Seed bone grade identification method and device, computer device and storage medium

Abstract

The application relates to a sesamoid grade identification method and device, computer equipment and a storage medium. The method comprises the following steps: acquiring a plurality of medical images to be corrected of a wrist part; the medical images to be corrected contain images of a target sesamoid; inputting each medical image to be corrected into a pre-trained hand position identification model to obtain a hand position identification result corresponding to each medical image to be corrected; determining a correct hand position image in each medical image to be corrected according to the hand position identification result corresponding to each medical image to be corrected; and obtaining a sesamoid grade classification result of the target sesamoid according to the correct hand position image and a pre-trained sesamoid grade classification model. The method can no longer mainly rely on the personal experience of doctors accumulated for a long time to determine the sesamoid grade, and the efficiency of sesamoid grade identification is improved.

Description

Technical Field

[0001] This application relates to the field of computer software technology, and in particular to a method, apparatus, computer device, and storage medium for identifying seed bone grades. Background Technology

[0002] Adolescence is the transitional period from childhood to adulthood. It is a crucial period of rapid growth and development, and the second peak of growth and development after infancy. Therefore, adolescence is an important stage in the process of growth and development. Zhang Shaoyan et al., through analyzing the relationship between the appearance of specific bone maturity indicators in the wrist of Chinese children and the stage of puberty, discovered a corresponding relationship between the age of appearance of specific maturity indicators of the sesamoid bones in the wrist of both boys and girls and the stage of puberty. Therefore, using the Chinese method for assessing the bone age of the radius, ulna, and short bones (RC), the puberty stage can be classified through sesamoid bone grading, thus providing better guidance for the growth and development of adolescents and offering effective decision-making basis for clinical diagnosis by doctors.

[0003] In practice, the determination of sesamoid bone grading relies primarily on the physician's personal experience. However, before determining sesamoid bone grading, physicians require extensive study and practice, making it difficult to master. Therefore, this method of sesamoid bone grading is inefficient. Summary of the Invention

[0004] Therefore, it is necessary to provide a method, apparatus, computer equipment, and storage medium for identifying sesamoid bone grades that can improve efficiency in addressing the aforementioned technical problems.

[0005] A method for identifying sesamoid bone grade, the method comprising:

[0006] Acquire multiple medical images of the wrist to be corrected; the medical images to be corrected include images of the target sesamoid bone;

[0007] Each of the medical images to be corrected is input into a pre-trained hand position recognition model to obtain the hand position recognition result corresponding to each of the medical images to be corrected.

[0008] Based on the hand position recognition results corresponding to each of the medical images to be corrected, determine the positive hand position image in each of the medical images to be corrected;

[0009] Based on the forehand image and the pre-trained sesamoid bone classification model, the sesamoid bone classification result of the target sesamoid bone is obtained.

[0010] In one embodiment, the training process of the sesamoid bone classification model includes:

[0011] Obtain a first training dataset and a first validation dataset; the first training dataset includes a first sample image and a first sample-level classification result of the sesamoid bones contained in the first sample image; the first validation dataset includes a second sample image and a second sample-level classification result of the sesamoid bones contained in the second sample image; both the first sample image and the second sample image are medical images of the wrist;

[0012] The preset target neural network is trained using the first training dataset to obtain multiple candidate sesamoid bone classification models.

[0013] The candidate sesamoid bone classification models are validated using the first validation dataset to obtain the accuracy of each candidate sesamoid bone classification model.

[0014] Based on the accuracy of each of the candidate sesamoid bone grading models, the sesamoid bone grading model is determined from among the candidate sesamoid bone grading models.

[0015] In one embodiment, the step of training a preset target neural network using the first training dataset to obtain multiple candidate seed bone level classification models includes:

[0016] The first training dataset is divided into a predetermined number of subsets;

[0017] According to the preset polling order, the second validation dataset is selected from the first training dataset, and the remaining subset is used as the second training dataset.

[0018] The second training dataset and the second validation dataset are respectively input into the target neural network to obtain the first sesamoid bone level classification prediction result corresponding to the second training dataset and the second sesamoid bone level classification prediction result corresponding to the second validation dataset;

[0019] Based on the first sesamoid bone classification prediction result and the first sample classification result corresponding to the second training dataset, determine and record the first target loss function; based on the second sesamoid bone classification prediction result and the first sample classification result corresponding to the second validation dataset, determine and record the second target loss function.

[0020] Based on the first target loss function, the parameters of the target neural network are updated, and the process returns to the step of determining the second validation dataset in the first training dataset according to a preset polling order, and using the remaining subset as the second training dataset.

[0021] If the first target loss function corresponding to the target neural network after this update is less than the first target loss function corresponding to the target neural network before this update, then save the target neural network after this update.

[0022] Based on the recorded first and second target loss functions, it is determined whether the preset iteration stopping condition has been met. If the preset iteration stopping condition is met, each stored target neural network is used as a candidate seed bone level classification model.

[0023] In one embodiment, determining whether a preset iteration stopping condition has been met based on the recorded first target loss function and second target loss function includes:

[0024] Determine whether the first target loss function corresponding to the target neural network after this update is less than the first target loss function corresponding to the target neural network before this update, and whether the second target loss function corresponding to the target neural network after this update is greater than or equal to the second target loss function corresponding to the target neural network before this update;

[0025] If so, then it is determined that the target neural network after this update does not meet the preset fitting conditions;

[0026] If the number of times the updated target neural network fails to meet the preset fitting conditions exceeds a preset number, the currently stored target neural network will be used as a candidate seed bone level classification model.

[0027] In one embodiment, determining and recording the first target loss function based on the first sesamoid bone level classification prediction result and the first sample level classification result corresponding to the second training dataset includes:

[0028] Based on the first sesamoid bone classification prediction result and the first sample classification result corresponding to the second training dataset, a soft loss function is determined;

[0029] The center loss function is determined based on the first sesamoid bone classification prediction result and the first sample classification result corresponding to the second training dataset.

[0030] The soft loss function and the central loss function are weighted and summed to obtain and record the first target loss function.

[0031] In one embodiment, acquiring multiple medical images of the wrist to be corrected includes:

[0032] Acquire medical images of the wrist;

[0033] Adjust the medical image to a preset size;

[0034] The preset-size image is rotated by multiple preset angles to obtain multiple rotated images, and the preset-size image and the multiple rotated images are determined as the medical images to be corrected.

[0035] In one embodiment, obtaining the sesamoid bone classification result of the target sesamoid bone based on the forehand image and a pre-trained sesamoid bone classification model includes:

[0036] The image of the proximal hand position is input into a pre-trained epiphyseal localization model to obtain the coordinate information of the first proximal phalanx in the image of the proximal hand position; the first proximal phalanx is the proximal phalanx of the thumb.

[0037] Based on the positional relationship between the first proximal phalanx and the sesamoid bone, and the coordinate information of the first proximal phalanx, the coordinate information of the target sesamoid bone in the orthopedic image is obtained;

[0038] Based on the coordinate information of the target sesamoid bone, the target sesamoid bone image is determined in the forehand position image;

[0039] The target sesamoid image is input into the pre-trained sesamoid classification model to obtain the sesamoid classification result of the target sesamoid.

[0040] A sesamoid bone grading identification device, the device comprising:

[0041] The first acquisition module is used to acquire multiple medical images of the wrist to be corrected; the medical images to be corrected include images of the target sesamoid bone.

[0042] The hand position recognition module is used to input each of the medical images to be corrected into a pre-trained hand position recognition model to obtain the hand position recognition result corresponding to each of the medical images to be corrected.

[0043] The first determining module is used to determine the positive hand position image in each of the medical images to be corrected based on the hand position recognition results corresponding to each of the medical images to be corrected;

[0044] The sesamoid bone classification module is used to obtain the sesamoid bone classification result of the target sesamoid bone based on the forehand image and the pre-trained sesamoid bone classification model.

[0045] In one embodiment, the device further includes:

[0046] The second acquisition module is used to acquire a first training dataset and a first validation dataset; the first training dataset includes a first sample image and a first sample level classification result of the sesamoid bones contained in the first sample image; the first validation dataset includes a second sample image and a second sample level classification result of the sesamoid bones contained in the second sample image; both the first sample image and the second sample image are medical images of the wrist.

[0047] The training module is used to train a preset target neural network using the first training dataset to obtain multiple candidate seed bone level classification models.

[0048] The verification module is used to verify each of the candidate sesamoid bone level classification models using the first verification dataset, and to obtain the accuracy of each of the candidate sesamoid bone level classification models.

[0049] The second determining module is used to determine the sesamoid bone classification model from among the candidate sesamoid bone classification models based on the accuracy of each candidate sesamoid bone classification model.

[0050] In one embodiment, the training module is specifically used for:

[0051] The first training dataset is divided into a predetermined number of subsets;

[0052] According to the preset polling order, the second validation dataset is selected from the first training dataset, and the remaining subset is used as the second training dataset.

[0053] The second training dataset and the second validation dataset are respectively input into the target neural network to obtain the first sesamoid bone level classification prediction result corresponding to the second training dataset and the second sesamoid bone level classification prediction result corresponding to the second validation dataset;

[0054] Based on the first sesamoid bone classification prediction result and the first sample classification result corresponding to the second training dataset, determine and record the first target loss function; based on the second sesamoid bone classification prediction result and the first sample classification result corresponding to the second validation dataset, determine and record the second target loss function.

[0055] Based on the first target loss function, the parameters of the target neural network are updated, and the process returns to the step of determining the second validation dataset in the first training dataset according to a preset polling order, and using the remaining subset as the second training dataset.

[0056] If the first target loss function corresponding to the target neural network after this update is less than the first target loss function corresponding to the target neural network before this update, then save the target neural network after this update.

[0057] Based on the recorded first and second target loss functions, it is determined whether the preset iteration stopping condition has been met. If the preset iteration stopping condition is met, each stored target neural network is used as a candidate seed bone level classification model.

[0058] In one embodiment, the training module is specifically used for:

[0059] Determine whether the first target loss function corresponding to the target neural network after this update is less than the first target loss function corresponding to the target neural network before this update, and whether the second target loss function corresponding to the target neural network after this update is greater than or equal to the second target loss function corresponding to the target neural network before this update;

[0060] If so, then it is determined that the target neural network after this update does not meet the preset fitting conditions;

[0061] If the number of times the updated target neural network fails to meet the preset fitting conditions exceeds a preset number, the currently stored target neural network will be used as a candidate seed bone level classification model.

[0062] In one embodiment, the training module is specifically used for:

[0063] Based on the first sesamoid bone classification prediction result and the first sample classification result corresponding to the second training dataset, a soft loss function is determined;

[0064] The center loss function is determined based on the first sesamoid bone classification prediction result and the first sample classification result corresponding to the second training dataset.

[0065] The soft loss function and the central loss function are weighted and summed to obtain and record the first target loss function.

[0066] In one embodiment, the first acquisition module is specifically used for:

[0067] Acquire medical images of the wrist;

[0068] Adjust the medical image to a preset size;

[0069] The preset-size image is rotated by multiple preset angles to obtain multiple rotated images, and the preset-size image and the multiple rotated images are determined as the medical images to be corrected.

[0070] In one embodiment, the sesamoid bone grading module is specifically used for:

[0071] The image of the proximal hand position is input into a pre-trained epiphyseal localization model to obtain the coordinate information of the first proximal phalanx in the image of the proximal hand position; the first proximal phalanx is the proximal phalanx of the thumb.

[0072] Based on the positional relationship between the first proximal phalanx and the sesamoid bone, and the coordinate information of the first proximal phalanx, the coordinate information of the target sesamoid bone in the orthopedic image is obtained;

[0073] Based on the coordinate information of the target sesamoid bone, the target sesamoid bone image is determined in the forehand position image;

[0074] The target sesamoid image is input into the pre-trained sesamoid classification model to obtain the sesamoid classification result of the target sesamoid.

[0075] A computer device includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program to perform the steps described in the first aspect above.

[0076] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, performs the steps described in the first aspect above.

[0077] The aforementioned sesamoid bone grading identification method, device, computer equipment, and storage medium acquire multiple medical images of the wrist containing the target sesamoid bone; input each medical image to be grading into a pre-trained hand position recognition model to obtain the hand position recognition result corresponding to each medical image to be grading; determine the orthogonal hand position image among the medical images to be grading based on the hand position recognition result corresponding to each medical image to be grading; and obtain the sesamoid bone grading classification result of the target sesamoid bone based on the orthogonal hand position image and the pre-trained sesamoid bone grading classification model. This method no longer relies primarily on the personal experience accumulated by doctors over a long period of time to determine the sesamoid bone grading, thus improving the efficiency of sesamoid bone grading identification. Attached Figure Description

[0078] Figure 1 This is a flowchart illustrating a method for identifying sesamoid bone grades in one embodiment;

[0079] Figure 2 This is a flowchart illustrating the training process of a sesamoid bone grading classification model in one embodiment.

[0080] Figure 3 This is a flowchart illustrating the steps of training a preset target neural network in one embodiment.

[0081] Figure 4 This is a flowchart illustrating the process of determining whether a preset iteration stop condition has been met in one embodiment.

[0082] Figure 5This is a flowchart illustrating the process of determining and recording the first target loss function in one embodiment;

[0083] Figure 6 This is a flowchart illustrating the process of acquiring multiple medical images of the wrist to be corrected in one embodiment;

[0084] Figure 7 This is a flowchart illustrating the process of obtaining the sesamoid bone grading results of the target sesamoid bone in one embodiment;

[0085] Figure 8 This is a schematic diagram of the first proximal phalanx;

[0086] Figure 9 This is a structural block diagram of a sesamoid bone grade identification device in one embodiment;

[0087] Figure 10 This is an internal structural diagram of a computer device in one embodiment. Detailed Implementation

[0088] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0089] In one embodiment, such as Figure 1 As shown, a method for identifying sesamoid bone levels is provided. This embodiment illustrates the application of this method to a terminal. It is understood that this method can also be applied to a server, and to a system including both a terminal and a server, and is implemented through interaction between the terminal and the server. The terminal can be, but is not limited to, various personal computers, laptops, smartphones, tablets, and portable wearable devices. The server can be a standalone server or a server cluster consisting of multiple servers. In this embodiment, the method includes the following steps:

[0090] Step 101: Acquire multiple medical images of the wrist to be calibrated.

[0091] Among them, the medical images to be corrected include images of the target sesamoid bone.

[0092] In this embodiment, the terminal can acquire multiple medical images of the wrist of the target object being examined via medical examination instruments. These medical images represent the internal structure or function of an anatomical region and may include, but are not limited to, one or more of X-ray images, CT images, and MRI images. The medical images to be corrected are those that have not undergone hand position recognition; for example, they may be medical images obtained directly through the medical examination instruments, or medical images obtained through the medical examination instruments that have undergone certain image processing (such as image processing to improve clarity) but have not undergone hand position recognition.

[0093] Step 102: Input each medical image to be corrected into the pre-trained hand position recognition model to obtain the hand position recognition result corresponding to each medical image to be corrected.

[0094] In this embodiment, the terminal can input each medical image to be corrected into a pre-trained hand position recognition model to obtain the hand position recognition result corresponding to each medical image to be corrected. The pre-trained hand position recognition model is mainly used to recognize the normal hand position. The hand position recognition result corresponding to each medical image to be corrected may include the score of each image.

[0095] In one example, the terminal can obtain a hand position recognition model by training an SSD (Single Shot MultiBox Detector) model, with MobileNetV3 as the backbone network. The loss function of the hand position recognition model is a weighted sum of confidence loss and localization loss, which can be expressed as:

[0096]

[0097] Where L(x,c,l,g) is the loss function of the hand position recognition model, L conf For confidence level error, L loc The positional error is represented by α; N is the number of prior boxes that match the ground truth; the α parameter is used to adjust the ratio between the confidence error and the positional error, with a default α = 1; c is the predicted class confidence value; l is the predicted positional value of the bounding box corresponding to the prior box; g is the positional parameter of the ground truth; and x is the indicator parameter.

[0098] Step 103: Based on the hand position recognition results corresponding to each medical image to be corrected, determine the positive hand position image in each medical image to be corrected.

[0099] In this embodiment of the application, the terminal can compare the scores of each medical image to be corrected and determine the image with the highest score as the forehand image.

[0100] In one example, the terminal can determine whether any of the medical images to be calibrated have a score not lower than a preset threshold based on the scores of each image. If an image has a score not lower than the preset threshold, the terminal identifies the image with the highest score as the forehand image. If no image has a score not lower than the preset threshold, the terminal re-executes the step of acquiring multiple medical images of the wrist, for example, the terminal may indicate that the image was rejected.

[0101] Step 104: Based on the forehand image and the pre-trained sesamoid bone classification model, obtain the sesamoid bone classification result of the target sesamoid bone.

[0102] In this embodiment, the terminal can input a forehand image into a pre-trained sesamoid bone classification model to obtain the sesamoid bone classification result of the target sesamoid bone. The sesamoid bone classification result represents the sesamoid bone grade of the target sesamoid bone. There is a correspondence between the sesamoid bone grade and the maturity indicators of the target sesamoid bone. Using the Chinese method for assessing the bone age of the radius, ulna, and short bone (RC), the sesamoid bone grade can be used to classify the adolescent stage, thus better guiding the growth and development of adolescents and providing effective decision-making basis for doctors' clinical diagnosis.

[0103] In one implementation, the terminal can first determine the target sesamoid image in the forehand position image, and then input the target sesamoid image into a pre-trained sesamoid grading classification model to obtain the sesamoid grading classification result of the target sesamoid. The above-described sesamoid grading recognition method further processes the forehand position image to obtain a clearer target sesamoid image, and inputs it into the pre-trained sesamoid grading classification model to improve the accuracy of sesamoid grading recognition.

[0104] In the aforementioned sesamoid bone grading method, the terminal first acquires multiple medical images of the wrist containing the target sesamoid bone, and inputs each image into a pre-trained hand position recognition model to obtain the hand position recognition result corresponding to each image. Then, based on the hand position recognition result, the terminal determines the orthogonal hand position image from among the images. Finally, based on the orthogonal hand position image and the pre-trained sesamoid bone grading classification model, the terminal obtains the sesamoid bone grading classification result for the target sesamoid bone. This sesamoid bone grading method no longer relies primarily on the doctor's long-term accumulated personal experience for sesamoid bone grading, thus improving the efficiency of sesamoid bone grading recognition.

[0105] In one embodiment, such as Figure 2 As shown, the specific training process of the sesamoid bone grading model includes the following steps:

[0106] Step 201: Obtain the first training dataset and the first validation dataset.

[0107] The first training dataset includes a first sample image and the first-sample hierarchical classification results of the sesamoid bones contained in the first sample image. The first validation dataset includes a second sample image and the second-sample hierarchical classification results of the sesamoid bones contained in the second sample image. Both the first and second sample images are medical images of the wrist.

[0108] In this embodiment, the terminal can first select a certain number of orthopedic medical image samples of the wrist, and then manually label the sesamoid bone grades of these orthopedic medical image samples. Finally, the terminal divides these orthopedic medical image samples of the wrist into a first training dataset and a first validation dataset.

[0109] In one example, the first and second sample images are medical images of the same type of wrist, which can both be X-ray images, CT images, or MRI images of the wrist.

[0110] Step 202: Train the preset target neural network using the first training dataset to obtain multiple candidate seed bone level classification models.

[0111] In this embodiment, the terminal can input a first training dataset into a preset target neural network to obtain a sesamoid bone classification prediction result for the sesamoid bones contained in the first sample image (for ease of distinction, this can be referred to as the training sesamoid bone classification prediction result). Then, the terminal can determine a training loss function based on the training sesamoid bone classification prediction result and the first sample classification result. Finally, the terminal trains the preset target neural network according to the training loss function to obtain multiple candidate sesamoid bone classification models.

[0112] In one example, the terminal can employ a K-fold cross-training method, training a pre-defined target neural network using a first training dataset to obtain multiple candidate sesamoid bone classification models. In this way, the aforementioned sesamoid bone classification method, by repeatedly dividing the training dataset and training the target data network multiple times, allows the model to encounter diverse data types, improving the generalization ability of the sesamoid bone classification model and increasing the efficiency of data utilization.

[0113] In another example, the terminal can employ a hierarchical K-fold cross-training method, training a pre-defined target neural network using a first training dataset to obtain multiple candidate sesamoid classification models. This sesamoid classification method not only allows the model to encounter diverse data types through multiple divisions of the training dataset and repeated training of the target data network, improving the generalization ability and data utilization efficiency of the sesamoid classification model, but also ensures that the model learns the characteristics of the data in the validation dataset during training, thereby guaranteeing the reliability of the validation results.

[0114] Step 203: Validate each candidate sesamoid bone level classification model using the first validation dataset to obtain the accuracy of each candidate sesamoid bone level classification model.

[0115] In this embodiment, the terminal can input the first verification dataset into each candidate sesamoid classification model to obtain the sesamoid classification prediction result corresponding to each candidate sesamoid classification model. Then, the terminal can calculate the accuracy of each candidate sesamoid classification model based on the sesamoid classification prediction result corresponding to each candidate sesamoid classification model and the second sample classification result.

[0116] Step 204: Based on the accuracy of each candidate sesamoid bone classification model, determine the sesamoid bone classification model from among the candidate sesamoid bone classification models.

[0117] In this embodiment, the terminal can determine the sesamoid bone classification model from among the candidate sesamoid bone classification models based on their accuracy rates. In one example, the terminal can compare the accuracy rates of each candidate sesamoid bone classification model and determine the candidate sesamoid bone classification model with the highest accuracy rate as the sesamoid bone classification model.

[0118] In the above-mentioned sesamoid bone level recognition method, the terminal first trains the preset target neural network through the first training dataset to obtain multiple candidate sesamoid bone level classification models; the terminal then determines the accuracy of each candidate sesamoid bone level classification model through the first verification dataset, and determines the sesamoid bone level classification model from among the candidate sesamoid bone level classification models based on the accuracy of each candidate sesamoid bone level classification model to ensure the accuracy of the sesamoid bone level recognition model.

[0119] In one embodiment, such as Figure 3 As shown, the specific process of training a preset target neural network using the first training dataset to obtain multiple candidate seed bone classification models includes the following steps:

[0120] Step 301: Divide the first training dataset into a preset number of subsets.

[0121] In this embodiment of the application, the terminal may first randomly sort the first training dataset, and then divide the randomly sorted first training dataset into a preset number of subsets.

[0122] In one example, the terminal can divide the first training dataset into subsets based on the proportion of each category in the first training dataset, ensuring that each subset maintains the proportional relationship between the categories in the first training dataset. Here, each category in the first training dataset represents the training data corresponding to each seed bone level. This ensures that during model training, the model learns the characteristics of the data corresponding to each seed bone level in the validation dataset, thereby guaranteeing the reliability of the validation results.

[0123] Step 302: Select the second validation dataset from the first training dataset according to the preset polling order, and use the remaining subset as the second training dataset.

[0124] In this embodiment, the terminal can select the first subset of the first training dataset as the second validation dataset and use the remaining subsets as the second training dataset during the first training iteration, following a preset polling order. During the second training iteration, the terminal can select the second subset of the first training dataset as the second validation dataset and use the remaining subsets as the second training dataset. During the third training iteration, the terminal can select the third subset of the first training dataset as the second validation dataset and use the remaining subsets as the second training dataset. This process continues until the Kth training iteration, at which point the terminal can select the Kth subset of the first training dataset as the second validation dataset and use the remaining subsets as the second training dataset.

[0125] In one example, the terminal can divide the first training dataset into K subsets. Then, following a preset polling order, during the first training iteration, the terminal can select the Kth subset of the first training dataset as the second validation dataset, and use the remaining subsets as the second training dataset. During the second training iteration, the terminal can select the (K-1)th subset of the first training dataset as the second validation dataset, and use the remaining subsets as the second training dataset. During the third training iteration, the terminal can select the (K-2)th subset of the first training dataset as the second validation dataset, and use the remaining subsets as the second training dataset. And so on, until the Kth training iteration, where the terminal can select the first subset of the first training dataset as the second validation dataset, and use the remaining subsets as the second training dataset.

[0126] Step 303: Input the second training dataset and the second validation dataset into the target neural network to obtain the first sesamoid bone level classification prediction result corresponding to the second training dataset and the second sesamoid bone level classification prediction result corresponding to the second validation dataset.

[0127] In this embodiment, the terminal can input a second training dataset into the target neural network to obtain a first sesamoid bone level classification prediction result corresponding to the second training dataset. Simultaneously, the terminal can input a second verification dataset into the target neural network to obtain a second sesamoid bone level classification prediction result corresponding to the second verification dataset. The first sesamoid bone level classification prediction result may include the sesamoid bone level prediction feature vector (referred to as the first sesamoid bone level prediction feature vector for easy distinction) of the sesamoid bones contained in the first sample image corresponding to the second training dataset; the second sesamoid bone level classification prediction result may include the sesamoid bone level prediction feature vector (referred to as the second sesamoid bone level prediction feature vector) of the sesamoid bones contained in the first sample image corresponding to the second verification dataset.

[0128] Step 304: Determine and record the first objective loss function based on the first seed bone level classification prediction result and the first sample level classification result corresponding to the second training dataset; determine and record the second objective loss function based on the second seed bone level classification prediction result and the first sample level classification result corresponding to the second validation dataset.

[0129] In this embodiment, the terminal can first determine the true value of the sesamoid bone level of the first sample image corresponding to the second training dataset (which can be referred to as the true value of the first sesamoid bone level for easy distinction) based on the classification result of the first sample level corresponding to the second training dataset. Then, the terminal can use the predicted feature vector of the first sesamoid bone level and the true value of the first sesamoid bone level as parameters of the first target loss function, and determine and record the first target loss function.

[0130] Simultaneously, the terminal can first determine the true value of the sesamoid bone level of the first sample image corresponding to the second validation dataset based on the classification result of the first sample level corresponding to the second validation dataset (which can be referred to as the true value of the second sesamoid bone level for easy distinction). Then, the terminal can use the predicted feature vector of the second sesamoid bone level and the true value of the second sesamoid bone level as parameters of the second target loss function, and determine and record the second target loss function.

[0131] In one example, the true value of the first sesamoid bone level and the true value of the first sesamoid bone level can both be one-hot encoded.

[0132] Step 305: Update the parameters of the target neural network according to the first target loss function, and return to execute the step of determining the second validation dataset in the first training dataset according to the preset polling order, and using the remaining subset as the second training dataset.

[0133] In this embodiment, the terminal first calculates the gradient of the target neural network based on the first target loss function. Then, the terminal updates the parameters of the target neural network based on the gradient. After updating the parameters of the target neural network, the terminal returns to execute the step of determining the second validation dataset from the first training dataset according to a preset polling order, and using the remaining subset as the second training dataset, to perform a new round of training on the target neural network.

[0134] Step 306: If the first target loss function corresponding to the updated target neural network is less than the first target loss function corresponding to the target neural network before the update, save the updated target neural network.

[0135] In this embodiment, the terminal can determine whether the first target loss function corresponding to the target neural network after each update is less than the first target loss function corresponding to the target neural network before each update, based on the recorded first target loss function. Specifically, after an update, the terminal can first obtain the first target loss function corresponding to the target neural network after the update and the first target loss function corresponding to the target neural network before the update, both recorded by the terminal. Then, the terminal compares the first target loss function corresponding to the target neural network after the update with the first target loss function corresponding to the target neural network before the update to determine whether the first target loss function corresponding to the target neural network after the update is less than the first target loss function corresponding to the target neural network before the update. If the first target loss function corresponding to the target neural network after the update is less than the first target loss function corresponding to the target neural network before the update, the terminal saves the target neural network after the update.

[0136] Step 307: Determine whether the preset iteration stopping condition has been met based on the recorded first target loss function and second target loss function. If the preset iteration stopping condition has been met, use each stored target neural network as a candidate seed bone level classification model.

[0137] In this embodiment, the terminal can determine whether the training process has reached a preset iteration stopping condition based on the recorded first target loss function and second target loss function. If the preset iteration stopping condition is reached, the terminal can stop training and use the stored target neural networks as candidate seed bone level classification models.

[0138] In the above-mentioned seed bone level recognition method, the terminal divides the first training dataset into a preset number of subsets and selects the validation dataset and training dataset from the first training dataset according to a preset polling order, thereby realizing multiple updates of the parameters of the target neural network. Moreover, based on the above selection method, the richness of the samples can be improved, thereby improving the generalization ability of the candidate seed bone level classification model and the efficiency of data utilization.

[0139] In one embodiment, such as Figure 4 As shown, the specific process of determining whether the preset iteration stopping condition has been met based on the recorded first and second objective loss functions includes the following steps:

[0140] Step 401: Determine whether the first target loss function corresponding to the target neural network after this update is less than the first target loss function corresponding to the target neural network before this update, and whether the second target loss function corresponding to the target neural network after this update is greater than or equal to the second target loss function corresponding to the target neural network before this update.

[0141] In this embodiment, after an update, the terminal can first obtain the first and second target loss functions corresponding to the target neural network after the update, as well as the first and second target loss functions corresponding to the target neural network before the update, which are recorded by the terminal. Then, the terminal can compare the first target loss function corresponding to the target neural network after the update with the first target loss function corresponding to the target neural network before the update, and determine whether the first target loss function corresponding to the target neural network after the update is less than the first target loss function corresponding to the target neural network before the update. Simultaneously, the terminal can compare the second target loss function corresponding to the target neural network after the update with the second target loss function corresponding to the target neural network before the update, and determine whether the second target loss function corresponding to the target neural network after the update is greater than or equal to the second target loss function corresponding to the target neural network before the update.

[0142] Step 402: If yes, then determine that the target neural network after this update does not meet the preset fitting conditions.

[0143] In this embodiment, the terminal can determine whether the updated target neural network meets the preset fitting conditions based on the judgment results of whether the first target loss function corresponding to the updated target neural network is less than the first target loss function corresponding to the target neural network before the update, and whether the second target loss function corresponding to the updated target neural network is greater than or equal to the second target loss function corresponding to the target neural network before the update. If the first target loss function corresponding to the updated target neural network is less than the first target loss function corresponding to the target neural network before the update, and the second target loss function corresponding to the updated target neural network is greater than or equal to the second target loss function corresponding to the target neural network before the update, then it is determined that the updated target neural network does not meet the preset fitting conditions.

[0144] Step 403: If the number of times the updated target neural network fails to meet the preset fitting conditions exceeds a preset number, the currently stored target neural network is used as a candidate seed bone level classification model.

[0145] In this embodiment, the terminal can record the number of times the updated target neural network fails to meet the preset fitting conditions. Then, the terminal compares the number of times the updated target neural network fails to meet the preset fitting conditions with a preset number. If the number of times the updated target neural network fails to meet the preset fitting conditions exceeds the preset number, the terminal can use the currently stored target neural network as a candidate seed-bone level classification model.

[0146] In one implementation, the terminal can also record the number of times the updated target neural network consecutively fails to meet the preset fitting conditions. Then, the terminal compares the number of times the updated target neural network consecutively fails to meet the preset fitting conditions with a preset number. If the number of times the updated target neural network consecutively fails to meet the preset fitting conditions exceeds the preset number, the terminal can use the currently stored target neural network as a candidate seed-bone level classification model.

[0147] In the above-mentioned sesamoid bone level recognition method, the terminal sets a second target loss function to set a preset fitting condition. If the number of times the updated target neural network fails to meet the preset fitting condition exceeds a preset number, the currently stored target neural network is used as a candidate sesamoid bone level classification model. This ensures that the first loss function and the second loss function corresponding to the updated target neural network maintain similar change directions, guaranteeing that the model can fit the data without overfitting during the training process.

[0148] In one embodiment, such as Figure 5As shown, the specific process of determining and recording the first objective loss function based on the first sesamoid bone classification prediction result and the first sample classification result corresponding to the second training dataset includes the following steps:

[0149] Step 501: Determine the soft loss function based on the first sesamoid bone level classification prediction result and the first sample level classification result corresponding to the second training dataset.

[0150] In this embodiment, the terminal can first determine the true value of the first sesamoid level based on the classification result of the first sample level corresponding to the second training dataset. Then, the terminal can use the predicted feature vector of the first sesamoid level and the true value of the first sesamoid level as parameters of the soft target loss function to determine the soft loss function.

[0151] Step 502: Determine the center loss function based on the first sesamoid bone level classification prediction result and the first sample level classification result corresponding to the second training dataset.

[0152] In this embodiment, the terminal can first determine the true value of the first sesamoid level based on the classification result of the first sample level corresponding to the second training dataset. Then, the terminal can use the predicted feature vector of the first sesamoid level and the true value of the first sesamoid level as parameters of the central target loss function to determine the central loss function.

[0153] Step 503: Perform a weighted summation of the soft loss function and the central loss function to obtain and record the first target loss function.

[0154] In this embodiment of the application, the terminal can perform a weighted summation of the soft loss function and the central loss function to obtain the first target loss function, and record the first target loss function.

[0155] In one example, the first objective loss function can be expressed as:

[0156]

[0157] Where L is the first objective loss function, L S For a soft loss function, L c Let λ be the center loss function, and λ be the weight of the weighted sum of the soft loss function and the center loss function; i and j represent the i-th and j-th samples in this training, respectively; x is the predicted feature vector of the first sesamoid level, and y is the true value of the first sesamoid level; m is the number of samples contained in the mini-batch, n is the total number of categories of the sesamoid level, T is the number of categories of the sesamoid level contained in the mini-batch, c is the class depth feature center, W is the weight of the current neural network output layer, and b is the offset.

[0158] In the above-mentioned sesamoid bone level recognition method, the terminal performs a weighted summation of the soft loss function and the center loss function to obtain and record the first target loss function, and updates the parameters of the target neural network according to the first loss function. This method takes into account both intra-class aggregation and inter-class separation, making the feature differences between different categories more obvious and the differences within the same category smaller, thereby improving the accuracy of sesamoid bone level recognition.

[0159] In one embodiment, such as Figure 6 As shown, the specific process of acquiring multiple medical images of the wrist to be corrected includes the following steps:

[0160] Step 601: Obtain a medical image of the wrist.

[0161] In this embodiment, the terminal can acquire medical images of the wrist of the target object being examined using medical examination instruments. The medical images represent the internal structure or function of an anatomical region and may include X-ray images, CT images, and magnetic resonance images.

[0162] Step 602: Adjust the medical image to a preset size.

[0163] In this embodiment, the terminal can adjust the medical image to a preset size. For example, the preset size can be 300*300.

[0164] Step 603: Rotate the preset size image by multiple preset angles to obtain multiple rotated images, and determine the preset size image and the multiple rotated images as the medical images to be corrected.

[0165] In this embodiment, the terminal can rotate an image of a preset size by multiple preset angles to obtain multiple rotated images. For example, the multiple preset angles can be 90 degrees, 180 degrees, and 270 degrees, respectively. Then, the terminal can determine the image of the preset size and the multiple rotated images as the medical image to be corrected.

[0166] In the aforementioned sesamoid bone grading recognition method, the terminal acquires a medical image of the wrist and adjusts it to a preset size. This ensures that all images input to the pre-trained sesamoid bone grading classification model are of uniform size, improving the accuracy of sesamoid bone grading recognition. Furthermore, the terminal rotates the preset-size image by multiple preset angles to obtain multiple rotated images. These preset-size images and the multiple rotated images are then identified as the medical images to be corrected. This achieves the acquisition of multiple medical images of the wrist to be corrected, ensuring that these images differ only in rotation angle, reducing interference factors and improving the accuracy of hand position recognition.

[0167] In one embodiment, such as Figure 7As shown, the specific process of obtaining the sesamoid bone classification result of the target sesamoid bone based on the forehand image and a pre-trained sesamoid bone classification model includes the following steps:

[0168] Step 701: Input the forehand image into the pre-trained epiphyseal localization model to obtain the coordinate information of the first proximal phalanx in the forehand image.

[0169] The first proximal phalanx is the proximal phalanx of the thumb.

[0170] In this embodiment, the terminal can input the forehand image into a pre-trained epiphyseal localization model to obtain the coordinate information of the first proximal phalanx in the forehand image.

[0171] Among them, the pre-trained epiphyseal localization model is mainly used to locate the positions of the 13 epiphyses in the wrist. For example... Figure 8 As shown, the thumb has two phalanges, while the other fingers each have three, arranged from proximal to distal as the proximal phalanx, middle phalanx, and distal phalanx. The first proximal phalanx is the proximal phalanx of the thumb. The coordinate information of the first proximal phalanx refers to its coordinates in the upright hand image. In one example, the coordinate information of the first proximal phalanx may include the coordinates of its topmost (top), leftmost (left), bottommost (bottom), and rightmost (right). For example, the coordinate information of the first proximal phalanx can be represented as (top, left, bottom, right).

[0172] In one example, the terminal can obtain an epiphyseal localization model by training an m2det network. The backbone network of the epiphyseal localization model can be a 101-layer residual neural network (resnet101).

[0173] Step 702: Based on the positional relationship between the first proximal phalanx and the sesamoid bone, and the coordinate information of the first proximal phalanx, obtain the coordinate information of the target sesamoid bone in the forehand image.

[0174] In this embodiment, the terminal can determine whether the forehand image is a left-hand forehand image based on the coordinate information of the first proximal phalanx. If the forehand image is a left-hand forehand image, the terminal can obtain the coordinate information of the target sesamoid bone in the forehand image based on the positional relationship between the first proximal phalanx and the sesamoid bone of the left hand, and the coordinate information of the first proximal phalanx. If the forehand image is not a left-hand forehand image, the terminal can obtain the coordinate information of the target sesamoid bone in the forehand image based on the positional relationship between the first proximal phalanx and the sesamoid bone of the right hand, and the coordinate information of the first proximal phalanx.

[0175] In one example, the terminal can use Python to calculate the coordinates of the target sesamoid bone in the forehand image based on the positional relationship between the first proximal phalanx and the sesamoid bone, as well as the coordinate information of the first proximal phalanx.

[0176] If the forehand image is a left-hand forehand image, the coordinate information of the target sesamoid bone is calculated as follows:

[0177] left += int((right-left) / 2)

[0178] left-=int(((bottom-top)-(right-left)) / 2)

[0179] right += int((right-left) / 2)

[0180] right+=(bottom-top)-(right-left)

[0181] If the forehand image is not a left-hand forehand image, the coordinate information of the target sesamoid bone is calculated as follows:

[0182] left - = int((right - left) / 2)

[0183] right - = int((right-left) / 2)

[0184] right+=int(((bottom-top)-(right-left)) / 2)

[0185] left-=(bottom-top)-(right-left)

[0186] The coordinate information of the first proximal phalanx can be represented as (top, left, bottom, right).

[0187] Step 703: Determine the target sesamoid bone image in the forehand position image based on the coordinate information of the target sesamoid bone.

[0188] In this embodiment of the application, the terminal can determine the target sesamoid image in the forehand position image based on the coordinate information of the target sesamoid.

[0189] Step 704: Input the target sesamoid image into the pre-trained sesamoid classification model to obtain the sesamoid classification result of the target sesamoid.

[0190] In this embodiment, the terminal can input the target sesamoid image into a pre-trained sesamoid classification model to obtain the sesamoid classification result of the target sesamoid. The sesamoid classification result of the target sesamoid can represent the sesamoid level of the target sesamoid.

[0191] In the aforementioned sesamoid bone grading method, the terminal inputs a forehand image into a pre-trained epiphyseal localization model to obtain the coordinate information of the first proximal phalanx in the forehand image. Based on the positional relationship between the first proximal phalanx and the sesamoid bones, and the coordinate information of the first proximal phalanx, the coordinate information of the target sesamoid bone in the forehand image is obtained. Based on the coordinate information of the target sesamoid bone, the target sesamoid bone image is determined in the forehand image, improving the accuracy of the target sesamoid bone image. Furthermore, in the aforementioned sesamoid bone grading method, the terminal inputs the target sesamoid bone image into a pre-trained sesamoid bone grading classification model to obtain the sesamoid bone grading classification result. The forehand image is further processed to obtain a more clearly defined target sesamoid bone image, which is then input into the pre-trained sesamoid bone grading classification model, further improving the accuracy of sesamoid bone grading recognition.

[0192] It should be understood that, although Figure 1-7 The steps in the flowchart are shown sequentially as indicated by the arrows, but these steps are not necessarily executed in the order indicated by the arrows. Unless otherwise specified herein, there is no strict order in which these steps are executed, and they can be performed in other orders. Figure 1-7 At least some of the steps in the process may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but may be executed at different times. The execution order of these steps or stages is not necessarily sequential, but may be executed in turn or alternately with other steps or at least some of the steps or stages in other steps.

[0193] In one embodiment, such as Figure 9 As shown, a sesamoid bone grading identification device 900 is provided, comprising: a first acquisition module 910, a hand position recognition module 920, a first determination module 930, and a sesamoid bone grading classification module 940, wherein:

[0194] The first acquisition module 910 is used to acquire multiple medical images of the wrist to be corrected; the medical images to be corrected include images of the target sesamoid bone.

[0195] The hand position recognition module 920 is used to input each of the medical images to be corrected into a pre-trained hand position recognition model to obtain the hand position recognition result corresponding to each of the medical images to be corrected.

[0196] The first determining module 930 is used to determine the positive hand position image in each of the medical images to be corrected based on the hand position recognition results corresponding to each of the medical images to be corrected;

[0197] The sesamoid bone classification module 940 is used to obtain the sesamoid bone classification result of the target sesamoid bone based on the forehand image and the pre-trained sesamoid bone classification model.

[0198] Optionally, the device 900 further includes:

[0199] The second acquisition module is used to acquire a first training dataset and a first validation dataset; the first training dataset includes a first sample image and a first sample level classification result of the sesamoid bones contained in the first sample image; the first validation dataset includes a second sample image and a second sample level classification result of the sesamoid bones contained in the second sample image; both the first sample image and the second sample image are medical images of the wrist.

[0200] The training module is used to train a preset target neural network using the first training dataset to obtain multiple candidate seed bone level classification models.

[0201] The verification module is used to verify each of the candidate sesamoid bone level classification models using the first verification dataset, and to obtain the accuracy of each of the candidate sesamoid bone level classification models.

[0202] The second determining module is used to determine the sesamoid bone classification model from among the candidate sesamoid bone classification models based on the accuracy of each candidate sesamoid bone classification model.

[0203] Optionally, the training module is specifically used for:

[0204] The first training dataset is divided into a predetermined number of subsets;

[0205] According to the preset polling order, the second validation dataset is selected from the first training dataset, and the remaining subset is used as the second training dataset.

[0206] The second training dataset and the second validation dataset are respectively input into the target neural network to obtain the first sesamoid bone level classification prediction result corresponding to the second training dataset and the second sesamoid bone level classification prediction result corresponding to the second validation dataset;

[0207] Based on the first sesamoid bone classification prediction result and the first sample classification result corresponding to the second training dataset, determine and record the first target loss function; based on the second sesamoid bone classification prediction result and the first sample classification result corresponding to the second validation dataset, determine and record the second target loss function.

[0208] Based on the first target loss function, the parameters of the target neural network are updated, and the process returns to the step of determining the second validation dataset in the first training dataset according to a preset polling order, and using the remaining subset as the second training dataset.

[0209] If the first target loss function corresponding to the target neural network after this update is less than the first target loss function corresponding to the target neural network before this update, then save the target neural network after this update.

[0210] Based on the recorded first and second target loss functions, it is determined whether the preset iteration stopping condition has been met. If the preset iteration stopping condition is met, each stored target neural network is used as a candidate seed bone level classification model.

[0211] Optionally, the training module is specifically used for:

[0212] Determine whether the first target loss function corresponding to the target neural network after this update is less than the first target loss function corresponding to the target neural network before this update, and whether the second target loss function corresponding to the target neural network after this update is greater than or equal to the second target loss function corresponding to the target neural network before this update;

[0213] If so, then it is determined that the target neural network after this update does not meet the preset fitting conditions;

[0214] If the number of times the updated target neural network fails to meet the preset fitting conditions exceeds a preset number, the currently stored target neural network will be used as a candidate seed bone level classification model.

[0215] Optionally, the training module is specifically used for:

[0216] Based on the first sesamoid bone classification prediction result and the first sample classification result corresponding to the second training dataset, a soft loss function is determined;

[0217] The center loss function is determined based on the first sesamoid bone classification prediction result and the first sample classification result corresponding to the second training dataset.

[0218] The soft loss function and the central loss function are weighted and summed to obtain and record the first target loss function.

[0219] Optionally, the first acquisition module 910 is specifically used for:

[0220] Acquire medical images of the wrist;

[0221] Adjust the medical image to a preset size;

[0222] The preset-size image is rotated by multiple preset angles to obtain multiple rotated images, and the preset-size image and the multiple rotated images are determined as the medical images to be corrected.

[0223] Optionally, the sesamoid bone grading module 940 is specifically used for:

[0224] The image of the proximal hand position is input into a pre-trained epiphyseal localization model to obtain the coordinate information of the first proximal phalanx in the image of the proximal hand position; the first proximal phalanx is the proximal phalanx of the thumb.

[0225] Based on the positional relationship between the first proximal phalanx and the sesamoid bone, and the coordinate information of the first proximal phalanx, the coordinate information of the target sesamoid bone in the orthopedic image is obtained;

[0226] Based on the coordinate information of the target sesamoid bone, the target sesamoid bone image is determined in the forehand position image;

[0227] The target sesamoid image is input into the pre-trained sesamoid classification model to obtain the sesamoid classification result of the target sesamoid.

[0228] Specific limitations regarding the sesamoid bone grading device can be found in the limitations of the sesamoid bone grading method described above, and will not be repeated here. Each module in the aforementioned sesamoid bone grading device can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device, or stored in the memory of a computer device as software, so that the processor can call and execute the corresponding operations of each module.

[0229] In one embodiment, a computer device is provided, which may be a terminal, and its internal structure diagram may be as follows: Figure 10As shown, the computer device includes a processor, memory, communication interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage media. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, carrier networks, NFC (Near Field Communication), or other technologies. When executed by the processor, the computer program implements a sesamoid-level identification method. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad on the computer device's casing, or an external keyboard, touchpad, or mouse.

[0230] Those skilled in the art will understand that Figure 10 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0231] In one embodiment, a computer device is provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of the above-described sesamoid grade identification method.

[0232] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the steps of the above-described sesamoid grade identification method.

[0233] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the methods described above. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, or optical storage, etc. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM), etc.

[0234] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0235] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this patent application should be determined by the appended claims.

Claims

1. A method for identifying the grade of sesamoid bones, characterized in that, The method includes: Acquire a medical image of the wrist; adjust the medical image to a preset size; rotate the preset size image by multiple preset angles to obtain multiple rotated images, and determine the preset size image and the multiple rotated images as medical images to be corrected; the medical images to be corrected include images of the target sesamoid bone, and the medical images to be corrected are medical images that have not undergone hand position recognition; the hand position recognition results corresponding to each medical image to be corrected may include the score of each medical image to be corrected; Each of the medical images to be corrected is input into a pre-trained hand position recognition model to obtain the hand position recognition result corresponding to each of the medical images to be corrected. Based on the hand position recognition results corresponding to each of the medical images to be corrected, determine the positive hand position image in each of the medical images to be corrected; The image of the proximal hand position is input into a pre-trained epiphyseal localization model to obtain the coordinate information of the first proximal phalanx in the image of the proximal hand position; the first proximal phalanx is the proximal phalanx of the thumb. Based on the coordinate information of the first proximal phalanx, the forehand position type of the forehand image is determined; based on the positional relationship between the first proximal phalanx and the sesamoid bone, the coordinate information of the first proximal phalanx, and the forehand position type, the coordinate information of the target sesamoid bone in the forehand image is obtained, the forehand position type includes left-hand forehand and non-left-hand forehand, and the coordinate information of the first proximal phalanx includes the coordinates of the top, left, bottom, and rightmost points; Based on the coordinate information of the target sesamoid bone, the target sesamoid bone image is determined in the forehand position image; The target sesamoid image is input into the pre-trained sesamoid classification model to obtain the sesamoid classification result of the target sesamoid.

2. The method according to claim 1, characterized in that, The training process of the sesamoid bone grading model includes: Obtain a first training dataset and a first validation dataset; the first training dataset includes a first sample image and a first sample-level classification result of the sesamoid bones contained in the first sample image; the first validation dataset includes a second sample image and a second sample-level classification result of the sesamoid bones contained in the second sample image; both the first sample image and the second sample image are medical images of the wrist; The preset target neural network is trained using the first training dataset to obtain multiple candidate sesamoid bone classification models. The candidate sesamoid bone classification models are validated using the first validation dataset to obtain the accuracy of each candidate sesamoid bone classification model. Based on the accuracy of each of the candidate sesamoid bone grading models, the sesamoid bone grading model is determined from among the candidate sesamoid bone grading models.

3. The method according to claim 2, characterized in that, The process involves training a preset target neural network using the first training dataset to obtain multiple candidate seed bone classification models, including: The first training dataset is divided into a predetermined number of subsets; According to the preset polling order, the second validation dataset is selected from the first training dataset, and the remaining subset is used as the second training dataset. The second training dataset and the second validation dataset are respectively input into the target neural network to obtain the first sesamoid bone level classification prediction result corresponding to the second training dataset and the second sesamoid bone level classification prediction result corresponding to the second validation dataset; Based on the first sesamoid bone classification prediction result and the first sample classification result corresponding to the second training dataset, determine and record the first target loss function; based on the second sesamoid bone classification prediction result and the first sample classification result corresponding to the second validation dataset, determine and record the second target loss function. Based on the first target loss function, the parameters of the target neural network are updated, and the process returns to the step of determining the second validation dataset in the first training dataset according to a preset polling order, and using the remaining subset as the second training dataset. If the first target loss function corresponding to the target neural network after this update is less than the first target loss function corresponding to the target neural network before this update, then save the target neural network after this update. Based on the recorded first and second target loss functions, it is determined whether the preset iteration stopping condition has been met. If the preset iteration stopping condition is met, each stored target neural network is used as a candidate seed bone level classification model.

4. The method according to claim 3, characterized in that, The step of determining whether the preset iteration stopping condition has been met based on the recorded first target loss function and second target loss function includes: Determine whether the first target loss function corresponding to the target neural network after this update is less than the first target loss function corresponding to the target neural network before this update, and whether the second target loss function corresponding to the target neural network after this update is greater than or equal to the second target loss function corresponding to the target neural network before this update; If so, then it is determined that the target neural network after this update does not meet the preset fitting conditions; If the number of times the updated target neural network fails to meet the preset fitting conditions exceeds a preset number, the currently stored target neural network will be used as a candidate seed bone level classification model.

5. The method according to claim 3, characterized in that, The step of determining and recording the first target loss function based on the first sesamoid bone level classification prediction result and the first sample level classification result corresponding to the second training dataset includes: Based on the first sesamoid bone classification prediction result and the first sample classification result corresponding to the second training dataset, a soft loss function is determined; The center loss function is determined based on the first sesamoid bone classification prediction result and the first sample classification result corresponding to the second training dataset. The soft loss function and the central loss function are weighted and summed to obtain and record the first target loss function.

6. A bone marrow grading identification device, characterized in that, The device includes: The first acquisition module is used to acquire a medical image of the wrist; adjust the medical image to a preset size; rotate the preset size image by multiple preset angles to obtain multiple rotated images, and determine the preset size image and the multiple rotated images as medical images to be corrected; the medical images to be corrected include images of the target sesamoid bone, and the medical images to be corrected are medical images that have not undergone hand position recognition; the hand position recognition results corresponding to each medical image to be corrected may include the score of each medical image to be corrected; The hand position recognition module is used to input each of the medical images to be corrected into a pre-trained hand position recognition model to obtain the hand position recognition result corresponding to each of the medical images to be corrected. The first determining module is used to determine the positive hand position image in each of the medical images to be corrected based on the hand position recognition results corresponding to each of the medical images to be corrected; The sesamoid bone classification module is used to input the orthopedic image into a pre-trained epiphyseal localization model to obtain the coordinate information of the first proximal phalanx in the orthopedic image; the first proximal phalanx is the proximal phalanx of the thumb; Based on the coordinate information of the first proximal phalanx, the forehand position type of the forehand image is determined; based on the positional relationship between the first proximal phalanx and the sesamoid bone, the coordinate information of the first proximal phalanx, and the forehand position type, the coordinate information of the target sesamoid bone in the forehand image is obtained, the forehand position type includes left-hand forehand and non-left-hand forehand, and the coordinate information of the first proximal phalanx includes the coordinates of the top, left, bottom, and rightmost points; Based on the coordinate information of the target sesamoid bone, the target sesamoid bone image is determined in the forehand position image; The target sesamoid image is input into the pre-trained sesamoid classification model to obtain the sesamoid classification result of the target sesamoid.

7. The apparatus according to claim 6, characterized in that, The device further includes: The second acquisition module is used to acquire a first training dataset and a first validation dataset; the first training dataset includes a first sample image and a first sample level classification result of the sesamoid bones contained in the first sample image; the first validation dataset includes a second sample image and a second sample level classification result of the sesamoid bones contained in the second sample image; both the first sample image and the second sample image are medical images of the wrist. The training module is used to train a preset target neural network using the first training dataset to obtain multiple candidate seed bone level classification models. The verification module is used to verify each of the candidate sesamoid bone level classification models using the first verification dataset, and to obtain the accuracy of each of the candidate sesamoid bone level classification models. The second determining module is used to determine the sesamoid bone classification model from among the candidate sesamoid bone classification models based on the accuracy of each candidate sesamoid bone classification model.

8. The apparatus according to claim 7, characterized in that, The training module is specifically used to divide the first training dataset into a preset number of subsets; According to the preset polling order, the second validation dataset is selected from the first training dataset, and the remaining subset is used as the second training dataset. The second training dataset and the second validation dataset are respectively input into the target neural network to obtain the first sesamoid bone level classification prediction result corresponding to the second training dataset and the second sesamoid bone level classification prediction result corresponding to the second validation dataset; Based on the first sesamoid bone classification prediction result and the first sample classification result corresponding to the second training dataset, determine and record the first target loss function; based on the second sesamoid bone classification prediction result and the first sample classification result corresponding to the second validation dataset, determine and record the second target loss function. Based on the first target loss function, the parameters of the target neural network are updated, and the process returns to the step of determining the second validation dataset in the first training dataset according to a preset polling order, and using the remaining subset as the second training dataset. If the first target loss function corresponding to the target neural network after this update is less than the first target loss function corresponding to the target neural network before this update, then save the target neural network after this update. Based on the recorded first and second target loss functions, it is determined whether the preset iteration stopping condition has been met. If the preset iteration stopping condition is met, each stored target neural network is used as a candidate seed bone level classification model.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 5.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 5.

Images