Automatic assessment and correction method of hand bone age based on TW3 bone development correlation law
By constructing a maturity sequence of bone association rules and using deep learning networks to evaluate hand bone bone age, the problem of insufficient accuracy of the automated method of hand bone bone age assessment in the prior art is solved, and more efficient and accurate hand bone age assessment is achieved.
Patent Information
- Application Number
- CN202210246025.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-03-14
- Publication Date
- 2024-11-08
- Estimated Expiration
- 2042-03-14
AI Technical Summary
The existing TW3 method automated method for bone bone age evaluation of hand bones based on TW3 method has error problems caused by lack of data, small maturity differences and manual labeling errors, which reduces the accuracy of bone age evaluation.
Construct the array of regular maturity sequences of bone association, and combine the Faster R-CNN target detection network and ResNet-50 convolutional network to calculate the similarity and deviation of predicted maturity sequences and regular maturity sequences through bone region detection and maturity classification. , make corrections to improve evaluation accuracy.
It effectively improves the accuracy and efficiency of bone age assessment, reduces the dependence on expert knowledge, and avoids the influence of singular value errors in predicting maturity sequences.
Smart Images

Figure CN114822837B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of medical image analysis, and in particular to a method for automatically assessing and correcting hand bone age based on TW3 bone development correlation rules. Background Art
[0002] Bone age assessment based on hand bone X-ray images is often used to determine the growth and development of children. There are currently two main manual assessment methods: GP method and TW method. The TW method assesses the maturity level of specific bones in the hand, converts the bone maturity into scores, adds up the scores of all bones to obtain a total maturity score, and finally calculates the bone age through the total maturity score. The TW method has been revised many times, and the latest TW3 method has been widely used. Compared with the GP method, although the evaluation process of the TW3 method is more complicated and time-consuming, it is robust to incorrect predictions in certain bones, so this method is more accurate and reproducible.
[0003] In recent years, deep learning technology based on neural networks has developed rapidly, and deep learning technology has achieved advanced results in some medical image applications. The hand bone age assessment task is a typical object detection and classification problem that can benefit from deep learning. Therefore, neural networks and their variants are increasingly used to automate bone age assessment and have achieved excellent assessment results.
[0004] Bone age assessment based on the TW3 method first requires the use of an object detection network to segment different bones, and then a convolutional neural network to assess the maturity of each bone. However, the current assessment results are not ideal. Due to the lack of data, small differences between different maturity levels, and errors in manual annotation, the maturity assessment errors of some bones in some hand bone images are large, which seriously reduces the accuracy of bone age assessment. Summary of the invention
[0005] In order to overcome the shortcomings of the prior art, the present invention provides a method for automatic assessment and correction of hand bone age based on the TW3 bone development association law, so as to correct the predicted maturity sequence through the bone development association law in the TW3 method, thereby improving the accuracy of the automatic assessment method.
[0006] In order to achieve the above-mentioned purpose, the present invention adopts the following technical scheme:
[0007] The automatic assessment and correction method of hand bone age based on TW3 bone development correlation law of the present invention is characterized by following the steps:
[0008] Step 1: Construct the bone association rule maturity sequence array S = {S1, S2, S3, ..., S i ,...,S n}, where Si represents the i-th bone association regularity maturity sequence, n is the total number of bone association regularity maturity sequences; and Represents the i-th bone association law maturity sequence S i The maturity of the kth skeletal region in ; and k∈[1,K], K represents the type of bone region;
[0009] Step 2: Construct an automatic bone age assessment network based on the TW3 method, including: bone region detection network and maturity classification network;
[0010] The skeleton region detection network uses the Faster R-CNN target detection network;
[0011] The maturity classification network uses a ResNet-50 convolutional network;
[0012] Step 2.1: Obtain hand bone X-ray images, and mark the bone regions of the hand bone X-ray images, and use a batch of hand bone X-ray image data and the marked bone region data as a training data set;
[0013] Step 2.2: offline training the skeleton region detection network using the training data set to obtain a trained skeleton region detection network;
[0014] Step 2.3: The labeled bone regions are cropped according to the types of bone regions K to obtain K types of bone region images, and corresponding maturity labels are set. The same type of bone region images and their corresponding maturity labels are used as a bone region dataset, thereby constructing K types of bone region datasets;
[0015] Step 2.4: Use K kinds of bone region data sets to train the maturity classification network offline respectively, and obtain K kinds of maturity classification models;
[0016] Step 3: The hand bone age assessment network is constructed by the trained bone region detection network and K maturity classification models, and is used to assess the bone age of the hand bone X-ray images to be detected to obtain the predicted maturity sequence. in, represents the predicted maturity of the kth skeletal region;
[0017] Step 4: Calculate the predicted maturity sequence using formula (1) and the i-th regular maturity sequence S i , thus obtaining the similarity score array SCORE={score1, score2,..., score i ,...,score n}:
[0018]
[0019] In formula (1), score i Represents the predicted maturity sequence and the i-th regular maturity sequence S i Similarity score of ;
[0020] Step 5: Based on the minimum score in the similarity score array SCORE min , find the minimum score from the regular maturity sequence S min The corresponding similarity rule maturity sequence S min ;
[0021] Step 6: Calculate the predicted maturity sequence using formula (2) and similar regularity maturity sequence S min The maturity deviation of each bone region in , thus obtaining the deviation array D = {d1, d2, ..., d k ,...,d K};
[0022]
[0023] In formula (2), d k represents the predicted maturity sequence of the kth skeletal region and similar regularity maturity sequence S min The deviation Represents similar regularity maturity sequence S min The maturity corresponding to the kth skeletal region of the sequence;
[0024] Step 7: Select the deviations in the deviation array D that are greater than the threshold θ to obtain the large deviation array D * , by the large deviation array D * The subscript corresponding to each deviation in the deviation array D is obtained to obtain the correction area number array;
[0025] Step 8: Correct the predicted maturity sequence according to the corrected region number array The corresponding subscripts in the table correspond to the predicted maturity of the bone region, thus obtaining the corrected predicted maturity sequence
[0026] Step 9: Sequence the revised predicted maturity according to the TW3 scoring method It is converted into a score sequence, and all the scores in the score sequence are added together to obtain the total maturity score, thereby predicting the bone age of the hand bone X-ray image to be tested.
[0027] Compared with the prior art, the beneficial effects of the present invention are as follows:
[0028] 1. The present invention uses a convolutional neural network to automate the bone age assessment process of the TW3 method. First, a target detection network is used to locate the bone area used by the TW3 method, and then an image classification network is used to classify the maturity of the located bone area, thereby avoiding the reliance of bone age assessment on expert knowledge, thereby effectively improving the efficiency of bone age assessment.
[0029] 2. The present invention uses for the first time the hand bone development correlation rules summarized in TW3 to correct the results of the deep learning method for automatic assessment of hand bone maturity, avoiding the negative impact of singular values in the predicted maturity sequence on the bone age assessment error, thereby effectively improving the accuracy of bone age assessment. BRIEF DESCRIPTION OF THE DRAWINGS
[0030] Figure 1 A diagram of the correlation between the development of hand bones used in the present invention;
[0031] Figure 2 It is a schematic diagram of the process of the present invention;
[0032] Figure 3 This is the network diagram for automatic bone age assessment of the present invention. DETAILED DESCRIPTION
[0033] In this embodiment, a hand bone age automatic assessment and correction method based on the TW3 bone development correlation law is used to correct the bone age automatic assessment result using the TW3 bone development correlation law, which effectively improves the assessment accuracy. Specifically, Figure 2 As shown, the steps are as follows:
[0034] Step 1: The TW3 bone age scoring method summarizes the developmental correlation of different bones in the hand, such as Figure 1 As shown. Figure 1 Construct the bone association law maturity sequence array S = {S1, S2, S3, ..., S i ,...,S n}, where S i represents the i-th bone-associated regular maturity sequence, n is the total number of regular maturity sequences; and Represents the i-th bone association law maturity sequence S i The maturity of the kth skeletal region in ; and k∈[1,K], K represents the type of bone region;
[0035] Step 2: Construct an automatic bone age assessment network based on the TW3 method, such as Figure 3 As shown, it includes: bone region detection network, maturity classification network;
[0036] The skeleton region detection network uses the Faster R-CNN target detection network;
[0037] The maturity classification network uses the ResNet-50 convolutional network;
[0038] Step 2.1: Obtain hand bone X-ray images, and mark the bone regions of the hand bone X-ray images, and use a batch of hand bone X-ray image data and the marked bone region data as a training data set;
[0039] Step 2.2: Use the training data set to offline train the skeleton region detection network to obtain a trained skeleton region detection network;
[0040] Step 2.3: The labeled bone regions are cropped according to the types of bone regions K to obtain K types of bone region images, and corresponding maturity labels are set. The same type of bone region images and their corresponding maturity labels are used as a bone region dataset, thereby constructing K types of bone region datasets;
[0041] Step 2.4: Use K kinds of bone region data sets to train the maturity classification network offline respectively, and obtain K kinds of maturity classification models;
[0042] Step 3: The hand bone age assessment network is constructed by the trained bone region detection network and K maturity classification models, and is used to assess the bone age of the hand bone X-ray images to be detected to obtain the predicted maturity sequence. in, represents the predicted maturity of the kth skeletal region;
[0043] The TW3 method uses 13 different bone regions, so K = 13. The TW3 method assesses the maturity of each region.
[0044] Step 4: Calculate the predicted maturity sequence using formula (1) and the i-th regular maturity sequence S i , thus obtaining the similarity score array SCORE={score1, score2,..., score i ,...,score n}:
[0045]
[0046] In formula (1), score i Represents the predicted maturity sequence and the i-th regular maturity sequence S i Similarity score of ;
[0047] In this embodiment, different maturity levels are represented in the form of integer numbers, and similar matching maturity sequences are selected by calculating the distance between the predicted maturity sequence and the regular maturity sequence.
[0048] Step 5: Based on the minimum score in the similarity score array SCORE min , find the minimum score from the regular maturity sequence S min The corresponding similarity rule maturity sequence S min ;
[0049] Step 6: Calculate the predicted maturity sequence using formula (2) and similar regularity maturity sequence S min The maturity deviation of each bone region in , thus obtaining the deviation array D = {d1, d2, ..., d k ,...,d K};
[0050]
[0051] In formula (2), d k represents the predicted maturity sequence of the kth skeletal region and similar regularity maturity sequence S min The deviation Represents similar regularity maturity sequence S min The maturity corresponding to the kth skeletal region of the sequence;
[0052] Step 7: Select the deviations in the deviation array D that are greater than the threshold θ to obtain the large deviation array D * , by the large deviation array D * The subscript corresponding to each deviation in the deviation array D is obtained to obtain the correction area number array;
[0053] Due to individual differences, the law of hand bone development association is only an average reflection of all individuals. Therefore, in this embodiment, only the singular values in the predicted maturity sequence are modified, and the predicted maturity with a small difference from the regular maturity sequence is retained. While retaining individual differences, the influence of singular values on bone age assessment is also avoided.
[0054] Step 8: Correct the predicted maturity sequence based on the corrected region number array The corresponding subscripts in the table correspond to the predicted maturity of the bone region, thus obtaining the corrected predicted maturity sequence
[0055] Step 9: Sequence the revised predicted maturity according to the TW3 scoring method It is converted into a score sequence, and all the scores in the score sequence are added together to obtain the total maturity score, thereby predicting the bone age of the hand bone X-ray image to be tested.
Claims
1. A method for automatically assessing and correcting hand bone age based on the TW3 bone development correlation law, characterized by: Proceed as follows: Step 1: Construct the bone association rule maturity sequence array S = {S1, S2, S3, ..., S i ,...,S n }, where S i represents the i-th bone association regularity maturity sequence, n is the total number of bone association regularity maturity sequences; and Represents the i-th bone association law maturity sequence S i The maturity of the kth skeletal region in ; and k∈[1,K], K represents the type of bone region; Step 2: Construct an automatic bone age assessment network based on the TW3 method, including: bone region detection network and maturity classification network; The skeleton region detection network uses the Faster R-CNN target detection network; The maturity classification network uses a ResNet-50 convolutional network; Step 2.1: Obtain hand bone X-ray images, and mark the bone regions of the hand bone X-ray images, and use a batch of hand bone X-ray image data and the marked bone region data as a training data set; Step 2.2: offline training the skeleton region detection network using the training data set to obtain a trained skeleton region detection network; Step 2.3: The labeled bone regions are cropped according to the types of bone regions K to obtain K types of bone region images, and corresponding maturity labels are set. The same type of bone region images and their corresponding maturity labels are used as a bone region dataset, thereby constructing K types of bone region datasets; Step 2.4: Use K kinds of bone region data sets to train the maturity classification network offline respectively, and obtain K kinds of maturity classification models; Step 3: The hand bone age assessment network is constructed by the trained bone region detection network and K maturity classification models, and is used to assess the bone age of the hand bone X-ray images to be detected to obtain the predicted maturity sequence. in, represents the predicted maturity of the kth skeletal region; Step 4: Calculate the predicted maturity sequence using formula (1) and the i-th regular maturity sequence S i , thus obtaining the similarity score array SCORE={score1, score2,..., score i ,...,score n }: In formula (1), score i Represents the predicted maturity sequence and the i-th regular maturity sequence S i Similarity score of Step 5: Based on the minimum score in the similarity score array SCORE min , find the minimum score from the regular maturity sequence S min The corresponding similarity rule maturity sequence S min ; Step 6: Calculate the predicted maturity sequence using formula (2) and similar regularity maturity sequence S min The maturity deviation of each bone region in , thus obtaining the deviation array D = {d1, d2, ..., d k ,...,d K }; In formula (2), d k represents the predicted maturity sequence of the kth skeletal region and similar regularity maturity sequence S min The deviation Represents similar regularity maturity sequence S min The maturity corresponding to the kth skeletal region of the sequence; Step 7: Select the deviations in the deviation array D that are greater than the threshold θ to obtain the large deviation array D * , by the large deviation array D * The subscript corresponding to each deviation in the deviation array D is obtained to obtain the correction area number array; Step 8: Correct the predicted maturity sequence according to the corrected region number array The corresponding subscripts in the table correspond to the predicted maturity of the bone region, thus obtaining the corrected predicted maturity sequence Step 9: Sequence the revised predicted maturity according to the TW3 scoring method It is converted into a score sequence, and all the scores in the score sequence are added together to obtain the total maturity score, thereby predicting the bone age of the hand bone X-ray image to be tested.
