Every n examination cycles d, the bone age difference sequence of the children is counted, and the mean value Y of the bone age difference sequence is calculated a , the proportion of bone age differences exceeding the preset threshold d in the sequence P, and draw a line graph of the bone age difference over time to obtain the slope k between any two adjacent coordinate points; calculate the inspection cycle adjustment coefficient γ according to the following formula:

[0013]

[0014] Among them, i∈[1,n-1], k i It represents the i-th slope value in the line graph. According to the calculated inspection cycle adjustment coefficient γ, the number of days of the next n inspection cycles is adjusted to d'=d / γ.

[0015] As a further solution of the present invention: every n inspection cycles, the corresponding indicators of the bone age difference sequence are counted, and the indicators include the mean Y a , proportion P and slope k, and calculate the number of days of a single inspection cycle of the next n inspection cycles based on the indicators and formula.

[0016] As a further solution of the present invention: the process of extracting the key words of bone age characteristics is:

[0017] The annotated standard bone age feature dictionary is divided into a training set and a validation set. The training set is mapped into a vector and input into the BERT model after encoding. The content features in the text are extracted. The BERT model is back-propagated and updated through the validation set until convergence. The examination records of the children to be tested are encoded and input into the BERT model. The examination records include the actual age of the child and a bone age feature description text. The bone age features in the bone age feature description text are extracted, clustered to generate pending keywords, and the pending keywords are used to back-propagate and update the model. The pending keywords are obtained after repeated updates for several times and marked as keywords.

[0018] As a further solution of the present invention: the training process of the auxiliary diagnosis model is:

[0019] The historical data set includes standard bone age feature texts corresponding to children of different ages, and non-standard bone age feature texts of children of different ages marked in historical examination records. All bone age feature texts are sorted according to age to generate a time series of bone age features. The matrix formed by the time series of bone age features is regarded as an image and is input into the convolutional neural network in sequence. The convolutional neural network includes several convolutional neural layers, each convolutional neural layer includes several filters. Whenever the matrix image passes through a convolutional neural layer, the filter will slide and scan the matrix and generate a multidimensional array, reduce the size of the matrix image, and increase the dimension of the matrix image. The increased dimension is the same as the number of filters in the convolutional neural network. Each convolutional neural layer reduces the image size and increases the dimension of the matrix image in turn. The multidimensional array finally generated by the convolutional neural network is flattened and converted into a vector, and the vector is used as the input of the subsequent recurrent neural network.

[0020] As a further solution of the present invention: the training process of the auxiliary diagnosis model also includes:

[0021] The dimension of the matrix image output by the convolutional neural network remains unchanged, and the matrix image is flattened in the time dimension and eigenvector. The final multidimensional arrays of all convolutional neural layers generate a new matrix, and the new matrix vector is used as the input of the recurrent neural network LSTM. The flattened vector is used as the time dimension of the LSTM. The new matrix is ​​trained by the time dimension and eigenvector, and the corresponding relationship between bone age characteristics and children's age is output.

[0022] As a further solution of the present invention: the recurrent neural network also includes several additional Transformer layers, all Transformer layers except the penultimate Transformer layer adopt a return sequence mode, and generate a vector sequence corresponding to the continuous iterations of the recurrent neural network, and the vector sequence is grouped by time series as the input of the penultimate Transformer layer.

[0023] As a further solution of the present invention, the examination record of the child currently being examined is also stored in the historical data set.

[0024] The present invention also includes a child bone age image acquisition and positioning device, which is used to implement the above-mentioned child bone age image acquisition and detection method, and is characterized by comprising:

[0025] A memory, used to set a standard inspection cycle d, where d represents the number of days, collect the bone age image of the child every inspection cycle d, and obtain the inspection record written by the medical staff based on the bone age image;

[0026] A text processor, used for inputting the inspection record into a preset BERT model to extract keywords of bone age characteristics;

[0027] A graphics processor is used to build an auxiliary detection model based on a convolutional neural network and a recurrent neural network, train the model through a labeled historical data set, extract the child's age and keywords from the examination record, convert the child's age and keywords into vectors and input them into the trained auxiliary inspection model, and output the bone age difference between the child's actual age and bone age;

[0028] The central processing unit is used to count the bone age difference sequence of the child every n examination cycles d, and calculate the mean value Y of the bone age difference sequence a , the proportion of bone age differences exceeding the preset threshold d in the sequence P, and draw a line graph of the bone age difference over time to obtain the slope k between any two adjacent coordinate points; calculate the inspection cycle adjustment coefficient γ according to the following formula:

[0029]

[0030] Among them, i∈[1,n-1], k i It represents the i-th slope value in the line graph. According to the calculated inspection cycle adjustment coefficient γ, the number of days of the next n inspection cycles is adjusted to d'=d / γ.

[0031] Beneficial effects of the present invention:

[0032] The present invention provides a method for collecting and detecting children's bone age images. By setting a standard inspection cycle, children's bone age images are regularly collected and inspection records written by medical staff are obtained. The BERT model is used to extract bone age feature keywords, and an auxiliary detection model is constructed by combining convolutional neural networks and recurrent neural networks. The historical data set is trained to achieve accurate output of the difference between the child's actual age and bone age. This method dynamically adjusts the inspection cycle by statistically analyzing the mean, proportion and slope of the bone age difference sequence, thereby improving the evaluation efficiency and accuracy. By analyzing the inspection record text, the dependence on higher computing power equipment is reduced, the cost is reduced, and the popularity and effectiveness of bone age detection are improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] The present invention will be further described below in conjunction with the accompanying drawings.

[0034] Figure 1 It is a flow chart of a method for collecting and detecting children's bone age images according to the present invention. DETAILED DESCRIPTION

[0035] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0036] See also Figure 1 As shown, the present invention is a method for collecting and detecting children's bone age images, comprising the following steps:

[0037] Set the inspection cycle: First, set a standard inspection cycle d (in days). Every other inspection cycle, collect bone age images of children. These images are then analyzed by professional medical staff, and detailed inspection records are written.

[0038] Text feature extraction: The examination records written by medical staff are input into a preset BERT model. The BERT model is an advanced natural language processing model that can deeply understand the semantic information of the text. Through the BERT model, we can extract keywords related to bone age characteristics from the examination records, which are crucial for subsequent bone age assessment.

[0039] Building an auxiliary detection model: Based on convolutional neural networks (CNN) and recurrent neural networks (RNN), we built an auxiliary detection model. This model is trained using annotated historical data sets to learn how to extract children's age and bone age feature keywords from examination records and convert this information into vector form. These vectors are then input into the trained auxiliary detection model, which outputs the difference between the child's actual age and bone age.

[0040] Dynamically adjust the inspection cycle: Every n inspection cycles, we count and analyze the bone age difference sequence of children. Specifically, we calculate the mean of the bone age difference sequence Ya , and the proportion of bone age differences exceeding the preset threshold d in the sequence P. At the same time, we also draw a line graph of the bone age difference over time, and obtain the slope k between any two adjacent coordinate points by calculation. Based on these statistical data and slope information, we use a specific formula to calculate the adjustment coefficient γ of the inspection cycle. This adjustment coefficient is used to dynamically adjust the number of days in the next n inspection cycles. The formula is:

[0041]

[0042] Among them, i∈[1,n-1], k iIt represents the i-th slope value in the line graph. According to the calculated inspection cycle adjustment coefficient γ, the number of days of the next n inspection cycles is adjusted to d'=d / γ.

[0043] In a preferred embodiment of the present invention, the indicators corresponding to the bone age difference sequence are counted every n inspection cycles, and the indicators include the mean Y a , proportion P and slope k, and calculate the number of days of a single inspection cycle of the next n inspection cycles based on the indicators and formula.

[0044] In another preferred embodiment of the present invention, the process of extracting keywords of bone age characteristics is:

[0045] 1. Data preparation: First, we prepare a labeled standard bone age feature dictionary, which contains a rich vocabulary of bone age features and their corresponding annotation information. We divide this dictionary into two parts: training set and validation set. The training set is used to train the model, while the validation set is used to evaluate the performance of the model and make necessary adjustments.

[0046] 2. Vector Mapping and Encoding: Next, we map the data in the training set into vector form and encode it. These encoded vectors are then input into a pre-trained BERT model. The BERT model is a powerful natural language processing model that can deeply understand the semantic information of the text and extract useful features from it.

[0047] 3. Feature extraction and model training: In the BERT model, we train the model through the training set to learn how to extract content features related to bone age from the text. At the same time, we also use the validation set to evaluate the model and update the model through the back propagation algorithm until the model converges, that is, reaches a stable performance state.

[0048] 4. Preprocessing of the test data: After the model training is completed, we preprocess the examination records of the children to be tested. These examination records contain the children’s actual age and detailed descriptions of bone age characteristics. We perform the same encoding process on the test data and input it into the trained BERT model.

[0049] 5. Keyword extraction and clustering: In the BERT model, we conduct in-depth analysis of the test data and extract content features related to the bone age feature description text. These features are then used to generate pending keywords. In order to further improve the accuracy and reliability of the keywords, we also perform cluster analysis on these pending keywords to group similar or related keywords together.

[0050] 6. Model iteration and update: Finally, we use the pending keywords generated by clustering to back-propagate the BERT model. This is an iterative process, and we will update it repeatedly until we get satisfactory keywords. These keywords that have been repeatedly updated and optimized are finally marked as keywords related to bone age characteristics and used in the subsequent bone age assessment process.

[0051] In another preferred embodiment of the present invention, the training process of the auxiliary diagnosis model is:

[0052] 1. Historical data set preparation:

[0053] We collected standard bone age feature texts corresponding to children of different ages, as well as non-standard bone age feature texts of children of different ages that have been annotated in historical examination records. These data formed the basis for our training model.

[0054] 2. Time series generation:

[0055] According to the age of the child, all bone age feature texts are sorted to generate a time series of bone age features. This time series reflects the changing trend of children's bone age as they grow older.

[0056] 3. Matrix image construction:

[0057] The matrix formed by the time series of bone age characteristics is regarded as a special "image" and is input into the convolutional neural network (CNN) in sequence. This representation method enables CNN to capture the change pattern of bone age characteristics in the time dimension.

[0058] 4. Convolutional Neural Network Processing:

[0059] CNN consists of several convolutional neural layers, each of which contains multiple filters. Whenever the matrix image passes through a convolutional neural layer, the filter slides across the matrix and generates a multidimensional array. This process reduces the size of the matrix image while increasing the dimensionality of the matrix image by the same amount as the number of filters in the convolutional neural network.

[0060] Each convolutional neural layer successively reduces the image size and increases the dimension size of the matrix image. Finally, the multidimensional array generated by CNN is flattened and converted into a vector as the input of the subsequent recurrent neural network (RNN).

[0061] 5. Recurrent Neural Network LSTM Training:

[0062] In a preferred embodiment, we keep the dimension of the matrix image output by the convolutional neural network unchanged and flatten the matrix image in the time dimension and feature vector. In this way, the final multidimensional array of all convolutional neural layers generates a new matrix.

[0063] This new matrix vector is used as the input of the recurrent neural network LSTM. We use the flattened vector as the time dimension of LSTM, train the new matrix based on the time dimension and feature vector, and output the corresponding relationship between bone age features and child age.

[0064] 6. Transformer layer enhancement:

[0065] In another preferred embodiment, we add several additional Transformer layers to the recurrent neural network. Except for the last Transformer layer, all Transformer layers adopt the return sequence mode and generate vector sequences corresponding to the continuous iterations of the recurrent neural network.

[0066] These vector sequences are grouped according to time series and used as the input of the penultimate Transformer layer. This structure further enhances the model's ability to process time series data and improves the accuracy and stability of bone age assessment.

[0067] It is worth noting that the examination record of the child currently being examined is also stored in the historical data set.

[0068] The present invention also includes a child bone age image acquisition and positioning device, which is used to implement the above-mentioned child bone age image acquisition and detection method, and is characterized by comprising:

[0069] A memory, used to set a standard inspection cycle d, where d represents the number of days, collect the bone age image of the child every inspection cycle d, and obtain the inspection record written by the medical staff based on the bone age image;

[0070] A text processor, used for inputting the inspection record into a preset BERT model to extract keywords of bone age characteristics;

[0071] A graphics processor is used to build an auxiliary detection model based on a convolutional neural network and a recurrent neural network, train the model through a labeled historical data set, extract the child's age and keywords from the examination record, convert the child's age and keywords into vectors and input them into the trained auxiliary inspection model, and output the bone age difference between the child's actual age and bone age;

[0072] The central processing unit is used to count the bone age difference sequence of the child every n examination cycles d, and calculate the mean value Y of the bone age difference sequence a , the proportion of bone age differences exceeding the preset threshold d in the sequence P, and draw a line graph of the bone age difference over time to obtain the slope k between any two adjacent coordinate points; calculate the inspection cycle adjustment coefficient γ according to the following formula:

[0073]

[0074] Among them, i∈[1,n-1], k i It represents the i-th slope value in the line graph. According to the calculated inspection cycle adjustment coefficient γ, the number of days of the next n inspection cycles is adjusted to d'=d / γ.

[0075] The above is a detailed description of an embodiment of the present invention, but the content is only a preferred embodiment of the present invention and cannot be considered to limit the scope of implementation of the present invention. All equivalent changes and improvements made within the scope of the present invention should still fall within the scope of the patent coverage of the present invention.

Claims

1. A method for collecting and detecting children's bone age images, characterized in that: The following steps are involved: A standard inspection cycle d is set, where d represents the number of days, and a bone age image of the child is collected every inspection cycle d, and an inspection record written by a medical staff based on the bone age image is obtained; Inputting the inspection record into a preset BERT model to extract keywords of bone age characteristics; An auxiliary detection model is constructed based on a convolutional neural network and a recurrent neural network, the model is trained through annotated historical data sets, the child's age and keywords in the examination record are extracted, the child's age and keywords are converted into vectors and input into the trained auxiliary inspection model, and the bone age difference between the child's actual age and bone age is output; Every n examination cycles d, the bone age difference sequence of the children is counted, and the mean value Y of the bone age difference sequence is calculated a , the proportion P of the bone age difference that exceeds the preset threshold d in the sequence, and a line graph of the bone age difference changing over time is drawn to obtain the slope k between any two adjacent coordinate points; The inspection cycle adjustment factor γ is calculated according to the following formula: Among them, i∈[1,n-1], k i It represents the i-th slope value in the line graph. According to the calculated inspection cycle adjustment coefficient γ, the number of days of the next n inspection cycles is adjusted to d'=d / γ.

2. A method for collecting and detecting children's bone age images according to claim 1, characterized in that: The indicators corresponding to the bone age difference sequence are counted every n inspection cycles, and the indicators include the mean Y a , proportion P and slope k, and calculate the number of days of a single inspection cycle of the next n inspection cycles based on the indicators and formula.

3. A method for collecting and detecting children's bone age images according to claim 1, characterized in that: The process of extracting keywords for bone age characteristics is: The annotated standard bone age feature dictionary is divided into a training set and a validation set. The training set is mapped into a vector and input into the BERT model after encoding. The content features in the text are extracted. The BERT model is back-propagated and updated through the validation set until convergence. The examination records of the children to be tested are encoded and input into the BERT model. The examination records include the actual age of the child and a bone age feature description text. The bone age features in the bone age feature description text are extracted, clustered to generate pending keywords, and the pending keywords are used to back-propagate and update the model. The pending keywords are obtained after repeated updates for several times and marked as keywords.

4. A method for collecting and detecting children's bone age images according to claim 1, characterized in that: The training process of the auxiliary diagnosis model is as follows: The historical data set includes standard bone age feature texts corresponding to children of different ages, and non-standard bone age feature texts of children of different ages marked in historical examination records. All bone age feature texts are sorted according to age to generate a time series of bone age features. The matrix formed by the time series of bone age features is regarded as an image and is input into the convolutional neural network in sequence. The convolutional neural network includes several convolutional neural layers, each convolutional neural layer includes several filters. Whenever the matrix image passes through a convolutional neural layer, the filter will slide and scan the matrix and generate a multidimensional array, reduce the size of the matrix image, and increase the dimension of the matrix image. The increased dimension is the same as the number of filters in the convolutional neural network. Each convolutional neural layer reduces the image size and increases the dimension of the matrix image in turn. The multidimensional array finally generated by the convolutional neural network is flattened and converted into a vector, and the vector is used as the input of the subsequent recurrent neural network.

5. A method for collecting and detecting children's bone age images according to claim 4, characterized in that: The training process of the auxiliary diagnosis model also includes: The dimension of the matrix image output by the convolutional neural network remains unchanged, and the matrix image is flattened in the time dimension and eigenvector. The final multidimensional arrays of all convolutional neural layers generate a new matrix, and the new matrix vector is used as the input of the recurrent neural network LSTM. The flattened vector is used as the time dimension of the LSTM. The new matrix is ​​trained by the time dimension and eigenvector, and the corresponding relationship between bone age characteristics and children's age is output.

6. A method for collecting and detecting children's bone age images according to claim 5, characterized in that: The recurrent neural network also includes several additional Transformer layers. All Transformer layers except the last Transformer layer adopt a return sequence mode and generate vector sequences corresponding to continuous iterations of the recurrent neural network. The vector sequences are grouped in time series as inputs of the last Transformer layer.

7. A method for collecting and detecting children's bone age images according to claim 4, characterized in that: The examination record of the child currently being examined is also stored in the historical data set.

8. A child bone age image acquisition and positioning device, used to implement a child bone age image acquisition and detection method according to any one of claims 1 to 7, characterized in that: include: A memory, used to set a standard inspection cycle d, where d represents the number of days, collect the bone age image of the child every inspection cycle d, and obtain the inspection record written by the medical staff based on the bone age image; A text processor, used for inputting the inspection record into a preset BERT model to extract keywords of bone age characteristics; A graphics processor is used to build an auxiliary detection model based on a convolutional neural network and a recurrent neural network, train the model through a labeled historical data set, extract the child's age and keywords from the examination record, convert the child's age and keywords into vectors and input them into the trained auxiliary inspection model, and output the bone age difference between the child's actual age and bone age; The central processing unit is used to count the bone age difference sequence of the child every n examination cycles d, and calculate the mean value Y of the bone age difference sequence a , the proportion P of the bone age difference that exceeds the preset threshold d in the sequence, and a line graph of the bone age difference changing over time is drawn to obtain the slope k between any two adjacent coordinate points; The inspection cycle adjustment factor γ is calculated according to the following formula: Among them, i∈[1,n-1], k i It represents the i-th slope value in the line graph. According to the calculated inspection cycle adjustment coefficient γ, the number of days of the next n inspection cycles is adjusted to d'=d / γ.

Images