CN114724720B - Prenatal electronic fetal heart monitoring automatic identification system based on deep learning - Google Patents
Prenatal electronic fetal heart monitoring automatic identification system based on deep learning Download PDFInfo
- Publication number
- CN114724720B CN114724720B CN202210649767.1A CN202210649767A CN114724720B CN 114724720 B CN114724720 B CN 114724720B CN 202210649767 A CN202210649767 A CN 202210649767A CN 114724720 B CN114724720 B CN 114724720B
- Authority
- CN
- China
- Prior art keywords
- data
- segment
- result
- fetal heart
- monitoring
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
Images
Classifications
-
- G—PHYSICS
- G16—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR SPECIFIC APPLICATION FIELDS
- G16H—HEALTHCARE INFORMATICS, i.e. INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR THE HANDLING OR PROCESSING OF MEDICAL OR HEALTHCARE DATA
- G16H50/00—ICT specially adapted for medical diagnosis, medical simulation or medical data mining; ICT specially adapted for detecting, monitoring or modelling epidemics or pandemics
- G16H50/30—ICT specially adapted for medical diagnosis, medical simulation or medical data mining; ICT specially adapted for detecting, monitoring or modelling epidemics or pandemics for calculating health indices; for individual health risk assessment
-
- A—HUMAN NECESSITIES
- A61—MEDICAL OR VETERINARY SCIENCE; HYGIENE
- A61B—DIAGNOSIS; SURGERY; IDENTIFICATION
- A61B8/00—Diagnosis using ultrasonic, sonic or infrasonic waves
- A61B8/02—Measuring pulse or heart rate
-
- A—HUMAN NECESSITIES
- A61—MEDICAL OR VETERINARY SCIENCE; HYGIENE
- A61B—DIAGNOSIS; SURGERY; IDENTIFICATION
- A61B8/00—Diagnosis using ultrasonic, sonic or infrasonic waves
- A61B8/52—Devices using data or image processing specially adapted for diagnosis using ultrasonic, sonic or infrasonic waves
- A61B8/5215—Devices using data or image processing specially adapted for diagnosis using ultrasonic, sonic or infrasonic waves involving processing of medical diagnostic data
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/25—Fusion techniques
- G06F18/253—Fusion techniques of extracted features
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
-
- G—PHYSICS
- G16—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR SPECIFIC APPLICATION FIELDS
- G16H—HEALTHCARE INFORMATICS, i.e. INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR THE HANDLING OR PROCESSING OF MEDICAL OR HEALTHCARE DATA
- G16H50/00—ICT specially adapted for medical diagnosis, medical simulation or medical data mining; ICT specially adapted for detecting, monitoring or modelling epidemics or pandemics
- G16H50/20—ICT specially adapted for medical diagnosis, medical simulation or medical data mining; ICT specially adapted for detecting, monitoring or modelling epidemics or pandemics for computer-aided diagnosis, e.g. based on medical expert systems
-
- A—HUMAN NECESSITIES
- A61—MEDICAL OR VETERINARY SCIENCE; HYGIENE
- A61B—DIAGNOSIS; SURGERY; IDENTIFICATION
- A61B2503/00—Evaluating a particular growth phase or type of persons or animals
- A61B2503/02—Foetus
Landscapes
- Engineering & Computer Science (AREA)
- Health & Medical Sciences (AREA)
- Life Sciences & Earth Sciences (AREA)
- Theoretical Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Physics & Mathematics (AREA)
- Biomedical Technology (AREA)
- General Health & Medical Sciences (AREA)
- Public Health (AREA)
- Medical Informatics (AREA)
- Artificial Intelligence (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Evolutionary Computation (AREA)
- Molecular Biology (AREA)
- Pathology (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Biophysics (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Evolutionary Biology (AREA)
- Bioinformatics & Computational Biology (AREA)
- Primary Health Care (AREA)
- Nuclear Medicine, Radiotherapy & Molecular Imaging (AREA)
- Mathematical Physics (AREA)
- Computing Systems (AREA)
- Computational Linguistics (AREA)
- Epidemiology (AREA)
- Software Systems (AREA)
- Databases & Information Systems (AREA)
- Radiology & Medical Imaging (AREA)
- Heart & Thoracic Surgery (AREA)
- Surgery (AREA)
- Animal Behavior & Ethology (AREA)
- Veterinary Medicine (AREA)
- Cardiology (AREA)
- Measuring And Recording Apparatus For Diagnosis (AREA)
Abstract
The invention provides a prenatal electronic fetal heart monitoring automatic identification system based on deep learning, which comprises: the data preprocessing module is used for preprocessing the data of the electronic fetal heart monitoring data of the monitored target patient; the single-segment fetal heart monitoring evaluation module is used for acquiring single-segment data corresponding to the preprocessed data, obtaining current scores of the single-segment data based on a backbone network and an order regression model, and judging whether the target patient needs to continue monitoring; and the multi-segment fetal heart monitoring comprehensive evaluation module is used for acquiring corresponding multi-segment data when monitoring is required to be continued, carrying out comprehensive diagnosis, acquiring a fetal heart interpretation result, carrying out accuracy evaluation on the fetal heart interpretation result and acquiring a monitoring suggestion. The method can realize the evaluation of the health condition of the fetal heartbeat, measure the accuracy of the evaluation result, and provide reliable result evaluation for doctors while improving the efficiency and the accuracy.
Description
Technical Field
The invention relates to the technical field of deep learning, in particular to an automatic prenatal electronic fetal monitoring identification system based on deep learning.
Background
Electronic extra-Fetal cardiac Monitoring (External Electronic Fetal Rate Monitoring) is the only real-time Fetal noninvasive Monitoring means widely used worldwide. The scheme mainly obtains a fetal heart uterine contraction graph (CTG) according to a fetal heart and uterine contraction data tracing graph. The advantages of the monitoring means include convenient, safe and real-time dynamic monitoring, and wide application in different regions and different levels of medical institutions. At present, the dynamic assessment of intrauterine safety risk of a fetus through the interpretation of prenatal CTG mainly comprises two modes: visual CTG interpretation and computerized CTG interpretation. The visual CTG interpretation is mainly subjective visual assessment of medical workers, but due to the fact that CTG graphs are varied, wrong interpretation, different interpreters and even inconsistent results before and after the interpretation result of the interpreters occur frequently, and the consistency and reproducibility of image results are poor, the visual CTG interpretation is an important problem which troubles clinicians.
The development of computer technology has made computer-aided CTG interpretation a new interpretation approach. Computerized CTGs were the first to mathematically process the four basic features of visual assessment to develop a computer recognition system that converts pure visual assessment into computer-aided known parameter assessment. Then, the scholars extract new parameters related to the fetal condition and carry out quantitative evaluation to judge the fetal pathological condition. However, some prior studies have also shown that conventional computerized CTG interpretation is not better than conventional visual CTG interpretation indicators. Recently, some researches have made certain progress by adopting a mode of combining traditional feature extraction and machine learning. However, these models still rely heavily on the effectiveness of manually extracting features, and at the same time, these models can only output the interpretation result of the segment, and cannot provide the judgment accuracy prediction of the prediction result by the models. In recent years, with the emergence of deep learning algorithms, fetal heart interpretation with the assistance of artificial intelligence becomes possible. Compared with the traditional computer-aided CTG interpretation, the interpretation information obtained by deep learning does not depend on the characteristics of traditional manual extraction, so that the method has wider application scenes. Meanwhile, the features obtained through deep learning can often focus on features ignored by the traditional visual CTG, and therefore the interpretation effect of traditional doctors can be achieved and even surpassed.
Therefore, the invention provides an automatic identification system for prenatal electronic fetal heart monitoring based on deep learning.
Disclosure of Invention
The invention provides an automatic prenatal electronic fetal heart monitoring identification system based on deep learning, which is used for solving the technical problems.
The invention provides a prenatal electronic fetal heart monitoring automatic identification system based on deep learning, which comprises:
the data preprocessing module is used for preprocessing the data of the electronic fetal heart monitoring data of the monitored target patient;
the single-segment fetal heart monitoring evaluation module is used for acquiring single-segment data corresponding to the preprocessed data, obtaining current scores of the single-segment data based on a backbone network and a sequence regression model, and judging whether a target patient needs to continue monitoring;
and the multi-segment fetal heart monitoring comprehensive evaluation module is used for acquiring corresponding multi-segment data when monitoring is required to be continued, performing comprehensive diagnosis, acquiring a fetal heart interpretation result, performing accuracy evaluation on the fetal heart interpretation result and acquiring a monitoring suggestion.
Preferably, the data preprocessing module includes:
the data cleaning module is used for acquiring the electronic fetal heart monitoring data, eliminating information missing segments, and acquiring the monitoring data corresponding to the information missing segments in the same time again to obtain first monitoring data when the information missing segments meet a missing acquisition standard;
the data normalization module is used for counting the mean value and the variance of the fetal heart rate corresponding to the first monitoring data and carrying out normalization processing on the first monitoring data by utilizing the mean value and the variance to obtain second monitoring data;
and the second monitoring data is preprocessed data.
Preferably, the single-segment fetal heart monitoring evaluation module includes:
the main network is used for extracting fetal heart monitoring characteristics from bottom to top in the preprocessed data by adopting a one-dimensional residual convolution network as a basic frame;
the sequence regression model is used for carrying out linear transformation and normalization processing on the fetal heart monitoring characteristics to obtain normalization parameters;
on the basis of the normalization parameters, comparing the normalization parameters with a preset number of learnable reference feature vectors respectively to obtain a preset number of comparison scores, wherein each comparison score corresponds to one classification grade;
optimizing two classification loss functions of a preset number of comparison scores, wherein the supervision information of the ith classifier is determined by whether the result labeled by a doctor is greater than or equal to the grade corresponding to the corresponding reference feature, and the weight corresponding to each two classification loss function is determined by the ratio of the number of samples in the data set, which are greater than the classification grade, to the number of samples, which are smaller than the classification grade;
accumulating the preset number of comparison scores, and judging that the corresponding fetal heart monitoring segment is classified into the ith segment when the accumulated score is larger than the adjustable threshold corresponding to the ith class and smaller than the adjustable threshold corresponding to the (i + 1) th class, so as to obtain a current score;
wherein the fetal heart monitoring segment corresponds to the single segment data.
Preferably, the multi-segment fetal heart monitoring comprehensive evaluation module includes:
the multi-fragment interpretation module is used for interpreting the multi-fragment characteristics corresponding to the multi-fragment data to obtain a multi-fragment interpretation result;
and the interpretation accuracy prediction module is used for training an autoregressive model and predicting the accuracy of the fetal heart judgment result.
Preferably, the multi-fragment interpretation module includes:
the single-segment feature extraction unit is used for sequentially extracting features of single-segment data in the multi-segment data based on a one-dimensional residual error network;
the characteristic fusion unit is used for sequentially inputting the fetal heart characteristics of each single segment into the encoder structure and obtaining a plurality of correspondingly fused new segment characteristics through the self-attention layer, the first normalization layer, the feedforward neural network layer and the second normalization layer;
the overall feature extraction unit is used for splicing the new segment features and extracting the overall feature through a multilayer perception mechanism;
the classification unit is used for classifying the overall characteristics by using the sequence regression model to obtain a multi-segment interpretation result;
wherein short links are added at the self-attention layer and the feedforward neural network layer.
Preferably, the interpretation accuracy prediction module includes:
the feature prediction unit is used for acquiring the multi-segment data and the blank segment data, fusing the multi-segment data and the plurality of new segment features acquired from the multi-segment interpretation module through an attention mechanism, and predicting the next segment feature;
the accuracy rate determining unit is used for determining the accuracy rate of the predicted next segment characteristics based on the difference between the predicted next segment characteristics and the segment characteristics of the actual monitored data;
and taking the accuracy as an accuracy evaluation result of the fetal heart judgment result.
Preferably, the method further comprises the following steps:
the result judgment module is used for determining whether the judgment accuracy prediction module needs to be adjusted or not based on the accuracy evaluation result;
when the accuracy evaluation result is determined to be larger than the corresponding preset result time, judging that the judgment accuracy prediction module is not required to be adjusted;
otherwise, judging that the judgment accuracy prediction module needs to be adjusted;
the accuracy adjusting module is used for acquiring a first feature difference set of the predicted next segment feature and the corresponding first actual monitoring segment feature when the judgment accuracy predicting module is judged to need to adjust;
reversely predicting a previous segment feature based on a next segment feature, and comparing the previous segment feature with a corresponding second actual monitoring segment feature to obtain a second feature difference set;
constructing a first difference map based on the first set of feature differences, and simultaneously constructing a second difference map based on the second set of feature differences;
performing first framing on the difference range of the difference part in the first difference map, and simultaneously performing second framing on the difference range of the difference part in the second difference map;
judging whether the first frame selection result and the second frame selection result meet the rule consistency, if so, judging that the prediction abnormity is only related to a judgment accuracy prediction module;
determining an index to be adjusted according to the rule consistency, and optimizing the judgment accuracy prediction module according to the index to be adjusted;
otherwise, judging whether the prediction abnormity is related to the judgment accuracy prediction module and the index to be mined;
acquiring the frame selection consistent area and the frame selection inconsistent area of the first frame selection result and the second frame selection result, performing first calibration on the frame selection consistent area, performing second calibration on the frame selection inconsistent area matched with the first frame selection result, and performing third calibration on the frame selection inconsistent area matched with the second frame selection result;
constructing a calibration map according to the first calibration result, the second calibration result and the third calibration structure;
determining a first calibration area and a first calibration position related to a second calibration result in the calibration map to obtain a corresponding first reference weight;
wherein Y1 represents a first reference weight; s1 denotes a first calibration junctionArea value corresponding to the fruit; s2 represents a first calibration area corresponding to the second calibration result; s3 represents a second calibration area corresponding to the third calibration result; r represents a total position set of a corresponding calibration map; r2 represents a position set corresponding to the second calibration result;
represents the location weight assigned for R2;
indicating the position weight assigned to the position set R2 corresponding to the second calibration result and the total position set R;
representing the first nominal area versus the total areas1+s2+s3 assigned area weight;
determining a second calibration area and a second calibration position related to a third calibration result in the calibration map to obtain a corresponding second reference weight;
wherein Y2 represents a second reference weight; r3 represents a position set corresponding to the third calibration result;
based on the first reference weight and a second reference weight;
acquiring first transition factors of the second calibration result and the first calibration result, and acquiring second transition factors of the third calibration result and the first calibration result;
matching to obtain an index to be mined from the weight-factor-index database based on the reference weight and the transition factor;
determining an index to be optimized according to the second calibration result and the third calibration result;
and processing the judgment accuracy prediction module based on the index to be optimized and the index to be mined.
Preferably, the data normalization module further includes:
the data analysis unit is used for performing normalization processing on the first monitoring data by using the mean value and the variance, performing pre-analysis on the normalization processing data and judging a zero point existing in the normalization processing data;
the curve drawing unit is used for taking the normalized data formed by the adjacent zero points as a data segment and drawing a data curve of the data segment;
a first determination unit configured to determine a valley point and a peak point existing in the data curve, and perform a first determination on a length and a slope of a connection line formed by each of the valley point and the peak point;
a first obtaining unit, configured to obtain, according to the first determination result and based on a timestamp sequence, a first line sequence from a valley point to a peak point and a second line sequence from the peak point to the valley point;
the screening unit is used for carrying out regular consistency detection on the first line sequence and the second line sequence and screening to obtain a first inconsistent sequence and a second inconsistent sequence;
the second determining unit is used for determining a first position of the first inconsistent sequence and a second position of the second inconsistent sequence, and determining the probability to be adjusted according to the first position and the second position;
l1 represents the number of the first inconsistent sequences; l2 indicates the number of second inconsistent sequences;
indicating the position adjustable weights assigned for all first positions;
indicating the position adjustable weights assigned for all second positions;h k1 a sequence value representing the k1 th first inconsistent sequence;h k1,0 standard values representing the k1 th first inconsistent sequence;h k2 a sequence value representing the k2 th second inconsistent sequence;h k2,0 standard values representing the k2 th second inconsistent sequence;
the inserting unit is used for deleting the valley points and the peak points of the first line sequence and the second line sequence when the probability to be adjusted is larger than the preset adjusting probability, determining a first inserting position for inserting the valley points and a second inserting position for inserting the peak points in the corresponding data segment, and simultaneously determining a first inserting position for inserting the valley points and a second inserting position for inserting the peak points to obtain a qualified sequence so as to obtain corresponding second monitoring data;
when the probability to be adjusted is not greater than the preset adjustment probability, based on
Determining a corresponding primary and secondary adjustment sequence;
a supplement unit, configured to delete the valley point and the peak point corresponding to the main adjustment sequence, and sequentially supplement the deleted sequences with the received data of the receiving probe corresponding to the monitored data with the largest signal-to-noise ratio, and establish a mapping relationship between the supplement sequences and the optimal position;
meanwhile, acquiring a valley point and a peak point corresponding to the secondary adjustment sequence from fetal heart monitoring data acquired from the optimal fetal heart monitoring position to perform numerical value expansion or numerical value reduction processing;
and the second acquisition unit is used for acquiring a qualified sequence based on the mapping relation and the data processing result so as to acquire corresponding second monitoring data.
Additional features and advantages of the invention will be set forth in the description which follows, and in part will be obvious from the description, or may be learned by the practice of the invention. The objectives and other advantages of the invention will be realized and attained by the structure particularly pointed out in the written description and drawings.
The technical solution of the present invention is further described in detail by the accompanying drawings and embodiments.
Drawings
The accompanying drawings, which are included to provide a further understanding of the invention and are incorporated in and constitute a part of this specification, illustrate embodiments of the invention and together with the description serve to explain the principles of the invention and not to limit the invention. In the drawings:
FIG. 1 is a block diagram of an embodiment of an automatic prenatal electronic fetal monitoring identification system based on deep learning;
FIG. 2 is a block diagram of acquiring fetal heart monitoring characteristics in an embodiment of the present invention;
FIG. 3 is a block diagram illustrating steps performed by the sequential regression model according to an embodiment of the present invention;
FIG. 4 is a block diagram of a multi-segment interpretation model according to an embodiment of the present invention;
FIG. 5 is a diagram of input and output information during operation of a module according to an embodiment of the present invention.
Detailed Description
The preferred embodiments of the present invention will be described in conjunction with the accompanying drawings, and it will be understood that they are described herein for the purpose of illustration and explanation and not limitation.
Example 1:
the invention provides a prenatal electronic fetal heart monitoring automatic identification system based on deep learning, as shown in figure 1, comprising:
the data preprocessing module is used for preprocessing the data of the electronic fetal heart monitoring data of the monitored target patient;
the single-segment fetal heart monitoring evaluation module is used for acquiring single-segment data corresponding to the preprocessed data, obtaining current scores of the single-segment data based on a backbone network and a sequence regression model, and judging whether a target patient needs to continue monitoring;
and the multi-segment fetal heart monitoring comprehensive evaluation module is used for acquiring corresponding multi-segment data when monitoring is required to be continued, performing comprehensive diagnosis, acquiring a fetal heart interpretation result, performing accuracy evaluation on the fetal heart interpretation result and acquiring a monitoring suggestion.
In the embodiment, the conventional prenatal electronic fetal heart monitoring image (namely the acquired electronic fetal heart monitoring data) can be fully utilized to accurately interpret the fetal condition, and the CTG interpretation with higher sensitivity and specificity is obtained by integrating multi-time and multi-segment CTG data and combining historical information of fetal heartbeat. Meanwhile, the specially designed multi-fragment interpretation accuracy rate prediction module can predict the interpretation accuracy rate and provide richer interpretation information. The interpretation is generated by the same computer program, so that the consistency and the mobility are good, and the accurate diagnosis can be made by a doctor.
The beneficial effects of the above technical scheme are: the invention can realize the evaluation of the health condition of the fetal heartbeat and the measurement of the accuracy of the evaluation result by pre-processing the data, and performing the single-segment data analysis and the multi-segment data analysis subsequently based on the existing fetal heart data acquisition process through deep learning auxiliary electronic fetal heart monitoring data, thereby improving the efficiency and the accuracy and providing reliable result evaluation for doctors.
Example 2:
based on embodiment 1, the data preprocessing module includes:
the data cleaning module is used for acquiring the electronic fetal heart monitoring data, eliminating information missing segments, and acquiring the monitoring data corresponding to the information missing segments in the same time again to obtain first monitoring data when the information missing segments meet a missing acquisition standard;
the data normalization module is used for counting the mean value and the variance of the fetal heart rate corresponding to the first monitoring data and performing normalization processing on the first monitoring data by utilizing the mean value and the variance to obtain second monitoring data;
and the second monitoring data is preprocessed data.
In the embodiment, in an actual application scenario, due to the problems of maternal movement, poor contact of equipment and the like, the acquired electronic fetal heart monitoring data often has the problems of segment information loss and the like. The adoption of such segments for training and prediction can bring interference to the model and also can not generate an accurate model. Therefore, data processing requires such data to be culled. Specifically, for fragments with deletion accumulation reaching 50% or a deletion interval exceeding 30%, the fragments need to be removed from a training set and a testing set in the training process; in actual application, data of the same time should be collected again to ensure model accuracy.
In this embodiment, since the fluctuation range of the electronic fetal heart monitoring data is large and the whole is far away from the zero point, the data needs to be normalized to meet the requirement of the neural network. And uniformly normalizing the electronic fetal heart monitoring data. Specifically, the mean and variance of the fetal heart rate corresponding to all the data are counted, and the integral data are normalized by using the numerical value.
The beneficial effects of the above technical scheme are: by providing the information missing segments and utilizing the mean value and the variance to carry out normalization processing, reliable basic data are convenient to obtain, and the evaluation accuracy of subsequent single-segment data is convenient.
Example 3:
based on embodiment 1, the single-segment fetal heart monitoring evaluation module includes:
the main network is used for extracting fetal heart monitoring characteristics from bottom to top in the preprocessed data by adopting a one-dimensional residual convolution network as a basic frame;
the sequence regression model is used for carrying out linear transformation and normalization processing on the fetal heart monitoring characteristics to obtain normalization parameters;
on the basis of the normalization parameters, comparing the normalization parameters with a preset number of learnable reference feature vectors respectively to obtain a preset number of comparison scores, wherein each comparison score corresponds to one classification grade;
optimizing two classification loss functions of a preset number of comparison scores, wherein the supervision information of the ith classifier is determined by whether the result labeled by a doctor is greater than or equal to the grade corresponding to the corresponding reference feature, and the weight corresponding to each two classification loss function is determined by the ratio of the number of samples in the data set, which are greater than the classification grade, to the number of samples, which are smaller than the classification grade;
accumulating the preset number of comparison scores, and judging that the corresponding fetal heart monitoring segment is classified into the ith segment when the accumulated score is larger than the adjustable threshold corresponding to the ith class and smaller than the adjustable threshold corresponding to the (i + 1) th class, so as to obtain a current score;
wherein the fetal heart monitoring segment corresponds to the single segment data.
In this embodiment, the typical duration of the electronic fetal heart monitoring is 20 minutes, and based on the 20-minute detection result, the doctor can determine whether the patient needs to continue the detection. This requirement is addressed by providing the capability of single-fragment (20-minute) detection. The training of the module relies on the scoring of a professionally trained physician. The doctor scores 0-5 points for each CTG image used for training according to the experience of the doctor, wherein the score 0 represents the extreme danger of fetal heart monitoring and needs direct intervention; score 5 indicates that the fetal heartbeat is completely normal.
In the embodiment, the electronic fetal heart monitoring data has the characteristics of one-dimensional translation invariance and a large number of local features, so that the fetal heart monitoring features are extracted from bottom to top by adopting a one-dimensional residual convolution network as a basic frame. The main structure is shown in figure 2:
since the existing large model can generate the overfitting problem, an 18-layer convolution network is adopted as the framework of the backbone network. The first layer of the network employs a convolution kernel of size 7. In order to enlarge the field of view of the backbone network, a hole convolution technique is adopted in each layer of the network except the first layer. The hole convolution can increase the receptive field of the model under the condition of not expanding the parameter number, and further more accurate information characteristics can be obtained.
In this embodiment, the features obtained by the direct one-dimensional residual convolution network will not achieve the most accurate classification effect, which is caused by the imbalance of fetal heart data acquisition. A large amount of fetal heart data behaves normally, while only a small portion of fetal heart data is interpreted as abnormal. Meanwhile, the one-dimensional residual convolution structure is mainly applied to the classification problem, and fetal heart interpretation has very obvious sequence characteristics, namely, different classification results have obvious superior-inferior relation. The problem is solved by introducing an order regression model by combining two considerations. The order regression model represents the order relationship between classes by translating a multi-class problem into a comparison of multiple different learnable samples, i.e., multiple bi-class problems. In each dichotomy problem, a method of redistributing weights may be used to offset the effect of sample imbalance. The results of these two classifications also imply varying degrees of anomaly ranking. Based on this model, for example, 5 representative samples are learned to be generated, corresponding to 1-5 points of criterion. And comparing the input features with the feature samples, and integrating the comparison results of the samples to obtain a final classification judgment result.
In this embodiment, if there are 5 learnable reference feature vectors and 5 comparison scores are obtained, the above-mentioned steps of the sequential regression model are performed according to this embodiment, as shown in fig. 3.
The beneficial effects of the above technical scheme are: by adopting the one-dimensional residual convolution network and bottom-up extraction operation, accurate monitoring characteristics can be conveniently obtained, the over-fitting problem can be reduced through the adopted pork liver network, and more accurate information characteristics can be obtained.
Example 4:
based on embodiment 1, the multi-segment fetal heart monitoring comprehensive evaluation module includes:
the multi-segment interpretation module is used for interpreting the multi-segment characteristics corresponding to the multi-segment data to obtain a multi-segment interpretation result;
and the interpretation accuracy rate prediction module is used for training an autoregressive model and predicting the accuracy of the fetal heart judgment result.
In this embodiment, although a single segment fetal heart monitoring evaluation module is sufficient to cope with a portion of fetal heart monitoring scenarios, there are a significant number of fetal heart monitoring features that are not detectable by a 20 minute monitor. Meanwhile, in practical application, the accuracy of interpretation of doctors can be effectively improved by multi-segment CTG images, and missed diagnosis and misdiagnosis are avoided. Therefore, in case of a difficult case, the doctor may choose to increase the fetal heart monitoring duration to acquire more electronic fetal heart monitoring data. Based on the practical requirements, the multi-segment fetal heart monitoring comprehensive evaluation module is designed and realized. The multi-segment fetal heart monitoring and evaluating comprehensive module carries out comprehensive diagnosis on the multi-segment fetal heart data and provides more accurate fetal heart interpretation. In addition, the multi-segment comprehensive fetal heart monitoring evaluation module can evaluate the accuracy of current judgment and provide suggestions for judging whether more fetal heart monitoring data are needed.
The beneficial effects of the above technical scheme are: by the aid of the autoregressive model for interpretation and training of the multi-segment features, accuracy assessment is facilitated, accuracy of assessment results can be measured, efficiency and accuracy are improved, and reliable result assessment is provided for doctors.
Example 5:
based on embodiment 4, the multi-fragment interpretation module includes:
the single-segment feature extraction unit is used for sequentially extracting features of single-segment data in the multi-segment data based on a one-dimensional residual error network;
the characteristic fusion unit is used for sequentially inputting the fetal heart characteristics of each single segment into the encoder structure and obtaining a plurality of correspondingly fused new segment characteristics through the self-attention layer, the first normalization layer, the feedforward neural network layer and the second normalization layer;
the overall feature extraction unit is used for splicing the new segment features and extracting the overall feature through a multilayer perception mechanism;
the classification unit is used for classifying the overall characteristics by using the sequence regression model to obtain a multi-segment interpretation result;
wherein short links are added at the self-attention layer and the feed-forward neural network layer.
In this embodiment, for the features extracted by the single-segment one-dimensional residual error network, the features are input into an encoder structure of a transform, and the features of a plurality of input segments are fused by a self-attention mechanism, so as to obtain more accurate long-time scale information. The structure is composed of a plurality of substructures with consistent structures which are combined step by step. The specific form of each substructure is shown in fig. 4, the input of the substructure is different segment features output by the previous layer, and the fused segment features of the output of the structure are obtained after passing through the self-attention layer, the normalization layer, the feedforward neural network layer and the other normalization layer respectively. A short link is added in the middle to ensure the convergence of the model. After the plurality of segments are fully fused, the features of the segments are spliced together, and the overall features are extracted through a multilayer perceptron. And then, classifying the features by using a sequence regression model to obtain a multi-fragment interpretation result.
The beneficial effects of the above technical scheme are: by extracting the features of the data of multiple segments, the judgment accuracy is convenient to increase, missed diagnosis and misdiagnosis are avoided, the reliability of obtaining the features can be improved through different layers, and finally, the accurate interpretation result is convenient to obtain through feature classification.
Example 6:
based on embodiment 4, the interpretation accuracy prediction module includes:
the feature prediction unit is used for acquiring the multi-segment data and the blank segment data, fusing the multi-segment data and the plurality of new segment features acquired from the multi-segment interpretation module through an attention mechanism, and predicting the next segment feature;
the accuracy rate determining unit is used for determining the accuracy rate of the predicted next segment characteristics based on the difference between the predicted next segment characteristics and the segment characteristics of the actual monitored data;
and taking the accuracy as an accuracy evaluation result of the fetal heart judgment result.
In this embodiment, the interpretation accuracy prediction module predicts the interpretation result of the whole segment module and the accuracy of the interpretation result based on the fetal heart monitoring characteristics after the attention mechanism fusion. In this section, the sequential regression model is still used in the classification to avoid the influence of sample imbalance. Since it is difficult for doctors to quantify the accuracy of their evaluations, it is necessary to use an auto-supervision approach to predict the accuracy, i.e. to automatically generate learnable labels through machine learning models. The tag is constructed in an autoregressive manner. Specifically, an autoregressive model is first trained to predict the tags of the next segment of the fetal heart rate sequence. The model accuracy is measured through the difference between the label and the doctor interpretation result of the actual monitoring data. By this method, the labels required for accurate rate prediction are obtained.
In this embodiment, the input and output information of the module during operation is shown in fig. 5.
In this embodiment, in this module, the input is the known segment feature + the blank mask feature. Besides the decoder structure, the known segment features also pass through the encoder part of the multi-segment interpretation module, and the features extracted by the encoder and the attention mechanism are utilized to provide more accurate prediction results. The part outputs two parts, namely a classification result corresponding to the mask, and the accuracy of the classification result corresponding to the mask. In the training process, the real next segment classification result is used as the label of the first output item. The accuracy of the classification result corresponding to the mask is obtained by performing a linear transformation + softmax operation after the feature corresponding to the classification result. Because the sequential regression model is used for classification, the probability of neural network prediction cannot be directly provided, and in the training process, whether the classification result is a real result or not is used as a training label to obtain the label of the prediction probability.
The beneficial effects of the above technical scheme are: by predicting the next segment, the prediction accuracy is convenient to determine, the fetal heart can be judged more accurately in the follow-up process, more effective data can be obtained, and reliable result evaluation is provided for doctors.
Embodiment 7, on the basis of basic embodiment 1, further includes:
the result judgment module is used for determining whether the judgment accuracy prediction module needs to be adjusted or not based on the accuracy evaluation result;
when the accuracy evaluation result is determined to be larger than the corresponding preset result time, judging that the judgment accuracy prediction module is not required to be adjusted;
otherwise, judging that the judgment accuracy prediction module needs to be adjusted;
the accuracy adjusting module is used for acquiring a first feature difference set of the predicted next segment feature and the corresponding first actual monitoring segment feature when the judgment accuracy rate predicting module is judged to need to adjust;
reversely predicting the previous segment feature based on the next segment feature, and comparing the previous segment feature with the corresponding second actual monitored segment feature to obtain a second feature difference set;
constructing a first difference map based on the first set of feature differences, and simultaneously constructing a second difference map based on the second set of feature differences;
performing first framing on the difference range of the difference part in the first difference map, and simultaneously performing second framing on the difference range of the difference part in the second difference map;
judging whether the first framing result and the second framing result meet the rule consistency, if so, judging that the prediction abnormity is only related to a judgment accuracy prediction module;
determining an index to be adjusted according to the rule consistency, and optimizing the judgment accuracy prediction module according to the index to be adjusted;
otherwise, judging whether the prediction abnormity is related to the judgment accuracy prediction module and the index to be mined;
acquiring frame selection consistent areas and frame selection inconsistent areas of the first frame selection result and the second frame selection result, performing first calibration on the frame selection consistent areas, performing second calibration on the frame selection inconsistent areas matched with the first frame selection result, and performing third calibration on the frame selection inconsistent areas matched with the second frame selection result;
constructing a calibration map according to the first calibration result, the second calibration result and the third calibration structure;
determining a first calibration area and a first calibration position related to a second calibration result in the calibration map to obtain a corresponding first reference weight;
wherein Y1 represents a first reference weight; s1 represents the area value corresponding to the first calibration result; s2 represents a first calibration area corresponding to the second calibration result; s3 represents a second calibration area corresponding to the third calibration result; r represents a total position set of a corresponding calibration map; r2 represents a position set corresponding to the second calibration result;
represents the location weight assigned for R2;
indicating the position weights assigned to the position set R2 corresponding to the second calibration result and the total position set R;
representing the first nominal area versus the total areas1+s2+s3 assigned area weight;
determining a second calibration area and a second calibration position related to a third calibration result in the calibration map to obtain a corresponding second reference weight;
wherein Y2 represents a second reference weight; r3 represents a position set corresponding to the third calibration result;
based on the first reference weight and a second reference weight;
acquiring first transition factors of the second calibration result and the first calibration result, and acquiring second transition factors of the third calibration result and the first calibration result;
matching to obtain an index to be mined from the weight-factor-index database based on the reference weight and the transition factor;
determining an index to be optimized according to the second calibration result and the third calibration result;
and processing the judgment accuracy prediction module based on the index to be optimized and the index to be mined.
In this embodiment, the model is adjusted mainly to ensure the accuracy of the subsequent prediction.
Such as: the time 1 corresponds to the second actual monitoring segment characteristic, the time 2 corresponds to the first actual monitoring segment characteristic, the predicted characteristic corresponding to the time 1 is the previous segment characteristic, and the predicted characteristic corresponding to the time 2 is the next segment characteristic.
In this embodiment, the first feature difference set and the second feature difference set refer to differences between features, such as differences of a certain part of a curve, and the like, that are determined according to the block selection result.
In this embodiment, the term "regular consistency" refers to whether corresponding prediction differences are consistent in the process of performing forward prediction and backward prediction based on the same module, and if so, the regular consistency is considered to be satisfied, that is, the difference at this time satisfies a certain rule, and is considered to be only related to the module.
In this embodiment, the inconsistent area refers to inconsistency based on the first frame selection result, inconsistency based on the second frame selection result, and the like.
In the embodiment, the calibration map can be effectively obtained by adopting different calibrations on different areas.
In this embodiment, the result of both the first reference weight and the second reference weight is less than 1.
In this embodiment, the transition factors are the transition from consistent to inconsistent and from inconsistent to consistent, the transition factors resulting from this process, such as noise carried by the module itself.
In this embodiment, the weight-factor-index database includes different weight combinations, transition factors, and corresponding indexes to be mined, which facilitates optimization of the module.
In this embodiment, the index to be optimized may be optimized for a certain numerical value, so as to achieve optimization.
In this embodiment, the index to be mined is caused by, for example, excessive noise of a signal acquired by the ultrasound monitoring probe during the process of acquiring the monitored data.
The beneficial effects of the above technical scheme are: through forward prediction and backward prediction and comparison, the problems of the module can be effectively determined, the module is optimized and adjusted according to different modes, the prediction precision of the module is ensured, a calibration graph is established by determining a consistent area and an inconsistent area according to a frame selection result, the reference weight is convenient to calculate subsequently, and through obtaining a transition factor, the number of indexes to be excavated is convenient to obtain, the module is optimized, and the accuracy of fetal heart monitoring prediction is ensured.
Example 8:
based on embodiment 1, the data normalization module further includes:
the data analysis unit is used for performing normalization processing on the first monitoring data by using the mean value and the variance, performing pre-analysis on the normalization processing data and judging a zero point existing in the normalization processing data;
the curve drawing unit is used for taking the normalized data formed by the adjacent zero points as a data segment and drawing a data curve of the data segment;
a first determination unit configured to determine a valley point and a peak point existing in the data curve, and perform a first determination on a length and a slope of a connection line formed by each of the valley point and the peak point;
a first obtaining unit, configured to obtain, according to the first determination result and based on a timestamp sequence, a first line sequence from a valley point to a peak point and a second line sequence from the peak point to the valley point;
the screening unit is used for carrying out regular consistency detection on the first line sequence and the second line sequence and screening to obtain a first inconsistent sequence and a second inconsistent sequence;
the second determining unit is used for determining a first position of the first inconsistent sequence and a second position of the second inconsistent sequence, and determining the probability to be adjusted according to the first position and the second position;
l1 represents the number of the first inconsistent sequences; l2 indicates the number of second inconsistent sequences;
indicating the position adjustable weights assigned for all first positions;
indicating the position adjustable weights assigned for all second positions;h k1 a sequence value representing the k1 th first disparity sequence;h k1,0 a standard value representing the k1 th first discordance sequence;h k2 a sequence value representing the k2 th second inconsistent sequence;h k2,0 a standard value indicating the k2 th second discordance sequence;
the inserting unit is used for deleting the valley points and the peak points of the first line sequence and the second line sequence when the probability to be adjusted is larger than the preset adjusting probability, determining a first inserting position for inserting the valley points and a second inserting position for inserting the peak points in the corresponding data segment, and simultaneously determining a first inserting position for inserting the valley points and a second inserting position for inserting the peak points to obtain a qualified sequence so as to obtain corresponding second monitoring data;
when the probability to be adjusted is not greater than the preset adjustment probability, based on
Determining a corresponding primary and secondary adjustment sequence;
a supplement unit, configured to delete the valley point and the peak point corresponding to the main adjustment sequence, and sequentially supplement the deleted sequences with the received data of the receiving probe corresponding to the monitored data with the largest signal-to-noise ratio, and establish a mapping relationship between the supplement sequences and the optimal position;
meanwhile, acquiring valley points and peak points corresponding to the secondary adjustment sequence from fetal heart monitoring data acquired from the optimal fetal heart monitoring position to perform numerical value expansion or numerical value reduction processing;
and the second acquisition unit is used for acquiring a qualified sequence based on the mapping relation and the data processing result so as to acquire corresponding second monitoring data.
In this embodiment, one data segment is acquired based on two zeros, and data analysis can be performed effectively.
In this embodiment, the data curve is analyzed to determine the valley point, peak point, line length and slope of the sequence, so that the corresponding sequence can be effectively determined.
In this embodiment, by performing a regular consistency check, a first inconsistent sequence can be screened from a first line sequence, and a second inconsistent sequence can be screened from a second line sequence.
In this embodiment, the preset adjustment probability is, for example, 0.5, and the probability to be adjusted obtained by corresponding calculation is smaller than 1.
In this embodiment, the first plugging number is associated with the number of the deleted valley points, the second plugging number is associated with the number of the deleted peak points, and the first plugging position and the second plugging position may be determined according to fetal heart monitoring data obtained from the optimal fetal heart monitoring position.
In this embodiment, effective recording is facilitated by obtaining the mapping relationship, and in this embodiment, the numerical value expansion or numerical value reduction processing may be adjustment of the length of a line, the slope, and the like corresponding to the sequence.
The beneficial effects of the above technical scheme are: the data curve is drawn by acquiring the zero point, the sequence is constructed by determining the length, the slope and the like, different sequences are conveniently screened by detecting the regularity consistency, and then the adjustment probability is calculated, different modes can be adopted, the sequence is adjusted, the qualification of the sequence is ensured, the accuracy of second monitoring data is ensured, and an effective data basis is provided for the subsequent analysis of fetal heart characteristics.
It will be apparent to those skilled in the art that various changes and modifications may be made in the present invention without departing from the spirit and scope of the invention. Thus, if such modifications and variations of the present invention fall within the scope of the claims of the present invention and their equivalents, the present invention is also intended to include such modifications and variations.
Claims (5)
1. A prenatal electronic fetal heart monitoring automatic identification system based on deep learning is characterized by comprising:
the data preprocessing module is used for preprocessing the data of the electronic fetal heart monitoring data of the monitored target patient;
the single-segment fetal heart monitoring evaluation module is used for acquiring single-segment data corresponding to the preprocessed data, obtaining current scores of the single-segment data based on a backbone network and a sequence regression model, and judging whether a target patient needs to continue monitoring;
the multi-segment fetal heart monitoring comprehensive evaluation module is used for acquiring corresponding multi-segment data when monitoring is required to be continued, performing comprehensive diagnosis, acquiring a fetal heart interpretation result, and performing accuracy evaluation on the fetal heart interpretation result to acquire a monitoring suggestion;
wherein, the comprehensive evaluation module of the multi-segment fetal heart monitoring comprises:
the multi-segment interpretation module is used for interpreting the multi-segment characteristics corresponding to the multi-segment data to obtain a multi-segment interpretation result;
the interpretation accuracy rate prediction module is used for training an autoregressive model and predicting the accuracy of the fetal heart interpretation result;
wherein the multi-segment interpretation module comprises:
the single-segment feature extraction unit is used for sequentially extracting features of single-segment data in the multi-segment data based on a one-dimensional residual error network;
the characteristic fusion unit is used for sequentially inputting the fetal heart characteristics of each single segment into the encoder structure and obtaining a plurality of correspondingly fused new segment characteristics through the self-attention layer, the first normalization layer, the feedforward neural network layer and the second normalization layer;
the overall feature extraction unit is used for splicing the new segment features and extracting the overall feature through a multilayer perception mechanism;
the classification unit is used for classifying the overall characteristics by using the sequence regression model to obtain a multi-segment interpretation result;
wherein short links are added at the self-attention layer and the feedforward neural network layer;
wherein the interpretation accuracy prediction module comprises:
the feature prediction unit is used for acquiring the multi-segment data and the blank segment data, fusing the multi-segment data and the plurality of new segment features acquired from the multi-segment interpretation module through an attention mechanism, and predicting the next segment feature;
the accuracy rate determining unit is used for determining the accuracy rate of the predicted next segment characteristics based on the difference between the predicted next segment characteristics and the segment characteristics of the actual monitored data;
and taking the accuracy as an accuracy evaluation result of the fetal heart interpretation result.
2. The deep learning based prenatal electronic fetal monitoring automatic identification system as claimed in claim 1, wherein the data preprocessing module comprises:
the data cleaning module is used for acquiring the electronic fetal heart monitoring data, eliminating information missing segments, and acquiring monitoring data corresponding to the information missing segments in the same time again to obtain first monitoring data when the information missing segments meet missing acquisition standards;
the data normalization module is used for counting the mean value and the variance of the fetal heart rate corresponding to the first monitoring data and carrying out normalization processing on the first monitoring data by utilizing the mean value and the variance to obtain second monitoring data;
and the second monitoring data is preprocessed data.
3. The deep learning-based prenatal electronic fetal heart monitoring automatic identification system as claimed in claim 1, wherein the single-segment fetal heart monitoring evaluation module comprises:
the backbone network is used for extracting fetal heart monitoring characteristics in the preprocessed data from bottom to top by adopting a one-dimensional residual convolution network as a basic frame;
the sequence regression model is used for carrying out linear transformation and normalization processing on the fetal heart monitoring characteristics to obtain normalization parameters;
on the basis of the normalization parameters, comparing the normalization parameters with a preset number of learnable reference feature vectors respectively to obtain a preset number of comparison scores, wherein each comparison score corresponds to one classification grade;
optimizing two classification loss functions of a preset number of comparison scores, wherein the supervision information of the ith classifier is determined by whether the result labeled by a doctor is greater than or equal to the grade corresponding to the corresponding reference feature, and the weight corresponding to each two classification loss function is determined by the ratio of the number of samples in the data set, which are greater than the classification grade, to the number of samples, which are smaller than the classification grade;
accumulating a preset number of comparison scores, and judging that the corresponding fetal heart monitoring segment is classified into the ith segment when the accumulated score is larger than the adjustable threshold corresponding to the ith class and smaller than the adjustable threshold corresponding to the (i + 1) th class, so as to obtain a current score;
wherein the fetal heart monitoring segment corresponds to the single segment data.
4. The deep learning based prenatal electronic fetal monitoring automatic identification system as claimed in claim 1, further comprising:
the result judgment module is used for determining whether the interpretation accuracy prediction module needs to be adjusted or not based on the accuracy evaluation result;
when the accuracy evaluation result is determined to be larger than the corresponding preset result time, judging that the interpretation accuracy prediction module is not required to be adjusted;
otherwise, judging that the interpretation accuracy prediction module needs to be adjusted;
the accuracy adjusting module is used for acquiring a first feature difference set of the predicted next segment feature and the corresponding first actual monitoring segment feature when the interpretation accuracy predicting module is judged to need to adjust;
reversely predicting a previous segment feature based on a next segment feature, and comparing the previous segment feature with a corresponding second actual monitoring segment feature to obtain a second feature difference set;
constructing a first difference map based on the first set of feature differences, and simultaneously constructing a second difference map based on the second set of feature differences;
performing first framing on the difference range of the difference part in the first difference map, and simultaneously performing second framing on the difference range of the difference part in the second difference map;
judging whether the first frame selection result and the second frame selection result meet the rule consistency, if so, judging that the prediction abnormity is only related to the interpretation accuracy prediction module;
determining an index to be adjusted according to the rule consistency, and optimizing the interpretation accuracy prediction module according to the index to be adjusted;
otherwise, judging whether the prediction abnormity is related to the interpretation accuracy rate prediction module and the index to be mined;
acquiring frame selection consistent areas and frame selection inconsistent areas of the first frame selection result and the second frame selection result, performing first calibration on the frame selection consistent areas, performing second calibration on the frame selection inconsistent areas matched with the first frame selection result, and performing third calibration on the frame selection inconsistent areas matched with the second frame selection result;
constructing a calibration map according to the first calibration result, the second calibration result and the third calibration structure;
determining a first calibration area and a first calibration position related to a second calibration result in the calibration map to obtain a corresponding first reference weight;
wherein Y1 represents a first reference weight; s1 represents the area value corresponding to the first calibration result; s2 represents the first calibration area corresponding to the second calibration result; s3 represents a second calibration area corresponding to the third calibration result; r represents a total position set of a corresponding calibration map; r2 represents a position set corresponding to the second calibration result;
represents the location weight assigned for R2;
indicating the position weight assigned to the position set R2 corresponding to the second calibration result and the total position set R;
representing the first nominal area versus the total areas1+s2+s3 assigned area weight;
determining a second calibration area and a second calibration position related to a third calibration result in the calibration map to obtain a corresponding second reference weight;
wherein Y2 represents a second reference weight; r3 represents a position set corresponding to the third calibration result;
based on the first reference weight and a second reference weight;
acquiring first transition factors of the second calibration result and the first calibration result, and acquiring second transition factors of the third calibration result and the first calibration result;
matching to obtain an index to be mined from the weight-factor-index database based on the reference weight and the transition factor;
determining an index to be optimized according to the second calibration result and the third calibration result;
and processing the interpretation accuracy prediction module based on the index to be optimized and the index to be mined.
5. The deep learning-based prenatal electronic fetal monitoring automatic identification system as claimed in claim 2, wherein the data normalization module further comprises:
the data analysis unit is used for performing normalization processing on the first monitoring data by using the mean value and the variance, performing pre-analysis on the normalization processing data and judging a zero point existing in the normalization processing data;
the curve drawing unit is used for taking the normalized data formed by the adjacent zero points as a data segment and drawing a data curve of the data segment;
a first determination unit configured to determine a valley point and a peak point existing in the data curve, and perform a first determination on a length and a slope of a connection line formed by each of the valley point and the peak point;
a first obtaining unit, configured to obtain, according to the first determination result and based on a timestamp sequence, a first line sequence from a valley point to a peak point and a second line sequence from the peak point to the valley point;
the screening unit is used for carrying out regular consistency detection on the first line sequence and the second line sequence and screening to obtain a first inconsistent sequence and a second inconsistent sequence;
a second determining unit, configured to determine a first position of the first inconsistent sequence and a second position of the second inconsistent sequence, and determine a probability to be adjusted according to the first position and the second position;
l1 represents the number of the first inconsistent sequences; l2 indicates the number of second inconsistent sequences;
indicating the position adjustable weights assigned for all first positions;
indicating the position adjustable weights assigned for all second positions;h k1 a sequence value representing the k1 th first inconsistent sequence;h k1,0 standard values representing the k1 th first inconsistent sequence;h k2 a sequence value representing the k2 th second inconsistent sequence;h k2,0 a standard value indicating the k2 th second discordance sequence;
the inserting unit is used for deleting the valley points and the peak points of the first line sequence and the second line sequence when the probability to be adjusted is larger than the preset adjusting probability, determining a first inserting position for inserting the valley points and a second inserting position for inserting the peak points in the corresponding data segment, and simultaneously determining a first inserting position for inserting the valley points and a second inserting position for inserting the peak points to obtain a qualified sequence so as to obtain corresponding second monitoring data;
when the probability to be adjusted is not greater than the preset adjustment probability, based on
Determining a corresponding primary and secondary adjustment sequence;
a supplement unit, configured to delete the valley point and the peak point corresponding to the main adjustment sequence, and sequentially supplement the deleted sequences with the received data of the receiving probe corresponding to the monitored data with the largest signal-to-noise ratio, and establish a mapping relationship between the supplement sequences and the optimal position;
meanwhile, acquiring valley points and peak points corresponding to the secondary adjustment sequence from fetal heart monitoring data acquired from the optimal fetal heart monitoring position to perform numerical value expansion or numerical value reduction processing;
and the second acquisition unit is used for acquiring a qualified sequence based on the mapping relation and the data processing result so as to acquire corresponding second monitoring data.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202210649767.1A CN114724720B (en) | 2022-06-10 | 2022-06-10 | Prenatal electronic fetal heart monitoring automatic identification system based on deep learning |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202210649767.1A CN114724720B (en) | 2022-06-10 | 2022-06-10 | Prenatal electronic fetal heart monitoring automatic identification system based on deep learning |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN114724720A CN114724720A (en) | 2022-07-08 |
| CN114724720B true CN114724720B (en) | 2022-09-09 |
Family
ID=82233030
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202210649767.1A Active CN114724720B (en) | 2022-06-10 | 2022-06-10 | Prenatal electronic fetal heart monitoring automatic identification system based on deep learning |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN114724720B (en) |
Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN102319063A (en) * | 2011-07-26 | 2012-01-18 | 深圳市理邦精密仪器股份有限公司 | Device and method for improving acceleration identification accuracy of fetal heart rate data |
| CN108261176A (en) * | 2017-12-28 | 2018-07-10 | 深圳京柏医疗科技股份有限公司 | Fetal heart monitoring data processing method, device, system, storage medium and computer equipment |
Family Cites Families (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109567868A (en) * | 2018-10-15 | 2019-04-05 | 广东宝莱特医用科技股份有限公司 | A kind of CTG fetal rhythm methods of marking and system |
| CN111696670B (en) * | 2020-06-16 | 2023-06-20 | 广州三瑞医疗器械有限公司 | Intelligent interpretation method for prenatal fetal monitoring based on deep forest |
| CN112971853B (en) * | 2021-01-21 | 2022-08-02 | 中山大学附属第一医院 | Method and device for analyzing electronic fetal heart monitoring data based on artificial intelligence |
| CN112971799A (en) * | 2021-02-04 | 2021-06-18 | 北京理工大学 | Non-stimulation fetal heart monitoring classification method based on machine learning |
| CN113456064B (en) * | 2021-07-15 | 2024-04-16 | 广州三瑞医疗器械有限公司 | Intelligent interpretation method for prenatal fetal heart monitoring signals |
-
2022
- 2022-06-10 CN CN202210649767.1A patent/CN114724720B/en active Active
Patent Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN102319063A (en) * | 2011-07-26 | 2012-01-18 | 深圳市理邦精密仪器股份有限公司 | Device and method for improving acceleration identification accuracy of fetal heart rate data |
| CN108261176A (en) * | 2017-12-28 | 2018-07-10 | 深圳京柏医疗科技股份有限公司 | Fetal heart monitoring data processing method, device, system, storage medium and computer equipment |
Also Published As
| Publication number | Publication date |
|---|---|
| CN114724720A (en) | 2022-07-08 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN108464840B (en) | Automatic detection method and system for breast lumps | |
| TWI684997B (en) | Establishing method of bone age assessment and height prediction model, bone age assessment and height prediction system, and bone age assessment and height prediction method | |
| CN109117890B (en) | Image classification method and device and storage medium | |
| CN111862020B (en) | Method and device for predicting physiological age of anterior ocular segment, server and storage medium | |
| CN111557672A (en) | Nicotinic acid skin reaction image analysis method and equipment | |
| CN116864104A (en) | Chronic thromboembolic pulmonary artery high-pressure risk classification system based on artificial intelligence | |
| CN113903082A (en) | Human body gait monitoring algorithm based on dynamic time planning | |
| CN111598868B (en) | Lung ultrasonic image identification method and system | |
| CN112256754A (en) | Ultrasonic detection analysis system and method based on standard model | |
| CN116740426A (en) | Classification prediction system for functional magnetic resonance images | |
| KR101603308B1 (en) | Biological age calculation model generation method and system thereof, biological age calculation method and system thereof | |
| CN114864086A (en) | Disease prediction method based on lung function report template | |
| CN117322865B (en) | Temporal-mandibular joint disc shift MRI (magnetic resonance imaging) examination and diagnosis system based on deep learning | |
| CN114724720B (en) | Prenatal electronic fetal heart monitoring automatic identification system based on deep learning | |
| US6941288B2 (en) | Online learning method in a decision system | |
| CN117038091A (en) | Rapid diagnosis system and method for diabetic nephropathy through predictive model | |
| KR20210054140A (en) | Medical image diagnosis assistance apparatus and method using a plurality of medical image diagnosis algorithm for endoscope images | |
| CN116738352B (en) | Method and device for classifying abnormal rod cells of retinal vascular occlusion disease | |
| TW202219977A (en) | Method for assessing cardiothoracic ratio and cardiothoracic ratio assessment system | |
| Gahiwad et al. | Brain Stroke Detection Using CNN Algorithm | |
| CN115393353B (en) | Medical image data quality monitoring and management method and system | |
| CN117954116B (en) | Breathing severe patient state monitoring and early warning method based on artificial intelligence | |
| KR20230125509A (en) | Apparatus for determining meniscus tear using MRI images and method thereof | |
| TWI765420B (en) | Assembly of heart failure prediction | |
| KR20230132138A (en) | Method and system for providing health information based on bio-signal measured by non-contact manner |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |