Method and system for segmenting ossification center and cartilage area based on ultrasonic image

Abstract

The invention relates to an ossification center and cartilage area segmentation method and system based on an ultrasonic image, and the method comprises the steps: obtaining an ulna ultrasonic image of a sample patient, carrying out the preprocessing of the ultrasonic image, and generating a sample set; constructing a classification network model of a multi-scale coding structure based on the sample set, and inputting the sample set and the corresponding ulna closure type into the classification network model for training; inputting a target sample set obtained based on the ulna ultrasonic image of the target patient into the trained classification network model to obtain a corresponding target ulna closure type; and according to the target ulna closure type, inputting the corresponding ulna ultrasonic image into a segmentation model for segmentation, and respectively obtaining corresponding segmentation results of the ossification center and the cartilage region. According to the method, a multi-scale feature extraction classification algorithm and an automatic segmentation method are combined, the ossification center region can be better segmented under the condition of ensuring good classification precision, and the segmentation efficiency is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of ultrasonic image recognition, and in particular to a method and system for segmenting an ossification center and a cartilage area based on ultrasonic images. Background Art

[0002] At present, the segmentation of ossification center and cartilage has important application value in clinical diagnosis, bone age assessment, surgical planning and scientific research. In the incompletely closed ulna, there is usually a significant contrast difference between the ossification center of the bone and the surrounding cartilage tissue, which enables the segmentation algorithm to well identify and distinguish the ossification center and the cartilage area in medical images; however, when the ulna is completely closed, the high contrast originally caused by the difference between cartilage and bone tissue disappears, and the ossification center and the surrounding bone tissue are difficult to distinguish in the image. Automated segmentation algorithms may find it difficult to accurately identify and segment the ossification center area after the ulna is completely closed. Most ossification center segmentation algorithms rely on the boundary clarity between cartilage and bone tissue. In the closed image, the blurred or disappeared boundary may lead to a decrease in the segmentation efficiency of the ossification center area and an increase in the algorithm error. The existing technology is difficult to distinguish between the completely closed ossification center and the surrounding bone tissue, resulting in low ossification center segmentation efficiency, which affects subsequent diagnosis and treatment. Summary of the invention

[0003] The technical problem to be solved by the present invention is to provide a method and system for segmenting the ossification center and the cartilage area based on ultrasound images in view of the above-mentioned deficiencies in the prior art.

[0004] The technical solution of the present invention to solve the above technical problem is as follows: a method for segmenting the ossification center and the cartilage area based on ultrasound images, comprising the following steps:

[0005] Acquire ulna ultrasound images of sample patients, pre-process the ultrasound images and generate a sample set;

[0006] Constructing a classification network model with a multi-scale coding structure based on the sample set, and inputting the sample set and the corresponding ulnar closure type into the classification network model for training;

[0007] Inputting a target sample set obtained based on the ulna ultrasound image of the target patient into the trained classification network model to obtain a corresponding target ulna closure type;

[0008] According to the target ulna closure type, the corresponding ulna ultrasound image is input into the segmentation model for segmentation, and the segmentation results of the corresponding ossification center and cartilage area are obtained respectively.

[0009] On the basis of the above technical solution, the present invention can also be improved as follows:

[0010] Further: the preprocessing of ultrasound images and generating a sample set specifically includes the following steps:

[0011] Performing image recognition processing on the ultrasound image, determining that the ulna region in the ultrasound image is the target region, and generating a boundary contour of the target region;

[0012] Performing center cropping on the pixel values ​​of the ultrasound image according to a preset pixel threshold range, and scaling the pixel values ​​of the ultrasound image to within [0, 1];

[0013] The cropped target area image in the ultrasound image is normalized according to a preset normalization ratio to generate a sample set.

[0014] The beneficial effect of the above further scheme is: by performing image recognition processing on the ultrasound image, the ulna area can be accurately identified as the target area, thereby facilitating subsequent cropping and pixel value scaling to improve the generalization ability of the model, and by normalizing the target area image in the cropped ultrasound image, the grayscale value of the image can have a zero mean and a standardized distribution, thereby reducing the convergence time in model training and improving the accuracy of the model.

[0015] Further: the step of inputting the sample set and the corresponding ulnar closure type into the classification network model for training specifically includes the following steps:

[0016] The CDB module in the classification network model performs convolution processing on the ultrasound images in the sample set, obtains a feature map after standardization, and outputs it to the DAB module;

[0017] The DAB module performs convolution, normalization, correction linearization and regularization on the feature map output by the CDB module, and then splices it with the feature map input by the DAB module, and performs convolution, normalization, correction linearization and regularization on the feature map obtained by splicing, and then splices it with the feature map input by the DAB module for the next iteration, until the set number of splicing times is reached, and then outputs the feature map obtained by the last splicing to the MSCM module;

[0018] The MSCM module normalizes and corrects the linear processing of the feature map output by the DAB module, divides it into four paths for multi-scale convolution stacking processing, and outputs it to the DA module;

[0019] The DA module performs convolution and normalization processing on the feature map output by the MSCM module, and calculates the ulna closure probability corresponding to the feature map using an activation function, so as to determine the ulna predicted closure type corresponding to the feature map according to a preset ulna closure probability threshold;

[0020] A cross entropy loss function is constructed, and the loss value between the predicted ulna closure type and the corresponding ulna closure type of all sample patients is calculated using the loss function, and back propagation is performed to adjust the parameters of the classification network model, and the predicted ulna closure type and the corresponding ulna closure type of the sample patient are used as input to calculate the accuracy of the classification network model after adjusting the parameters, and the model parameters with the highest accuracy are used as the parameters of the classification network model to complete the training.

[0021] The beneficial effect of the above further scheme is: the feature map is convolved and standardized by the CDB module to facilitate unified processing by the DAB module, the MSCM module and the DA module, and it is iteratively spliced ​​by the DAB module, and multi-scale convolution superposition processing is performed by the MSCM module, and then the ulna closure probability corresponding to the feature map is calculated by the activation function, the ulna predicted closure type is determined, and back propagation is performed. Finally, the accuracy of the ulna predicted closure type predicted by the classification network model is calculated, and the model parameters with the highest accuracy are used as the optimal model parameters to complete the training, so that the model has a higher recognition accuracy.

[0022] Further: the MSCM module normalizes and corrects the linear processing of the feature map output by the DAB module and then performs scale convolution stacking processing, which specifically includes the following steps:

[0023] Perform 1*1 convolution processing on the feature map of the first output;

[0024] Perform two consecutive 3x3 convolutions on the feature map of the second output;

[0025] The feature map of the third output is processed by 3x3 convolution and 5*5 convolution in sequence;

[0026] Perform global average pooling on the feature map output by the fourth channel;

[0027] The feature maps of the four outputs are added element by element and output to the DA module.

[0028] The beneficial effect of the above further scheme is that by performing convolution and global average pooling on the four-way feature maps respectively, features of different scales can be extracted, and subtle features of ossification centers and cartilage areas can be captured at different scales, thereby ensuring classification accuracy.

[0029] Further: the step of inputting the corresponding ulna ultrasound image into the segmentation model for segmentation according to the target ulna closure type, and obtaining the segmentation results of the corresponding ossification center and cartilage area respectively specifically includes the following steps:

[0030] When the target ulna closure type is incomplete ulna closure, the segmentation model obtains two segmentation regions according to the ulna ultrasound image, takes the intersection of the two regions to obtain the ossification center region, and takes the union of the two regions to obtain the cartilage region;

[0031] When the target ulna closure type is complete ulna closure, the segmentation model obtains a segmentation area according to the ulna ultrasound image, and the area is both an ossification center area and a cartilage area.

[0032] The beneficial effect of the above further scheme is that by performing segmentation for different target ulna closure types respectively, the ossification center area can be better segmented, effectively solving the problem of low ossification center segmentation efficiency when the ulna is fully closed, while also maintaining the segmentation efficiency of the cartilage area.

[0033] Further: the method for segmenting the ossification center and the cartilage area based on ultrasound imaging also includes the following steps:

[0034] The Dice coefficient is used to evaluate the segmentation results of the ossification center and the cartilage area. The calculation formula is:

[0035]

[0036] Among them, X is the segmentation result, Y is the standard result, |X∩Y| represents the number of elements in the intersection of X and Y; |X| and |Y| represent the number of elements in X and Y respectively.

[0037] The beneficial effect of the above further scheme is that by calculating the Dice coefficient of the segmentation result, the accuracy of the segmentation result can be accurately evaluated, which facilitates the evaluation of the prediction accuracy of the classification network model and the accuracy of the segmentation model, so as to facilitate the subsequent optimization of the classification network model and the segmentation model.

[0038] The present invention also provides a segmentation system for ossification center and cartilage area based on ultrasound images, which is characterized by comprising a preprocessing module, a classification network model module and a segmentation module;

[0039] The preprocessing module is used to obtain the ulna ultrasound image of the sample patient, preprocess the ultrasound image and generate a sample set;

[0040] The classification network model module is used to construct a classification network model with a multi-scale coding structure based on the sample set, and input the sample set and the corresponding ulnar closure type into the classification network model for training;

[0041] The classification network model module is further used to input a target sample set obtained based on the ulnar ultrasound image of the target patient into the trained classification network model to obtain a corresponding target ulnar closure type;

[0042] The segmentation module is used to input the corresponding ulna ultrasound image into the segmentation model for segmentation according to the target ulna closure type, and obtain the segmentation results of the corresponding ossification center and cartilage area respectively.

[0043] On the basis of the above technical solution, the present invention can also be improved as follows:

[0044] Further: the classification network model module inputs the ultrasound images in the sample set and the corresponding ulnar closure types into the classification network model for training in a specific implementation as follows:

[0045] The CDB module in the classification network model performs convolution processing on the feature map, obtains the feature map after the first normalization, and outputs it to the DAB module;

[0046] The DAB module performs convolution, normalization, correction linearization and regularization on the feature map output by the CDB module, and then splices it with the feature map input by the DAB module, and performs convolution, normalization, correction linearization and regularization on the feature map obtained by splicing, and then splices it with the feature map input by the DAB module for the next iteration, until the set number of splicing times is reached, and then outputs the feature map obtained by the last splicing to the MSCM module;

[0047] The MSCM module normalizes and corrects the linear processing of the feature map output by the DAB module, divides it into four paths for multi-scale convolution stacking processing, and outputs it to the DA module;

[0048] The DA module performs convolution and normalization processing on the feature map output by the MSCM module, and calculates the ulna closure probability corresponding to the feature map using an activation function, so as to determine the ulna predicted closure type corresponding to the feature map according to a preset ulna closure probability threshold;

[0049] A cross entropy loss function is constructed, and the loss value between the predicted ulna closure type and the corresponding ulna closure type of all sample patients is calculated using the loss function, and back propagation is performed to adjust the parameters of the classification network model, and the predicted ulna closure type and the corresponding ulna closure type of the sample patient are used as input to calculate the accuracy of the classification network model after adjusting the parameters, and the model parameters with the highest accuracy are used as the parameters of the classification network model to complete the training.

[0050] The beneficial effect of the above further scheme is: the feature map is convolved and standardized by the CDB module to facilitate unified processing by the DAB module, the MSCM module and the DA module, and it is iteratively spliced ​​by the DAB module, and multi-scale convolution superposition processing is performed by the MSCM module, and then the ulna closure probability corresponding to the feature map is calculated by the activation function, the ulna predicted closure type is determined, and back propagation is performed. Finally, the accuracy of the ulna predicted closure type predicted by the classification network model is calculated, and the model parameters with the highest accuracy are used as the optimal model parameters to complete the training, so that the model has a higher recognition accuracy.

[0051] Further: the MSCM module performs scale convolution and stacking processing after normalizing and correcting the feature map output by the DAB module as follows:

[0052] Perform 1*1 convolution processing on the feature map of the first output;

[0053] Perform two consecutive 3x3 convolutions on the feature map of the second output;

[0054] The feature map of the third output is processed by 3x3 convolution and 5*5 convolution in sequence;

[0055] Perform global average pooling on the feature map output by the fourth channel;

[0056] The feature maps of the four outputs are added element by element and output to the DA module.

[0057] The present invention also provides a computer-readable storage medium storing a computer program, and when the computer program is executed by a processor, the method for segmenting the ossification center and the cartilage area based on ultrasound images is implemented.

[0058] The beneficial effects of the present invention are as follows: the method and system for segmenting the ossification center and cartilage area based on ultrasound images of the present invention obtain a sample set by preprocessing the ultrasound image, and input it into a classification network model with the corresponding ulna closure type for training. The trained classification network model identifies the target ultrasound image of the target patient to obtain the target ulna closure type, and then combines the segmentation model to segment the ulna ultrasound image of the target ulna closure type to obtain the segmentation result of the ossification center and the cartilage area. The classification algorithm of multi-scale feature extraction and the automatic segmentation method are combined to better segment the ossification center area while ensuring good classification accuracy, thereby improving the segmentation efficiency and facilitating medical staff to perform subsequent medical treatment. BRIEF DESCRIPTION OF THE DRAWINGS

[0059] Figure 1It is a schematic flow chart of a method for segmenting ossification center and cartilage area based on ultrasound image according to an embodiment of the present invention;

[0060] Figure 2 The figure is a flow chart of a system for segmenting ossification centers and cartilage regions based on ultrasound images according to an embodiment of the present invention. DETAILED DESCRIPTION

[0061] The principles and features of the present invention are described below in conjunction with the accompanying drawings. The examples given are only used to explain the present invention and are not used to limit the scope of the present invention.

[0062] like Figure 1 As shown, a method for segmenting ossification center and cartilage area based on ultrasound image includes the following steps:

[0063] S1: Obtain ulna ultrasound images of sample patients, preprocess the ultrasound images and generate a sample set;

[0064] S2: constructing a classification network model with a multi-scale coding structure based on the sample set, and inputting the sample set and the corresponding ulnar closure type into the classification network model for training;

[0065] S3: inputting the target ultrasound image obtained based on the ulna ultrasound image of the target patient into the trained classification network model to obtain the corresponding target ulna closure type;

[0066] S4: According to the target ulna closure type, the corresponding ulna ultrasound image is input into the segmentation model for segmentation, and the segmentation results of the corresponding ossification center and cartilage area are obtained respectively.

[0067] The ultrasound image-based ossification center and cartilage area segmentation method of the present invention obtains a sample set by preprocessing the ultrasound image, and inputs it into a classification network model with the corresponding ulna closure type for training. The trained classification network model identifies the target ultrasound image of the target patient to obtain the target ulna closure type, and then combines the segmentation model to segment the ulna ultrasound image of the target ulna closure type to obtain the segmentation result of the ossification center and the cartilage area. The classification algorithm of multi-scale feature extraction and the automatic segmentation method are combined to better segment the ossification center area while ensuring good classification accuracy, thereby improving the segmentation efficiency and facilitating medical staff to perform subsequent medical treatment.

[0068] In one or more embodiments of the present invention, the preprocessing of ultrasound images and generating a sample set specifically includes the following steps:

[0069] S11: performing image recognition processing on the ultrasound image, determining that the ulna region in the ultrasound image is the target region, and generating a boundary contour of the target region;

[0070] S12: performing center cropping on the pixel values ​​of the ultrasound image according to a preset pixel threshold range, and scaling the pixel values ​​of the ultrasound image to within [0, 1];

[0071] Here, in order to reduce the impact of the background area on the classification and scale the pixel value range from the original [a, b] to [0, 1], so that the image data remains in a reasonable and easy-to-process numerical range, the specific scaling formula is as follows:

[0072]

[0073] Where I represents the initialized image, a and b are the minimum and maximum values ​​of the original pixel values, respectively. normalized is the scaled image.

[0074] S13: Normalizing the cropped target area image in the ultrasound image according to a preset normalization ratio to generate a sample set.

[0075] Here, the normalized calculation formula is:

[0076]

[0077] Among them, I normalized Represents the grayscale value of the original image pixel, μ and σ are the mean and standard deviation of the grayscale values ​​of all training set image pixels, respectively.

[0078] By performing image recognition processing on the ultrasound image, the ulna area can be accurately identified as the target area, thereby facilitating subsequent cropping and pixel value scaling to improve the generalization ability of the model. By normalizing the target area image in the cropped ultrasound image, the grayscale value of the image can have a zero mean and a standardized distribution, thereby reducing the convergence time in model training and improving the accuracy of the model.

[0079] In one or more embodiments of the present invention, the step of inputting the sample set and the corresponding ulnar closure type into the classification network model for training specifically includes the following steps:

[0080] S21: the CDB module in the classification network model performs 7*7 convolution processing on the ultrasound images in the sample set, obtains a feature map after standardization, and outputs it to the DAB module;

[0081] S22: The DAB module performs 3*3 convolution, normalization, corrected linearization and regularization on the feature map output by the CDB module, and then splices it with the feature map input by the DAB module, and performs 3*3 convolution, normalization, corrected linearization and regularization on the feature map obtained by splicing, and then splices it with the feature map input by the DAB module for the next iteration, until the set number of splicing times is reached (in the embodiment of the present invention, the number of splicing times is set to 7 times), and then outputs the feature map obtained by the last splicing to the MSCM module;

[0082] Here, the dense convolution operation formula of the DAB module is:

[0083] H0=X

[0084] H l =D(LR(BN(Conv([H0,H1,H2,...,H l-1 ]))))l=[1,7]

[0085] H8=BN(LR(Conv([H0,H1,H2,...,H7])))

[0086] Among them, [H0,H1,H2,...,H l-1 ] means concatenating the output feature maps of all previous layers, Conv means convolution operation, LR means LeakyReLU, and BN means batch normalization.

[0087] S23: the MSCM module normalizes and corrects the linear processing of the feature map output by the DAB module, divides the feature map into four paths, performs multi-scale convolution and superposition processing, and outputs the result to the DA module;

[0088] S24: The DA module performs 1*1 convolution and normalization processing on the feature map output by the MSCM module, and uses the activation function to calculate the ulna closure probability corresponding to the feature map, so as to determine the ulna predicted closure type corresponding to the feature map according to the preset ulna closure probability threshold; S25: Construct a cross entropy loss function, use the loss function to calculate the loss value between the ulna predicted closure type and the corresponding ulna closure type of all sample patients, and perform back propagation to adjust the parameters of the classification network model, and use the ulna predicted closure type and the corresponding ulna closure type of the sample patient as input to calculate the accuracy of the classification network model after adjusting the parameters, and use the model parameters with the highest accuracy as the parameters of the classification network model to complete the training. During training, the present invention uses the ADAM optimizer, and the initial learning rate is 1×10 -2 The learning rate attenuation coefficient used in this invention is 3×10 -5, bach_size is set to 16, the present invention sets 200 epochs, and each epoch performs 200 iterations. The present invention uses cross entropy loss to train our network, and the cross entropy loss function is:

[0089]

[0090] Where N is the number of samples, y i is the true label of the sample, p i is the predicted probability of the sample.

[0091] During validation, we use accuracy (ACC) as our evaluation metric. ACC indicates the proportion of samples correctly predicted by the model to the total number of samples. The closer the ACC value is to 1, the better the overall performance of the model. We will save the model with the best ACC during validation for inference. The formula for ACC is as follows:

[0092]

[0093] Among them, TP (True Positive): true positive example, that is, the number of samples correctly predicted as positive, TN (True Negative): true negative example, that is, the number of samples correctly predicted as negative, FP (False Positive): false positive example, that is, the number of negative samples incorrectly predicted as positive, FN (False Negative): false negative example, that is, the number of positive samples incorrectly predicted as negative.

[0094] In the inference stage, the ultrasound image is preprocessed and input into the trained classification network model to obtain the classification probability, and then the classification probability is converted into the classification result to classify the ultrasound image and obtain the corresponding ulnar closure type.

[0095] The CDB module is used to perform convolution and standardization on the feature map to facilitate unified processing by the DAB module, the MSCM module and the DA module. At the same time, the DAB module is used for iterative splicing, and the MSCM module is used for multi-scale convolution superposition processing. The activation function is then used to calculate the ulna closure probability corresponding to the feature map, and the ulna predicted closure type is determined. Back propagation is performed, and finally the accuracy of the ulna predicted closure type predicted by the classification network model is calculated. The model parameters with the highest accuracy are used as the optimal model parameters to complete the training, so that the model has a higher recognition accuracy.

[0096] In one or more embodiments of the present invention, the MSCM module normalizes and corrects the linear processing of the feature map output by the DAB module and then performs scale convolution stacking processing, which specifically includes the following steps:

[0097] S231: Perform 1*1 convolution processing on the feature map output by the first channel;

[0098] S232: Perform two consecutive 3x3 convolution processes on the feature map output by the second path;

[0099] S233: performing 3x3 convolution processing and 5*5 convolution processing on the feature map output by the third path in sequence;

[0100] S234: Perform global average pooling processing on the feature map output by the fourth channel;

[0101] S235: Add the feature maps of the four outputs element by element and output them to the DA module.

[0102] The formulas for the above convolution operations of different scales are:

[0103] y1=C 1×1 (LR(BN(x)))

[0104] y2=C 3×3 (C 3×3 (LR(BN(x))))

[0105] y3=C 5×5 (C 3×3 (LR(BN(x))))

[0106] y4=Avg(LR(BN(x)))

[0107]

[0108] Among them, x is the input feature map data, BN is the batch normalization processing layer, LR is the LeakyReLU activation function, C k×k is a k×k convolution operation, AvgPool is an average pooling operation, It is element-by-element addition processing;

[0109] By performing convolution and global average pooling on the four feature maps respectively, features of different scales can be extracted, and subtle features of ossification centers and cartilage areas can be captured at different scales, thereby ensuring classification accuracy.

[0110] In one or more embodiments of the present invention, the step of inputting the corresponding ulna ultrasound image into the segmentation model for segmentation according to the target ulna closure type, and obtaining the segmentation results of the corresponding ossification center and cartilage area respectively specifically includes the following steps:

[0111] S41: when the target ulna closure type is incomplete ulna closure, the segmentation model obtains two segmentation regions according to the ulna ultrasound image, takes the intersection of the two regions to obtain the ossification center region, and takes the union of the two regions to obtain the cartilage region;

[0112] S42: When the target ulna closure type is complete ulna closure, the segmentation model obtains a segmentation area according to the ulna ultrasound image, and the area is both an ossification center area and a cartilage area.

[0113] By performing segmentation for different target ulna closure types, the ossification center area can be better segmented, effectively solving the problem of low segmentation efficiency of the ossification center when the ulna is fully closed, while also maintaining the segmentation efficiency of the cartilage area.

[0114] It should be pointed out that in the embodiment of the present invention, the segmentation model is based on the ulna ultrasound image, which is a prior art and will not be described in detail in the present invention.

[0115] Optionally, in one or more embodiments of the present invention, the method for segmenting the ossification center and the cartilage area based on ultrasound imaging further includes the following steps:

[0116] S5: The Dice coefficient is used to evaluate the segmentation results of the ossification center and the cartilage area. The calculation formula is:

[0117]

[0118] Among them, X is the segmentation result, Y is the standard result, |X∩Y| represents the number of elements in the intersection of X and Y; |X| and |Y| represent the number of elements in X and Y, respectively. Here, the standard result is the result of comprehensive evaluation by multiple experts in the preset clinical practice.

[0119] By calculating the Dice coefficient of the segmentation result, the accuracy of the segmentation result can be accurately evaluated, which is convenient for evaluating the prediction accuracy of the classification network model and the accuracy of the segmentation model, so as to facilitate the subsequent optimization of the classification network model and the segmentation model.

[0120] The segmentation method of the ossification center and cartilage area based on ultrasound images of the present invention was used to perform segmentation, and the following experimental results were obtained:

[0121] Table 1

[0122]

[0123] Table 2

[0124]

[0125] It can be seen from Tables 1 and 2 that the segmentation results obtained by the segmentation method of the present invention are significantly improved compared with the prior art method of inputting all images into the segmentation model for segmentation. At the same time, the Dice coefficient of the cartilage area segmentation remains almost unchanged compared with the original method, indicating that our method does effectively solve the problem of low segmentation efficiency of the ossification center in the fully closed state of the ulna, while also maintaining the segmentation efficiency of the cartilage area.

[0126] The present invention also provides a segmentation system for ossification center and cartilage area based on ultrasound images, which is characterized by comprising a preprocessing module, a classification network model module and a segmentation module;

[0127] The preprocessing module is used to obtain the ulna ultrasound image of the sample patient, preprocess the ultrasound image and generate a sample set;

[0128] The classification network model module is used to construct a classification network model with a multi-scale coding structure based on the sample set, and input the sample set and the corresponding ulnar closure type into the classification network model for training;

[0129] The classification network model module is further used to input a target sample set obtained based on the ulnar ultrasound image of the target patient into the trained classification network model to obtain a corresponding target ulnar closure type;

[0130] The segmentation module is used to input the corresponding ulna ultrasound image into the segmentation model for segmentation according to the target ulna closure type, and obtain the segmentation results of the corresponding ossification center and cartilage area respectively.

[0131] In one or more embodiments of the present invention, the preprocessing module preprocesses the ultrasound image and generates a sample set, specifically comprising the following steps:

[0132] Performing image recognition processing on the ultrasound image, determining that the ulna region in the ultrasound image is the target region, and generating a boundary contour of the target region;

[0133] Performing center cropping on the pixel values ​​of the ultrasound image according to a preset pixel threshold range, and scaling the pixel values ​​of the ultrasound image to within [0, 1];

[0134] The cropped target area image in the ultrasound image is normalized according to a preset normalization ratio to generate a sample set.

[0135] By performing image recognition processing on the ultrasound image, the ulna area can be accurately identified as the target area, thereby facilitating subsequent cropping and pixel value scaling to improve the generalization ability of the model. By normalizing the target area image in the cropped ultrasound image, the grayscale value of the image can have a zero mean and a standardized distribution, thereby reducing the convergence time in model training and improving the accuracy of the model.

[0136] In one or more embodiments of the present invention, the specific implementation of inputting the ultrasound images in the sample set and the corresponding ulnar closure types into the classification network model for training is as follows:

[0137] The CDB module in the classification network model performs convolution processing on the feature map, obtains the feature map after the first normalization, and outputs it to the DAB module;

[0138] The DAB module performs convolution, normalization, correction linearization and regularization on the feature map output by the CDB module, and then splices it with the feature map input by the DAB module, and performs convolution, normalization, correction linearization and regularization on the feature map obtained by splicing, and then splices it with the feature map input by the DAB module for the next iteration, until the set number of splicing times is reached, and then outputs the feature map obtained by the last splicing to the MSCM module;

[0139] The MSCM module normalizes and corrects the linear processing of the feature map output by the DAB module, divides it into four paths for multi-scale convolution stacking processing, and outputs it to the DA module;

[0140] The DA module performs convolution and normalization processing on the feature map output by the MSCM module, and calculates the ulna closure probability corresponding to the feature map using an activation function, so as to determine the ulna predicted closure type corresponding to the feature map according to a preset ulna closure probability threshold;

[0141] A cross entropy loss function is constructed, and the loss value between the predicted ulna closure type and the corresponding ulna closure type of all sample patients is calculated using the loss function, and back propagation is performed to adjust the parameters of the classification network model, and the predicted ulna closure type and the corresponding ulna closure type of the sample patient are used as input to calculate the accuracy of the classification network model after adjusting the parameters, and the model parameters with the highest accuracy are used as the parameters of the classification network model to complete the training.

[0142] The CDB module is used to perform convolution and standardization on the feature map to facilitate unified processing by the DAB module, the MSCM module and the DA module. At the same time, the DAB module is used for iterative splicing, and the MSCM module is used for multi-scale convolution superposition processing. The activation function is then used to calculate the ulna closure probability corresponding to the feature map, and the ulna predicted closure type is determined. Back propagation is performed, and finally the accuracy of the ulna predicted closure type predicted by the classification network model is calculated. The model parameters with the highest accuracy are used as the optimal model parameters to complete the training, so that the model has a higher recognition accuracy.

[0143] In one or more embodiments of the present invention, the MSCM module performs scale convolution and stacking processing after normalizing and correcting the feature map output by the DAB module as follows:

[0144] Perform 1*1 convolution processing on the feature map of the first output;

[0145] Perform two consecutive 3x3 convolutions on the feature map of the second output;

[0146] The feature map of the third output is processed by 3x3 convolution and 5*5 convolution in sequence;

[0147] Perform global average pooling on the feature map output by the fourth channel;

[0148] The feature maps of the four outputs are added element by element and output to the DA module.

[0149] By performing convolution and global average pooling on the four feature maps respectively, features of different scales can be extracted, and subtle features of ossification centers and cartilage areas can be captured at different scales, thereby ensuring classification accuracy.

[0150] In one or more embodiments of the present invention, the segmentation module inputs the corresponding ulna ultrasound image into the segmentation model for segmentation according to the target ulna closure type, and obtains the segmentation results of the corresponding ossification center and cartilage area respectively, specifically including the following steps:

[0151] When the target ulna closure type is incomplete ulna closure, the segmentation model obtains two segmentation regions according to the ulna ultrasound image, takes the intersection of the two regions to obtain the ossification center region, and takes the union of the two regions to obtain the cartilage region;

[0152] When the target ulna closure type is complete ulna closure, the segmentation model obtains a segmentation area according to the ulna ultrasound image, and the area is both an ossification center area and a cartilage area.

[0153] By performing segmentation for different target ulna closure types, the ossification center area can be better segmented, effectively solving the problem of low segmentation efficiency of the ossification center when the ulna is fully closed, while also maintaining the segmentation efficiency of the cartilage area.

[0154] In one or more embodiments of the present invention, the segmentation of the ossification center and the cartilage area based on ultrasound imaging further includes an evaluation module for evaluating the segmentation results of the ossification center and the cartilage area using the Dice coefficient, and the calculation formula is:

[0155]

[0156] Among them, X is the segmentation result, Y is the standard result, |X∩Y| represents the number of elements in the intersection of X and Y; |X| and |Y| represent the number of elements in X and Y respectively.

[0157] By calculating the Dice coefficient of the segmentation result, the accuracy of the segmentation result can be accurately evaluated, which is convenient for evaluating the prediction accuracy of the classification network model and the accuracy of the segmentation model, so as to facilitate the subsequent optimization of the classification network model and the segmentation model.

[0158] The present invention also provides a computer-readable storage medium storing a computer program, and when the computer program is executed by a processor, the method for segmenting the ossification center and the cartilage area based on ultrasound images is implemented.

[0159] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

Claims

1. A method for segmenting ossification center and cartilage area based on ultrasound image, characterized in that: The steps include: Acquire ulna ultrasound images of sample patients, pre-process the ultrasound images and generate a sample set; Constructing a classification network model with a multi-scale coding structure based on the sample set, and inputting the sample set and the corresponding ulnar closure type into the classification network model for training; Inputting a target ultrasound image obtained based on the ulna ultrasound image of the target patient into the trained classification network model to obtain a corresponding target ulna closure type; According to the target ulna closure type, the corresponding ulna ultrasound image is input into the segmentation model for segmentation, and the segmentation results of the corresponding ossification center and cartilage area are obtained respectively.

2. The method for segmenting ossification center and cartilage area based on ultrasound image according to claim 1, characterized in that: The preprocessing of ultrasound images and generating a sample set specifically includes the following steps: Performing image recognition processing on the ultrasound image, determining that the ulna region in the ultrasound image is the target region, and generating a boundary contour of the target region; Performing center cropping on the pixel values ​​of the ultrasound image according to a preset pixel threshold range, and scaling the pixel values ​​of the ultrasound image to within [0, 1]; The cropped target area image in the ultrasound image is normalized according to a preset normalization ratio to generate a sample set.

3. The method for segmenting ossification center and cartilage area based on ultrasound image according to claim 1, characterized in that: The step of inputting the sample set and the corresponding ulnar closure type into the classification network model for training specifically includes the following steps: The CDB module in the classification network model performs convolution processing on the ultrasound images in the sample set, obtains a feature map after standardization, and outputs it to the DAB module; The DAB module performs convolution, normalization, correction linearization and regularization on the feature map output by the CDB module, and then splices it with the feature map input by the DAB module, and performs convolution, normalization, correction linearization and regularization on the feature map obtained by splicing, and then splices it with the feature map input by the DAB module for the next iteration, until the set number of splicing times is reached, and then outputs the feature map obtained by the last splicing to the MSCM module; The MSCM module normalizes and corrects the linear processing of the feature map output by the DAB module, divides it into four paths for multi-scale convolution stacking processing, and outputs it to the DA module; The DA module performs convolution and normalization processing on the feature map output by the MSCM module, and calculates the ulna closure probability corresponding to the feature map using an activation function, so as to determine the ulna predicted closure type corresponding to the feature map according to a preset ulna closure probability threshold; A cross entropy loss function is constructed, and the loss value between the predicted ulna closure type and the corresponding ulna closure type of all sample patients is calculated using the loss function, and back propagation is performed to adjust the parameters of the classification network model, and the predicted ulna closure type and the corresponding ulna closure type of the sample patient are used as input to calculate the accuracy of the classification network model after adjusting the parameters, and the model parameters with the highest accuracy are used as the parameters of the classification network model to complete the training.

4. The method for segmenting ossification center and cartilage area based on ultrasound image according to claim 3, characterized in that: The MSCM module performs scale convolution and stacking processing on the feature map output by the DAB module after normalization and correction linear processing, which specifically includes the following steps: Perform 1*1 convolution processing on the feature map of the first output; Perform two consecutive 3x3 convolutions on the feature map of the second output; The feature map of the third output is processed by 3x3 convolution and 5*5 convolution in sequence; Perform global average pooling on the feature map output by the fourth channel; The feature maps of the four outputs are added element by element and output to the DA module.

5. The method for segmenting ossification center and cartilage area based on ultrasound image according to claim 1, characterized in that: The step of inputting the corresponding ulna ultrasound image into the segmentation model for segmentation according to the target ulna closure type, and obtaining the segmentation results of the corresponding ossification center and cartilage area respectively specifically includes the following steps: When the target ulna closure type is incomplete ulna closure, the segmentation model obtains two segmentation regions according to the ulna ultrasound image, takes the intersection of the two regions to obtain the ossification center region, and takes the union of the two regions to obtain the cartilage region; When the target ulna closure type is complete ulna closure, the segmentation model obtains a segmentation area according to the ulna ultrasound image, and the area is both an ossification center area and a cartilage area.

6. The method for segmenting ossification center and cartilage area based on ultrasound image according to any one of claims 1 to 5, characterized in that: The following steps are also included: The Dice coefficient is used to evaluate the segmentation results of the ossification center and the cartilage area. The calculation formula is: Among them, X is the segmentation result, Y is the standard result, |X∩Y| represents the number of elements in the intersection of X and Y; |X| and |Y| represent the number of elements in X and Y respectively.

7. A system for segmenting ossification centers and cartilage regions based on ultrasound images, characterized in that: It includes a preprocessing module, a classification network model module and a segmentation module; The preprocessing module is used to obtain the ulna ultrasound image of the sample patient, preprocess the ultrasound image and generate a sample set; The classification network model module is used to construct a classification network model with a multi-scale coding structure based on the sample set, and input the sample set and the corresponding ulnar closure type into the classification network model for training; The classification network model module is further used to input a target sample set obtained based on the ulnar ultrasound image of the target patient into the trained classification network model to obtain a corresponding target ulnar closure type; The segmentation module is used to input the corresponding ulna ultrasound image into the segmentation model for segmentation according to the target ulna closure type, and obtain the segmentation results of the corresponding ossification center and cartilage area respectively.

8. The ultrasound image-based ossification center and cartilage area segmentation system according to claim 7, characterized in that: The classification network model module inputs the sample set and the corresponding ulnar closure type into the classification network model for training in a specific manner as follows: The CDB module in the classification network model performs convolution processing on the ultrasound images in the sample set, obtains a feature map after standardization, and outputs it to the DAB module; The DAB module performs convolution, normalization, correction linearization and regularization on the feature map output by the CDB module, and then splices it with the feature map input by the DAB module, and performs convolution, normalization, correction linearization and regularization on the feature map obtained by splicing, and then splices it with the feature map input by the DAB module for the next iteration, until the set number of splicing times is reached, and then outputs the feature map obtained by the last splicing to the MSCM module; The MSCM module normalizes and corrects the linear processing of the feature map output by the DAB module, divides it into four paths for multi-scale convolution stacking processing, and outputs it to the DA module; The DA module performs convolution and normalization processing on the feature map output by the MSCM module, and calculates the ulna closure probability corresponding to the feature map using an activation function, so as to determine the ulna predicted closure type corresponding to the feature map according to a preset ulna closure probability threshold; A cross entropy loss function is constructed, and the loss value between the predicted ulna closure type and the corresponding ulna closure type of all sample patients is calculated using the loss function, and back propagation is performed to adjust the parameters of the classification network model, and the predicted ulna closure type of the sample patient and the corresponding ulna closure type are used as input to calculate the accuracy of the classification network model, and the model parameters with the highest accuracy are used as the parameters of the classification network model to complete the training.

9. The ultrasound image-based ossification center and cartilage area segmentation system according to claim 7, characterized in that: The specific implementation of the MSCM module performing scale convolution and stacking processing after normalizing and correcting the feature map output by the DAB module is as follows: Perform 1*1 convolution processing on the feature map of the first output; Perform two consecutive 3x3 convolutions on the feature map of the second output; The feature map of the third output is processed by 3x3 convolution and 5*5 convolution in sequence; Perform global average pooling on the feature map output by the fourth channel; The feature maps of the four outputs are added element by element and output to the DA module.

10. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, the method for segmenting the ossification center and the cartilage area based on ultrasound images according to any one of claims 1 to 6 is implemented.

Images