Standard section automatic identification method based on 3D ultrasonic data

Abstract

The invention discloses a standard section automatic identification method based on 3D ultrasonic data, and belongs to the technical field of medical imaging, and the method comprises the following steps: obtaining a 2D ultrasonic standard section image and 3D ultrasonic data; the obtained 2D standard section image is preprocessed, and corresponding parameters are set according to classification of different ultrasonic acquisition devices for batch automatic cutting; and meanwhile, automatic cutting is realized by adopting a maximum connected domain method, so that interference of a pure color background and an equipment interface is removed, and pure standard section data is obtained. According to the method, anatomical information is provided by utilizing 3D ultrasonic data, the neural network is guided to realize efficient standard section automatic identification, interference of a background and an equipment interface is removed by designing an automatic cutting method and a 2D data expansion method, a large number of 2D slice images are generated, and the richness of training data is increased. And a VGG16 and OC-CNN design network is combined, and the preprocessed and expanded 2D image is trained to obtain a model for automatically identifying the standard section.

Description

Technical Field

[0001] The present invention belongs to the field of medical imaging technology, and more specifically, relates to a method for automatically identifying standard sections based on 3D ultrasound data. Background Art

[0002] Standard sections refer to those anatomical planes that have been widely accepted and clearly defined. These planes can provide key information about organs or lesions and help doctors make accurate diagnoses. In the ultrasound examination of developmental dislocation of the hip (DDH), the identification, angle measurement and diagnostic classification of standard sections are particularly important. The standard section image of the hip joint must include key anatomical structures such as the straight ilium, the lower edge of the ilium, the labrum and the bone-cartilage junction. Obtaining standard sections helps doctors evaluate the structure and functional status of the hip joint. Therefore, the identification of standard sections requires not only the operator to have rich experience and knowledge, but also highly accurate hand-eye coordination. Given the importance of standard sections to diagnosis, ensuring that these sections can be obtained stably and reliably for each examination has become a basic requirement of medical imaging technology.

[0003] However, in practical applications, the identification and extraction of standard sections face many difficulties: First, standard sections are difficult to obtain, and misjudgment and missed judgments are prone to occur. This is mainly because during ultrasound screening, infants are prone to movement, which increases the difficulty of scanning the standard section; in addition, the judgment criteria for standard sections are strict, requiring high professional knowledge and clinical experience of doctors, and it takes a long time to obtain standard sections. Secondly, parameter measurement has a strong subjective dependence. Due to the presence of noise and artifacts in ultrasound images, the quality is poor, and doctors rely on their own experience to judge the morphology of anatomical structures, resulting in large differences in measurement results between different observers.

[0004] With the development of 3D ultrasound technology, the field of medical imaging has been able to quickly obtain 3D data of the target area, which is a huge improvement for clinical diagnosis. 3D ultrasound not only provides more intuitive and three-dimensional image information, but also allows doctors to observe lesions or organs from multiple angles, enhancing the comprehensiveness and accuracy of diagnosis. However, despite the many advantages brought by 3D ultrasound technology, it is still a challenge to efficiently and accurately identify standard sections from these 3D data. Standard sections are key references in the diagnostic process, and their accurate identification is crucial for subsequent analysis and treatment planning. Summary of the invention

[0005] In response to the defects of the prior art and the need for improvement, the present invention proposes a deep learning method for automatically identifying standard sections using 3D ultrasound data. The method can directly input 3D ultrasound data and train a neural network model to automatically classify standard sections and non-standard sections in the ultrasound data, and identify and extract the required standard sections. This automated method can not only assist doctors in conducting examinations more quickly and reduce the time required for manual positioning, but also improve the consistency and accuracy of diagnosis, especially when dealing with complex cases or when multiple repeated examinations are required.

[0006] To achieve the above object, the present invention provides the following technical solution: a deep learning method for automatically identifying standard sections using 3D ultrasound data, comprising the following steps:

[0007] (1-1) Acquire 2D ultrasound standard section images and 3D ultrasound data;

[0008] (1-2) Preprocessing the 2D standard section images obtained in step (1-1), setting corresponding parameters according to different ultrasonic acquisition equipment classifications for automatic batch cutting; at the same time, using the maximum connected domain method to achieve automatic cutting, thereby removing the interference of pure color background and device interface, thereby obtaining pure standard section data;

[0009] (1-3) A method for expanding 2D data using 3D data, wherein a series of 2D images are collected at certain intervals for the 3D ultrasound data obtained in step (1-1), and after screening, as comprehensive as possible available non-standard section data is obtained;

[0010] (1-4) using a binary classification network to process the data obtained in steps (1-2) and (1-3), dividing the data into a training set and a test set, inputting the 2D standard cross-sectional images and the 2D non-standard cross-sectional images and their corresponding labels into the network, and training the network after preprocessing and data enhancement to obtain a network model that can output a predicted label for the input 2D image, testing the model using a test set, and evaluating the performance of the model;

[0011] (1-5) In the application stage, the 3D data of the target area is directly collected using a 3D ultrasound device, and then 2D slice processing is performed according to the method described in step (1-3). These slices are input into the trained model for classification, and finally the goal of automatically identifying standard sections from the 3D ultrasound data is achieved.

[0012] As a further preferred embodiment of the present invention, in step (1-2), the 2D standard section image data is derived from the ultrasound PACS system, and the data is a screenshot image with a solid color background and a device UI interface; the specific steps of the automatic cutting method are as follows:

[0013] (2-1) Read all 2D image files in the specified folder;

[0014] (2-2) Convert the RGB image to a grayscale image using the rgb2gray function to simplify subsequent processing;

[0015] (2-3) Calculate the minimum pixel value in the grayscale image and add a fixed threshold increment to determine the threshold for binarization. Use the imbinarize function to perform binarization and convert the grayscale image into a black and white image.

[0016] (2-4) Apply connected component analysis to find all connected regions in the image, and use the regionprops function to obtain the area and bounding box information of each connected region. Select the connected region with the largest area as the target region and record its bounding box.

[0017] (2-5) Perform morphological processing on the largest connected region, including filling holes and opening operations, to remove noise from the image;

[0018] (2-6) Use the find_largest_square function to find the largest square area. This function determines the boundary of the target area by filling holes, removing noise, and finding the largest connected area. It converts the determined boundary into a coordinate format suitable for the imcrop function, namely [x, y, width, height], and crops the original image. The cropped image is saved to the specified output folder.

[0019] As a further preferred embodiment of the present invention, in the step (1-3), the 2D data expansion method of the 3D data is based on the corresponding spatial relationship between the pixel points in the section two-dimensional image and the voxels in the original three-dimensional image, and uses mathematical tools such as rotation matrix, quaternion and trilinear interpolation to determine the gray value of each pixel in the two-dimensional image of any section, so as to obtain the section image stored in BMP format. The specific steps of the data expansion method are as follows:

[0020] (3-1) Determine the size of the slice;

[0021] (3-2) Determine the reference surface;

[0022] (3-3) Determine the spatial relationship between the reference plane and the original three-dimensional image and introduce the parameter r;

[0023] (3-4) Determine the spatial relationship between any other plane and the original three-dimensional image. Any other plane can be obtained by translating and rotating the reference plane. The translation is represented by the parameter r, and the rotation is represented by the other two parameters, the tilt angle θ and the azimuth angle φ, thereby obtaining a quaternion representation, which is then converted into a rotation matrix.

[0024] (3-5) The grayscale value of the pixel on the slice is determined based on the correspondence between the voxel and the grayscale value. Trilinear interpolation is used to traverse row by row and column by column starting from the first point on the slice to obtain the information of the entire two-dimensional image.

[0025] As a further preferred embodiment of the present invention, in step (1-4), the binary classification network combines the methods of VGG16 and OC-CNN to process the data obtained in steps (1-2) and (1-3), and the specific network structure is as follows:

[0026] (4-1) Use the feature extraction part of VGG16 to vector encode the input data;

[0027] (4-2) concatenating the vector in step (4-1) with the pseudo-negative vector obtained by adding Gaussian noise;

[0028] (4-3) Input the vector obtained in step (4-2) into the custom classifier for classification.

[0029] In general, the above technical solution conceived by the present invention has the following beneficial effects compared with the prior art:

[0030] 1. The automatic cutting method provided by the present invention can effectively remove irrelevant information from the original 2D standard section image with background and device user interface, and retain key anatomical structures. Compared with manual cutting, this method not only greatly improves efficiency and reduces human errors, but also ensures the consistency and reliability of cutting results through an automated process.

[0031] 2. The 2D data expansion method proposed in the present invention can generate a large number of 2D slice images from existing 3D ultrasound data without increasing additional acquisition costs, significantly increasing the diversity and richness of the training data set. This method not only improves the effect of subsequent model training and ensures the quality of the generated 2D image data, but can also be directly applied in the subsequent automatic identification of standard sections in 3D ultrasound data.

[0032] 3. By combining the methods of VGG16 and OC-CNN, the binary classification network designed by the present invention effectively improves the ability to distinguish between standard and non-standard sections. The pre-trained VGG16 feature extraction part can learn rich image features and enhance the generalization ability of the model in new tasks. OC-CNN solves the problem of imbalance between positive and negative samples, improves the performance of the classifier on unbalanced data sets, increases the generalization ability of the model, and has good robustness.

[0033] 4. Introducing neural networks and 3D ultrasound data into the problem of automatic identification of medical standard sections provides a deep learning method for automatically identifying standard sections from 3D ultrasound data. Through efficient preprocessing, rich training data and powerful classification capabilities, it can provide more reliable technical support for ultrasound examination of hip dysplasia, thereby improving the accuracy and reliability of diagnosis, improving the work efficiency of doctors, and providing a solid scientific basis for clinical diagnosis. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] The present invention will now be described by way of example with reference to the accompanying drawings, in which:

[0035] Figure 1 A flow chart of a method for automatically identifying standard sections using 3D ultrasound data provided by the present invention;

[0036] Figure 2 A schematic diagram of a method for expanding 2D data using 3D data provided by the present invention;

[0037] Figure 3 A user interface diagram of the 2D slice generation program provided by the present invention;

[0038] Figure 4 This is a diagram showing the application results of the standard section automatic recognition program based on 3D ultrasound data provided by the present invention. DETAILED DESCRIPTION

[0039] Any feature disclosed in this specification, unless otherwise stated, can be replaced by other equivalent or alternative features with similar purposes. That is, unless otherwise stated, each feature is only an example of a series of equivalent or similar features.

[0040] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0041] The process of the method for automatic identification of standard sections based on 3D ultrasound data provided by the present invention is as follows: Figure 1 As shown, the specific steps include:

[0042] (1-1) Obtaining 2D ultrasound standard section images and 3D ultrasound data:

[0043] Obtain 2D standard cross-sectional images from the hospital's ultrasound PACS system, which are usually screenshots with a solid color background and device UI interface. Use a 3D ultrasound device to acquire 3D ultrasound data of the target area, ensuring that the data covers the required anatomical structure area.

[0044] (1-2) Preprocessing the 2D standard section images obtained in step (1-1), setting corresponding parameters according to different ultrasonic acquisition equipment classifications for batch automatic cutting; at the same time, using the maximum connected domain method to achieve automatic cutting, thereby removing the interference of pure color background and device interface, thereby obtaining pure standard section data:

[0045] For 2D standard section images obtained from the ultrasound PACS system, an automatic cutting method is used to remove irrelevant information. The automatic cutting method reads all image files in the specified folder and performs preprocessing operations on each 2D image, including conversion to grayscale images, binarization, edge detection, target area determination, maximum connected area extraction and square area search, and target area cropping and saving. Specifically, the RGB image is first converted to a grayscale image, and then binarized based on the minimum pixel value of the grayscale image plus a fixed threshold; then, the connected component analysis is used to find the largest connected area and determine the bounding box of the area; then, unnecessary noise is removed through morphological operations; finally, the target area is cropped by finding the largest square area, and the processed image is saved to the specified folder. The specific steps are as follows:

[0046] (2-1) Read all 2D image files in the specified folder;

[0047] (2-2) Convert the RGB image to a grayscale image using the rgb2gray function to simplify subsequent processing;

[0048] (2-3) Calculate the minimum pixel value in the grayscale image and add a fixed threshold increment (the default is 5) to determine the threshold for binarization. Use the imbinarize function to perform binarization and convert the grayscale image into a black and white image;

[0049] (2-4) Apply connected component analysis (bwconncomp) to find all connected regions in the image, and use the regionprops function to obtain the area and bounding box information of each connected region. Select the connected region with the largest area as the target region and record its bounding box;

[0050] (2-5) Perform morphological processing on the largest connected area, including filling holes (imfill) and opening operations (imopen) to remove noise in the image;

[0051] (2-6) Use the find_largest_square function to find the largest square area. This function determines the boundary of the target area by filling holes, removing noise, and finding the largest connected area. Convert the determined boundary into a coordinate format suitable for the imcrop function, that is, [x, y, width, height], and crop the original image. Save the cropped image to the specified output folder.

[0052] The automatic cutting method provided by the present invention can effectively remove irrelevant information from the original 2D standard section image with background and device user interface, and retain key anatomical structures. The method ensures accurate extraction of the target area through steps such as image binarization, maximum connected domain detection and morphological operations. Compared with manual cutting, this method not only greatly improves efficiency and reduces human errors, but also ensures the consistency and reliability of cutting results through automated processes. Especially when processing large amounts of image data, the advantages of the automatic cutting method are more obvious. It can provide doctors with consistent and reliable images, thereby supporting more accurate diagnostic decisions.

[0053] (1-3) A method for expanding 2D data using 3D data, wherein a series of 2D images are collected at certain intervals for the 3D ultrasound data obtained in step (1-1), and after screening, as comprehensive as possible available non-standard section data is obtained;

[0054] The data method is based on the corresponding spatial relationship between the pixel points in the two-dimensional image of the section and the voxels in the original three-dimensional image, and uses mathematical tools such as rotation matrix, quaternion and trilinear interpolation to determine the grayscale value of each pixel in the two-dimensional image of any section, so as to obtain the section image stored in BMP format, and the plane with the positive direction of the Z axis as the normal vector and passing through the origin of the three-dimensional coordinate system is taken as the initial plane, that is, the reference plane. The original three-dimensional image voxel corresponding to each pixel on this reference plane is obtained through the corresponding spatial relationship between the two-dimensional data and the three-dimensional data, and then the rotation matrix of the target plane is determined by the given parameters, and then the original three-dimensional image voxel corresponding to each pixel on the target plane can be obtained, so that the grayscale value of any pixel on any section can be confirmed, and finally the 2D image data is obtained.

[0055] 2D data expansion, according to Figure 2 The schematic diagram shown in the figure determines the size and reference plane of the slice, as well as how to generate other arbitrary planes through rotation and translation. Analyzing the numbers in the figure, ① the cube represents a cube containing a three-dimensional solid; ② and ③ dotted boxes are slice planes of set size; ④ the dotted box is the reference plane, and the corresponding normal vector is the positive Z axis (⑷ vector); ⑤ the dotted box is a rotated plane, and the corresponding normal vector is ⑸ vector; the three-dimensional coordinate axis is set at the center of the cube, and the specific steps are as follows:

[0056] (3-1) Determine the size of the slice (usually a square with equal height and width). To ensure that the effective information of the target slice is not lost and that it is easy to observe, a moderate size needs to be selected. In this example, the largest diagonal of the three faces of the rectangular block containing the three-dimensional image is selected;

[0057] (3-2) Determine the reference plane. The choice of the reference plane is arbitrary, but it is best to be one of the three vertical planes, because the positions of these planes are relatively regular, and it is easy to find the spatial relationship between them and the three-dimensional image. In this example, the plane with the Z axis (0,0,1) as the normal vector is selected as the reference plane. The advantage of this is that the position of the reference plane is relatively simple and fixed, which is convenient for subsequent calculations and processing;

[0058] (3-3) Determine the spatial relationship between the reference plane and the original three-dimensional image. The three-dimensional coordinates of the three-dimensional image corresponding to any point P(i,j) on the reference plane (0≤i≤H, 0≤j≤W, H is the height of the slice, W is the width of the slice) are (iH / 2,jW / 2,r), where r is the distance from the plane to the origin, which is 0 on the reference plane, thus introducing the first parameter r;

[0059] (3-4) Determine the spatial relationship between any other plane and the original three-dimensional image. Any other plane can be obtained by translation and rotation of the reference plane. Translation is represented by parameter r, and rotation is represented by the other two parameters. The rotation of a plane is actually the rotation of its normal vector. The reference vector is the positive direction of the Z axis, and the target vector (sinθcosφ, sinθsinφ, cosθ) can be obtained from the last two parameters, the tilt angle (elevation angle) θ and the azimuth angle φ. Calculate the outer product of the reference vector and the target vector to get the rotation axis, and the rotation angle is the tilt angle θ, thereby obtaining a quaternion representation, which is then converted into a rotation matrix;

[0060] (3-5) Determine the grayscale value of the pixel on the slice based on the correspondence between the voxel and the grayscale value. This example uses trilinear interpolation to traverse row by row and column by column starting from the first point on the slice to obtain the information of the entire two-dimensional image, generate a two-dimensional image and save it;

[0061] (3-6) In order to generate multiple sets of continuously changing slices, the present invention sets a "cut a series of pictures" function mode. In this mode, nine parameters need to be input: the initial value, step size, and end value of the distance r, the tilt angle θ, and the azimuth angle φ. Through these parameters, multiple sets of continuously changing slices can be generated, thereby realizing the expansion of 2D data. Figure 3 As shown, the user can operate through the user interface of the 2D slice generation program.

[0062] The 2D data expansion method provided by the present invention can generate a large number of 2D slice images from existing 3D ultrasound data without increasing additional acquisition costs, significantly increasing the diversity and richness of the training data set. This method not only improves the effect of subsequent model training, but also ensures the quality of the generated 2D image data, making it more suitable for standard section recognition tasks. In addition, through precise mathematical tools and algorithms, it is ensured that each generated 2D image can accurately reflect the information of the original 3D data, thereby improving the accuracy and reliability of diagnosis.

[0063] (1-4) using a binary classification network to process the data obtained in steps (1-2) and (1-3), dividing the data into a training set and a test set, inputting the 2D standard cross-sectional images and the 2D non-standard cross-sectional images and their corresponding labels into the network, and training the network after preprocessing and data enhancement to obtain a network model that can output a predicted label for the input 2D image, testing the model using a test set, and evaluating the performance of the model;

[0064] The binary classification network combines the methods of VGG16 and OC-CNN. Specifically, the feature extraction part of VGG16 is used to vectorize the input data, which is then concatenated with the pseudo-negative vector encoding obtained by adding Gaussian noise, and then input into the classifier for classification. The VGG16 part has been pre-trained on the ImageNet 1K dataset.

[0065] In order to realize the automatic identification of standard and non-standard sections, the present invention designs a binary classification network that combines the VGG16 (Visual Geometry Group 16-layer model) feature extraction part and the OC-CNN (One-Class Convolutional Neural Network) method. The specific steps are as follows:

[0066] (4-1) Use the pre-trained VGG16 model as a feature extractor. In order to meet the needs of this task, the classification layer of VGG16 is removed, and only the first two fully connected layers are retained. Based on the feature extraction part of VGG16, a custom classifier is added. The classifier consists of multiple fully connected layers, which gradually reduce the feature dimension and finally output the classification result;

[0067] (4-2) During the training process, in order to enhance the robustness of the model, pseudo-negative vectors are generated. The specific method is to add Gaussian noise to each positive sample feature vector to generate the same number of pseudo-negative vectors. These pseudo-negative vectors are concatenated with the positive sample feature vector to form a new input;

[0068] (4-3) Read and preprocess the preprocessed 2D standard cross-section images and 2D non-standard cross-section images and their corresponding labels. Preprocessing includes image normalization and data enhancement (such as random rotation, flipping, etc.). Divide the processed data into training sets and test sets to ensure the diversity and representativeness of the data set;

[0069] (4-4) Define the model, loss function, and optimizer for training. Use the Adam optimizer and set the initial learning rate to 0.0001. The learning rate scheduler uses StepLR, which decays by 0.94 times every epoch. Use PyTorch's DataLoader to load the training dataset, set the batch size, enable data shuffling, and discard the last incomplete batch. In each epoch, call the function for training and update the model parameters. After training is completed, call the learning rate scheduler to update the learning rate;

[0070] The binary classification network designed by the present invention effectively improves the ability to distinguish between standard sections and non-standard sections by combining the methods of VGG16 and OC-CNN. The feature extraction part of VGG16 has been pre-trained on the ImageNet1K data set, which enables the model to learn rich image features and generalize better in new tasks. OC-CNN enables the network to learn good decision boundaries when the positive and negative training samples are unbalanced. This method not only enhances the model's ability to recognize standard sections, but also improves the classifier's ability to discriminate non-standard sections, improves the performance of the model on unbalanced data sets, and can achieve effective classification of new samples while maintaining high accuracy, thereby providing more reliable technical support for ultrasonic examination of hip dysplasia.

[0071] (1-5) In the application stage, the 3D data of the target area is directly collected using a 3D ultrasound device, and then 2D slice processing is performed according to the method described in step (1-3). These slices are input into the trained model for classification, and finally the goal of automatically identifying standard sections from 3D ultrasound data is achieved. Figure 4 The results of the application of the automatic recognition procedure of standard slices based on 3D ultrasound data are presented.

[0072] The present invention is not limited to the above-mentioned specific embodiments, but extends to any new features or any new combination disclosed in this specification, as well as any new method or process steps or any new combination disclosed.

Claims

1. A method for automatically identifying standard sections based on 3D ultrasound data, characterized in that: The method includes: (1-1) Acquire 2D ultrasound standard section images and 3D ultrasound data; (1-2) Preprocessing the 2D standard section images obtained in step (1-1), setting corresponding parameters according to different ultrasonic acquisition equipment classifications for batch automatic cutting; at the same time, using the maximum connected domain method to achieve automatic cutting, thereby removing interference from pure color background and device interface, thereby obtaining pure standard section data; (1-3) A method for expanding 2D data using 3D data, wherein a series of 2D images are collected at certain intervals for the 3D ultrasound data obtained in step (1-1), and after screening, as comprehensive as possible available non-standard section data is obtained; (1-4) using a binary classification network to process the data obtained in steps (1-2) and (1-3), dividing the data into a training set and a test set, inputting the 2D standard cross-sectional images and the 2D non-standard cross-sectional images and their corresponding labels into the network, and training the network after preprocessing and data enhancement to obtain a network model that can output a predicted label for the input 2D image, testing the model using a test set, and evaluating the performance of the model; (1-5) In the application stage, the 3D data of the target area is directly collected using a 3D ultrasound device, and then 2D slice processing is performed according to the method described in step (1-3). These slices are input into the trained model for classification, and finally the goal of automatically identifying standard sections from the 3D ultrasound data is achieved.

2. The method for automatically identifying standard sections based on 3D ultrasound data according to claim 1, characterized in that: In the step (1-2), the 2D standard section image data is derived from the ultrasound PACS system, and the data is a screenshot image with a solid color background and a device UI interface; the specific steps of the automatic cutting method are as follows: (2-1) Read all 2D image files in the specified folder; (2-2) Convert the RGB image to a grayscale image using the rgb2gray function to simplify subsequent processing; (2-3) Calculate the minimum pixel value in the grayscale image and add a fixed threshold increment to determine the threshold for binarization. Use the imbinarize function to perform binarization and convert the grayscale image into a black and white image. (2-4) Apply connected component analysis to find all connected regions in the image, and use the regionprops function to obtain the area and bounding box information of each connected region. Select the connected region with the largest area as the target region and record its bounding box. (2-5) Perform morphological processing on the largest connected region, including filling holes and opening operations, to remove noise from the image; (2-6) Use the find_largest_square function to find the largest square area. This function determines the boundary of the target area by filling holes, removing noise, and finding the largest connected area. It converts the determined boundary into a coordinate format suitable for the imcrop function, namely [x, y, width, height], and crops the original image. The cropped image is saved to the specified output folder.

3. The method for automatically identifying standard sections based on 3D ultrasound data according to claim 1, characterized in that: In the step (1-3), the 2D data expansion method of the 3D data is based on the corresponding spatial relationship between the pixel points in the section two-dimensional image and the voxels in the original three-dimensional image, and uses mathematical tools such as rotation matrix, quaternion and trilinear interpolation to determine the gray value of each pixel in the two-dimensional image of any section, so as to obtain the section image stored in BMP format. The specific steps of the data expansion method are as follows: (3-1) Determine the size of the slice; (3-2) Determine the reference surface; (3-3) Determine the spatial relationship between the reference plane and the original three-dimensional image and introduce the parameter r; (3-4) Determine the spatial relationship between any other plane and the original three-dimensional image. Any other plane can be obtained by translating and rotating the reference plane. The translation is represented by the parameter r, and the rotation is represented by the other two parameters, the tilt angle θ and the azimuth angle φ, thereby obtaining a quaternion representation, which is then converted into a rotation matrix. (3-5) The grayscale value of the pixel on the slice is determined based on the correspondence between the voxel and the grayscale value. Trilinear interpolation is used to traverse row by row and column by column starting from the first point on the slice to obtain the information of the entire two-dimensional image.

4. The method for automatically identifying standard sections based on 3D ultrasound data according to claim 1, characterized in that: In the step (1-4), the binary classification network combines the methods of VGG16 and OC-CNN to process the data obtained in steps (1-2) and (1-3). The specific network structure is as follows: (4-1) Use the feature extraction part of VGG16 to vector encode the input data; (4-2) concatenating the vector in step (4-1) with the pseudo-negative vector obtained by adding Gaussian noise; (4-3) Input the vector obtained in step (4-2) into the custom classifier for classification.

Images