Intelligent identification and abnormality judgment method and device for heart section in embryonic development period

Abstract

The application discloses a kind of based on deep learning's embryo development period heart section intelligent identification and abnormality judging method and device.The method first constructs embryo heart development four-dimensional image database, carries out multidimensional matching technology to the real-time import embryo heart section image, and to embryo heart section structure identification, labeling and three-dimensional display;Again, compared with normal embryo heart development data and abnormal embryo development data, the evaluation of embryo development and the judgment of abnormality are realized.The method can help clinical perinatal examination physician and embryo related researchers to identify the section structure of embryo heart at different periods and different angles, improve work efficiency and accuracy of prenatal examination, and promote the research of embryo heart development mechanism and congenital malformation disease.

Description

Technical Field

[0001] The present invention relates to a disease identification method based on deep learning, and in particular to a method and device for intelligent identification and abnormality judgment of heart sections during embryonic development. Background Art

[0002] Congenital heart malformations account for over half of all birth defects. In-depth research into the pathogenesis of congenital heart disease, enabling early diagnosis and intervention, has significant practical significance and economic value for perinatal care and the prevention and treatment of congenital malformations.

[0003] With the continuous advancement of imaging technologies such as B-mode ultrasound and magnetic resonance imaging (MRI), as well as the in-depth development of research methods such as developmental biology and histoembryology, it is now possible to overcome the limitations of the uterine wall and amniotic membrane to obtain clear images of the heart during embryonic development, providing direct, visible data for fetal development assessment and research on the causes of cardiac malformations. However, due to the rapid pace of embryonic heart development, the extremely complex changes in internal and external structures, and the constant changes in spatial and temporal position within the uterus, the two-dimensional data obtained from embryonic heart sections are diverse and complex, making it difficult for clinical perinatal physicians and researchers involved in embryonic development to identify and judge the cross-sectional structure of the developing heart.

[0004] While resources such as atlases of embryonic heart development and websites showcasing images of embryonic heart development exist both domestically and internationally, the complexity and niche nature of embryonic development knowledge, as well as the monotonous and unsystematic nature of its presentation, make acquiring and proficiently applying development-related cardiac cross-sectional knowledge difficult and cost-effective. Furthermore, the diversity and complexity of two-dimensional images of embryonic heart cross-sections, caused by factors such as intrauterine rotation of the embryo and probe (blade) angle deviation, further complicate identification and judgment. This significantly impacts the efficiency of clinical perinatal physicians in evaluating embryonic development and diagnosing malformations, complicates basic research on embryonic development and malformation, and limits the development and progress of related disciplines. Summary of the Invention

[0005] The purpose of the present invention is to address the deficiencies of the above-mentioned existing technologies and provide a method and device for intelligent recognition and abnormality judgment of heart sections during embryonic development.

[0006] This method utilizes the large amount of data accumulated in the study of embryonic heart development and malformations, and perinatal fetal heart development imaging examinations. The obtained two-dimensional image data of the heart section during embryonic development are calibrated, superimposed, annotated, segmented and three-dimensionally visualized using the three-dimensional graphics software Amira. A four-dimensional image database of embryonic heart development is constructed based on the developmental timeline, allowing for the subsequent continuous addition and improvement of species, malformation and other data.

[0007] After the four-dimensional image database of embryonic heart development is basically established, the method uses multi-dimensional matching technologies such as developmental time fuzzy matching, biomedical key structure matching, and computer image grayscale matching on the embryonic heart cross-sectional images obtained from real-time imported examinations or research to achieve automatic intelligent matching with the heart development data in the library, and recognition, labeling and three-dimensional display of the embryonic heart cross-sectional structure.

[0008] After the multi-dimensional matching is completed, the method compares the normal embryonic heart development data and the deformed embryonic development data in the four-dimensional image database to achieve the evaluation of the embryonic development status and the judgment of the deformity.

[0009] The method described in the present invention can help clinical perinatal physicians and embryo-related researchers identify the cross-sectional structure of the embryonic heart at different angles at different stages, improve work efficiency and prenatal examination accuracy, and promote research on the mechanism of embryonic heart development and congenital malformation diseases.

[0010] The purpose of the present invention is achieved through the following technical solutions:

[0011] A method for intelligent recognition and abnormality judgment of heart sections during embryonic development comprises the following steps: (1) collecting and constructing a four-dimensional image database of embryonic heart development, (2) matching target embryonic heart section image data with data in the database, and performing recognition, positioning, and measurement on the target embryonic heart sections, (3) creating a Faster-RCNN recognition model, and (4) using the Faster-RCNN recognition model to obtain detection results of the target embryonic heart.

[0012] Preferably, the step (1) of collecting and constructing a four-dimensional image database of embryonic heart development comprises:

[0013] (a) arranging the collected or obtained two-dimensional images of the embryonic heart sections by stage from the smallest to the largest embryo;

[0014] (b) processing the arranged two-dimensional image of the embryonic heart section into a three-dimensional image, and outlining, superimposing and annotating important structural parts therein;

[0015] (c) Arrange the images according to the timeline of embryonic development and construct a four-dimensional image database of embryonic heart development.

[0016] Preferably, the step (2) matches the target embryo heart cross-sectional image data with the data in the database, and identifies, locates and measures the data, including:

[0017] (a) removing useless information from the target embryonic heart cross-sectional image data through an attention mechanism to obtain an image of a region of interest;

[0018] (b) matching the region of interest image with corresponding structures of the four-dimensional image data in the database by grayscale matching and geometric matching;

[0019] (c) marking the outflow tract, atrioventricular canal, heart valve, atrial septum and ventricular septum of the target embryonic heart cross-sectional image data, and performing data measurement.

[0020] Preferably, the step (3) creates a Faster-RCNN recognition model, which includes a feature extraction module, an RPN module, a Proposal Layer module, a Roi Align module, a classification and regression module, and a C-NMS module.

[0021] Preferably, the step (4) inputs the processed target embryo heart cross-sectional image data into the Faster-RCNN recognition model to generate a test result, and further includes:

[0022] (a) The feature extraction module uses a set of basic conv+relu+pooling layers to extract feature maps, including Convlayers, RPN layers and fully connected layers;

[0023] (b) The RPN module inputs the feature map into the RPN network for recognition and screening, obtains the approximate location of the target from the feaure map, and obtains a recommended box for the target location to be measured;

[0024] (c) The Proposal Layer module uses the approximate position obtained by the RPN to continue training and obtain a more accurate position from the proposed frame of the target position, that is, obtains a proposed frame with a more accurate position of the target position;

[0025] (d) The Roi Align module inputs the recommended box of the more accurate position of the target and the feature map into the ROI Align network for mapping and pooling to obtain a feature recommendation map;

[0026] (e) The classification and regression module inputs the obtained feature recommendation map into the classification and regression network for classification and regression processing to obtain the initial coordinates, initial category and initial confidence of the prediction box;

[0027] (f) The C-NMS module inputs the initial coordinates, initial category, and initial confidence of the prediction box into the C-NMS module for screening and processing to obtain the target coordinates, target category, and target confidence of the prediction box.

[0028] The initial coordinates, initial category, and initial confidence of the predicted frame are input into the C-NMS module for screening and processing to obtain the target coordinates, target category, and target confidence of the predicted frame. The steps include:

[0029] (a) The initial coordinates, initial category and initial confidence of the predicted box are input into the C-NMS network;

[0030] (b) taking the prediction frame with the highest initial confidence as the reference prediction frame;

[0031] (c) calculating the overlap ratio of each prediction frame with the reference prediction frame according to the initial coordinates;

[0032] (d) taking the prediction frame whose overlap ratio is greater than or equal to a preset confidence threshold as the prediction frame to be adjusted;

[0033] (e) updating the initial confidence of the prediction frame to be adjusted according to the overlap ratio and area ratio between the prediction frame to be adjusted and the reference prediction frame;

[0034] (f) The prediction box uses the initial coordinates as the target coordinates, the initial category as the target category, and the updated initial confidence as the target confidence.

[0035] The present invention also adopts the following scheme: an intelligent recognition and abnormality judgment device for heart sections during embryonic development, the device comprising a scanner, an intelligent image processing device, a computing and processing device, a storage platform and a display, wherein the scanner is connected to the intelligent image processing device so that the heart section image of the target embryo can be quickly preprocessed after acquisition; the storage platform is connected to the computing and processing device to facilitate the conversion of the embryo specimen to be tested into three-dimensional structural data for analysis; the display is connected to the intelligent image processing device to present the processed image in real time; the intelligent image processing device is connected to the computing and processing device to transmit the preprocessed image to a computer for image data matching, and the target embryo detection result is obtained using the Faster-RCNN recognition model.

[0036] Preferably, the scanner is used to digitally process embryonic heart slices or heart cross-sectional images; the intelligent image processing device is used to identify the images processed by the scanner through an algorithm and form a two-dimensional image that meets the intelligent matching requirements; the computing and processing device is used to provide the required computing power for the entire device; the display is used to control the entire device through interactive operations on the screen, and display the obtained images and recognition results; the storage table is used to place the embryo specimen to be tested so as to perform tomography scanning and obtain its three-dimensional structural data.

[0037] Preferably, the intelligent image processing device is composed of an image preprocessing module, an image feature extraction module, an image recognition module, an image intelligent matching module, and a result display module.

[0038] Preferably, the image preprocessing module is responsible for converting, adjusting and aligning the incoming images; the image feature extraction module is responsible for extracting and collecting the feature structures, colors, textures and other features in the two-dimensional images and transmitting them to the image recognition module; the image recognition module is responsible for reprocessing according to the image feature information and presenting images that meet the intelligent matching requirements; the image intelligent matching module is responsible for matching through the gray scale matching method and the geometric matching method according to the information provided by the image feature extraction module and the image recognition module; and the result display module is responsible for intelligently selecting the most suitable result according to the above matching principles and algorithms and past learning experience.

[0039] The present application has the following advantages:

[0040] The present application helps to quickly identify the structure of the embryonic heart section and determine its position by establishing a four-dimensional image database of embryonic heart development and matching the two-dimensional image of the embryonic heart section to the four-dimensional image database; at the same time, according to the characteristics of the four-dimensional database timeline, the possible future development of the heart section is displayed, which is beneficial to the diagnosis of clinicians and the identification of embryonic heart section structure by researchers, reduces their workload and provides their work efficiency.

[0041] In addition, by comparing the measurement data of the key structures of the embryonic heart with the data of the typical samples already included in the four-dimensional image database, it can be preliminarily judged whether there is a deformity and abnormal structure, which has important significance for the prevention and treatment of fetal perinatal health care and congenital deformity diseases. BRIEF DESCRIPTION OF DRAWINGS

[0042] Figure 1 . Structure schematic diagram of the intelligent identification and positioning device for heart section in embryonic development period;

[0043] Figure 2 . Structure schematic diagram of the intelligent image processing device;

[0044] Figure 3 . Process flow chart of the intelligent image processing device;

[0045] Figure 4 . Method schematic diagram of collecting and constructing the four-dimensional image database of embryonic heart development;

[0046] Figure 5 . Method flow chart of intelligently identifying and preliminarily interpreting the embryonic heart section based on deep learning. DETAILED DESCRIPTION

[0047] The preferred embodiment of the present invention automatically identifies, labels and judges abnormalities of embryonic heart cross-section images at different developmental stages, different position levels and different angles. First, by constructing a four-dimensional image database of embryonic heart development, it contains embryonic heart data of normal development and embryonic heart data of heart malformation, and allows for the subsequent continuous addition and improvement of species and malformation data. Secondly, an intelligent identification and positioning device for embryonic heart cross-sections is set up. By adopting matching technology based on multiple dimensions and principles, the heart cross-section images obtained from clinical prenatal examinations or embryo experiments are matched with the data in the library to achieve the identification, positioning and labeling of the target embryonic heart cross-section during development. Finally, by setting up a comprehensive evaluation and judgment device for embryonic heart development, after the target embryonic heart cross-section is three-dimensionally reconstructed, key structural data is identified and measured, and by comparing it with the normal embryonic heart data and the malformed embryonic heart data in the library, the embryonic heart development is comprehensively evaluated, and a preliminary judgment is made as to whether it is deformed and what kind of deformity it is.

[0048] The following examples and experimental examples are used to illustrate the present invention, but are not intended to limit the scope of the present invention. The present invention will be further described below in conjunction with specific examples and experimental examples.

[0049] Example 1: Intelligent identification and positioning device for heart sections during embryonic development.

[0050] like Figure 1 As shown, the intelligent recognition and positioning device for heart sections during embryonic development includes a complete set of interactive equipment - a scanner, an intelligent image processing device, a computing device, a storage table and a display.

[0051] Among them, the scanner is connected to the intelligent image processing device, so that the target embryo heart cross-sectional image is quickly preprocessed after acquisition, and is used to digitize the embryo heart slices or heart cross-sectional images; the storage table is connected to the computing and processing device, which is convenient for converting the embryo specimen to be tested into three-dimensional structural data for analysis, and is used to place the embryo specimen to be tested so as to perform tomography and obtain its three-dimensional structural data; the display is connected to the intelligent image processing device, and the processed image is presented in real time, and is used to control the entire device through interactive operations on the screen, and display the obtained image and recognition results; the intelligent image processing device is used to identify the image processed by the scanner through an algorithm, and form a two-dimensional image that meets the intelligent matching requirements; the computing and processing device is used to provide the required computing power for the entire device; the intelligent image processing device is connected to the computing and processing device, and the preprocessed image is transmitted to the computer for image data matching, and the target embryo detection result is obtained using the Faster-RCNN recognition model.

[0052] like Figure 2As shown, the intelligent image processing device is composed of an image preprocessing module, an image feature extraction module, an image recognition module, an image intelligent matching module, and a result display module.

[0053] like Figure 3 As shown, the processing steps of the intelligent image processing device are as follows.

[0054] First, the image preprocessing module is responsible for converting, adjusting, and aligning the incoming image.

[0055] Secondly, the image feature extraction module is responsible for extracting and collecting the characteristic structure, color, texture and other features in the two-dimensional image and transmitting them to the image recognition module.

[0056] Then, the image recognition module is responsible for reprocessing the image feature information and presenting an image that meets the requirements of the intelligent matching system, whose content structure is consistent with the database image, to obtain an accurate image and its structural features.

[0057] The image intelligent matching module is responsible for matching the corresponding structure of the four-dimensional image in the database through grayscale matching and geometric matching methods based on the information provided by the image feature extraction module and the image recognition module, in accordance with the principles of fuzzy matching of embryonic development time, matching of key biomedical structures, and computer image matching.

[0058] The specific implementation principle of the image intelligent matching module is as follows: each pixel point is divided into different grayscale values ​​through the image color and image texture features extracted by the image feature extraction module, and is searched and matched with each pixel point of the four-dimensional image according to a certain similarity measurement method. At the same time, several adjacent pixel points are automatically integrated into a pixel block, and the same matching is performed with the image in the database. The one with the highest matching degree is the matching result; geometric matching refers to the structural features extracted from the image, by analyzing the actual composition meaning of this structural feature, and performing edge detection to obtain the point chain composed of each edge inflection point, and then statistically matching the image in the image database according to a certain deviation value.

[0059] The intelligent image matching module then applies these two methods—search matching and geometric matching—to match the 2D image against the database, selecting the most appropriate result. While both methods have a certain degree of error, the system intelligently provides the optimal result for each method and presents the matching percentage of the result to the original 2D image. Those skilled in the art can assess the value of an image based on this matching percentage.

[0060] Finally, the result display module is responsible for intelligently selecting the most suitable result based on the above matching principles and algorithms as well as previous learning experience.

[0061] Example 2: Method for intelligent recognition and abnormality judgment of heart sections during embryonic development.

[0062] 1. Collect and construct a four-dimensional image database of embryonic heart development.

[0063] like Figure 4 As shown, a sequence of two-dimensional images of embryonic heart sections is collected and arranged: the two-dimensional images of embryonic heart sections acquired by an existing system, or the two-dimensional images of embryonic heart sections obtained by four-dimensional color ultrasound or MRI scanning of the embryo during prenatal examinations in a hospital, are arranged in stages from small to large, and the arranged results are named consecutively.

[0064] The two-dimensional image of the embryonic heart section is preprocessed, and the three-dimensional image of the embryonic heart after preprocessing is labeled:

[0065] 1) Import the arranged and collated two-dimensional image data of the embryonic heart section into Amira three-dimensional graphics software, and adjust and align all the heart section images of the same embryo by setting the channel and pixel value.

[0066] 2) The important structural parts in each pre-processed three-dimensional image of the embryonic heart are outlined and marked with different colors to show the correspondence between different structures.

[0067] 3) Then, the superimposed and annotated three-dimensional spatial structure is displayed, and the constructed three-dimensional structure of the embryonic heart can also be cut into different planes and re-superimposed.

[0068] 4) Preserve the three-dimensional images and stereoscopic structural images of the embryonic heart and arrange them according to the timeline of embryonic development to construct a four-dimensional (three-dimensional space + time dimension) image database of embryonic heart development.

[0069] The four-dimensional image database of embryonic heart development continues to record and update various cases in clinical or scientific research in order to improve data such as species and malformation cases, and dynamically adjust to reduce errors.

[0070] 2. Evaluate and judge heart development during embryonic development.

[0071] 1) Label and measure key structures of the embryonic heart;

[0072] The processed three-dimensional image of the embryonic heart is cut using Amira three-dimensional graphics software to highlight important structures in the development of the embryonic heart, such as the outflow tract, atrioventricular canal, heart valves, atrial septum and ventricular septum, and perform data measurement.

[0073] 2) Calibration and comparison measurement data;

[0074] The measured data is calibrated by comparing it with the normal development structure data and typical deformity structure data included in the embryonic heart development four-dimensional image database.

[0075] 3) Result judgment;

[0076] Multi-structure analysis measures and compares data results to comprehensively evaluate the heart development of the target embryo.

[0077] Example 3: A method for intelligent identification and preliminary interpretation of embryonic heart sections based on deep learning.

[0078] like Figure 5 As shown, the method for intelligently identifying and preliminarily interpreting embryonic heart sections based on deep learning includes the following steps:

[0079] First, image acquisition and image preprocessing were performed on the embryonic heart section.

[0080] Methods for acquiring and receiving cross-sectional images of the heart of the patient to be tested include ultrasound, CT, MRI, etc. The cross-sectional images are preprocessed, and the preprocessed embryonic images are annotated. Taking CT images as an example for the preprocessing of the images, since the original CT images are 12-bit data, there will be information loss when they are converted to 8-bit images displayed by the computer, and the density information of organ tissues, bones, body fluids, etc. in the original images is close. If the image information is mapped to the computer without preprocessing, a large amount of information will be lost, so the window method invented by Hansenfeld is used to preprocess the images. Preprocessing of the embryonic heart images can improve the contrast of the required information to achieve the purpose of reducing useless information.

[0081] Secondly, after removing useless information through the attention mechanism, the ROI area (region of interest) image is obtained and compared with the normal heart embryo cross-sectional image database to detect whether the heart morphology and structure are abnormal, and obtain the tissue structure image containing the abnormal part.

[0082] Among them, the attention mechanism uses the label data of each data set to create an empirical template, that is, the probability distribution map of the target segmentation in the original image, and then combines this mask with the original image to remove a large amount of useless information, that is, negative samples, to obtain the ROI area (region of interest) image, which is convenient for the next step to be sent to the target detection neural network of the Faster-RCNN recognition model. It can reduce the training difficulty caused by category imbalance, reduce network output redundancy, and reduce the false positive rate.

[0083] Among them, the method for establishing the cross-sectional image database of normal heart embryos at various stages includes but is not limited to selecting a specimen to be prepared, confirming the integrity of the embryo preservation through MRI scanning, embedding, freezing and other processes, fixing it on a CNC milling machine, and performing continuous cross-sectional cutting on the embryo specimen. During this period, a digital camera continuously photographs the cross-sectional structure layer by layer to obtain a high-definition two-dimensional image of the embryo's cross-sectional structure, and performing structure identification and annotation, and storing the image data in a computer memory.

[0084] The image comparison method uses a deep learning-based image recognition algorithm, which has a higher image recognition accuracy than traditional template-based recognition and comparison methods. When compared with a normal heart embryo cross-sectional image, highly consistent image areas are removed, and image areas showing abnormalities are retained to determine whether the heart is developing normally.

[0085] Then, the abnormal area images retained after comparison are input into the pre-trained Faster-RCNN recognition model to generate tissue abnormality prediction results.

[0086] Both the normal embryonic heart cross-sectional image data and the abnormal embryonic heart images used for comparison with the acquired images in the Faster-RCNN recognition model are derived from a pre-established embryonic heart cross-sectional image database.

[0087] The Faster-RCNN recognition model includes a feature extraction module, an RPN module, a Proposal Layer module, a Roi Align module, a classification and regression module, and a C-NMS module.

[0088] Inputting the abnormal embryo cross-sectional image data into a pre-trained Faster-RCNN recognition model to generate a prediction result comprises the following steps:

[0089] 1) Extract feature maps from the original image using a set of basic conv+relu+pooling layers, including Convlayers, RPN layers, and fully connected layers;

[0090] 2) Input the feature map into the RPN network for recognition and screening, obtain the approximate location of the target from the feaure map, and obtain the recommended box;

[0091] 3) In the Proposal Layer, the approximate position obtained by the RPN is used to continue training to obtain a more accurate position, that is, to obtain the recommended box;

[0092] 4) Input the obtained recommendation box and feature map into the ROI Align network for mapping and pooling to obtain the feature recommendation map;

[0093] 5) Inputting the feature recommendation graph into the classification regression network for classification regression processing to obtain the initial coordinates, initial category and initial confidence of the prediction box;

[0094] 6) Inputting the initial coordinates, initial category, and initial confidence of the predicted frame into the C-NMS module for screening and processing to obtain the target coordinates, target category, and target confidence of the predicted frame;

[0095] The specific steps are as follows: input the initial coordinates, initial category and initial confidence of the prediction frame into the C-NMS network; take the prediction frame with the largest initial confidence as the reference prediction frame; calculate the overlap ratio of each prediction frame with the reference prediction frame according to the initial coordinates; take the prediction frame with the overlap ratio greater than or equal to the preset confidence threshold as the prediction frame to be adjusted; update the initial confidence of the prediction frame to be adjusted according to the overlap ratio and area ratio between the prediction frame to be adjusted and the reference prediction frame; the prediction frame takes the initial coordinates as the target coordinates, the initial category as the target category, and the updated initial confidence as the target confidence.

[0096] Finally, the tissue lesion type is identified based on the prediction results.

[0097] The beneficial effects of this embodiment are as follows: the Faster-RCNN recognition model used in this method can effectively improve the detection rate of cardiac morphological structural abnormalities and reduce the false positive rate. This method utilizes an attention mechanism to preprocess the cross-sectional image to be tested, and uses existing data to create an empirical mask, that is, a probability distribution map of the target segmentation in the original image. Combined with the image to be tested, this mask can remove a large amount of useless information (negative samples) in the original image, reduce the increased training difficulty caused by class imbalance during deep learning, and reduce network output redundancy.

[0098] The beneficial effects of the present invention are: when fetal gene testing indicates a high rate of a single-gene genetic disease related to heart development, or a mouse model is specially made to silence the gene, image data of the heart section of the developing embryonic specimen can be obtained through histological or imaging technology. The use of the present invention can help clinicians or researchers quickly confirm the various structures of the embryonic heart section, and can also assist in the preliminary judgment and evaluation of the target embryonic heart development.

Claims

1. A method for intelligent recognition and abnormality determination of heart sections during embryonic development, characterized by: The method comprises the following steps: (1) collecting and constructing a four-dimensional image database of embryonic heart development, (2) matching target embryonic heart cross-sectional image data with data in the database, and identifying, locating, and measuring the cross-sectional image data, (3) creating a Faster-RCNN recognition model, and (4) obtaining a detection result of the target embryonic heart using the Faster-RCNN recognition model; The step (1) of collecting and constructing a four-dimensional image database of embryonic heart development includes: (a) Arranging the collected or obtained two-dimensional images of the embryonic heart sections from the target embryo into stages in order of embryo size; (b) processing the arranged two-dimensional image of the embryonic heart section into a three-dimensional image, and outlining, superimposing and annotating important structural parts therein; (c) Arrange the images according to the timeline of embryonic development and construct a four-dimensional image database of embryonic heart development; The step (2) matches the target embryo heart cross-sectional image data with the data in the database, and identifies, locates and measures the target embryo heart cross-sectional image data, including: (a) removing useless information from the target embryonic heart cross-sectional image data through an attention mechanism to obtain an image of a region of interest; (b) matching the region of interest image with corresponding structures of the four-dimensional image data in the database by grayscale matching and geometric matching; marking the outflow tract, atrioventricular canal, heart valve, atrial septum and ventricular septum of the cross-sectional image data of the target embryonic heart, and performing data measurement; (c) calibrating the measured data by comparing it with normal developmental structure data and typical deformed structure data included in the four-dimensional image database of embryonic heart development; and comprehensively evaluating the heart development of the target embryo using multi-structure analysis measurement and comparison data results.

2. The method for intelligent recognition and abnormality determination of heart sections during embryonic development according to claim 1, wherein: The step (3) creates a Faster-RCNN recognition model, which includes a feature extraction module, an RPN module, a Proposal Layer module, a Roi Align module, a classification and regression module, and a C-NMS module.

3. The method for intelligent recognition and abnormality determination of heart sections during embryonic development according to claim 2, wherein: Step (4) includes: inputting the processed target embryo heart cross-sectional image data into the Faster-RCNN recognition model to generate a detection result, and further includes: (a) The feature extraction module uses a set of basic conv+relu+pooling layers to extract feature maps, including Convlayers, RPN layers and fully connected layers; (b) The RPN module inputs the feature map into the RPN network for recognition and screening, obtains the approximate location of the target from the feaure map, and obtains a recommended box for the target location to be measured; (c) The Proposal Layer module uses the approximate position obtained by the RPN to continue training and obtain a more accurate position from the proposed frame of the target position, that is, obtains a proposed frame with a more accurate position of the target position; (d) the Roi Align module inputs the acquired recommended box of the more accurate position of the target and the feature map into the ROI Align network for mapping and pooling to obtain a feature recommendation map; (e) The classification and regression module inputs the obtained feature recommendation graph into the classification and regression network for classification and regression processing to obtain the initial coordinates, initial category and initial confidence of the prediction box; (f) The C-NMS module inputs the initial coordinates, initial category, and initial confidence of the prediction box into the C-NMS module for screening and processing to obtain the target coordinates, target category, and target confidence of the prediction box.

4. The method for intelligent recognition and abnormality determination of heart sections during embryonic development according to claim 3, wherein: The initial coordinates, initial category, and initial confidence of the predicted frame are input into the C-NMS module for screening and processing to obtain the target coordinates, target category, and target confidence of the predicted frame. The steps include: (a) The initial coordinates, initial category and initial confidence of the predicted box are input into the C-NMS network; (b) taking the prediction frame with the highest initial confidence as the reference prediction frame; (c) calculating the overlap ratio of each prediction frame with the reference prediction frame according to the initial coordinates; (d) taking the prediction frame whose overlap ratio is greater than or equal to a preset confidence threshold as the prediction frame to be adjusted; (e) updating the initial confidence of the prediction frame to be adjusted according to the overlap ratio and area ratio between the prediction frame to be adjusted and the reference prediction frame; (f) The prediction box uses the initial coordinates as the target coordinates, the initial category as the target category, and the updated initial confidence as the target confidence.

5. A device for intelligent recognition and abnormality judgment of heart sections during embryonic development, which executes the method for intelligent recognition and abnormality judgment of heart sections during embryonic development according to claim 1, characterized in that: The device includes a scanner, an intelligent image processing device, a computing and processing device, a storage platform and a display, wherein the scanner is connected to the intelligent image processing device so that the target embryo's heart cross-sectional image can be quickly preprocessed after acquisition; the storage platform is connected to the computing and processing device to facilitate conversion of the embryo specimen to be tested into three-dimensional structural data for analysis; the display is connected to the intelligent image processing device to instantly present the processed image; the intelligent image processing device is connected to the computing and processing device to transmit the preprocessed image to a computer for image data matching, and the target embryo detection result is obtained using the Faster-RCNN recognition model.

6. The device for intelligent recognition and abnormality judgment of heart sections during embryonic development as claimed in claim 5, characterized in that: The scanner is used to digitally process embryonic heart slices or heart cross-sectional images; The intelligent image processing device is used to identify the image processed by the scanner through an algorithm and form a two-dimensional image that meets the intelligent matching requirements; The computing and processing device is used to provide the required computing power for the entire device; The display device is used to control the entire device through interactive operations on the screen and display the obtained image and recognition results; The object placement table is used to place the embryo specimen to be tested so as to perform tomography scanning and obtain its three-dimensional structure data.

7. The device for intelligent recognition and abnormality judgment of heart sections during embryonic development according to claim 5, characterized in that: The intelligent image processing device is composed of an image preprocessing module, an image feature extraction module, an image recognition module, an image intelligent matching module, and a result display module.

8. The device for intelligent recognition and abnormality judgment of heart sections during embryonic development according to claim 7, characterized in that: The image preprocessing module is responsible for converting, adjusting and aligning the incoming images; The image feature extraction module is responsible for extracting and collecting the characteristic structure, color, texture and other features in the two-dimensional image and transmitting them to the image recognition module; The image recognition module is responsible for reprocessing the image feature information and presenting an image that meets the intelligent matching requirements; The image intelligent matching module is responsible for matching based on the information provided by the image feature extraction module and the image recognition module through grayscale matching method and geometric matching method; The result display module is responsible for intelligently selecting the most suitable result based on the above matching principles and algorithms as well as previous learning experience.

Images