Middle-late pregnancy fetus ultrasonic cerebellum standard section image parallel recognition and measurement method

Abstract

The invention relates to a middle-late pregnancy fetus ultrasonic cerebellum standard section image parallel recognition and measurement method and device, computer equipment, a storage medium and a computer program product, and the method comprises the steps: obtaining a middle-late pregnancy fetus ultrasonic image and a deep learning model of cerebellum standard section recognition, the deep learning model is obtained based on multi-node parallel technology accelerated training; inputting the fetus ultrasonic image in the middle-late pregnancy period into a deep learning model to identify a detection frame of each structure in a cerebellar section and a corresponding confidence coefficient; screening each structure in the recognized cerebellar section based on the detection frame and the corresponding confidence to obtain a structure recognition result in the cerebellar section; determining a cerebellum center line according to a structure identification result in the cerebellum section; and measuring the transverse diameter of the cerebellum according to the midline of the cerebellum. According to the whole scheme, efficient and accurate parallel recognition and measurement of the ultrasonic cerebellum standard section image of the fetus in the middle-late pregnancy period can be achieved.

Description

Technical Field

[0001] The present application relates to the field of medical image processing technology, and in particular to a method, device, computer equipment, storage medium and computer program product for parallel recognition and measurement of standard section images of fetal cerebellum in mid- to late-pregnancy ultrasound. Background Art

[0002] Standard cross-sectional images of the cerebellum in mid- to late-pregnancy fetal ultrasound images are important tools for assessing fetal cerebellar development.

[0003] The current general process for measuring the standard section of the cerebellum in ultrasound images of fetuses in the middle and late stages of pregnancy is as follows: 1) First, the doctor needs to use a probe to scan and obtain a video; 2) Then, the doctor needs to observe the ultrasound video stream and manually find the standard section of the cerebellum; 3) Finally, after finding the standard section of the cerebellum, the doctor needs to manually select two measurement points on the outside of the cerebellum to measure the transverse diameter of the cerebellum.

[0004] Although the above method can measure the standard section of the cerebellum in ultrasound images of mid- to late-pregnancy fetuses, this method based on manual measurement has the following defects: 1) It requires high experience and technical skills of doctors, who need to be able to accurately find the standard section of the cerebellum and the measurement points; 2) It is time-consuming, and doctors need to watch the video for a long time and watch it repeatedly to obtain the standard section; 3) The selection of measurement points depends on the doctor's subjective judgment and the operation is complicated. Therefore, there is an urgent need for an efficient and accurate parallel recognition and measurement solution for the standard section image of the cerebellum of mid- to late-pregnancy fetuses. Summary of the invention

[0005] Based on this, it is necessary to provide an efficient and accurate method, device, computer equipment, computer-readable storage medium and computer program product for parallel recognition and measurement of standard cross-section images of fetal cerebellum in mid- and late-pregnancy ultrasound in order to address the above-mentioned technical problems.

[0006] In a first aspect, the present application provides a method for parallel recognition and measurement of standard section images of the cerebellum of a fetus in mid- to late-pregnancy ultrasound. The method comprises:

[0007] Obtaining a deep learning model for mid- to late-pregnancy fetal ultrasound images and cerebellum standard section recognition, wherein the deep learning model is accelerated through training based on multi-node parallel technology;

[0008] Inputting the mid- to late-pregnancy fetal ultrasound image into the deep learning model to identify the detection frame of each structure in the cerebellar section and the corresponding confidence level;

[0009] Screening each structure in the identified cerebellar section based on the detection frame and the corresponding confidence level to obtain a structure recognition result in the cerebellar section;

[0010] Determining the cerebellum midline according to the structure recognition result in the cerebellum section;

[0011] The transverse diameter of the cerebellum was measured according to the midline of the cerebellum.

[0012] In one embodiment, obtaining a deep learning model for identifying standard sections of the cerebellum includes:

[0013] Obtain training data and an initial deep learning model for standard sections of the cerebellum in mid- to late-pregnancy fetal ultrasound images;

[0014] Dividing the training data to obtain multiple batches of training data;

[0015] The multiple batches of training data are respectively sent to multiple computing nodes for training calculations, and during the training calculation process, each computing node adopts a wait-free reverse propagation strategy to update the model parameters of the initial deep learning model to obtain a deep learning model for cerebellum standard section recognition.

[0016] In one embodiment, the multiple batches of training data are sent to multiple computing nodes for training calculations, and during the training calculation process, each computing node uses a wait-free backpropagation strategy to update the model parameters of the initial deep learning model to obtain a deep learning model for cerebellum standard section recognition, including:

[0017] Sending the multiple batches of training data to multiple computing nodes respectively;

[0018] Each of the computing nodes independently calculates a gradient according to the received training data;

[0019] Aggregating the gradients calculated by all the computing nodes at preset intervals;

[0020] The average value of the aggregated gradient is calculated, and the model parameters of the initial deep learning model are updated based on the gradient average value to obtain a deep learning model for identifying the standard section of the cerebellum.

[0021] In one embodiment, inputting the mid- to late-pregnancy fetal ultrasound image into the deep learning model to identify the detection frames of various structures in the cerebellar section and the corresponding confidences includes:

[0022] Copying the deep learning model to multiple computing units, with each computing unit serving as an independent computing node;

[0023] Performing data segmentation on the mid- to late-pregnancy fetal ultrasound images to obtain multiple small batches of data;

[0024] Each of the small batches of data is distributed to different independent computing nodes for multi-node parallel computing, and the distributed computing framework Horovod is used to coordinate the computing tasks of each computing node during the computing process to obtain the detection box for identifying each structure in the cerebellar section and the corresponding confidence.

[0025] In one embodiment, the screening of each structure in the identified cerebellar section based on the detection frame and the corresponding confidence level to obtain the structure recognition result in the cerebellar section includes:

[0026] Determine whether there is a skull halo detection frame in the detection frame;

[0027] If so, traverse all structures in the cerebellar section, determine whether the center point of each detection frame is within the skull halo detection frame, delete all detection frames corresponding to structures whose center points are outside the skull halo detection frame, and delete detection frames whose confidence is lower than a preset confidence threshold;

[0028] If it does not exist, the detection box with a confidence lower than the preset confidence threshold is deleted;

[0029] Obtaining the weight of each structure in the cerebellar section, multiplying the weight by the confidence level corresponding to each structure, and accumulating the multiplication results to obtain the total score of the structure;

[0030] Determining whether the cerebellar section is a standard section, and obtaining a section recognition result;

[0031] The filtered detection frames, the structure total scores and the section recognition results are collected to obtain the structure recognition results in the cerebellar section.

[0032] In one embodiment, determining the cerebellum midline according to the structure recognition result in the cerebellum section includes:

[0033] According to the structure recognition result in the cerebellar section, the center points of the segmentation results of the third ventricle and the cerebral aqueduct are calculated to obtain a first center point set;

[0034] According to the structure recognition results in the cerebellum section, the center points of the cerebellum, falx cerebrum and thalamus target detection results are calculated to obtain a second center point set;

[0035] A straight line passing through all points in the first center point set and the second center point set is fitted using a fitting function to obtain the cerebellum midline.

[0036] In one embodiment, the step of calculating the center points of the segmentation results of the third ventricle and the cerebral aqueduct according to the structure recognition results in the cerebellar section to obtain the first center point set includes:

[0037] According to the structure recognition results in the cerebellar section, a third ventricle mask and a cerebral aqueduct mask are obtained respectively;

[0038] The first maximum contour of the third ventricle mask and the second maximum contour of the cerebral aqueduct mask were calculated respectively;

[0039] calculating a first centroid of the first maximum contour and a second centroid of the second maximum contour;

[0040] The first centroid and the second centroid are aggregated to obtain a first center point set.

[0041] In one embodiment, measuring the transverse diameter of the cerebellum according to the cerebellar midline comprises:

[0042] Identify the number of contours containing the cerebellum segmentation mask in the cerebellum section image;

[0043] If the number of contours is two, the minimum circumscribed rectangle containing the two cerebellum segmentation mask contours is calculated, and the center point of the minimum circumscribed rectangle is obtained; based on the center point and the cerebellum midline, the transverse diameter of the cerebellum is measured;

[0044] If the number of contours is one, all contour points on the cerebellum segmentation mask are traversed, the distance from each contour point to the cerebellum midline is calculated, the contour point farthest from the cerebellum midline is found, a first measurement point is obtained, a point symmetrical to the first measurement point about the cerebellum midline is calculated, a second measurement point is obtained, and the transverse diameter of the cerebellum is measured based on the first measurement point and the second measurement point.

[0045] In a second aspect, the present application also provides a device for parallel recognition and measurement of standard section images of the cerebellum of a fetus in the middle and late pregnancy using ultrasound. The device comprises:

[0046] A data acquisition module, used to acquire ultrasound images of fetuses in the middle and late stages of pregnancy, and a deep learning model for identifying standard sections of the cerebellum, wherein the deep learning model is obtained by accelerated training based on multi-node parallel technology;

[0047] A recognition module, used to input the mid- to late-pregnancy fetal ultrasound image into the deep learning model to identify the detection frame of each structure in the cerebellar section and the corresponding confidence level;

[0048] A screening module, used for screening each structure in the identified cerebellar section based on the detection frame and the corresponding confidence level to obtain a structure recognition result in the cerebellar section;

[0049] The measurement module is used to determine the cerebellum midline according to the structure recognition result in the cerebellum section; and to measure the transverse diameter of the cerebellum according to the cerebellum midline.

[0050] In a third aspect, the present application further provides a computer device. The computer device includes a memory and a processor, the memory stores a computer program, and the processor implements the following steps when executing the computer program:

[0051] Obtaining a deep learning model for mid- to late-pregnancy fetal ultrasound images and cerebellum standard section recognition, wherein the deep learning model is accelerated through training based on multi-node parallel technology;

[0052] Inputting the mid- to late-pregnancy fetal ultrasound image into the deep learning model to identify the detection frame of each structure in the cerebellar section and the corresponding confidence level;

[0053] Screening each structure in the identified cerebellar section based on the detection frame and the corresponding confidence level to obtain a structure recognition result in the cerebellar section;

[0054] Determining the cerebellum midline according to the structure recognition result in the cerebellum section;

[0055] The transverse diameter of the cerebellum was measured according to the midline of the cerebellum.

[0056] In a fourth aspect, the present application further provides a computer-readable storage medium. The computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the following steps are implemented:

[0057] Obtaining a deep learning model for mid- to late-pregnancy fetal ultrasound images and cerebellum standard section recognition, wherein the deep learning model is accelerated through training based on multi-node parallel technology;

[0058] Inputting the mid- to late-pregnancy fetal ultrasound image into the deep learning model to identify the detection frame of each structure in the cerebellar section and the corresponding confidence level;

[0059] Screening each structure in the identified cerebellar section based on the detection frame and the corresponding confidence level to obtain a structure recognition result in the cerebellar section;

[0060] Determining the cerebellum midline according to the structure recognition result in the cerebellum section;

[0061] The transverse diameter of the cerebellum was measured according to the midline of the cerebellum.

[0062] In a fifth aspect, the present application further provides a computer program product. The computer program product includes a computer program, and when the computer program is executed by a processor, the following steps are implemented:

[0063] Obtaining a deep learning model for mid- to late-pregnancy fetal ultrasound images and cerebellum standard section recognition, wherein the deep learning model is accelerated through training based on multi-node parallel technology;

[0064] Inputting the mid- to late-pregnancy fetal ultrasound image into the deep learning model to identify the detection frame of each structure in the cerebellar section and the corresponding confidence level;

[0065] Screening each structure in the identified cerebellar section based on the detection frame and the corresponding confidence level to obtain a structure recognition result in the cerebellar section;

[0066] Determining the cerebellum midline according to the structure recognition result in the cerebellum section;

[0067] According to the cerebellar midline, the transverse diameter of the cerebellum was measured.

[0068] The above-mentioned method, device, computer equipment, storage medium and computer program product for parallel recognition and measurement of standard section images of cerebellum in fetal ultrasound in the middle and late pregnancy period obtains deep learning models for ultrasound images of fetuses in the middle and late pregnancy period and recognition of standard sections of cerebellum, and the deep learning models are obtained based on multi-node parallel technology for accelerated training; the ultrasound images of fetuses in the middle and late pregnancy period are input into the deep learning model to identify the detection frames and corresponding confidences of various structures in the cerebellar section; the various structures in the identified cerebellar section are screened based on the detection frames and corresponding confidences to obtain the recognition results of the structures in the cerebellar section; the cerebellum midline is determined based on the recognition results of the structures in the cerebellar section; and the transverse diameter of the cerebellum is measured based on the cerebellum midline. In the whole process, the deep learning model for recognition of standard sections of cerebellum obtained by training with parallel acceleration technology quickly realizes the recognition of various structures in the cerebellar section, and the obtained detection frames are screened based on confidences to improve the recognition accuracy, and then the cerebellum midline is further determined and the transverse diameter of the cerebellum is measured, so that efficient and accurate parallel recognition and measurement of standard section images of cerebellum in fetal ultrasound in the middle and late pregnancy period can be realized. BRIEF DESCRIPTION OF THE DRAWINGS

[0069] Figure 1 A diagram of the application environment of a method for parallel recognition and measurement of standard cross-sectional images of fetal cerebellum in mid- to late-pregnancy ultrasound in one embodiment;

[0070] Figure 2 It is a flow chart of a method for parallel recognition and measurement of standard section images of the cerebellum of a fetus in mid- to late-pregnancy ultrasound in one embodiment;

[0071] Figure 3 This is a horizontal cross-section of the cerebellum after post-processing and scoring;

[0072] Figure 4 It is a flowchart of another embodiment of a method for parallel recognition and measurement of standard section images of the cerebellum of a fetus in mid- to late-pregnancy using ultrasound;

[0073] Figure 5 This is a schematic diagram of the midline of the cerebellum horizontal cross section;

[0074] Figure 6 The mask and measurement line diagram of the cerebellar hemisphere;

[0075] Figure 7 This is the horizontal cross-section detection diagram of the cerebellum;

[0076] Figure 8 This is the horizontal cross-section measurement of the cerebellum;

[0077] Fig. 9 It is a structural block diagram of a device for parallel recognition and measurement of standard section images of cerebellum of a fetus in mid- to late-pregnancy ultrasound in one embodiment;

[0078] Fig.10 FIG. 4 is a diagram showing the internal structure of a computer device in one embodiment. DETAILED DESCRIPTION

[0079] In order to make the purpose, technical solution and advantages of the present application more clearly understood, the present application is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.

[0080] The method for parallel recognition and measurement of standard section images of the cerebellum of a fetus in the second and third trimester provided in the present application can be applied to Figure 1 In the application environment shown. Among them, the terminal 102 communicates with the server 104 through the network. The data storage system can store the data that the server 104 needs to process. The data storage system can be integrated on the server 104, or it can be placed on the cloud or other network servers. The terminal 102 collects mid-to-late pregnancy fetal ultrasound images, sends the mid-to-late pregnancy fetal ultrasound images to the server 104, and the server 104 obtains the mid-to-late pregnancy fetal ultrasound images and the deep learning model for the recognition of the standard cerebellar section. The deep learning model is obtained by accelerating the training based on multi-node parallel technology; the mid-to-late pregnancy fetal ultrasound images are input into the deep learning model to identify the detection frame and corresponding confidence of each structure in the cerebellar section; based on the detection frame and the corresponding confidence, the identified cerebellar section The structure is screened to obtain the structure recognition result in the cerebellar section; according to the structure recognition result in the cerebellar section, the cerebellum midline is determined; according to the cerebellum midline, the cerebellum transverse diameter is measured. The server 104 can be implemented with an independent server or a server cluster consisting of multiple servers.

[0081] In one embodiment, Figure 2 As shown in the figure, a method for parallel recognition and measurement of standard section images of fetal cerebellum in mid- to late-pregnancy ultrasound is provided. Figure 1 Taking the server 104 in the example as an example, the following steps are included:

[0082] S100: A deep learning model for obtaining mid- to late-pregnancy fetal ultrasound images and identifying standard cerebellum sections. The deep learning model is trained based on multi-node parallel technology.

[0083] The middle and late pregnancy period (generally refers to the period from 20 weeks of pregnancy to before delivery) is an important period for the development and maturation of various fetal systems, and ultrasound examination is particularly important at this stage. Ultrasound examination can clearly display the images of various fetal organs, which is of great significance for clarifying the diagnosis of malformations and reducing the birth rate of defective fetuses. At this stage, doctors usually collect multiple fetal ultrasound section images, including but not limited to the horizontal cross-section of the thalamus, the horizontal cross-section of the cerebellum, the nasolabial coronal section, the four-chamber heart section, the upper abdominal cross-section, etc., to comprehensively evaluate the development of the fetus. The standard section of the cerebellum is an important section in fetal ultrasound examination, which helps to evaluate the development of the cerebellum. Standard section images of the cerebellum usually include structures such as the cerebellar hemispheres, the cerebellar vermis, and the cerebellomedullary cistern. In deep learning, recognizing standard section images of the cerebellum requires building a model that can accurately distinguish between different section images.

[0084] The deep learning model for cerebellum standard section recognition is a pre-trained model, which can specifically detect targets in the cerebellum standard section of fetal ultrasound images in the middle and late pregnancy. Specifically, the deep learning model is accelerated by multi-node parallel technology, and it uses the target detection model YOLOv5 for training and detection.

[0085] S200: Inputting mid- to late-pregnancy fetal ultrasound images into a deep learning model to identify the detection frames of various structures in the cerebellar section and the corresponding confidence levels.

[0086] The ultrasound images of the fetus in the second and third trimesters are input into the deep learning model, which detects the cerebellum section, the important structures in the cerebellum section, and the corresponding confidence levels. Specifically, after the deep learning model is tested, it is possible to determine whether the cerebellum section is standard and the detection frames and corresponding confidence levels of the various structures in the cerebellum section.

[0087] S300: Screening each structure in the identified cerebellar section based on the detection frame and the corresponding confidence level to obtain a structure recognition result in the cerebellar section.

[0088] The structures in the identified cerebellar sections are screened based on the detection frame and the corresponding confidence. Specifically, a scoring mechanism can be adopted based on the confidence to screen the structures in the identified cerebellar sections, and obtain the structure recognition results in the cerebellar sections that can be further measured.

[0089] S400: Determine the cerebellar midline based on the structure recognition results in the cerebellar section.

[0090] In fetal ultrasound images, determining the cerebellar midline is an important step in assessing cerebellar development and symmetry. The cerebellar midline usually refers to the central part of the cerebellar vermis, which separates the two hemispheres of the cerebellum. Based on the results of structural identification in the cerebellar section, the positions of different structures in the cerebellar section are determined, and then the cerebellar midline is determined based on some key structures among these structures. Specifically, the cerebellar midline can be determined based on the positions of structures such as the third ventricle, the cerebral aqueduct, the cerebellum, the falx cerebrum, and the thalamus.

[0091] S500: Measure the transverse diameter of the cerebellum based on the midline of the cerebellum.

[0092] After determining the midline of the cerebellum, the transverse diameter of the cerebellum is analyzed and measured based on the midline of the cerebellum. Specifically, after obtaining the midline of the cerebellum, it is necessary to determine how many cerebellums there are in the current section segmentation result. There are two situations: two cerebellums; one cerebellum. There are certain differences in the process of measuring the transverse diameter of the cerebellum for the case of two cerebellums and one cerebellum. For the case of two cerebellums, it is necessary to calculate the minimum circumscribed rectangle containing the contours of the two cerebellar segmentation masks and obtain the center point of the minimum circumscribed rectangle; measure the transverse diameter of the cerebellum based on the center point and the midline of the cerebellum; for the case of only one cerebellum, the transverse diameter of the cerebellum is measured based on the contour points on the cerebellar contour and the midline of the cerebellum.

[0093] The above-mentioned method for parallel recognition and measurement of standard section images of cerebellum in fetal ultrasound in the middle and late pregnancy period obtains the deep learning model of the ultrasound image of the fetus in the middle and late pregnancy period and the recognition of the standard section of the cerebellum, and the deep learning model is obtained by accelerated training based on multi-node parallel technology; the ultrasound image of the fetus in the middle and late pregnancy period is input into the deep learning model to identify the detection frame of each structure in the cerebellar section and the corresponding confidence; the identified structures in the cerebellar section are screened based on the detection frame and the corresponding confidence to obtain the recognition result of the structure in the cerebellar section; the cerebellum midline is determined according to the recognition result of the structure in the cerebellar section; the transverse diameter of the cerebellum is measured according to the cerebellum midline. In the whole process, the deep learning model for the recognition of the standard section of the cerebellum obtained by training with the parallel acceleration technology quickly realizes the recognition of each structure in the cerebellar section, and the obtained detection frame is screened based on the confidence to improve the recognition accuracy, and then the cerebellum midline is further determined and the transverse diameter of the cerebellum is measured, so that efficient and accurate parallel recognition and measurement of standard section images of cerebellum in fetal ultrasound in the middle and late pregnancy period can be realized.

[0094] In one embodiment, obtaining a deep learning model for identifying standard sections of the cerebellum includes:

[0095] Step 1: Obtain training data of standard sections of the cerebellum in mid- to late-pregnancy fetal ultrasound images and an initial deep learning model.

[0096] The training data of the standard cross-section of the cerebellum in the ultrasound images of the fetus in the middle and late pregnancy is specifically the data that has completed data labeling and is used for model training. The data labeling here refers to the completion of the labeling of information such as the size, shape, and position of the cerebellum and its related structures.

[0097] Step 2: Divide the training data to obtain multiple batches of training data.

[0098] During the specific training, the ultrasound training data is divided into multiple batches. The division here can be performed in an equal manner to obtain multiple batches of training data.

[0099] Step 3: Send multiple batches of training data to multiple computing nodes for training calculations. During the training calculation process, each computing node uses a wait-free reverse propagation strategy to update the model parameters of the initial deep learning model to obtain a deep learning model for cerebellum standard section recognition.

[0100] Multiple batches of training data are sent to different computing nodes (GPU) for training calculations, and then the gradients are summarized to update the model. At the same time, in order to reduce communication overhead, a wait-free backpropagation strategy is adopted to interleave the calculation and communication of computing nodes (GPU) and processing nodes (CPU). The whole process is implemented using local update and global aggregation methods.

[0101] In one embodiment, multiple batches of training data are sent to multiple computing nodes for training calculations, and during the training calculation process, each computing node uses a wait-free reverse propagation strategy to update the model parameters of the initial deep learning model, and the deep learning model for cerebellum standard section recognition is obtained, including:

[0102] Multiple batches of training data are sent to multiple computing nodes respectively; each computing node independently calculates the gradient according to the received training data; at preset time intervals, the gradients calculated by all computing nodes are aggregated; the average value of the aggregated gradients is calculated, and the model parameters of the initial deep learning model are updated based on the gradient average value, so as to obtain a deep learning model for cerebellum standard section recognition.

[0103] First, each computing node (GPU) independently calculates the gradient on its own data partition. This means that each node can complete the gradient calculation without waiting for other nodes. Then each node updates the model parameters locally based on the gradient calculated by the node. After a set time interval, all nodes aggregate their calculated gradients through peer-to-peer communication to ensure that the model parameters are consistent among all nodes. Finally, the aggregation node calculates the average of all received gradients and updates the global model parameters based on this. These updated parameters are then sent back to each node, overwriting their local model parameters. Then prepare for the next round of local training. Wait-free reverse propagation reduces communication overhead, improves parallel efficiency, and accelerates model training.

[0104] Furthermore, due to the large size of the training dataset, the memory capacity of the computing node (GPU) becomes the main bottleneck. Data parallelism requires each GPU to store its own copy of the neural network. For the current ultrasound image dataset, the memory required to store the activations, gradients, and parameters of these neural networks usually exceeds the capacity of a single GPU DRAM, so memory optimization is also required. The Zero Redundancy Optimizer (ZeRO) is used to reduce the memory consumption of each GPU. The first step is to distribute the model parameters and optimizer states, distribute these states among multiple GPUs, and reduce the memory usage of a single GPU. At the same time, the gradients are further distributed, and each GPU only calculates and stores the gradients of the parameters it is responsible for, as well as the activation values ​​temporarily calculated during the forward and backward propagation process, to maximize memory savings.

[0105] In one embodiment, mid- to late-pregnancy fetal ultrasound images are input into a deep learning model to identify the detection boxes of various structures in the cerebellar section and the corresponding confidence levels, including:

[0106] Step 1: Copy the deep learning model to multiple computing units, with each computing unit as an independent computing node.

[0107] Step 2: segment the fetal ultrasound images in the middle and late pregnancy to obtain multiple small batches of data;

[0108] Step 3: Allocate each small batch of data to different independent computing nodes for multi-node parallel computing. During the computing process, the distributed computing framework Horovod is used to coordinate the computing tasks of each computing node to obtain the detection box and corresponding confidence level of each structure in the cerebellar section.

[0109] In the model initialization phase, the model is copied to multiple graphics cards, and then the data is split into multiple mini-batches, and each mini-batch is assigned to a different computing node for multi-node parallel processing. Each computing node processes the data assigned to it independently, thereby achieving parallel computing. Here, the distributed computing framework Horovod is used to coordinate the computing tasks of each node.

[0110] At the same time, heterogeneous computing methods are used during calculations to further accelerate detection. Specifically, each computing node contains two different types of computing units: GPU and FPGA. GPU is good at handling large-scale parallel computing tasks and is suitable for forward propagation and back propagation of deep learning models. Therefore, GPU is used to be responsible for the main computing part of model reasoning, including complex calculations of convolutional layers and fully connected layers. Its large-scale parallel processing capabilities are used to accelerate the reasoning of deep learning models. FPGA has high programmability and low latency characteristics, which is suitable for accelerated computing of specific algorithms. Therefore, FPGA is used to pre-process images, including image enhancement, noise filtering, and result post-processing, including non-maximum suppression and result filtering.

[0111] Specifically, the data is first preprocessed by FPGA, and preprocessing operations such as image enhancement, denoising, and cropping are performed. The processed image is passed to the GPU, and the GPU performs the main calculations of the deep learning model to generate preliminary detection results. Finally, the FPGA performs post-processing on the results generated by the GPU, including non-maximum suppression and bounding box adjustment, to generate the final detection results. In one embodiment, the various structures in the identified cerebellar section are screened based on the detection box and the corresponding confidence level, and the structure recognition results in the cerebellar section include:

[0112] Determine whether there is a skull halo detection frame in the detection frame; if so, traverse all structures in the cerebellar section, determine whether the center point of each detection frame is within the skull halo detection frame, delete all detection frames corresponding to structures whose center points are outside the skull halo detection frame, and delete detection frames whose confidences are lower than the preset confidence threshold; if not, delete detection frames whose confidences are lower than the preset confidence threshold; obtain the weights of each structure in the cerebellar section, multiply the weights by the confidences corresponding to each structure, and accumulate the multiplication results to obtain the total score of the structure; determine whether the cerebellar section is a standard section to obtain the section recognition result; aggregate the screened detection frames, total structure scores, and section recognition results to obtain the structure recognition result in the cerebellar section.

[0113] Specifically, after the deep learning model is tested, we can get whether the cerebellar section is standard, as well as the detection frame and corresponding confidence of each structure in the cerebellar section. After that, it needs to go through the post-processing scoring mechanism. The screening here includes filtering the detection frame, calculating the total score of the structure, and calculating the section score.

[0114] (1) Filter detection box

[0115] In order to prevent redundant detection frames from affecting the post-processing score, it is necessary to filter the detection frames. First, determine whether there is a skull halo. If so, traverse all structures to determine whether the center point of each frame is within the skull halo frame, and delete all structures whose center points are outside the skull halo. Then delete the detection frames with a confidence score less than 0.5. If there is no skull halo, directly delete the detection frames with a confidence score less than 0.5. In addition, it is necessary to limit the number of each structure. For example, there can only be two thalamus, so only the two thalamus detection frames with the highest confidence are retained, and the rest are filtered.

[0116] (2) Calculate the total structure score (struct_score)

[0117] The score of each small structure is the confidence of the small structure multiplied by the corresponding weight, and the total score of the structure is the sum of the scores of each small structure. The weights of each small structure are shown in Table 1 below.

[0118] Table 1. Corresponding weights of small structures in horizontal cross-section of cerebellum

[0119] Structure Name Weight Skull Halo 5 Transparent compartment 15 Choroid plexus 0 Falx cerebri 5 thalamus 5 Sylvian fissure 20 Cerebellum 35 Posterior horn of lateral ventricle 0 Posterior fossa 10

[0120] At the same time, points are deducted for missing structures. Except for structures with a weight of 0 (choroid plexus, posterior horn of the lateral ventricle), 5 points will be deducted for each missing structure.

[0121] (3) Calculate the slice score (auto_score)

[0122] If the section is a standard section, then auto_score = (struct_score + skull halo confidence * 100) / 2. If the section is a non-standard section, then auto_score = (1-skull halo confidence) * 100. The horizontal cross section of the cerebellum after post-processing and scoring is as follows Figure 3 shown.

[0123] like Figure 4 As shown, in one embodiment, S400 includes:

[0124] S420: According to the structure recognition result in the cerebellar section, the center points of the segmentation results of the third ventricle and the cerebral aqueduct are calculated to obtain a first center point set.

[0125] According to the structure recognition results in the cerebellar section, the segmentation results of the third ventricle and the cerebral aqueduct can be further analyzed, and these results may be given in the form of a binary image (or mask), in which the segmented area is marked as 1 (or a specific value) and the background is 0. For each segmented area (third ventricle and cerebral aqueduct), the corresponding centroid is first calculated, and then the center point of the segmentation results of the third ventricle and the cerebral aqueduct is obtained to obtain the first center point set.

[0126] Further, according to the structure recognition results in the cerebellar section, the center points of the segmentation results of the third ventricle and the cerebral aqueduct are calculated to obtain the first center point set, including: according to the structure recognition results in the cerebellar section, the third ventricle mask and the cerebral aqueduct mask are respectively obtained; the first maximum contour of the third ventricle mask and the second maximum contour of the cerebral aqueduct mask are respectively calculated; the first centroid of the first maximum contour and the second centroid of the second maximum contour are calculated; the first centroid and the second centroid are aggregated to obtain the first center point set.

[0127] In practical applications, to obtain the center point of the segmentation results of the third ventricle and the cerebral aqueduct, it is necessary to calculate the maximum contour of the mask of the third ventricle and the cerebral aqueduct respectively, and then calculate the center of mass by the obtained contour, using the moments function in the Opencv library. The specific method is as follows:

[0128] M = cv2.moments(contour);

[0129] center_x=M['m10'] / M['m00'];

[0130] center_y=M['m01'] / M['m00'];

[0131] Get the x- and y-coordinates of the center of mass respectively.

[0132] S440: According to the structure recognition results in the cerebellum section, the center points of the target detection results of the cerebellum, the falx cerebrum and the thalamus are calculated to obtain a second center point set.

[0133] For the cerebellum and falx cerebri, there is only one detection box, so the center point can be calculated using the coordinates of the upper left and lower right corners of the detection box (the midpoint coordinates are (the x coordinate of the upper left corner plus the x coordinate of the upper right corner divided by 2, and the y coordinate of the upper left corner plus the y coordinate of the upper right corner divided by 2)). For the thalamus, there are two detection boxes, so the midpoints of the two detection boxes are calculated first, and then the midpoint of the two midpoints is calculated.

[0134] S460: Use a fitting function to fit a straight line passing through all points in the first center point set and the second center point set to obtain the cerebellum midline.

[0135] The fitting function specifically uses the np.polyfit function, that is, the np.polyfit function is used to fit a straight line passing through these points as the midline of the cerebellum. The effect is as follows Figure 5 As shown ( Figure 5 The medium green line is the center line).

[0136] In one embodiment, measuring the transverse diameter of the cerebellum based on the midline of the cerebellum comprises:

[0137] Identify the number of cerebellar segmentation mask contours in the cerebellar section image; if the number of contours is two, calculate the minimum circumscribed rectangle containing the two cerebellar segmentation mask contours, and obtain the center point of the minimum circumscribed rectangle; measure the transverse diameter of the cerebellum based on the center point and the cerebellar midline; if the number of contours is one, traverse all contour points on the cerebellar segmentation mask, calculate the distance from each contour point to the cerebellar midline, find the contour point farthest from the cerebellar midline, obtain the first measurement point, calculate the point symmetrical to the first measurement point about the cerebellar midline, obtain the second measurement point, and measure the transverse diameter of the cerebellum based on the first measurement point and the second measurement point.

[0138] After obtaining the midline of the cerebellum, it is necessary to determine how many cerebellums there are in the current section segmentation result. There are two cases: two cerebellums and one cerebellum. Different methods need to be used for measurement in the two cases.

[0139] (1) Measurement method with two cerebellums

[0140] 1.1. First, obtain the contours of the segmentation masks of the two cerebellums respectively, use the cv2.approxPolyDP function to reduce the number of contour points to reduce the amount of calculation, and then use the cv2.minAreaRect function to calculate the minimum enclosing rectangle containing the two contour points and obtain the center point of the rectangle.

[0141] 1.2. Since the cerebellum midline generally has a slope, and the cerebellum transverse diameter measurement line needs to be perpendicular to the cerebellum midline, when the cerebellum midline is rotated into a horizontal line, it is only necessary to calculate the two points with the largest vertical distance on the two cerebellar hemispheres as the end points of the cerebellum transverse diameter. The following are the specific operations:

[0142] First, we need to calculate the slope of the cerebellum midline (k), and calculate re_angle (re_angle = -math.atan(k)). Then, we rotate the contour points by the angle re_angle (equivalent to the cerebellum midline being horizontal at this time) with the center point (center) of the smallest circumscribed rectangle containing the two contour points as the center. Then fill the rotated contour to form a new mask, and then find the two points with the largest vertical distance in the mask. Finally, we rotate the two points back to the original angle to obtain the final cerebellum hemisphere measurement points (points_CEREB). The specific measurement results are as follows: Figure 6 shown.

[0143] (2) Measurement method with only one cerebellum

[0144] Traverse all the contour points of the cerebellum in turn, calculate the distance from each point to the midline of the cerebellum, the distance calculation formula is: |kx-y+b| / sqrt(k^2+1), find the point farthest from the midline of the cerebellum, and use it as a measurement point. Then calculate the point symmetrical to the midline of the cerebellum and use it as another measurement point. Based on these two measurement points, the transverse diameter of the cerebellum is finally obtained.

[0145] By executing the method for parallel recognition and measurement of standard ultrasound cerebellum sections of the late pregnancy fetus in the present application in the above embodiment, the standard horizontal cross section of the cerebellum can be obtained from the video or picture in real time, and the transverse diameter of the cerebellum (TCD) can be accurately measured. The final effect is as follows Figure 7 as well as Figure 8 As shown, Figure 7 This is the horizontal cross-section of the cerebellum. Figure 7 The numbers marked in different detection boxes are the corresponding total structure scores. Figure 8 This is a horizontal cross-sectional measurement of the cerebellum, based on Figure 8 The transverse diameter of the cerebellum can be obtained.

[0146] It should be understood that, although the steps in the flowcharts involved in the above embodiments are displayed in sequence according to the indication of the arrows, these steps are not necessarily executed in sequence according to the order indicated by the arrows. Unless there is a clear explanation in this article, the execution of these steps is not strictly limited in order, and these steps can be executed in other orders. Moreover, at least a part of the steps in the flowcharts involved in the above embodiments may include multiple steps or multiple stages, and these steps or stages are not necessarily executed at the same time, but can be executed at different times, and the execution order of these steps or stages is not necessarily carried out in sequence, but can be executed in turn or alternately with other steps or at least a part of the steps or stages in other steps.

[0147] Based on the same inventive concept, the embodiment of the present application also provides a device for parallel recognition and measurement of standard section images of the cerebellum of a fetus in mid-to-late pregnancy for realizing the above-mentioned method for parallel recognition and measurement of standard section images of the cerebellum of a fetus in mid-to-late pregnancy. The implementation scheme for solving the problem provided by the device is similar to the implementation scheme recorded in the above-mentioned method, so the specific limitations in one or more embodiments of the device for parallel recognition and measurement of standard section images of the cerebellum of a fetus in mid-to-late pregnancy provided below can be referred to the limitations of the method for parallel recognition and measurement of standard section images of the cerebellum of a fetus in mid-to-late pregnancy above, and will not be repeated here.

[0148] In one embodiment, Fig. 9 As shown, a device for parallel recognition and measurement of standard section images of fetal cerebellum by ultrasound in the middle and late pregnancy period is provided, comprising:

[0149] A data acquisition module 200 is used to acquire ultrasound images of fetuses in the middle and late stages of pregnancy and a deep learning model for identifying standard sections of the cerebellum. The deep learning model is obtained by accelerated training based on multi-node parallel technology.

[0150] The recognition module 400 is used to input the mid- to late-pregnancy fetal ultrasound image into the deep learning model to identify the detection frame of each structure in the cerebellar section and the corresponding confidence level;

[0151] A screening module 600 is used to screen each structure in the identified cerebellar section based on the detection frame and the corresponding confidence level to obtain a structure recognition result in the cerebellar section;

[0152] The measurement module 800 is used to determine the cerebellum midline according to the structure recognition results in the cerebellum section; and to measure the transverse diameter of the cerebellum according to the cerebellum midline.

[0153] In one of the embodiments, the data acquisition module 200 is also used to obtain training data and an initial deep learning model of standard sections of the cerebellum in mid- to late-pregnancy fetal ultrasound images; divide the training data to obtain multiple batches of training data; send the multiple batches of training data to multiple computing nodes for training calculations, and during the training calculation process, each computing node uses a wait-free reverse propagation strategy to update the model parameters of the initial deep learning model to obtain a deep learning model for recognizing standard sections of the cerebellum.

[0154] In one embodiment, the data acquisition module 200 is also used to send multiple batches of training data to multiple computing nodes respectively; each computing node independently calculates the gradient according to the received training data; at a preset time interval, the gradients calculated by all computing nodes are aggregated; the average value of the aggregated gradients is calculated, and the model parameters of the initial deep learning model are updated based on the gradient average value, so as to obtain a deep learning model for recognizing standard cerebellum sections.

[0155] In one embodiment, the screening module 600 is also used to determine whether there is a skull halo detection frame in the detection frame; if so, traverse all structures in the cerebellar section, determine whether the center point of each detection frame is within the skull halo detection frame, delete all detection frames corresponding to structures whose center points are outside the skull halo detection frame, and delete detection frames whose confidences are lower than a preset confidence threshold; if not, delete detection frames whose confidences are lower than a preset confidence threshold; obtain the weights of each structure in the cerebellar section, multiply the weights by the confidences corresponding to each structure, and accumulate the multiplication results to obtain the total score of the structure; determine whether the cerebellar section is a standard section to obtain the section recognition result; aggregate the screened detection frames, structure total scores, and section recognition results to obtain the structure recognition result in the cerebellar section.

[0156] In one embodiment, the measurement module 800 is also used to calculate the center points of the segmentation results of the third ventricle and the cerebral aqueduct based on the structure recognition results in the cerebellar section to obtain a first center point set; calculate the center points of the cerebellum, falx cerebri and thalamus target detection results based on the structure recognition results in the cerebellar section to obtain a second center point set; use a fitting function to fit a straight line passing through all points in the first center point set and the second center point set to obtain the cerebellum midline.

[0157] In one embodiment, the measurement module 800 is also used to obtain the third ventricle mask and the cerebral aqueduct mask respectively according to the structure recognition results in the cerebellar section; calculate the first maximum contour of the third ventricle mask and the second maximum contour of the cerebral aqueduct mask respectively; calculate the first centroid of the first maximum contour and the second centroid of the second maximum contour; and aggregate the first centroid and the second centroid to obtain a first center point set.

[0158] In one embodiment, the measurement module 800 is also used to identify the number of cerebellar segmentation mask contours included in the cerebellar section image; if the number of contours is two, the minimum circumscribed rectangle including the two cerebellar segmentation mask contours is calculated, and the center point of the minimum circumscribed rectangle is obtained; based on the center point and the cerebellar midline, the transverse diameter of the cerebellum is measured; if the number of contours is one, all contour points on the cerebellar segmentation mask are traversed, the distance from each contour point to the cerebellar midline is calculated, the contour point farthest from the cerebellar midline is found, a first measurement point is obtained, a point symmetrical to the first measurement point about the cerebellar midline is calculated, a second measurement point is obtained, and the transverse diameter of the cerebellum is measured based on the first measurement point and the second measurement point.

[0159] Each module in the above-mentioned device for parallel recognition and measurement of standard section images of fetal cerebellum in mid- to late-pregnancy ultrasound can be implemented in whole or in part by software, hardware, and a combination thereof. Each of the above-mentioned modules can be embedded in or independent of a processor in a computer device in the form of hardware, or can be stored in a memory in a computer device in the form of software, so that the processor can call and execute operations corresponding to each of the above modules.

[0160] In one embodiment, a computer device is provided. The computer device may be a server, and its internal structure diagram may be as follows: Fig.10 As shown. The computer device includes a processor, a memory and a network interface connected through a system bus. Among them, the processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program and a database. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The database of the computer device is used to store preset data. The network interface of the computer device is used to communicate with an external terminal through a network connection. When the computer program is executed by the processor, a method for parallel recognition and measurement of standard cross-section images of the cerebellum of a fetus in the middle and late pregnancy is implemented.

[0161] Those skilled in the art will understand that Fig.10 The structure shown in the figure is only a block diagram of a part of the structure related to the solution of the present application, and does not constitute a limitation on the computer device to which the solution of the present application is applied. The specific computer device may include more or fewer components than those shown in the figure, or combine certain components, or have a different arrangement of components.

[0162] In one embodiment, a computer device is provided, including a memory and a processor, wherein a computer program is stored in the memory, and when the processor executes the computer program, the above-mentioned method for parallel recognition and measurement of standard cross-section images of the cerebellum of a fetus in the middle and late pregnancy is implemented.

[0163] In one embodiment, a computer-readable storage medium is provided, on which a computer program is stored. When the computer program is executed by a processor, the above-mentioned method for parallel recognition and measurement of standard cross-section images of the cerebellum of a fetus in mid- to late-pregnancy ultrasound is implemented.

[0164] In one embodiment, a computer program product is provided, including a computer program, which, when executed by a processor, implements the above-mentioned method for parallel recognition and measurement of standard cross-section images of the cerebellum of a fetus in mid- to late-pregnancy ultrasound.

[0165] Those of ordinary skill in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing the relevant hardware through a computer program, and the above-mentioned computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to the memory, database or other medium used in the embodiments provided in the present application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetoresistive random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. As an illustration and not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM). The database involved in each embodiment provided in this application may include at least one of a relational database and a non-relational database. Non-relational databases may include distributed databases based on blockchains, etc., but are not limited thereto. The processors involved in each embodiment provided in this application may be general-purpose processors, central processing units, graphics processors, digital signal processors, programmable logic devices, data processing logic devices based on quantum computing, etc., but are not limited thereto.

[0166] The technical features of the above embodiments may be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0167] The above embodiments only express several implementation methods of the present application, and the descriptions thereof are relatively specific and detailed, but they cannot be understood as limiting the scope of the present application. It should be pointed out that, for a person of ordinary skill in the art, several variations and improvements can be made without departing from the concept of the present application, and these all belong to the protection scope of the present application. Therefore, the protection scope of the present application shall be subject to the attached claims.

Claims

1. A method for parallel recognition and measurement of standard section images of the cerebellum of a mid- to late-pregnancy fetus using ultrasound, characterized in that: The method comprises: Obtaining a deep learning model for mid- to late-pregnancy fetal ultrasound images and cerebellum standard section recognition, wherein the deep learning model is accelerated through training based on multi-node parallel technology; Inputting the mid- to late-pregnancy fetal ultrasound image into the deep learning model to identify the detection frame of each structure in the cerebellar section and the corresponding confidence level; Screening each structure in the identified cerebellar section based on the detection frame and the corresponding confidence level to obtain a structure recognition result in the cerebellar section; Determining the cerebellum midline according to the structure recognition result in the cerebellum section; According to the cerebellar midline, the transverse diameter of the cerebellum was measured.

2. The method according to claim 1, characterized in that The deep learning model for obtaining standard cerebellum section recognition includes: Obtain training data and an initial deep learning model for standard sections of the cerebellum in mid- to late-pregnancy fetal ultrasound images; Dividing the training data to obtain multiple batches of training data; The multiple batches of training data are respectively sent to multiple computing nodes for training calculations, and during the training calculation process, each computing node adopts a wait-free reverse propagation strategy to update the model parameters of the initial deep learning model to obtain a deep learning model for cerebellum standard section recognition.

3. The method according to claim 2, characterized in that The multiple batches of training data are respectively sent to multiple computing nodes for training calculation, and during the training calculation process, each computing node adopts a wait-free reverse propagation strategy to update the model parameters of the initial deep learning model, and the deep learning model for cerebellum standard section recognition is obtained, which includes: Sending the multiple batches of training data to multiple computing nodes respectively; Each of the computing nodes independently calculates a gradient according to the received training data; Aggregating the gradients calculated by all the computing nodes at preset intervals; The average value of the aggregated gradient is calculated, and the model parameters of the initial deep learning model are updated based on the gradient average value to obtain a deep learning model for identifying the standard section of the cerebellum.

4. The method according to claim 1, characterized in that The step of inputting the mid- to late-pregnancy fetal ultrasound image into the deep learning model to identify the detection frames of various structures in the cerebellar section and the corresponding confidence levels includes: Copying the deep learning model to multiple computing units, with each computing unit serving as an independent computing node; Performing data segmentation on the mid- to late-pregnancy fetal ultrasound images to obtain multiple small batches of data; Each of the small batches of data is distributed to different independent computing nodes for multi-node parallel computing, and the distributed computing framework Horovod is used to coordinate the computing tasks of each computing node during the computing process to obtain the detection box for identifying each structure in the cerebellar section and the corresponding confidence.

5. The method according to claim 1, characterized in that: The screening of each structure in the identified cerebellar section based on the detection frame and the corresponding confidence level to obtain the structure recognition result in the cerebellar section includes: Determine whether there is a skull halo detection frame in the detection frame; If so, traverse all structures in the cerebellar section, determine whether the center point of each detection frame is within the skull halo detection frame, delete all detection frames corresponding to structures whose center points are outside the skull halo detection frame, and delete detection frames whose confidence is lower than a preset confidence threshold; If it does not exist, the detection box with a confidence lower than the preset confidence threshold is deleted; Obtaining the weight of each structure in the cerebellar section, multiplying the weight by the confidence level corresponding to each structure, and accumulating the multiplication results to obtain the total score of the structure; Determining whether the cerebellar section is a standard section, and obtaining a section recognition result; The filtered detection frames, the structure total scores and the section recognition results are collected to obtain the structure recognition results in the cerebellar section.

6. The method according to claim 1, characterized in that Determining the cerebellum midline according to the structure recognition result in the cerebellum section includes: According to the structure recognition result in the cerebellar section, the center points of the segmentation results of the third ventricle and the cerebral aqueduct are calculated to obtain a first center point set; According to the structure recognition results in the cerebellum section, the center points of the cerebellum, falx cerebrum and thalamus target detection results are calculated to obtain a second center point set; A straight line passing through all points in the first center point set and the second center point set is fitted using a fitting function to obtain the cerebellum midline.

7. The method according to claim 6, characterized in that The step of calculating the center points of the segmentation results of the third ventricle and the cerebral aqueduct according to the structure recognition results in the cerebellar section to obtain the first center point set includes: According to the structure recognition results in the cerebellar section, a third ventricle mask and a cerebral aqueduct mask are obtained respectively; The first maximum contour of the third ventricle mask and the second maximum contour of the cerebral aqueduct mask were calculated respectively; calculating a first centroid of the first maximum contour and a second centroid of the second maximum contour; The first centroid and the second centroid are aggregated to obtain a first center point set.

8. The method according to claim 1, characterized in that: The measuring of the transverse diameter of the cerebellum according to the cerebellar midline comprises: Identify the number of contours containing the cerebellum segmentation mask in the cerebellum section image; If the number of contours is two, the minimum circumscribed rectangle containing the two cerebellum segmentation mask contours is calculated, and the center point of the minimum circumscribed rectangle is obtained; based on the center point and the cerebellum midline, the transverse diameter of the cerebellum is measured; If the number of contours is one, all contour points on the cerebellum segmentation mask are traversed, the distance from each contour point to the cerebellum midline is calculated, the contour point farthest from the cerebellum midline is found, a first measurement point is obtained, a point symmetrical to the first measurement point about the cerebellum midline is calculated, a second measurement point is obtained, and the transverse diameter of the cerebellum is measured based on the first measurement point and the second measurement point.

9. A device for parallel recognition and measurement of standard section images of fetal cerebellum by ultrasound in mid- to late-pregnancy, characterized in that: The device comprises: A data acquisition module, used to acquire ultrasound images of fetuses in the middle and late stages of pregnancy, and a deep learning model for identifying standard sections of the cerebellum, wherein the deep learning model is obtained by accelerated training based on multi-node parallel technology; A recognition module, used to input the mid- to late-pregnancy fetal ultrasound image into the deep learning model to identify the detection frame of each structure in the cerebellar section and the corresponding confidence level; A screening module, used for screening each structure in the identified cerebellar section based on the detection frame and the corresponding confidence level to obtain a structure recognition result in the cerebellar section; The measurement module is used to determine the cerebellum midline according to the structure recognition result in the cerebellum section; and to measure the transverse diameter of the cerebellum according to the cerebellum midline.

10. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that: When the processor executes the computer program, the steps of the method according to any one of claims 1 to 8 are implemented.

Images