Bone age evaluation classification model training method and device, equipment and storage medium

Abstract

The invention discloses a bone age evaluation classification model training method, device and equipment and a storage medium, and the training method comprises the steps: obtaining a plurality of left palm and phalanx maps, recognizing a plurality of key points in the left palm and phalanx maps, correcting the palm and phalanx maps according to the key points, and obtaining a plurality of palm and phalanx part maps through interception, the metacarpal and phalanx part maps are classified and placed in a plurality of part folders, then the metacarpal and phalanx part maps in the part folders are classified and placed in sub-folders according to classification levels to form a plurality of part training sets, neural network models are set for the part training sets, and hyper-parameters, loss functions, optimizers and output levels of the neural network models are set. And carrying out data enhancement processing on the metacarpal and phalangeal part maps in the part training sets so as to train the part training sets by utilizing the set neural network model. According to the bone age assessment classification model training method, the classification model with high bone age assessment accuracy can be trained, and a large amount of ROI annotation data does not need to be used.

Description

Technical Field

[0001] The present invention relates to the technical field of bone age assessment classification model training, and in particular to a bone age assessment classification model training method, device, equipment and storage medium. Background Art

[0002] Bone age assessment is an important means of assessing children's growth and development. Traditional assessment methods rely on the analysis of X-ray images. Assessment methods include counting method, atlas method, and scoring method. At present, the more commonly used assessment methods in China are GP atlas method and scoring method. Among them, the GP atlas method obtains the bone age value by comparing the difference between the hand X-ray film and the standard atlas, which is easily affected by the experience of the assessor and the accuracy of the assessment result is insufficient. The scoring method is an assessment method that determines the maturity score of a single bone and then summarizes and converts it into bone age through a specific combination method, including: TW series method, percentage counting method, CHN method and "Zhonghua-05" method. The advantages of this assessment method are its intuitiveness and empiricism, and it also has some disadvantages, such as low efficiency, and it is also easily affected by the personal experience and judgment of experts, and there is a limitation of bias in the judgment results.

[0003] With the development of artificial intelligence technology, deep learning has shown great potential in the field of medical image analysis. Some studies have begun to try to use deep learning models to automatically assess bone age, which predicts and analyzes bone maturity by identifying and analyzing specific regions of interest (ROI) in X-ray images, which greatly improves the efficiency of bone age assessment and the consistency of assessment results. However, the bone age assessment method based on ROI prediction requires a large amount of labeled data to train the deep learning model, and the data collection and labeling work is cumbersome and costly. Summary of the invention

[0004] The purpose of the present invention is to provide a bone age assessment classification model training method, device, equipment and storage medium, which can train a classification model with high accuracy in bone age assessment without using a large amount of ROI annotation data, which is conducive to promoting the implementation of the bone age assessment classification model and solving the problem of time-consuming, labor-intensive and poor consistency in manual reading of films to assess bone age.

[0005] In order to achieve the above object, the present invention discloses a bone age assessment classification model training method, which comprises:

[0006] Obtain several left hand metacarpal phalangeal images;

[0007] Identify multiple key points of the acquired metacarpophalangeal bone image;

[0008] Align the metacarpophalangeal bone diagram according to the identified key points;

[0009] The straightened metacarpal and phalangeal bone images are intercepted according to the identified key points to obtain a plurality of metacarpal and phalangeal bone part images, and the images are classified and placed in a plurality of part folders;

[0010] According to the classification level of each part folder, multiple subfolders are created in each part folder, and the metacarpal and phalangeal part images in each part folder are classified and placed into corresponding subfolders to form multiple part training sets;

[0011] Setting multiple neural network models corresponding to multiple part training sets, and setting hyperparameters, loss functions and optimizers of each neural network model, wherein the hyperparameters include the number of iterations, the number of batch samples and the image size;

[0012] Randomly perform at least one data enhancement process on the metacarpal and phalangeal part images in each part training set;

[0013] Set the output level of the neural network model according to the classification level of each part folder;

[0014] The set neural network model is used to train the training sets of each part after data enhancement processing to obtain multiple classification models corresponding to the part folders.

[0015] Furthermore, the key points include the current point of the carpal bone, the current point of the distal phalanx III, and the current point of the distal phalanx I.

[0016] The “straightening the metacarpal bone diagram according to the identified key points” includes:

[0017] Determine whether the current point of the distal phalanx I is on the right side of the current point of the carpal bone. If it is not, flip the metacarpal phalanx image. If it is, keep the existing direction of the metacarpal phalanx image.

[0018] Calculate the deviation between the current point of the carpal bone and the standard point of the carpal bone to obtain the translation matrix;

[0019] Calculate the angle between the line between the current point of the carpal bone and the current point of the third distal phalanx and the horizontal axis to obtain a rotation matrix;

[0020] Calculate the ratio of the distance between the current point of the distal phalanx III and the current point of the carpal bone to the standard distance to obtain a scaling matrix;

[0021] The key points of the metacarpophalangeal bone diagram and the metacarpophalangeal bone diagram are aligned using the translation matrix, rotation matrix and scaling matrix.

[0022] Further, each metacarpal and phalangeal bone image captures 13 metacarpal and phalangeal bone part images, and 9 part folders are set, and the metacarpal and phalangeal bone part images include an ulna image, a radius image, a first metacarpal bone image, a third metacarpal bone image, a fifth metacarpal bone image, a first proximal phalanx image, a first distal phalanx image, a third proximal phalanx image, a fifth proximal phalanx image, a third middle phalanx image, a fifth middle phalanx image, a third distal phalanx image, and a fifth distal phalanx image; wherein the third metacarpal bone image and the fifth metacarpal bone image, the third proximal phalanx image and the fifth proximal phalanx image, the third middle phalanx image and the fifth middle phalanx image, the third distal phalanx image and the fifth distal phalanx image are placed in the same part folder.

[0023] Furthermore, there are 9 part folders, the ulna folder is divided into 12 levels, the radius folder is divided into 14 levels, the I proximal phalanx folder is divided into 12 levels, the III V proximal phalanx folder is divided into 12 levels, the III V middle phalanx folder is divided into 12 levels, the I metacarpal folder is divided into 11 levels, the III V metacarpal folder is divided into 10 levels, the I distal phalanx folder is divided into 11 levels, and the III V distal phalanx folder is divided into 11 levels.

[0024] Furthermore, a Resnet-50 neural network is constructed as the neural network model, and the number of iterations of the Resnet-50 neural network is configured to be 20, the number of batch samples is 16, and the image size is 224*224, the cross entropy loss function is set as the loss function of the Resnet-50 neural network, and the Adam optimizer is set as the optimizer algorithm of the Resnet-50 neural network.

[0025] Furthermore, the data enhancement process includes randomly rotating the metacarpophalangeal part map within a set angle, randomly adjusting the brightness, contrast, saturation and hue of the metacarpophalangeal part map, and normalizing and standardizing the pixel values ​​of the metacarpophalangeal part map.

[0026] Furthermore, after the step of “obtaining a plurality of classification models corresponding to the part folders”, the step further includes:

[0027] Collect new metacarpal and phalangeal part maps and create multiple new part folders;

[0028] Use the classification model to classify the corresponding new part folders;

[0029] The accuracy of the corresponding classification model is judged based on the classification results.

[0030] In order to achieve the above object, the present invention discloses a bone age assessment classification model training device, which comprises:

[0031] An acquisition module, used for acquiring a plurality of left hand metacarpal phalangeal images;

[0032] A recognition module, used for recognizing a plurality of key points of the acquired metacarpal and phalangeal bone images;

[0033] A straightening module is used to straighten the metacarpal and phalangeal bone images according to the identified key points;

[0034] A cutting module is used to cut the straightened metacarpal and phalangeal bone images according to the identified key points to obtain a plurality of metacarpal and phalangeal bone part images, and classify and place them in a plurality of part folders;

[0035] A building module is used to build multiple subfolders in each part folder according to the classification level of each part folder, and classify the metacarpal and phalangeal part images in each part folder into corresponding subfolders to form multiple part training sets;

[0036] A first setting module is used to set multiple neural network models corresponding to multiple part training sets, and set hyperparameters, loss functions and optimizers of each neural network model;

[0037] A processing module, used for randomly performing at least one data enhancement process on the metacarpal and phalangeal part images in each part training set;

[0038] The second setting module is used to set the output level of the neural network model according to the classification level of each part folder;

[0039] The training module is used to use the set neural network model to train the training sets of each part after data enhancement processing to obtain multiple classification models corresponding to the part folders.

[0040] In order to achieve the above object, the present invention discloses an electronic device, which includes:

[0041] one or more processors;

[0042] One or more memories are used to store one or more programs. When one or more of the programs are executed by the processor, the processor implements the bone age assessment classification model training method as described above.

[0043] In order to achieve the above-mentioned purpose, the present invention discloses a computer-readable storage medium having a program stored thereon, and when the program is executed by a processor, the bone age assessment classification model training method as described above is implemented.

[0044] In the present application, a neural network model is trained using a position training set formed by metacarpophalangeal position maps to obtain a classification model. First, a number of left-hand metacarpophalangeal images are obtained and multiple key points therein are identified. The metacarpophalangeal images are straightened according to the key points and multiple metacarpophalangeal position maps are obtained by interception. The metacarpophalangeal position maps are classified and placed in multiple position folders. Then, the metacarpophalangeal position maps in each position folder are classified and placed in subfolders according to the classification level to form multiple position training sets. A neural network model is set for these position training sets and its hyperparameters, loss function, optimizer and output level are set. Data enhancement processing is performed on the metacarpophalangeal position maps in the position training set, so that these position training sets are trained using the set neural network model, thereby obtaining a classification model with high accuracy in bone age assessment. There is no need to use a large amount of ROI annotation data, which is conducive to promoting the implementation of the bone age assessment classification model and solving the problem of time-consuming, labor-intensive and poor consistency in manual bone age assessment through film reading. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] Figure 1 This is a flowchart of the bone age assessment classification model training method according to an embodiment of the present invention.

[0046] Figure 2 This is a left hand metacarpophalangeal bone diagram and key points for identification in the bone age assessment classification model training method according to an embodiment of the present invention.

[0047] Figure 3 This is a diagram of the left metacarpophalangeal bones before being straightened in the bone age assessment classification model training method in an embodiment of the present invention.

[0048] Figure 4 This is the left hand metacarpophalangeal bone image after being straightened in the bone age assessment classification model training method according to an embodiment of the present invention.

[0049] Figure 5 This is a schematic diagram of the part folders and subfolders in the bone age assessment classification model training method according to an embodiment of the present invention.

[0050] Figure 6 This is a module diagram of a bone age assessment classification model training device according to an embodiment of the present invention.

[0051] Figure 7 4 is a system diagram of an electronic device according to an embodiment of the present invention. DETAILED DESCRIPTION

[0052] In order to explain the technical content, structural features, achieved objectives and effects of the present invention in detail, the following is a detailed description in conjunction with the implementation methods and the accompanying drawings.

[0053] Embodiment 1

[0054] See also Figures 1 to 5 The present invention discloses a bone age assessment classification model training method, which comprises:

[0055] 101. Obtain several left hand metacarpal phalangeal bone images;

[0056] Further, “obtaining a plurality of left hand metacarpal phalangeal images” includes:

[0057] 1011. Obtain several left hand X-ray images;

[0058] It is understandable that it is necessary to first build a training set required for model training. Therefore, several left-hand X-rays with clear images and reliable quality are obtained from open source databases and tertiary hospitals to build a left-hand X-ray dataset.

[0059] 1012. Take the left hand metacarpal phalangeal bone image from the left hand X-ray image;

[0060] 1013. Perform a mask operation on the intercepted left hand metacarpal phalangeal image;

[0061] 1014. Perform contrast enhancement on the left hand metacarpal phalangeal image;

[0062] 1015. Perform adaptive histogram equalization on the left hand metacarpophalangeal image.

[0063] It can be understood that each left hand X-ray in the left hand X-ray data set is intercepted to retain only the left metacarpal phalanges to obtain a number of left metacarpal phalanges images, and then the above operations 1013 to 1015 are performed to perform data preprocessing on all left metacarpal phalanges images.

[0064] 102. Identify multiple key points of the acquired metacarpal bone image;

[0065] Specifically, in this embodiment, the metacarpal bone posture recognition algorithm is used to identify the following Figure 2 The key points of the X-ray image of the left metacarpophalangeal bone diagram are shown, but are not limited to these.

[0066] 103. Align the metacarpal and phalangeal bones according to the identified key points;

[0067] Furthermore, the key points include the current point of the carpal bone, the current point of the distal phalanx III, and the current point of the distal phalanx I, and “straightening the metacarpal bone diagram according to the identified key points” includes:

[0068] 1031. Determine whether the current point of the distal phalanx I is on the right side of the current point of the carpal bone. If it is not, flip the metacarpal phalanx image. If it is yes, keep the existing direction of the metacarpal phalanx image.

[0069] 1032. Calculate the deviation between the current point of the carpal bone and the standard point of the carpal bone to obtain a translation matrix H1;

[0070] 1033. Calculate the angle between the line between the current point of the carpal bone and the current point of the third distal phalanx and the horizontal axis to obtain a rotation matrix H2;

[0071] 1034. Calculate the ratio of the distance between the current point of the distal phalanx III and the current point of the carpal bone to the standard distance to obtain a scaling matrix H3;

[0072] 1035. Use the translation matrix, rotation matrix and scaling matrix to align the key points of the metacarpophalangeal bone diagram and the metacarpophalangeal bone diagram.

[0073] It is understandable that if Figure 3 As shown, the image registration algorithm is used to register each metacarpophalangeal bone image according to the three key points of the metacarpophalangeal bone image to obtain the translation matrix H1, the rotation matrix H2, and the scaling matrix H3. Then the data of the metacarpophalangeal bone image is multiplied by the translation matrix H1*rotation matrix H2*scaling matrix H3, and the data of each key point in the metacarpophalangeal bone image is also multiplied by the translation matrix H1*rotation matrix H2*scaling matrix H3 to align all the metacarpophalangeal bone images and the key points in the metacarpophalangeal bone images, as shown in FIG. Figure 4 As shown, but not limited to.

[0074] 104. The straightened metacarpal and phalangeal bone images are intercepted according to the identified key points to obtain a plurality of metacarpal and phalangeal bone part images, and the images are classified and placed in a plurality of part folders;

[0075] Furthermore, each metacarpal and phalangeal bone image captures 13 metacarpal and phalangeal bone part images, and 9 part folders are set. The metacarpal and phalangeal bone part images include the ulna image, the radius image, the first metacarpal bone image, the third metacarpal bone image, the fifth metacarpal bone image, the first proximal phalanx image, the first distal phalanx image, the third proximal phalanx image, the fifth proximal phalanx image, the third middle phalanx image, the fifth middle phalanx image, the third distal phalanx image, and the fifth distal phalanx image; among them, the third metacarpal bone image and the fifth metacarpal bone image, the third proximal phalanx image and the fifth proximal phalanx image, the third middle phalanx image and the fifth middle phalanx image, the third distal phalanx image and the fifth distal phalanx image are placed in the same part folder.

[0076] It should be noted that the present invention is based on the TW III-RUS series scoring method which is currently widely used in clinical practice to train the bone age assessment classification model. Therefore, the 13 metacarpal and phalangeal bone part diagrams as described above are intercepted from the metacarpal and phalangeal bone diagram according to the corrected key point information and the distance between adjacent key points, but it is not limited to this.

[0077] 105. Create multiple subfolders in each part folder according to the classification level of each part folder, and classify the metacarpal and phalangeal part images in each part folder into corresponding subfolders to form multiple part training sets;

[0078] Furthermore, there are 9 location folders, such as Figure 5 As shown, the ulna folders are divided into 12 levels, the radius folders are divided into 14 levels, the I proximal phalanx folders are divided into 12 levels, the III V proximal phalanx folders are divided into 12 levels, the III V middle phalanx folders are divided into 12 levels, the I metacarpal folders are divided into 11 levels, the III V metacarpal folders are divided into 10 levels, the I distal phalanx folders are divided into 11 levels, and the III V distal phalanx folders are divided into 11 levels.

[0079] It is understandable that if Figure 5 As shown, the 13 metacarpal and phalangeal part images are classified into 9 part folders, and a corresponding number of subfolders are established for each part folder according to the classification level. Each classification level indicates that the bone corresponds to a certain preset age group. Therefore, according to the TW III-RUS series scoring method, the metacarpal and phalangeal part images of the 9 part folders are classified into corresponding subfolders to form a part training set.

[0080] 106. Setting multiple neural network models corresponding to multiple part training sets, and setting hyperparameters, loss functions and optimizers of each neural network model, wherein the hyperparameters include the number of iterations, the number of batch samples and the image size;

[0081] Furthermore, a Resnet-50 neural network is constructed as a neural network model, and the number of iterations of the Resnet-50 neural network is configured to be 20, the number of batch samples is 16, and the image size is 224*224. The cross entropy loss function is set as the loss function of the Resnet-50 neural network, and the Adam optimizer is set as the optimizer algorithm of the Resnet-50 neural network.

[0082] It can be understood that during the training process, the Resnet-50 neural network iterates a total of 20 times. Each time training is performed, a batch of 16 metacarpal and phalangeal part map data is input into the neural network model, and the size of the metacarpal and phalangeal part map is uniformly scaled to 224*224.

[0083] It should be noted that the cross-entropy loss function is often used as a loss function for a multi-classification model. The image is forward-propagated through the neural network model to calculate the prediction result, and then the loss function is used to calculate the difference between the predicted value and the actual value to obtain the model error. In this embodiment, the cross-entropy loss function is used to calculate the difference between the probability distribution of the predicted metacarpal and phalangeal part map output by the neural network model belonging to each classification level and the classification level to which the metacarpal and phalangeal part map actually belongs. The specific formula is as follows:

[0084]

[0085] Wherein, C is the number of categories. In this embodiment, C is the number of bone age grades of each metacarpal and phalangeal part;

[0086] yc is the nth element in the one-hot encoding vector of the actual label, indicating the true category, with a value of 1, and the rest are 0. In this embodiment, it means that when the metacarpal phalangeal part map actually belongs to a certain classification level, it is 1, and when the metacarpal phalangeal part map actually does not belong to a certain classification level, it is 0;

[0087] y^c is the nth element in the probability vector predicted by the model, indicating the probability that the model predicts that the sample belongs to category n. In this embodiment, it indicates the probability that the neural network model predicts that the metacarpal and phalangeal part map belongs to a certain classification level.

[0088] Specifically, the model parameter gradient is calculated according to the error through the back propagation algorithm, so that the optimizer algorithm updates the weight and bias of the model according to the model parameter gradient and the learning rate, thereby gradually improving the prediction accuracy of the model. In this embodiment, Adam Optimizer is used as the optimizer algorithm of the neural network model, which can automatically adjust the learning rate of the neural network model to adapt to different parameters to achieve faster convergence of the model. As the number of training rounds increases, the Adam optimizer will adjust the learning rate to a cosine function, which is conducive to the stable convergence of the neural network model. For example, in the early stage of training, the learning rate is set to 0.001, which is conducive to the model to quickly find the convergence direction. As the training proceeds to the later stage, the learning rate is gradually adjusted to 0.0005 to avoid oscillation caused by excessive step length, and help the neural network model to converge to a better solution stably, but not limited to this.

[0089] 107. Randomly perform at least one data enhancement process on the metacarpal and phalangeal part images in each part training set;

[0090] Furthermore, the data enhancement processing includes randomly rotating the metacarpophalangeal part map within a set angle, randomly adjusting the brightness, contrast, saturation and hue of the metacarpophalangeal part map, and normalizing and standardizing the pixel values ​​of the metacarpophalangeal part map.

[0091] It should be noted that, according to the training framework, random data enhancement processing is performed on each metacarpophalangeal part map to increase the diversity of training data; for example, any angle between 1 and 15 degrees is selected to rotate the metacarpophalangeal part map.

[0092] Specifically, standardization processing refers to converting the pixel values ​​of the metacarpophalangeal part map into a standard range of 0 to 1 to accelerate the convergence of the neural network model; normalization processing refers to subtracting the average pixel value of all channels from the pixel value of each channel of the metacarpophalangeal part map, and then dividing it by the standard deviation of all channels to make the pixel value of each channel of the metacarpophalangeal part map more evenly distributed.

[0093] 108. Set the output level of the neural network model according to the classification level of each part folder;

[0094] It should be noted that the standards for determining bone age at different metacarpal and phalangeal parts are different, and therefore the levels of classification output of the neural network model for identifying and determining different metacarpal and phalangeal parts are also different. As shown in operation 105, corresponding classification standards are set for the nine neural network models respectively, so that the output level of the neural network model corresponds to the number of levels of bone age determined at each metacarpal and phalangeal part, thereby improving the accuracy of the neural network model classification.

[0095] 109. The set neural network model is used to train the training sets of each part after data enhancement processing to obtain multiple classification models corresponding to the part folders.

[0096] It is understandable that 9 configured neural network models are used to train the metacarpal and phalangeal part images in 9 part folders respectively, so as to train 9 classification models that can identify and judge bone age for different metacarpal and phalangeal parts, but it is not limited to this.

[0097] Furthermore, after “obtaining a plurality of classification models corresponding to the part folders”, it also includes:

[0098] 110. Collect new metacarpal and phalangeal part maps and create multiple new part folders;

[0099] 111. Use the classification model to classify the corresponding new part folders;

[0100] 112. Determine the accuracy of the corresponding classification model based on the classification results.

[0101] It should be noted that, during the testing phase, a new batch of test data sets are needed to perform classification tests on the trained classification model, that is, operations 101 to 104 are executed again, the metacarpophalangeal bone images are collected and key points are identified, the images are aligned and 13 metacarpophalangeal part images are intercepted, and classified into 9 part folders to form a test data set, and finally, the corresponding trained classification model is called for the 9 part folders in the test data set to perform classification and identification of the metacarpophalangeal part images, and the classification level of the metacarpophalangeal part image is compared with the actual level of the metacarpophalangeal part image according to the classification model to judge the accuracy of the classification model, but this is not limited to this.

[0102] In the present application, a neural network model is trained using a position training set formed by metacarpophalangeal position maps to obtain a classification model. First, a number of left-hand metacarpophalangeal images are obtained and multiple key points therein are identified. The metacarpophalangeal images are straightened according to the key points and multiple metacarpophalangeal position maps are obtained by interception. The metacarpophalangeal position maps are classified and placed in multiple position folders. Then, the metacarpophalangeal position maps in each position folder are classified and placed in subfolders according to the classification level to form multiple position training sets. A neural network model is set for these position training sets and its hyperparameters, loss function, optimizer and output level are set. Data enhancement processing is performed on the metacarpophalangeal position maps in the position training set, so that these position training sets are trained using the set neural network model, thereby obtaining a classification model with high accuracy in bone age assessment. There is no need to use a large amount of ROI annotation data, which is conducive to promoting the implementation of the bone age assessment classification model and solving the problem of time-consuming, labor-intensive and poor consistency in manual bone age assessment through film reading.

[0103] Embodiment 2

[0104] See also Figure 1 and Figure 6 The present invention discloses a bone age assessment classification model training device, which comprises:

[0105] An acquisition module 201 is used to acquire a plurality of left hand metacarpal and phalangeal bone images;

[0106] The recognition module 202 is used to recognize a plurality of key points of the acquired metacarpophalangeal bone image;

[0107] A straightening module 203, used to straighten the metacarpophalangeal bone image according to the identified key points;

[0108] The interception module 204 is used to intercept the straightened metacarpal and phalangeal bone images according to the identified key points to obtain a plurality of metacarpal and phalangeal bone part images, and classify and place them in a plurality of part folders;

[0109] Establishing module 205, for establishing a plurality of sub-folders in each part folder according to the classification level of each part folder, and classifying the metacarpal and phalangeal part images in each part folder into corresponding sub-folders to form a plurality of part training sets;

[0110] A first setting module 206 is used to set multiple neural network models corresponding to multiple part training sets, and set hyperparameters, loss functions and optimizers of each neural network model;

[0111] A processing module 207 is used to randomly perform at least one data enhancement process on the metacarpal and phalangeal part images in each part training set;

[0112] The second setting module 208 is used to set the output level of the neural network model according to the classification level of each part folder;

[0113] The training module 209 is used to use the set neural network model to train the training sets of each part after data enhancement processing to obtain multiple classification models corresponding to the part folders.

[0114] Embodiment 3

[0115] See also Figure 1 and Figure 7 , the present invention discloses an electronic device, comprising:

[0116] One or more processors 301;

[0117] One or more memories 302 are used to store one or more programs. When the one or more programs are executed by the processor, the processor implements the bone age assessment classification model training method as described above.

[0118] Embodiment 4

[0119] An embodiment of the present application discloses a computer-readable storage medium having a program stored thereon, which, when executed by a processor, implements the aforementioned bone age assessment classification model training method.

[0120] Embodiment 5

[0121] The embodiment of the present application discloses a computer program product or a computer program, which includes computer instructions stored in a computer-readable storage medium. A processor of an electronic device reads the computer instructions from the computer-readable storage medium, and the processor executes the computer instructions, so that the electronic device executes the above-mentioned bone age assessment classification model training method.

[0122] It should be understood that in the embodiments of the present application, the processor referred to may be a central processing unit (CPU), and the processor may also be other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor may be a microprocessor or the processor may also be any conventional processor, etc.

[0123] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiments can be implemented by hardware related to computer program instructions, and the program can be stored in a computer-readable storage medium. When the program is executed, it can include the processes of the embodiments of the above-mentioned methods. The storage medium can be a disk, an optical disk, a read-only memory (ROM) or a random access memory (RAM).

[0124] The above disclosure is only the preferred embodiment of the present invention, which certainly cannot be used to limit the scope of rights of the present invention. Therefore, equivalent changes made according to the scope of the patent application of the present invention are still within the scope covered by the present invention.

Claims

1. A bone age assessment classification model training method, characterized in that: include: Obtain several left hand metacarpal phalangeal images; Identify multiple key points of the acquired metacarpophalangeal bone image; Align the metacarpophalangeal bone diagram according to the identified key points; The straightened metacarpal and phalangeal bone images are intercepted according to the identified key points to obtain a plurality of metacarpal and phalangeal bone part images, and the images are classified and placed in a plurality of part folders; According to the classification level of each part folder, multiple subfolders are created in each part folder, and the metacarpal and phalangeal part images in each part folder are classified and placed into corresponding subfolders to form multiple part training sets; Setting multiple neural network models corresponding to multiple part training sets, and setting hyperparameters, loss functions and optimizers of each neural network model, wherein the hyperparameters include the number of iterations, the number of batch samples and the image size; Randomly perform at least one data enhancement process on the metacarpal and phalangeal part images in each part training set; Set the output level of the neural network model according to the classification level of each part folder; The set neural network model is used to train the training sets of each part after data enhancement processing to obtain multiple classification models corresponding to the part folders.

2. The bone age assessment classification model training method according to claim 1, characterized in that: The key points include the current point of the carpal bone, the current point of the distal phalanx III, and the current point of the distal phalanx I. The "straightening the metacarpophalangeal bone diagram according to the identified key points" includes: Determine whether the current point of the distal phalanx I is on the right side of the current point of the carpal bone. If it is not, flip the metacarpal phalanx image. If it is, keep the existing direction of the metacarpal phalanx image. Calculate the deviation between the current point of the carpal bone and the standard point of the carpal bone to obtain the translation matrix; Calculate the angle between the line between the current point of the carpal bone and the current point of the third distal phalanx and the horizontal axis to obtain a rotation matrix; Calculate the ratio of the distance between the current point of the distal phalanx III and the current point of the carpal bone to the standard distance to obtain a scaling matrix; The key points of the metacarpophalangeal bone diagram and the metacarpophalangeal bone diagram are aligned using the translation matrix, rotation matrix and scaling matrix.

3. The bone age assessment classification model training method according to claim 1, characterized in that: Each metacarpal and phalangeal bone image captures 13 metacarpal and phalangeal bone part images, and 9 part folders are set, and the metacarpal and phalangeal bone part images include an ulna image, a radius image, a first metacarpal bone image, a third metacarpal bone image, a fifth metacarpal bone image, a first proximal phalanx image, a first distal phalanx image, a third proximal phalanx image, a fifth proximal phalanx image, a third middle phalanx image, a fifth middle phalanx image, a third distal phalanx image, and a fifth distal phalanx image; among them, the third metacarpal bone image and the fifth metacarpal bone image, the third proximal phalanx image and the fifth proximal phalanx image, the third middle phalanx image and the fifth middle phalanx image, the third distal phalanx image and the fifth distal phalanx image are placed in the same part folder.

4. The bone age assessment classification model training method according to claim 1, characterized in that: There are 9 part folders, ulna folders are divided into 12 levels, radius folders are divided into 14 levels, I proximal phalanx folders are divided into 12 levels, III V proximal phalanx folders are divided into 12 levels, III V middle phalanx folders are divided into 12 levels, I metacarpal folders are divided into 11 levels, III V metacarpal folders are divided into 10 levels, I distal phalanx folders are divided into 11 levels, and III V distal phalanx folders are divided into 11 levels.

5. The bone age assessment classification model training method according to claim 1, characterized in that: A Resnet-50 neural network is constructed as the neural network model, and the number of iterations of the Resnet-50 neural network is configured to be 20, the number of batch samples is 16, and the image size is 224*224. The cross entropy loss function is set as the loss function of the Resnet-50 neural network, and the Adam optimizer is set as the optimizer algorithm of the Resnet-50 neural network.

6. The bone age assessment classification model training method according to claim 1, characterized in that: The data enhancement process includes randomly rotating the metacarpal phalangeal part map within a set angle, randomly adjusting the brightness, contrast, saturation and hue of the metacarpal phalangeal part map, and normalizing and standardizing the pixel values ​​of the metacarpal phalangeal part map.

7. The bone age assessment classification model training method according to claim 1, characterized in that: After the step of "obtaining a plurality of classification models corresponding to the part folders", the method further includes: Collect new metacarpal and phalangeal part maps and create multiple new part folders; Use the classification model to classify the corresponding new part folders; The accuracy of the corresponding classification model is judged based on the classification results.

8. A bone age assessment classification model training device, characterized in that: include: An acquisition module, used for acquiring a plurality of left hand metacarpal phalangeal images; A recognition module, used for recognizing a plurality of key points of the acquired metacarpal and phalangeal bone images; A straightening module is used to straighten the metacarpal and phalangeal bone images according to the identified key points; A cutting module is used to cut the straightened metacarpal and phalangeal bone images according to the identified key points to obtain a plurality of metacarpal and phalangeal bone part images, and classify and place them in a plurality of part folders; A building module is used to build multiple subfolders in each part folder according to the classification level of each part folder, and classify the metacarpal and phalangeal part images in each part folder into corresponding subfolders to form multiple part training sets; A first setting module is used to set multiple neural network models corresponding to multiple part training sets, and set hyperparameters, loss functions and optimizers of each neural network model; A processing module, used for randomly performing at least one data enhancement process on the metacarpal and phalangeal part images in each part training set; The second setting module is used to set the output level of the neural network model according to the classification level of each part folder; The training module is used to use the set neural network model to train the training sets of each part after data enhancement processing to obtain multiple classification models corresponding to the part folders.

9. An electronic device, characterized in that: include: one or more processors; One or more memories are used to store one or more programs. When one or more of the programs are executed by the processor, the processor implements the bone age assessment classification model training method as described in any one of claims 1 to 7.

10. A computer-readable storage medium having a program stored thereon, characterized in that: When the program is executed by a processor, the bone age assessment classification model training method as described in any one of claims 1 to 7 is implemented.

Images