Automatic identification system and method for fetal delivery head position

Abstract

The invention provides an automatic recognition system and method for a fetal delivery head position, and relates to the technical field of fetal delivery head position recognition, and the method comprises the steps: obtaining a head ultrasonic image of a fetus through a perineum ultrasonic device in a production process; and inputting the head ultrasound image into a pre-constructed head position classification model group to identify and obtain the head position type of the fetus in the head ultrasound image. The method has the beneficial effects that the head position of the fetus is identified by acquiring the head ultrasonic image of the fetus through the perineum ultrasonic equipment, and compared with a method for identifying the head position according to abdominal ultrasound in the traditional technology, the method not only shows the same feasibility and accuracy, but also has unique advantages in evaluating the standing position of the head position of the fetus. Moreover, the head position classification model group is adopted in the system to automatically identify the head ultrasonic image, so that the system can be used even if an operator lacks fetus brain imaging experience, the use threshold is greatly reduced, and the use range is greatly expanded.

Description

Technical Field

[0001] The present invention relates to the technical field of fetal birth head position recognition, and in particular to an automatic recognition system and method for a fetal birth head position. Background Art

[0002] During delivery, correctly identifying the fetal head position is crucial for a smooth delivery. Malposition of the fetal head may lead to improper placement of midwifery tools and incorrect traction direction, thereby increasing the risk of failed delivery and adverse neonatal outcomes. Therefore, how to accurately identify the fetal head position in the second stage of delivery has become an urgent problem to be solved. Traditionally, clinicians determine the fetal head position by manual transvaginal palpation, but this method has a low accuracy rate and is often affected by the examiner's experience.

[0003] In recent years, transabdominal ultrasound has gradually become the gold standard for assessing the position of the fetal head during delivery. Compared with traditional manual palpation methods, ultrasound technology can provide higher accuracy. However, although transabdominal ultrasound is widely used in clinical practice, in some cases, transperineal ultrasound as an alternative method has shown the same feasibility and accuracy, especially in assessing the position of the fetal head. Transperineal ultrasound can not only simultaneously assess the rotation angle of the fetal head position, but also display the position of the fetal head in the birth canal in real time, which is crucial for determining whether midwifery intervention is needed.

[0004] However, the application of transperineal ultrasound is challenging, especially for operators who lack experience in fetal brain imaging. Accurately identifying anatomical landmarks in the fetal brain is a challenging task. Summary of the invention

[0005] In view of the problems existing in the prior art, the present invention provides an automatic identification system for the head position of a fetus during delivery, comprising:

[0006] An image acquisition module, used to obtain an ultrasound image of the fetus's head through a transperineal ultrasound device during the delivery process;

[0007] The head position recognition module is connected to the image acquisition module and is used to input the head ultrasound image into a pre-constructed head position classification model group to identify the head position type of the fetus in the head ultrasound image.

[0008] Preferably, it also includes a model training module, connected to the head position recognition module, including:

[0009] A training data acquisition unit, configured to acquire transabdominal ultrasound images of a fetus through a transabdominal ultrasound device and transperineal ultrasound images of a fetus through a transperineal ultrasound device during multiple production processes, and associate the transabdominal ultrasound images and the transperineal ultrasound images of the same fetus into a group of images;

[0010] a data annotation unit connected to the training data acquisition unit, and used for annotating, for each group of the image groups, the rotation angle of the fetal head in the transperineal ultrasound image and the position of the birth canal according to the fetal head position in the transabdominal ultrasound image, and adding each of the annotated image groups to the ultrasound training data set;

[0011] The model training unit is connected to the data annotation unit and is used to input the ultrasound training data set into a pre-built initial recognition model for training to obtain the head position classification model group.

[0012] Preferably, the structure of the initial recognition model includes:

[0013] An input preprocessing module is used to divide each image group in the input ultrasound training data set into image blocks of different scales and perform preprocessing;

[0014] A multi-scale feature extraction module, connected to the input preprocessing module, for extracting features of each of the preprocessed image blocks using a standard CNN model, and fusing all the features;

[0015] A dynamic weight adjustment module, connected to the multi-scale feature extraction module, is used to perform self-attention calculation on the fused features and introduce a dynamic weight adjustment mechanism to perform dynamic weight adjustment;

[0016] A feature enhancement module, connected to the dynamic weight adjustment module, for enhancing the features calculated by self-attention;

[0017] A classification module is connected to the feature enhancement module and is used to classify the features after enhancement processing.

[0018] Preferably, it also includes an image processing module, connected to the image acquisition module and the head position recognition module, including:

[0019] A denoising unit, configured to remove random noise in the head ultrasound image by using a denoising algorithm;

[0020] A standardization unit, connected to the denoising unit, for performing standardization processing on the denoised head ultrasound image;

[0021] The image enhancement unit is connected to the standardization unit and is used to perform data enhancement on the standardized head ultrasound image by using a data enhancement technology, and finally update the head ultrasound image to the head ultrasound image after data enhancement.

[0022] Preferably, the head position classification model group includes a first type recognition model, a second type recognition model and a third type recognition model, and the model training unit includes:

[0023] a classification subunit, for dividing each of the image groups that have been labeled in the ultrasound training data set into an occipital anterior image group, an occipital posterior image group and an occipital transverse image group according to the labeled birth canal position;

[0024] A first training subunit, connected to the classification subunit, is used to input all the occipital-anterior image groups into a pre-built initial recognition model for training to obtain the first type of recognition model;

[0025] A second training subunit, connected to the classification subunit, is used to input all the occipital-posterior image groups into a pre-built initial recognition model for training to obtain the second type of recognition model;

[0026] A third training subunit, connected to the classification subunit, is used to input all the occipital transverse image groups into a pre-built initial recognition model for training to obtain the third type of recognition model;

[0027] The model group construction subunit connects the first training subunit, the second training subunit and the third training subunit, and is used to add the first type of recognition model, the second type of recognition model and the third type of recognition model to the head position classification model group, and use the head ultrasound image input into the head position classification model group as the input of the first type of recognition model, the second type of recognition model and the third type of recognition model, and use the output of the first type of recognition model, the second type of recognition model and the third type of recognition model as the output of the head position classification model group.

[0028] The present invention also provides a method for automatically identifying the head position of a fetus at birth, which is applied to the above-mentioned automatic identification system and comprises:

[0029] Step S1, the automatic recognition system obtains an ultrasound image of the fetus's head through a transperineal ultrasound device during the delivery process;

[0030] Step S2: the automatic recognition system inputs the head ultrasound image into a pre-constructed head position classification model group to identify the head position type of the fetus in the head ultrasound image.

[0031] Preferably, before executing step S2, a model training process is also included, and the model training process includes:

[0032] Step A1, the automatic recognition system collects transabdominal ultrasound images of the fetus through a transabdominal ultrasound device and transperineal ultrasound images of the fetus through a transperineal ultrasound device during multiple production processes, and associates the transabdominal ultrasound images and the transperineal ultrasound images of the same fetus into one image group;

[0033] Step A2, for each group of the image groups, the physician in the automatic recognition system marks the rotation angle of the fetal head and the birth canal position in the transperineal ultrasound image according to the fetal head position in the transabdominal ultrasound image, and adds each of the marked image groups to the ultrasound training data set;

[0034] In step A3, the automatic recognition system inputs the ultrasound training data set into a pre-built initial recognition model for training to obtain the head position classification model group.

[0035] Preferably, the structure of the initial recognition model includes:

[0036] An input preprocessing module is used to divide each image group in the input ultrasound training data set into image blocks of different scales and perform preprocessing;

[0037] A multi-scale feature extraction module, connected to the input preprocessing module, for extracting features of each of the preprocessed image blocks using a standard CNN model, and fusing all the features;

[0038] A dynamic weight adjustment module, connected to the multi-scale feature extraction module, is used to perform self-attention calculation on the fused features and introduce a dynamic weight adjustment mechanism to perform dynamic weight adjustment;

[0039] A feature enhancement module, connected to the dynamic weight adjustment module, for enhancing the features calculated by self-attention;

[0040] A classification module is connected to the feature enhancement module and is used to classify the features after enhancement processing.

[0041] Preferably, before executing step S2, an image processing process is also included, and the image processing process includes:

[0042] Step B1, the automatic recognition system uses a denoising algorithm to remove random noise in the head ultrasound image;

[0043] Step B2, the automatic recognition system performs standardization processing on the denoised head ultrasound image;

[0044] Step B3, the automatic recognition system uses data enhancement technology to perform data enhancement on the standardized head ultrasound image, and finally updates the head ultrasound image to the data enhanced head ultrasound image.

[0045] Preferably, the head position classification model group includes a first type recognition model, a second type recognition model and a third type recognition model, and the step A3 includes:

[0046] Step A31, the automatic recognition system divides each of the image groups that have been marked in the ultrasound training data set into an occipital anterior image group, an occipital posterior image group and an occipital transverse image group according to the marked birth canal position;

[0047] Step A32, the automatic recognition system inputs all the occipital-anterior image groups into a pre-built initial recognition model for training to obtain the first type of recognition model;

[0048] Step A33, the automatic recognition system inputs all the occipital-posterior image groups into a pre-built initial recognition model for training to obtain the second type of recognition model;

[0049] Step A34, the automatic recognition system inputs all the occipital transverse image groups into a pre-built initial recognition model for training to obtain the third type of recognition model;

[0050] Step A35, the automatic recognition system adds the first type of recognition model, the second type of recognition model and the third type of recognition model to the head position classification model group, and uses the head ultrasound image input into the head position classification model group as the input of the first type of recognition model, the second type of recognition model and the third type of recognition model, and uses the output of the first type of recognition model, the second type of recognition model and the third type of recognition model as the output of the head position classification model group.

[0051] The above technical solution has the following advantages or beneficial effects: using a transperineal ultrasound device to obtain a fetal head ultrasound image to identify the fetal head position, compared with the traditional method of head position identification based on transabdominal ultrasound, it not only shows the same feasibility and accuracy, but also has unique advantages in assessing the position of the fetal head. In addition, the system uses a head position classification model group to automatically identify the head ultrasound image, so even if the operator lacks experience in fetal brain imaging, the system can be used, which greatly reduces the threshold for use and greatly expands the scope of use. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] Figure 1 This is a structural schematic diagram of an automatic identification system for the head position of a fetus during delivery in a preferred embodiment of the present invention;

[0053] Figure 2 It is a structural schematic diagram of an initial recognition model in a preferred embodiment of the present invention;

[0054] Figure 3 The figure is a flow chart of a method for automatically identifying the head position of a fetus at birth in a preferred embodiment of the present invention;

[0055] Figure 4 A flow chart of a model training process in a preferred embodiment of the present invention;

[0056] Figure 5 A schematic diagram of the flow of image processing in a preferred embodiment of the present invention;

[0057] Figure 6 This is a sub-flow diagram of step A3 in a preferred embodiment of the present invention. DETAILED DESCRIPTION

[0058] The present invention is described in detail below in conjunction with the accompanying drawings and specific embodiments. The present invention is not limited to this embodiment, and other embodiments may also fall within the scope of the present invention as long as they conform to the gist of the present invention.

[0059] In a preferred embodiment of the present invention, based on the above problems existing in the prior art, a system for automatically identifying the head position of a fetus during delivery is provided. Figure 1 As shown, including:

[0060] An image acquisition module 1 is used to obtain an ultrasound image of the fetus's head through a transperineal ultrasound device during the delivery process;

[0061] The head position recognition module 2 is connected to the image acquisition module 1 and is used to input the head ultrasound image into a pre-built head position classification model group to identify the head position type of the fetus in the head ultrasound image.

[0062] Specifically, in this embodiment, a system for automatically identifying the head position of a fetus during delivery is provided, which uses a perineal ultrasound device to obtain an ultrasound image of the fetus' head to identify the head position of the fetus. Compared with the traditional method of head position identification based on transabdominal ultrasound, it not only shows the same feasibility and accuracy, but also has unique advantages in assessing the position of the fetal head. In addition, the system uses a head position classification model group to automatically identify the head ultrasound image, so that even if the operator lacks experience in fetal brain imaging, the system can be used, which greatly reduces the threshold for use and greatly expands the scope of use.

[0063] In a preferred embodiment of the present invention, a model training module 3 is further included, connected to the head position recognition module 2, including:

[0064] The training data acquisition unit 31 is used to acquire a transabdominal ultrasound image of a fetus through a transabdominal ultrasound device and a transperineal ultrasound image of a fetus through a transperineal ultrasound device during multiple production processes, and associate the transabdominal ultrasound image and the transperineal ultrasound image of the same fetus into one image group;

[0065] The data annotation unit 32 is connected to the training data acquisition unit 31, and is used for the physician to annotate the rotation angle of the fetal head position and the birth canal position in the transperineal ultrasound image according to the fetal head position in the transabdominal ultrasound image for each group of images, and add each annotated image group to the ultrasound training data set;

[0066] The model training unit 33 is connected to the data annotation unit 32 and is used to input the ultrasound training data set into a pre-built initial recognition model for training to obtain a head position classification model group.

[0067] Specifically, in this embodiment, the training process of the initial model mainly adopts the process of first data labeling and then model training;

[0068] First, the image of the fetal head is obtained through transabdominal ultrasound equipment, which is mainly used to determine the fetal head position. The clarity and stability of the image should be ensured during the acquisition process to improve the accuracy of subsequent data processing and model training.

[0069] Subsequently, the axial plane image of the fetal head is obtained using a transperineal ultrasound device to record the position of the fetal head in the birth canal. This is used to combine with the transabdominal ultrasound image for data annotation to ensure the consistency of the image data with the actual fetal head position.

[0070] According to the above method, multiple groups of images in multiple production processes are obtained, and then experienced doctors perform manual data annotation. According to the fetal head position determined by transabdominal ultrasound, the transperineal ultrasound images are annotated. The annotation content includes the rotation angle of the fetal head position, the specific position of the fetal head in the birth canal, etc. After the annotation is completed, the ultrasound training data set is obtained. The ultrasound training data set is then used to train the initial model to obtain the head position classification model group.

[0071] Further, in this embodiment, the head position classification model group includes a first type recognition model, a second type recognition model and a third type recognition model, and the model training unit 33 includes:

[0072] The classification subunit 331 is used to classify the image groups that have been marked in the ultrasound training data set into an occipital anterior image group, an occipital posterior image group and an occipital transverse image group according to the marked birth canal position;

[0073] The first training subunit 332 is connected to the classification subunit 331 and is used to input all occipital-anterior image groups into a pre-built initial recognition model for training to obtain a first type of recognition model;

[0074] The second training subunit 333 is connected to the classification subunit 331 and is used to input all occipital-posterior image groups into a pre-built initial recognition model for training to obtain a second type of recognition model;

[0075] The third training subunit 334 is connected to the classification subunit 331 and is used to input all occipital transverse image groups into a pre-built initial recognition model for training to obtain a third type of recognition model;

[0076] The model group construction subunit 335 connects the first training subunit 332, the second training subunit 333 and the third training subunit 334, and is used to add the first type of recognition model, the second type of recognition model and the third type of recognition model to the head position classification model group, and use the head ultrasound image input to the head position classification model group as the input of the first type of recognition model, the second type of recognition model and the third type of recognition model, and use the output of the first type of recognition model, the second type of recognition model and the third type of recognition model as the output of the head position classification model group.

[0077] Specifically, in this embodiment, the initial recognition model is trained using a labeled ultrasound training data set. During the training process, a cross entropy loss function and an Adam optimizer are used to adjust model parameters and minimize classification errors.

[0078] The initial recognition model selects the ViT (Vision Transformer) model as the basic framework for feature extraction and classification. The ViT model can better capture the global features in the image through the self-attention mechanism and is suitable for the recognition of complex image features.

[0079] In this embodiment, three independent ViT models are designed to form a head position classification model group, which are used to classify different fetal head positions:

[0080] The first type of identification model ViT1: used to distinguish occipital anterior position (OA) from non-occipital anterior position.

[0081] The second type of recognition model ViT2: used to distinguish occipital posterior position (OP) from non-occipital posterior position.

[0082] The third type of recognition model ViT3: used to distinguish between occipital transverse position (OT) and non-occipital transverse position.

[0083] The classification accuracy can be improved by multiple ViT models working together.

[0084] Therefore, before training these three types of recognition models, it is necessary to classify the image groups in the ultrasound training data set. The classification criteria are divided into occipital anterior position (occipital anterior position refers to the fetus's head being bent forward during delivery, with the occipital bone as the first exposed part entering the pelvis and located in front of the mother's pelvis), occipital posterior position (occipital posterior position refers to the fetus's head also with the occipital bone as the first exposed part, but located at the back of the mother's pelvis, that is, the fetus's face is facing the front of the mother.), occipital transverse position (occipital transverse position refers to the fetus's head with the occipital bone as the first exposed part, but lying horizontally above the pelvic entrance, that is, the fetus's face is facing one side of the mother.), and then construct three unique positions. An initial recognition model is established, and then the three types of image groups are respectively input into an initial recognition model for training to obtain three independent first-category recognition models, second-category recognition models and third-category recognition models. The head ultrasound image input into the head position classification model group is used as the input of the first-category recognition model, the second-category recognition model and the third-category recognition model (that is, the head position classification model group inputs a head ultrasound image, which is also used as the input of the first-category recognition model, the second-category recognition model and the third-category recognition model), and the outputs of the first-category recognition model, the second-category recognition model and the third-category recognition model are used as the output of the head position classification model group.

[0085] Further, in this embodiment, if Figure 2 As shown, the structure of the initial recognition model 100 includes:

[0086] An input preprocessing module 110 is used to divide each image group in the input ultrasound training data set into image blocks of different scales and perform preprocessing;

[0087] The multi-scale feature extraction module 120 is connected to the input preprocessing module 110 and is used to extract the features of each preprocessed image block using a standard CNN model and fuse all the features;

[0088] The dynamic weight adjustment module 130 is connected to the multi-scale feature extraction module 120, and is used to perform self-attention calculation on the fused features and introduce a dynamic weight adjustment mechanism to perform dynamic weight adjustment;

[0089] The feature enhancement module 140 is connected to the dynamic weight adjustment module 130 and is used to enhance the features calculated by the self-attention.

[0090] The classification module 150 is connected to the feature enhancement module 140 and is used to classify the enhanced features.

[0091] Specifically, the initial recognition model uses the ViT model as the basic framework. The visual transformer (ViT) has obvious advantages over the convolutional neural network (CNN):

[0092]

[0093] Next, the Vit model is improved in the following aspects:

[0094] Multi-scale feature fusion: Traditional ViT models usually only process image patches of fixed size. By introducing a multi-scale feature fusion mechanism, different details and global information in the image can be better captured.

[0095] Dynamic weight adjustment: A dynamic weight adjustment mechanism is introduced into the self-attention mechanism to adaptively adjust the attention weight according to the characteristics of the input image, thereby improving the flexibility and adaptability of the model.

[0096] Hybrid structure: Combining the advantages of CNN and ViT, a hybrid structure is designed so that the model can efficiently extract local features while maintaining the ability to extract global features.

[0097] Attention-based feature enhancement: An attention-based feature enhancement module is introduced into the model to further improve the effect of image feature representation.

[0098] The overall structure of the improved ViT model includes the following parts:

[0099] Input preprocessing module: divide the input image into image blocks of different scales and perform preprocessing.

[0100] Multi-scale feature extraction module: CNN is used to extract features of image blocks of different scales and fuse the features.

[0101] Dynamic weight adjustment self-attention module: performs self-attention calculation on the fused features and introduces a dynamic weight adjustment mechanism.

[0102] Attention-based feature enhancement module: enhances the features calculated by self-attention.

[0103] Classification module: classifies the enhanced features.

[0104] The following examples are used to further describe the above modules:

[0105] 1. For input preprocessing module:

[0106] In this embodiment, the input preprocessing module is used to divide the input image into image blocks of different scales, for example:

[0107] Scale 1: 16×16 pixel image patch

[0108] Scale 2: 32×32 pixel image patches

[0109] Scale 3: 64×64 pixel image patches

[0110] The image patches at each scale are normalized to meet the model input requirements.

[0111] 2. For the multi-scale feature extraction module:

[0112] In this embodiment, a convolutional neural network (CNN) is used to extract features of image blocks of different scales. Assume that the feature representation of each scale is:

[0113] Scale 1 features: F1;

[0114] Scale 2 features: F2;

[0115] Scale 3 features: F3;

[0116] Features of different scales are fused to obtain multi-scale feature representation:

[0117] Fmulti=concat(F1,F2,F3)

[0118] 3. For the dynamic weight adjustment self-attention module:

[0119] In this embodiment, the dynamic weight adjustment self-attention module is used to introduce a dynamic weight adjustment mechanism into the self-attention mechanism. The specific operation is as follows:

[0120] Perform self-attention calculation on Fmulti to obtain the attention weight A and feature representation Fattn:

[0121] A = softmax(QKT / sqrt(dk))

[0122] Fattn=A×V

[0123] Introduce a dynamic weight adjustment mechanism to adaptively adjust the attention weight Adyn according to the input features:

[0124] Adyn=σ(W×A+b)

[0125] Among them, W and b are learning parameters, and σ is the activation function.

[0126] Apply the dynamically adjusted attention weights to the feature representation:

[0127] Fdyn_attn=Adyn×V

[0128] 4. For the attention-based feature enhancement module:

[0129] In this embodiment, the attention-based feature enhancement module is used to perform feature enhancement processing on the feature Fdyn_attn after dynamic weight adjustment. The specific operations are as follows:

[0130] The convolution operation is used to further extract features:

[0131] Fenhance=Conv(Fdyn_attn)

[0132] Introduce residual connections to improve the stability of feature representation:

[0133] Foutput=Fenhance+Fdyn_attn

[0134] 5. For the classification module:

[0135] In this embodiment, the classification module is used to classify the enhanced feature Foutput. The specific operation is as follows:

[0136] Perform full connection layer processing on the features:

[0137] Ffc=FC(Foutput)

[0138] Classification using the softmax function:

[0139] y=softmax(Ffc)

[0140] After the initial recognition model is constructed in the above manner, the model is trained: the model is trained using the labeled ultrasound training data set. During the training process, the cross entropy loss function and Adam optimizer are used to adjust the model parameters and minimize the classification error.

[0141] The preferred implementation example of the present invention also includes an image processing module 4, which is connected to the image acquisition module 1 and the head position recognition module 2, and includes:

[0142] A denoising unit 41, used to remove random noise in the head ultrasound image by using a denoising algorithm;

[0143] A standardization unit 42, connected to the denoising unit 41, is used to perform standardization processing on the denoised head ultrasound image;

[0144] The image enhancement unit 43 is connected to the standardization unit 42 and is used for performing data enhancement on the standardized head ultrasound image by using a data enhancement technology, and finally updating the head ultrasound image to a data-enhanced head ultrasound image.

[0145] Specifically, in this embodiment, denoising algorithms such as median filtering and mean filtering are used to remove random noise in the image and improve image quality. Standardization methods in standardization processing include scaling and normalization. In image enhancement processing, data enhancement technology is used to enhance the original image data, including operations such as rotation, translation, scaling, and flipping, to increase data diversity and prevent model overfitting.

[0146] During the delivery process, after obtaining the fetal head ultrasound image through the perineal ultrasound device, the head ultrasound image needs to be processed and then input into the head position classification model group for head position type recognition, which can improve the image quality of the head ultrasound image and improve the accuracy of head position recognition.

[0147] The following are two examples in actual use:

[0148] Example 1: An implementation example of the process for dataset construction and model training.

[0149] Dataset construction: 2,000 pregnant women who underwent transabdominal and transperineal ultrasound examinations in the second stage of labor were randomly selected from the obstetrics database of a hospital. One transabdominal ultrasound image and one transperineal ultrasound image were extracted from each case and annotated by three experienced obstetricians to ensure the accuracy of image annotation.

[0150] Image preprocessing: All images are denoised by combining median filtering and mean filtering to remove random noise, followed by standardization, scaling the images to a uniform size, and using image enhancement techniques to increase data diversity.

[0151] Model training: Select the Vit model as the basic framework and improve the model. Divide the labeled image data set into a training set (80%), a validation set (10%), and a test set (10%). Use the training set to train the model, using the cross entropy loss function and the Adam optimizer. During the training process, monitor the classification accuracy of the validation set in real time and adjust the model parameters.

[0152] Model testing: In the test set, the overall accuracy of the model reached 96.2%, among which the accuracy of Vit1 for distinguishing occipital anterior position (OA) from non-occipital anterior position reached 98.3%.

[0153] Example 2: An embodiment for clinical application.

[0154] Case description: A 32-year-old woman was admitted to the hospital at 39 weeks of gestation for delivery. The first stage of labor went smoothly, but the second stage of labor stalled, and the doctor decided to evaluate the fetal head position by transperineal ultrasound.

[0155] System application: The automatic recognition system of the present invention is used to perform transperineal ultrasound examination on pregnant women to obtain axial plane images of the fetal head. The system automatically preprocesses the images and classifies them using the trained Vit model.

[0156] The results showed that the system displayed the fetal head position classification results in real time, judged the fetus to be occiput posterior position (OP), and provided relevant decision support suggestions, including adjusting the pregnant woman's position and the use of midwifery tools.

[0157] Clinical decision: The doctor adjusted the midwifery strategy based on the results and suggestions provided by the system, and the delivery was ultimately completed smoothly, with both mother and baby being safe.

[0158] It can be seen from all the above embodiments that the automatic identification system of the present invention has the following advantages:

[0159] Improve accuracy: The automatic recognition system based on deep learning can accurately identify the fetal head position without relying on the doctor's experience.

[0160] Reduce the burden on clinicians: Through the automatic recognition system, the workload of manual palpation and judgment of clinicians is reduced, and work efficiency is improved.

[0161] Real-time decision support: The system can display the classification results of fetal head position in real time and provide relevant decision support suggestions, which helps clinicians make quick decisions in emergency situations.

[0162] Improve delivery safety: Accurately identify the fetal head position, reduce the risk of improper placement of delivery tools and wrong traction direction, and improve the success rate of delivery and the health outcomes of newborns.

[0163] The present invention also provides a method for automatically identifying the head position of a fetus during delivery, which is applied to the above-mentioned automatic identification system. Figure 3 As shown, including:

[0164] Step S1, the automatic recognition system obtains an ultrasound image of the fetus's head through a transperineal ultrasound device during the delivery process;

[0165] Step S2: the automatic recognition system inputs the head ultrasound image into a pre-constructed head position classification model group to identify the head position type of the fetus in the head ultrasound image.

[0166] In a preferred embodiment of the present invention, a model training process is also included before executing step S2. Figure 4 As shown in the figure, the model training process includes:

[0167] Step A1, the automatic recognition system collects transabdominal ultrasound images of the fetus through a transabdominal ultrasound device and transperineal ultrasound images of the fetus through a transperineal ultrasound device during multiple production processes, and associates the transabdominal ultrasound images and transperineal ultrasound images of the same fetus into one image group;

[0168] Step A2, the automatic recognition system labels, for each image group, the rotation angle of the fetal head and the birth canal position in the transperineal ultrasound image according to the fetal head position in the transabdominal ultrasound image, and adds each labeled image group to the ultrasound training data set;

[0169] In step A3, the automatic recognition system inputs the ultrasound training data set into a pre-built initial recognition model for training to obtain a head position classification model group.

[0170] In a preferred embodiment of the present invention, the structure of the initial recognition model includes:

[0171] An input preprocessing module is used to divide each image group in the input ultrasound training data set into image blocks of different scales and perform preprocessing;

[0172] The multi-scale feature extraction module is connected to the input preprocessing module and is used to extract the features of each preprocessed image block using a standard CNN model and fuse all the features;

[0173] The dynamic weight adjustment module is connected to the multi-scale feature extraction module to perform self-attention calculation on the fused features and introduce a dynamic weight adjustment mechanism for dynamic weight adjustment;

[0174] The feature enhancement module is connected to the dynamic weight adjustment module and is used to enhance the features calculated by self-attention;

[0175] The classification module is connected to the feature enhancement module and is used to classify the enhanced features.

[0176] In a preferred embodiment of the present invention, an image processing process is also included before executing step S2. Figure 5 As shown, the image processing process includes:

[0177] Step B1, the automatic recognition system uses a denoising algorithm to remove random noise in the head ultrasound image;

[0178] Step B2, the automatic recognition system performs standardization processing on the denoised head ultrasound image;

[0179] Step B3: the automatic recognition system uses data enhancement technology to perform data enhancement on the standardized head ultrasound image, and finally updates the head ultrasound image to the data enhanced head ultrasound image.

[0180] In a preferred embodiment of the present invention, the head position classification model group includes a first type recognition model, a second type recognition model and a third type recognition model. Figure 6 As shown, step A3 includes:

[0181] Step A31, the automatic recognition system divides each image group that has been marked in the ultrasound training data set into an occipital anterior image group, an occipital posterior image group and an occipital transverse image group according to the marked birth canal position;

[0182] Step A32, the automatic recognition system inputs all occipital-anterior image groups into a pre-built initial recognition model for training to obtain a first type of recognition model;

[0183] Step A33, the automatic recognition system inputs all occipital-posterior image groups into a pre-built initial recognition model for training to obtain a second type of recognition model;

[0184] Step A34, the automatic recognition system inputs all occipital transverse image groups into a pre-built initial recognition model for training to obtain a third type of recognition model;

[0185] In step A35, the automatic recognition system adds the first type recognition model, the second type recognition model and the third type recognition model to the head position classification model group, and uses the head ultrasound image input to the head position classification model group as the input of the first type recognition model, the second type recognition model and the third type recognition model, and uses the output of the first type recognition model, the second type recognition model and the third type recognition model as the output of the head position classification model group.

[0186] The above are only preferred embodiments of the present invention, and are not intended to limit the implementation methods and protection scope of the present invention. Those skilled in the art should be aware that all solutions obtained by equivalent substitutions and obvious changes made using the contents of this specification and illustrations should be included in the protection scope of the present invention.

Claims

1. An automatic identification system for the head position of a fetus at birth, characterized in that: include: An image acquisition module, used to obtain an ultrasound image of the fetus's head through a transperineal ultrasound device during the delivery process; The head position recognition module is connected to the image acquisition module and is used to input the head ultrasound image into a pre-constructed head position classification model group to identify the head position type of the fetus in the head ultrasound image.

2. The automatic identification system according to claim 1, characterized in that: It also includes a model training module, connected to the head position recognition module, including: A training data acquisition unit, configured to acquire transabdominal ultrasound images of a fetus through a transabdominal ultrasound device and transperineal ultrasound images of a fetus through a transperineal ultrasound device during multiple production processes, and associate the transabdominal ultrasound images and the transperineal ultrasound images of the same fetus into a group of images; a data annotation unit connected to the training data acquisition unit, and used for annotating, for each group of the image groups, the rotation angle of the fetal head in the transperineal ultrasound image and the position of the birth canal according to the fetal head position in the transabdominal ultrasound image, and adding each of the annotated image groups to the ultrasound training data set; The model training unit is connected to the data annotation unit and is used to input the ultrasound training data set into a pre-built initial recognition model for training to obtain the head position classification model group.

3. The automatic identification system according to claim 2, characterized in that: The structure of the initial recognition model includes: An input preprocessing module is used to divide each image group in the input ultrasound training data set into image blocks of different scales and perform preprocessing; A multi-scale feature extraction module, connected to the input preprocessing module, for extracting features of each of the preprocessed image blocks using a standard CNN model, and fusing all the features; A dynamic weight adjustment module, connected to the multi-scale feature extraction module, is used to perform self-attention calculation on the fused features and introduce a dynamic weight adjustment mechanism to perform dynamic weight adjustment; A feature enhancement module, connected to the dynamic weight adjustment module, for enhancing the features calculated by self-attention; A classification module is connected to the feature enhancement module and is used to classify the features after enhancement processing.

4. The automatic identification system according to claim 1, characterized in that: It also includes an image processing module, connected to the image acquisition module and the head position recognition module, including: A denoising unit, configured to remove random noise in the head ultrasound image by using a denoising algorithm; A standardization unit, connected to the denoising unit, for performing standardization processing on the denoised head ultrasound image; The image enhancement unit is connected to the standardization unit and is used to perform data enhancement on the standardized head ultrasound image by using a data enhancement technology, and finally update the head ultrasound image to the head ultrasound image after data enhancement.

5. The automatic identification system according to claim 2, characterized in that: The head position classification model group includes a first type recognition model, a second type recognition model and a third type recognition model, and the model training unit includes: a classification subunit, for dividing each of the image groups that have been labeled in the ultrasound training data set into an occipital anterior image group, an occipital posterior image group and an occipital transverse image group according to the labeled birth canal position; A first training subunit, connected to the classification subunit, is used to input all the occipital-anterior image groups into a pre-built initial recognition model for training to obtain the first type of recognition model; A second training subunit, connected to the classification subunit, is used to input all the occipital-posterior image groups into a pre-built initial recognition model for training to obtain the second type of recognition model; A third training subunit, connected to the classification subunit, is used to input all the occipital transverse image groups into a pre-built initial recognition model for training to obtain the third type of recognition model; The model group construction subunit connects the first training subunit, the second training subunit and the third training subunit, and is used to add the first type of recognition model, the second type of recognition model and the third type of recognition model to the head position classification model group, and use the head ultrasound image input into the head position classification model group as the input of the first type of recognition model, the second type of recognition model and the third type of recognition model, and use the output of the first type of recognition model, the second type of recognition model and the third type of recognition model as the output of the head position classification model group.

6. A method for automatically identifying the head position of a fetus at birth, characterized in that: An automatic identification system as claimed in any one of claims 1 to 5, comprising: Step S1, the automatic recognition system obtains an ultrasound image of the fetus's head through a transperineal ultrasound device during the delivery process; Step S2: the automatic recognition system inputs the head ultrasound image into a pre-constructed head position classification model group to identify the head position type of the fetus in the head ultrasound image.

7. The automatic identification method according to claim 6, characterized in that: Before executing step S2, a model training process is also included, and the model training process includes: Step A1, the automatic recognition system collects transabdominal ultrasound images of the fetus through a transabdominal ultrasound device and transperineal ultrasound images of the fetus through a transperineal ultrasound device during multiple production processes, and associates the transabdominal ultrasound images and the transperineal ultrasound images of the same fetus into one image group; Step A2, for each group of the image groups, the physician in the automatic recognition system marks the rotation angle of the fetal head and the birth canal position in the transperineal ultrasound image according to the fetal head position in the transabdominal ultrasound image, and adds each of the marked image groups to the ultrasound training data set; In step A3, the automatic recognition system inputs the ultrasound training data set into a pre-built initial recognition model for training to obtain the head position classification model group.

8. The automatic identification method according to claim 7, characterized in that: The structure of the initial recognition model includes: An input preprocessing module is used to divide each image group in the input ultrasound training data set into image blocks of different scales and perform preprocessing; A multi-scale feature extraction module, connected to the input preprocessing module, for extracting features of each of the preprocessed image blocks using a standard CNN model, and fusing all the features; A dynamic weight adjustment module, connected to the multi-scale feature extraction module, is used to perform self-attention calculation on the fused features and introduce a dynamic weight adjustment mechanism to perform dynamic weight adjustment; A feature enhancement module, connected to the dynamic weight adjustment module, for enhancing the features calculated by self-attention; A classification module is connected to the feature enhancement module and is used to classify the features after enhancement processing.

9. The automatic identification method according to claim 6, characterized in that: Before executing step S2, an image processing process is also included, and the image processing process includes: Step B1, the automatic recognition system uses a denoising algorithm to remove random noise in the head ultrasound image; Step B2, the automatic recognition system performs standardization processing on the denoised head ultrasound image; Step B3, the automatic recognition system uses data enhancement technology to perform data enhancement on the standardized head ultrasound image, and finally updates the head ultrasound image to the data enhanced head ultrasound image.

10. The automatic identification method according to claim 7, characterized in that: The head position classification model group includes a first type recognition model, a second type recognition model and a third type recognition model, and step A3 includes: Step A31, the automatic recognition system divides each of the image groups that have been marked in the ultrasound training data set into an occipital anterior image group, an occipital posterior image group and an occipital transverse image group according to the marked birth canal position; Step A32, the automatic recognition system inputs all the occipital-anterior image groups into a pre-built initial recognition model for training to obtain the first type of recognition model; Step A33, the automatic recognition system inputs all the occipital-posterior image groups into a pre-built initial recognition model for training to obtain the second type of recognition model; Step A34, the automatic recognition system inputs all the occipital transverse image groups into a pre-built initial recognition model for training to obtain the third type of recognition model; Step A35, the automatic recognition system adds the first type of recognition model, the second type of recognition model and the third type of recognition model to the head position classification model group, and uses the head ultrasound image input into the head position classification model group as the input of the first type of recognition model, the second type of recognition model and the third type of recognition model, and uses the output of the first type of recognition model, the second type of recognition model and the third type of recognition model as the output of the head position classification model group.

Images