[go: up one dir, main page]
More Web Proxy on the site http://driver.im/

CN116350231A - Semi-supervised electrocardiogram myocardial infarction positioning method based on hierarchical different granularity class constraint - Google Patents

Semi-supervised electrocardiogram myocardial infarction positioning method based on hierarchical different granularity class constraint Download PDF

Info

Publication number
CN116350231A
CN116350231A CN202310234705.9A CN202310234705A CN116350231A CN 116350231 A CN116350231 A CN 116350231A CN 202310234705 A CN202310234705 A CN 202310234705A CN 116350231 A CN116350231 A CN 116350231A
Authority
CN
China
Prior art keywords
category
class
level
electrocardiosignals
network
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.)
Pending
Application number
CN202310234705.9A
Other languages
Chinese (zh)
Inventor
刘华锋
李楚晨
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Zhejiang University ZJU
Original Assignee
Zhejiang University ZJU
Priority date (The priority date 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 date listed.)
Filing date
Publication date
Application filed by Zhejiang University ZJU filed Critical Zhejiang University ZJU
Priority to CN202310234705.9A priority Critical patent/CN116350231A/en
Publication of CN116350231A publication Critical patent/CN116350231A/en
Pending legal-status Critical Current

Links

Images

Classifications

    • A—HUMAN NECESSITIES
    • A61—MEDICAL OR VETERINARY SCIENCE; HYGIENE
    • A61B—DIAGNOSIS; SURGERY; IDENTIFICATION
    • A61B5/00—Measuring for diagnostic purposes; Identification of persons
    • A61B5/24—Detecting, measuring or recording bioelectric or biomagnetic signals of the body or parts thereof
    • A61B5/316—Modalities, i.e. specific diagnostic methods
    • A61B5/318—Heart-related electrical modalities, e.g. electrocardiography [ECG]
    • A—HUMAN NECESSITIES
    • A61—MEDICAL OR VETERINARY SCIENCE; HYGIENE
    • A61B—DIAGNOSIS; SURGERY; IDENTIFICATION
    • A61B5/00—Measuring for diagnostic purposes; Identification of persons
    • A61B5/24—Detecting, measuring or recording bioelectric or biomagnetic signals of the body or parts thereof
    • A61B5/316—Modalities, i.e. specific diagnostic methods
    • A61B5/318—Heart-related electrical modalities, e.g. electrocardiography [ECG]
    • A61B5/346—Analysis of electrocardiograms
    • A—HUMAN NECESSITIES
    • A61—MEDICAL OR VETERINARY SCIENCE; HYGIENE
    • A61B—DIAGNOSIS; SURGERY; IDENTIFICATION
    • A61B5/00—Measuring for diagnostic purposes; Identification of persons
    • A61B5/72—Signal processing specially adapted for physiological signals or for diagnostic purposes
    • A61B5/7235—Details of waveform analysis
    • A61B5/7264—Classification of physiological signals or data, e.g. using neural networks, statistical classifiers, expert systems or fuzzy systems
    • A61B5/7267—Classification of physiological signals or data, e.g. using neural networks, statistical classifiers, expert systems or fuzzy systems involving training the classification device
    • A—HUMAN NECESSITIES
    • A61—MEDICAL OR VETERINARY SCIENCE; HYGIENE
    • A61B—DIAGNOSIS; SURGERY; IDENTIFICATION
    • A61B5/00—Measuring for diagnostic purposes; Identification of persons
    • A61B5/72—Signal processing specially adapted for physiological signals or for diagnostic purposes
    • A61B5/7271—Specific aspects of physiological measurement analysis
    • A61B5/7275—Determining trends in physiological measurement data; Predicting development of a medical condition based on physiological measurements, e.g. determining a risk factor
    • 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
    • Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
    • Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
    • Y02A—TECHNOLOGIES FOR ADAPTATION TO CLIMATE CHANGE
    • Y02A90/00—Technologies having an indirect contribution to adaptation to climate change
    • Y02A90/10—Information and communication technologies [ICT] supporting adaptation to climate change, e.g. for weather forecasting or climate simulation

Landscapes

  • Health & Medical Sciences (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Engineering & Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Molecular Biology (AREA)
  • Artificial Intelligence (AREA)
  • Biophysics (AREA)
  • General Health & Medical Sciences (AREA)
  • Biomedical Technology (AREA)
  • Heart & Thoracic Surgery (AREA)
  • Medical Informatics (AREA)
  • Surgery (AREA)
  • Animal Behavior & Ethology (AREA)
  • Pathology (AREA)
  • Public Health (AREA)
  • Veterinary Medicine (AREA)
  • Cardiology (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Theoretical Computer Science (AREA)
  • Mathematical Physics (AREA)
  • Evolutionary Computation (AREA)
  • Physiology (AREA)
  • Psychiatry (AREA)
  • Signal Processing (AREA)
  • Computational Linguistics (AREA)
  • Data Mining & Analysis (AREA)
  • Computing Systems (AREA)
  • General Engineering & Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Software Systems (AREA)
  • Fuzzy Systems (AREA)
  • Measurement And Recording Of Electrical Phenomena And Electrical Characteristics Of The Living Body (AREA)

Abstract

The invention discloses a semi-supervised electrocardiogram myocardial infarction positioning method based on category constraints of different levels, which comprises the steps of constructing a self-learning deep learning neural network, performing positioning prediction on myocardial infarction of a 12-lead electrocardiogram with only a small number of labels, outputting three levels of prediction results by the network, and converting fine granularity, confusion granularity and coarse granularity by using level category conversion on images without labels; because of the non-uniformity of the data category, the invention uses the category-aware dynamic loss constraint based on cross entropy and the hierarchical constraint to pay less attention to learned categories and to take more attention to non-learned categories. The method can obtain a better identification effect under the condition of a small number of marks to a certain extent, thereby helping the analysis and diagnosis of experimental diseases of the intelligent medical system.

Description

Semi-supervised electrocardiogram myocardial infarction positioning method based on hierarchical different granularity class constraint
Technical Field
The invention belongs to the technical field of biomedical signal analysis and classification, and particularly relates to a semi-supervised electrocardiogram myocardial infarction positioning method based on hierarchical different granularity class constraint.
Background
Myocardial infarction is one of five killers threatening human health, and has the characteristics of quick onset, high mortality rate and the like. Electrocardiogram is used as a preliminary diagnostic evidence for diagnosing myocardial infarction which is most widely applied clinically, and plays an important role, and because of the complexity of heart blood vessels, blood vessel blockage at different positions can cause infarction at different heart positions, so that how to accurately locate the infarcted position by the electrocardiogram becomes an important point of research. Because the localization of myocardial infarction requires a great deal of clinical experience, tagging an electrocardiogram becomes a judgment that not only requires a great deal of time and cost, but also has an experienced clinician to make the localization.
It is known that deep learning networks require fine and accurate labeling as supervisory information to allow efficient convergence of the model, and the larger the labeling amount, the more likely the model will have this better training effect, so a more important problem is reflected in the eye curtains of researchers, i.e., if less labeling information is used while achieving an effect similar to full labeling. Along with the gradual development of deep learning, researchers are gradually researching fine granularity characteristics, namely distinguishing finer categories from coarser granularity categories to approach, so that diagnosis and treatment of diseases are perfected, and positioning classification of myocardial infarction is a common problem.
Since the heart is typically supplied with blood from a plurality of blood vessels belonging to different layers, blockage of these blood vessels can cause ischemia or necrosis of the different lumen walls that are affected by their blood supply. Generally, the myocardial infarction is classified into a fine classification based on the anterior wall, the outer wall, the posterior wall, and the like, depending on the heart position to which the blood vessels belong (ALMI, AMI, ASMI, ILMI, IMI, IPLMI, IPMI, PMI, LMI). However, since the specific information of the myocardial infarction often exists only in the subtle distinction between the oscillograms (Q wave, ST band), this classification task based on fine granularity has a certain difficulty for the model. In previous work, the literature [ Du N, cao Q, yu L, et al FM-ECG: afine-graded Multi-label framework for ECG image classification [ J ]. Information Sciences,2020,549] used image-based electrocardiographic input to explore fine-grained classes, the model was entirely using a transformer structure, and features were extracted by fusion of spatial attention mechanisms. The literature [ Wang R, fan J, li y. Deep multi-scale fusion neural network for multi-class arrhythmia detection [ J ]. IEEE journal of biomedical and health informatics,2020,24 (9): 2461-2472] uses two parallel network structures with different convolution kernels to convolve shallow features extracted from the backbone network, and then outputs the final result through a two feature concatenation and attention mechanism. Although these methods achieve good results, the model only extracts fine-grained features from the structural design aspect, and does not consider the hierarchical relationship existing in the tag itself, so that the information of mutual constraint between different hierarchies is ignored. Meanwhile, as the characteristics of the fine-granularity categories have the problem of high similarity, how to distinguish the categories which are easy to be confused with each other and have high similarity becomes one of the key points of academic research.
Fine-grained recognition has been an active area for many years, by powerful feature coding, locating distinct semantic parts, or using external information, but fine-grained annotation is often time consuming and requires experienced specialists because some subclasses of objects have marginal visual differences, and even are difficult for humans to recognize. One way to reduce the cost of fine granularity marking is to use weak labels, such as by image level labels to locate fine granularity portions; semi-supervised learning is another approach, typically with a small set of labeled data and many unlabeled images for efficient model training, and merging unlabeled data by employing unsupervised consistent regularization is a popular approach in current research.
In the semi-supervised algorithm, in detail, we have a large amount of unlabeled data, while only a small portion of the labeled data. To better utilize existing data and give the model enough information, we need to use unlabeled data as auxiliary information to improve the overall performance of the model. However, there are only a few attempts to introduce semi-supervised learning into fine granularity image analysis, analyzing the cause of which, there are mainly two challenges: (1) how should we combine detail information of different granularity in order to better exploit unlabeled data? (2) How should we deal with the confusing categories? Moreover, the scarcity of labeling in semi-supervised learning also exacerbates the above-mentioned dilemma.
Disclosure of Invention
In view of the above, the invention provides a semi-supervised electrocardiogram myocardial infarction positioning method based on hierarchical different granularity category constraints, which outputs three prediction results with different granularities through a deep learning neural network, and improves the recognition performance of a model through the combination of hierarchical fine granularity category constraints and category perception dynamic loss functions.
A semi-supervised electrocardiogram myocardial infarction positioning method based on hierarchical different granularity class constraints comprises the following steps:
(1) Collecting electrocardiosignals which come from different patients and have myocardial infarction category labels, and dividing the electrocardiosignals into training sets and testing sets according to the same signal length;
(2) Selecting a deep learning neural network to predict myocardial infarction categories, wherein prediction output is divided into three levels, namely a fine-granularity level, a confusion level and a coarse-granularity level;
(3) Inputting electrocardiosignals of a training set into the deep learning neural network for training according to batches, wherein one part of the electrocardiosignals in each batch adopts category marking information, and the other part does not adopt category marking information; further designing a loss function based on cross entropy loss, category perception dynamic loss and cross category self-constraint for network iterative updating;
(4) And inputting the electrocardiosignals of the test set into the trained neural network, and predicting the corresponding myocardial infarction type result.
Further, the deep learning neural network adopts a ResNet structure-based one-dimensional feature extraction network comprising three feature output heads.
Further, provided thatThe prediction results output by the deep learning neural network are P respectively l3 ,P l2 ,P l1 Wherein P is l3 Class probability prediction results corresponding to fine granularity levels are C multiplied by D; p (P) l2 Class probability prediction results corresponding to the confusion hierarchy are B multiplied by D; p (P) l1 The size of the class probability prediction result corresponding to the coarse granularity level is 2 xD, D is the number of input signals of each batch, C and B are the number of classes of the fine granularity level and the confusion level respectively, and C > B > 2.
Further, the specific process of training the deep learning neural network in the step (3) is as follows:
initializing model parameters including bias vectors and weight matrixes of each layer, learning rate and an optimizer;
3.2 inputting the electrocardiosignals of the training set into the network in batches, outputting three levels of prediction results by forward propagation of the network, and calculating a loss function L between the prediction results and the labels all ；
3.3 according to the loss function L all The model parameters are continuously and iteratively updated by using an optimizer through a gradient descent method until a loss function L all Convergence and training are completed; the loss function L all The expression of (2) is as follows:
L all ＝L su +L un
L su ＝L CAD +L ce +L cross
wherein: l (L) su As a loss function of the supervised part, L un As a loss function of the self-supervision part, L CAD For class-aware dynamic loss function, L ce As a cross entropy function, L cross Is a three-layer class constraint function.
Further, the cross entropy function L ce The expression of (2) is as follows:
Figure BDA0004121624300000041
wherein:
Figure BDA0004121624300000042
the nth group of electrocardiosignals which are used for indicating the category label information in one batch are input into a network to output category probability prediction results about the level l, and the n group of electrocardiosignals are input into the network to output category probability prediction results about the level l>
Figure BDA0004121624300000043
Representation->
Figure BDA0004121624300000044
And the corresponding category labeling information is that N is the number of electrocardiosignals adopting the category labeling information in one batch, and l1, l2 and l3 respectively represent a coarse granularity level, a confusion level and a fine granularity level.
Further, the class-aware dynamic loss function L CAD The expression of (2) is as follows:
Figure BDA0004121624300000045
wherein:
Figure BDA0004121624300000046
the nth group of electrocardiosignals which are used for indicating the category label information in one batch are input into a network to output category probability prediction results about the level l, and the n group of electrocardiosignals are input into the network to output category probability prediction results about the level l>
Figure BDA0004121624300000047
Representation->
Figure BDA0004121624300000048
Corresponding class marking information, wherein N is the number of electrocardiosignals adopting the class marking information in one batch, and l1, l2 and l3 respectively represent coarse granularity level, confusion level and fine granularity level, and w avg Representation->
Figure BDA0004121624300000049
And the corresponding category weight.
Further, for any category of hierarchy l, its category weight w avg The expression of (2) is as follows:
Figure BDA00041216243000000410
Figure BDA00041216243000000411
Figure BDA00041216243000000412
wherein: l (L) avg Represents the average loss function, L tar Representing a target loss function, L, with respect to the overall distribution of data m Representing the cross entropy function, K, of the class in the mth iteration m Represents the number of electrocardiographic signals belonging to the category in the mth iteration, M representing the total number of iterations.
Further, the three-layer class constraint function L cross The expression of (2) is as follows:
Figure BDA00041216243000000535
wherein:
Figure BDA0004121624300000051
an nth group of electrocardiosignals which are used for representing a batch and adopt category labeling information are input into a network to output category probability prediction results about a fine granularity level l3, < >>
Figure BDA0004121624300000052
Representation->
Figure BDA0004121624300000053
Corresponding category label information->
Figure BDA0004121624300000054
Representing the use of category annotation messages in a batchThe n-th group of electrocardiosignals of the information are input into the network to output a category probability prediction result about the confusion level l2, < >>
Figure BDA0004121624300000055
Representation->
Figure BDA0004121624300000056
Corresponding category label information->
Figure BDA0004121624300000057
An nth group of electrocardiosignals which are used for representing a batch and adopt category labeling information are input into a network to output category probability prediction results about coarse granularity level l1, < >>
Figure BDA0004121624300000058
Representation->
Figure BDA0004121624300000059
And the corresponding category label information is N, and the number of electrocardiosignals adopting the category label information in one batch is N.
Further, the loss function L un The expression for is as follows:
Figure BDA00041216243000000510
wherein:
Figure BDA00041216243000000511
the kth group of electrocardiosignals which are not used for marking information of categories in one batch are input into a network to output category probability prediction results about a fine granularity level l3, +/->
Figure BDA00041216243000000512
The kth group of electrocardiosignals which are not used for marking information of category in one batch are input into a network to output a category probability prediction result about the confusion level l2, < >>
Figure BDA00041216243000000513
The kth group of electrocardiosignals which are not used for marking information of category in one batch are input into a network to output category probability prediction results about coarse granularity level l1, +/->
Figure BDA00041216243000000514
Representation->
Figure BDA00041216243000000515
Class probability tag result converted into confusion hierarchy l2,>
Figure BDA00041216243000000516
representation->
Figure BDA00041216243000000517
Class probability tag result converted into coarse-grained level l1,/->
Figure BDA00041216243000000518
Representation->
Figure BDA00041216243000000519
And converting the result into a class probability label result of the coarse-granularity level l1, wherein K is the number of electrocardiosignals which do not adopt class label information in one batch.
Further, the class probability tag results
Figure BDA00041216243000000520
B is the class number of confusion hierarchy l2, which is a B-dimensional vector, and +.>
Figure BDA00041216243000000521
1≤b≤B，/>
Figure BDA00041216243000000522
Is->
Figure BDA00041216243000000523
The value of element b of (i.e.)>
Figure BDA00041216243000000524
The maximum value of all class probability predictors belonging to class b of the confusion hierarchy l 2; the class probability label results
Figure BDA00041216243000000525
Is a 2-dimensional vector, 2 is the category number of coarse-grained level l1, and +.>
Figure BDA00041216243000000526
Figure BDA00041216243000000527
Is->
Figure BDA00041216243000000528
The a-th element value of (a>
Figure BDA00041216243000000529
The maximum value of all class probability predictors belonging to class a of coarse-grained level l 1; said class probability tag result->
Figure BDA00041216243000000530
In the form of a 2-dimensional vector,
Figure BDA00041216243000000531
Figure BDA00041216243000000532
is->
Figure BDA00041216243000000533
The a-th element value of (a>
Figure BDA00041216243000000534
The maximum of all class probability predictors belonging to class a of coarse-grained level l 1.
The invention uses semi-supervised algorithm to solve the problem of fine granularity classification and identification of electrocardiogram, and inputs the normalized multi-lead electrocardiosignal into a deep learning neural network and outputs three granularity prediction results; the network integrally uses an end-to-end training mode to form an integral closed loop, and three-level granularity constraint and class sensitive loss functions are used for extracting corresponding information for marked and unmarked data; the model integrally uses a self-learning and self-constraint method, so that a better learning effect can be achieved under the condition of insufficient labeling quantity.
Drawings
FIG. 1 is a schematic diagram of a semi-supervised model of the present invention.
FIG. 2 is a flow chart of the model training method of the present invention.
FIG. 3 shows the confusion matrix results (label on horizontal axis and predictor on vertical axis) before the class-aware dynamic loss function and three-level granularity class constraint are not added using 20% training label for semi-supervised training.
FIG. 4 is a confusion matrix result (label on horizontal axis and predictor on vertical axis) after adding class-aware dynamic loss function and three-level granularity class constraint using 20% training label for semi-supervised training.
Detailed Description
In order to more particularly describe the present invention, the following detailed description of the technical scheme of the present invention is provided with reference to the accompanying drawings and the specific embodiments.
The whole network model of the invention is shown in figure 1, and the electrocardiosignal x with the label is provided a And electrocardiosignal x without label u Input to ResNet based deep learning neural network, model performs three levels of prediction P on output x l3 、P l2 、P l1 . For the supervised part, using cross entropy functions, class-aware dynamic loss functions and three-level loss simultaneous constraints; for unsupervised parts, first, a hierarchical class transformation expression is used to perform predictive transformations of different hierarchies, and then three-level loss is used to constrain. The probability prediction result P is defined as follows:
P l3 ,P l2 ,P l1 ＝f(X)
specifically:
Figure BDA0004121624300000061
wherein: c is the total number of categories of classification.
For the supervised part, supervision will be performed using a cross entropy function, a class-aware dynamic loss function, and three-layer class constraints, where the cross entropy function is expressed as:
Figure BDA0004121624300000071
wherein:
Figure BDA0004121624300000072
the nth group of electrocardiosignals which are used for indicating the category label information in one batch are input into a network to output category probability prediction results about the level l, and the n group of electrocardiosignals are input into the network to output category probability prediction results about the level l>
Figure BDA0004121624300000073
Representation->
Figure BDA0004121624300000074
And the corresponding category labeling information is that N is the number of electrocardiosignals adopting the category labeling information in one batch, and l1, l2 and l3 respectively represent a coarse granularity level, a confusion level and a fine granularity level.
The class-aware dynamic loss function is:
Figure BDA0004121624300000075
Figure BDA0004121624300000076
Figure BDA0004121624300000077
Figure BDA0004121624300000078
three-layer category constraints are:
Figure BDA0004121624300000079
the overall loss function is therefore:
L su ＝L CAD +L ce +L cross
for data without labels, the hierarchical class conversion is performed after the model is output, and the expression is:
y l3-l2 ＝[max(p l3 ∈b)]，1≤b≤B
y l3-l1 ＝[max(p l3 ∈a)]，a＝1,2，
y l2-l1 ＝[max(p l2 ∈a)]，a＝1,2
wherein: max (p) l3 E b) is y l3-l2 The value of the b-th element in (i.e. p) l3 The maximum value of all class probability predictors belonging to class b of the confusion hierarchy l2, max (p l3 E.a) is y l3-l1 The value of the a-th element, p l3 The maximum value of all class probability predictors belonging to class a of coarse-grained hierarchy l1, max (p l2 E.a) is y l2-l1 The value of the a-th element, p l2 The maximum of all class probability predictors belonging to class a of coarse-grained level l 1.
Three levels of constraints are used on this basis:
Figure BDA0004121624300000081
whereby the supervision information L of the whole model all Namely, the combination of two parts:
L all ＝L su +L un
the flow of the model training method of the present invention is shown in fig. 2, where a random method is used to firstly divide a part of the data set with labels from the data set, and the rest part will be the data set without labels, where the ratio is 0.2, 0.4, 0.6, and 0.8.
(x a ,y a )＝random_select(x,ratio)
And then, constructing a deep learning neural network ResNet, and for the learning of the network, solving three objective functions by adopting a random gradient descent algorithm on the whole model, wherein the iteration steps are as follows:
Figure BDA0004121624300000082
Figure BDA0004121624300000083
wherein:
Figure BDA0004121624300000084
for gradient operators, lr is the learning rate, w is the matrix parameters of the network, and C is the total number of classified categories.
The final conditions for model convergence are as follows:
|L″-L′|≤ε
wherein: l 'and L' are loss functions before and after iteration, ε is a converging threshold, and is set to 10 -5 。
To verify the effectiveness of the present invention, we performed experiments using cardiac 12 lead electrocardiographic signals. When the method is used for carrying out semi-supervised algorithm research, for experimental comparison, the same data set division is also applicable to carry out fully supervised learning on the part of data, and the proportion of the data is still 0.2, 0.4, 0.6 and 0.8; the overall input to the network is normalized so that all signal lengths are 1000 samples, and specific data settings are given in table 1:
TABLE 1
Figure BDA0004121624300000085
Figure BDA0004121624300000091
For the training model part, the deep learning algorithm network also uses a random gradient descent algorithm to converge the model. The experiment adopts GPU of NVIDIA 1080Ti to perform acceleration operation, the batch size of the whole training is 4, the signal is normalized to be 12 multiplied by 1000, the learning rate of the model is 0.0025, and the whole training process needs 40 iterations; the fully supervised training dataset is consistent with its corresponding semi-supervised annotation data and the same generation of the segmentation network is used. FIG. 3 is a confusion matrix result before the class-aware dynamic loss function and three-level granularity class constraint are not added using 20% training labels for semi-supervised training. FIG. 4 is a confusion matrix result after adding a class-aware dynamic loss function and three-level granularity class constraint, using 20% training labels for semi-supervised training, where the horizontal axis is label and the vertical axis is predictor, where the light representation is less in number and the dark representation is more in number. It can be seen from the figure that the problem of prediction errors (the problem of misprediction of the results of eg.ALMI, AMI and ASMI) often occurs in the same confusion set of prediction results before the class-aware dynamic loss function and the three-level granularity class constraint are not added, but after the two parts of supervision information proposed by us are added, the problem is greatly reduced, the prediction result of the whole model is more accurate, and the prediction is more concentrated on a diagonal line. According to the semi-supervised algorithm, the problem of misprediction of the confusion set with similar characteristics can be better processed under the condition of less labeling quantity, so that a better prediction result is achieved under the condition of less labeling, the problem that a large number of labels are needed by the traditional full-supervised algorithm is relieved to a certain extent, and the analysis of intelligent medical images is better assisted under the condition of less labeling quantity.
The previous description of the embodiments is provided to facilitate a person of ordinary skill in the art in order to make and use the present invention. It will be apparent to those having ordinary skill in the art that various modifications to the above-described embodiments may be readily made and the generic principles described herein may be applied to other embodiments without the use of inventive faculty. Therefore, the present invention is not limited to the above-described embodiments, and those skilled in the art, based on the present disclosure, should make improvements and modifications within the scope of the present invention.

Claims (10)

1. A semi-supervised electrocardiogram myocardial infarction positioning method based on hierarchical different granularity class constraints comprises the following steps:
(1) Collecting electrocardiosignals which come from different patients and have myocardial infarction category labels, and dividing the electrocardiosignals into training sets and testing sets according to the same signal length;
(2) Selecting a deep learning neural network to predict myocardial infarction categories, wherein prediction output is divided into three levels, namely a fine-granularity level, a confusion level and a coarse-granularity level;
(3) Inputting electrocardiosignals of a training set into the deep learning neural network for training according to batches, wherein one part of the electrocardiosignals in each batch adopts category marking information, and the other part does not adopt category marking information; further designing a loss function based on cross entropy loss, category perception dynamic loss and cross category self-constraint for network iterative updating;
(4) And inputting the electrocardiosignals of the test set into the trained neural network, and predicting the corresponding myocardial infarction type result.
2. The semi-supervised electrocardiographic myocardial infarction positioning method as set forth in claim 1, wherein: the deep learning neural network is a one-dimensional feature extraction network which is based on a ResNet structure and comprises three feature output heads.
3. The semi-supervised electrocardiographic myocardial infarction positioning method as set forth in claim 1, wherein: by a means ofThe prediction results output by the deep learning neural network are P respectively l3 ，P l2 ，P l1 Wherein P is l3 Class probability prediction results corresponding to fine granularity levels are C multiplied by D; p (P) l2 Class probability prediction results corresponding to the confusion hierarchy are B multiplied by D; p is p l1 The size of the class probability prediction result corresponding to the coarse granularity level is 2 xD, D is the number of input signals of each batch, C and B are the number of classes of the fine granularity level and the confusion level respectively, and C > B > 2.
4. The semi-supervised electrocardiographic myocardial infarction positioning method as set forth in claim 1, wherein: the specific process of training the deep learning neural network in the step (3) is as follows:
initializing model parameters including bias vectors and weight matrixes of each layer, learning rate and an optimizer;
3.2 inputting the electrocardiosignals of the training set into the network in batches, outputting three levels of prediction results by forward propagation of the network, and calculating a loss function L between the prediction results and the labels all ；
3.3 according to the loss function L all The model parameters are continuously and iteratively updated by using an optimizer through a gradient descent method until a loss function L all Convergence and training are completed; the loss function L all The expression of (2) is as follows:
L all ＝L su +L un
L su ＝L CAD +L ce +L cross
wherein: l (L) su As a loss function of the supervised part, L un As a loss function of the self-supervision part, L CAD For class-aware dynamic loss function, L ce As a cross entropy function, L cross Is a three-layer class constraint function.
5. The semi-supervised electrocardiographic myocardial infarction positioning method as set forth in claim 4, wherein: the cross entropy function L ce The expression of (2) is as follows:
Figure FDA0004121624290000021
wherein:
Figure FDA0004121624290000022
the nth group of electrocardiosignals which are used for indicating the category label information in one batch are input into a network to output category probability prediction results about the level l, and the n group of electrocardiosignals are input into the network to output category probability prediction results about the level l>
Figure FDA0004121624290000023
Representation->
Figure FDA0004121624290000024
And the corresponding category labeling information is that N is the number of electrocardiosignals adopting the category labeling information in one batch, and l1, l2 and l3 respectively represent a coarse granularity level, a confusion level and a fine granularity level.
6. The semi-supervised electrocardiographic myocardial infarction positioning method as set forth in claim 4, wherein: the class-aware dynamic loss function L CAD The expression of (2) is as follows:
Figure FDA0004121624290000025
wherein:
Figure FDA0004121624290000026
the nth group of electrocardiosignals which are used for indicating the category label information in one batch are input into a network to output category probability prediction results about the level l, and the n group of electrocardiosignals are input into the network to output category probability prediction results about the level l>
Figure FDA0004121624290000027
Representation->
Figure FDA0004121624290000028
Corresponding class marking information, wherein N is the number of electrocardiosignals adopting the class marking information in one batch, and l1, l2 and l3 respectively represent coarse granularity level, confusion level and fine granularity level, and w avg Representation of
Figure FDA0004121624290000029
And the corresponding category weight.
7. The semi-supervised electrocardiographic myocardial infarction positioning method of claim 6, wherein: for any class of hierarchy l, its class weight w avg The expression of (2) is as follows:
Figure FDA00041216242900000210
Figure FDA00041216242900000211
Figure FDA0004121624290000031
wherein: l (L) avg Represents the average loss function, L tar Representing a target loss function, L, with respect to the overall distribution of data m Representing the cross entropy function, K, of the class in the mth iteration m Represents the number of electrocardiographic signals belonging to the category in the mth iteration, M representing the total number of iterations.
8. The semi-supervised electrocardiographic myocardial infarction positioning method as set forth in claim 4, wherein: the three-layer category constraint function L cross The expression of (2) is as follows:
Figure FDA0004121624290000032
wherein:
Figure FDA0004121624290000033
an nth group of electrocardiosignals which are used for representing a batch and adopt category labeling information are input into a network to output category probability prediction results about a fine granularity level l3, < >>
Figure FDA0004121624290000034
Representation->
Figure FDA0004121624290000035
Corresponding category label information->
Figure FDA0004121624290000036
An nth group of electrocardiosignals which are used for representing a batch and adopt category labeling information are input into a network to output category probability prediction results about a confusion level l2, < >>
Figure FDA0004121624290000037
Representation->
Figure FDA0004121624290000038
Corresponding category label information->
Figure FDA0004121624290000039
An nth group of electrocardiosignals which are used for representing a batch and adopt category labeling information are input into a network to output category probability prediction results about coarse granularity level l1, < >>
Figure FDA00041216242900000310
Representation->
Figure FDA00041216242900000311
And the corresponding category label information is N, and the number of electrocardiosignals adopting the category label information in one batch is N.
9. The semi-supervised electrocardiographic myocardial infarction positioning method as set forth in claim 4, wherein: the loss function L un The expression for is as follows:
Figure FDA00041216242900000312
wherein:
Figure FDA00041216242900000313
the kth group of electrocardiosignals which are not used for marking information of categories in one batch are input into a network to output category probability prediction results about a fine granularity level l3, +/->
Figure FDA00041216242900000314
The kth group of electrocardiosignals which are not used for marking information of category in one batch are input into a network to output a category probability prediction result about the confusion level l2, < >>
Figure FDA00041216242900000315
The kth group of electrocardiosignals which are not used for marking information of category in one batch are input into a network to output category probability prediction results about coarse granularity level l1, +/->
Figure FDA00041216242900000316
Representation->
Figure FDA00041216242900000317
Class probability tag result converted into confusion hierarchy l2,>
Figure FDA00041216242900000318
representation->
Figure FDA00041216242900000319
Class probability tag result converted into coarse-grained level l1,/->
Figure FDA00041216242900000320
Representation->
Figure FDA00041216242900000321
And converting the result into a class probability label result of the coarse-granularity level l1, wherein K is the number of electrocardiosignals which do not adopt class label information in one batch.
10. The semi-supervised electrocardiographic myocardial infarction positioning method as set forth in claim 9, wherein: the class probability label results
Figure FDA0004121624290000041
B is the class number of confusion hierarchy l2, which is a B-dimensional vector, and +.>
Figure FDA0004121624290000042
Figure FDA0004121624290000043
Is->
Figure FDA0004121624290000044
The value of element b of (i.e.)>
Figure FDA0004121624290000045
The maximum value of all class probability predictors belonging to class b of the confusion hierarchy l 2; said class probability tag result->
Figure FDA0004121624290000046
Is a 2-dimensional vector, 2 is the category number of coarse-grained level l1, and
Figure FDA0004121624290000047
Figure FDA0004121624290000048
is->
Figure FDA0004121624290000049
The a-th element value of (a>
Figure FDA00041216242900000410
The maximum value of all class probability predictors belonging to the sigma class of the coarse-grained level l 1; said class probability tag result->
Figure FDA00041216242900000411
Is a 2-dimensional vector>
Figure FDA00041216242900000412
Is->
Figure FDA00041216242900000413
The a-th element value of (a>
Figure FDA00041216242900000414
The maximum of all class probability predictors belonging to class a of coarse-grained level l 1.
CN202310234705.9A 2023-03-13 2023-03-13 Semi-supervised electrocardiogram myocardial infarction positioning method based on hierarchical different granularity class constraint Pending CN116350231A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202310234705.9A CN116350231A (en) 2023-03-13 2023-03-13 Semi-supervised electrocardiogram myocardial infarction positioning method based on hierarchical different granularity class constraint

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202310234705.9A CN116350231A (en) 2023-03-13 2023-03-13 Semi-supervised electrocardiogram myocardial infarction positioning method based on hierarchical different granularity class constraint

Publications (1)

Publication Number Publication Date
CN116350231A true CN116350231A (en) 2023-06-30

Family

ID=86918114

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202310234705.9A Pending CN116350231A (en) 2023-03-13 2023-03-13 Semi-supervised electrocardiogram myocardial infarction positioning method based on hierarchical different granularity class constraint

Country Status (1)

Country Link
CN (1) CN116350231A (en)

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN117796817A (en) * 2024-02-29 2024-04-02 山东大学齐鲁医院 Method and system for rapid detection and early warning of acute myocardial infarction
CN118177827A (en) * 2024-04-22 2024-06-14 中国人民解放军南部战区总医院 Myocardial infarction positioning method based on electrocardiograph vector diagram

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN117796817A (en) * 2024-02-29 2024-04-02 山东大学齐鲁医院 Method and system for rapid detection and early warning of acute myocardial infarction
CN118177827A (en) * 2024-04-22 2024-06-14 中国人民解放军南部战区总医院 Myocardial infarction positioning method based on electrocardiograph vector diagram

Similar Documents

Publication Publication Date Title
CN111192245B (en) Brain tumor segmentation network and method based on U-Net network
Biffi et al. Explainable anatomical shape analysis through deep hierarchical generative models
CN110475505A (en) Utilize the automatic segmentation of full convolutional network
CN116350231A (en) Semi-supervised electrocardiogram myocardial infarction positioning method based on hierarchical different granularity class constraint
WO2021114130A1 (en) Unsupervised self-adaptive mammary gland lesion segmentation method
CN107766874B (en) Measuring method and measuring system for ultrasonic volume biological parameters
CN106599549A (en) Computer-aided diagnosis system and method, and medical system
Kazemi et al. Classifying tumor brain images using parallel deep learning algorithms
Wang et al. Metastatic cancer image binary classification based on resnet model
Abbas et al. Enhanced Skin Disease Diagnosis through Convolutional Neural Networks and Data Augmentation Techniques
CN116864140A (en) Intracardiac branch of academic or vocational study postoperative care monitoring data processing method and system thereof
CN114972291B (en) Medical image structured automatic labeling method and system based on hybrid enhanced intelligence
Zeeshan Aslam et al. AML‐Net: Attention‐based multi‐scale lightweight model for brain tumour segmentation in internet of medical things
Anshad et al. Recent methods for the detection of tumor using computer aided diagnosis—A review
Verma et al. Role of deep learning in classification of brain MRI images for prediction of disorders: a survey of emerging trends
Pavithra et al. An Overview of Convolutional Neural Network Architecture and Its Variants in Medical Diagnostics of Cancer and Covid-19
Xiao et al. Rcga-net: An improved multi-hybrid attention mechanism network in biomedical image segmentation
Reda et al. A transfer learning-enabled optimized extreme deep learning paradigm for diagnosis of covid-19
Muis et al. Medical image classification of brain tumor using convolutional neural network algorithm
CN117274667A (en) Fetal heart disease ultrasonic image quality detection method based on multitasking learning
CN116759076A (en) Unsupervised disease diagnosis method and system based on medical image
Zhang et al. Research on brain glioma segmentation algorithm
Jin et al. Multiple instance learning for lymph node metastasis prediction from cervical cancer MRI
Liu et al. Neural network for lung cancer diagnosis
Kantheti et al. Medical Image Classification for Disease Prediction with the aid of Deep Learning approaches

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