A digital electrocardiograph calibration method based on deep learning combined with corner detection

Abstract

The application discloses a digital electrocardiograph machine calibration method based on deep learning combined with corner point detection, which comprises the following steps: paper simulation calibration electrocardiogram data set is prepared, digital image processing is performed on the paper simulation calibration electrocardiogram data set, and SIFT corner point features are extracted; corner point detection is performed on the paper simulation calibration electrocardiogram data set after digital image processing through a SIFT corner point detection algorithm, and multiple SIFT corner point key points are selected as labels of a DeepLabCut paper simulation calibration electrocardiogram training data set; the paper simulation calibration electrocardiogram data set is tested by DeepLabCut, and the positions of the multiple key points are predicted; according to the predicted key point positions, amplitude parameters and time parameters are calculated, the amplitude parameters and the time parameters are used as amplitude-time parameters required for digital electrocardiograph machine calibration, and physical quantities are converted.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of metrology, in particular to a digital electrocardiograph machine calibration method based on deep learning combined with corner point detection. BACKGROUND

[0002] Digital electrocardiograph machine calibration is to determine whether the digital electrocardiograph machine meets the specified requirements by experiment according to the calibration regulation by the legal metrology department or the legally authorized organization. The electrocardiograph machine metrological standard supporting equipment required by JJG 1041-2008 Digital Electrocardiograph Machine Calibration Regulation includes signal generator, analog impedance circuit, dividing rule, and scale, and the calibration personnel uses the dividing rule and scale to measure the waveform pattern characteristic parameters for calibration. In actual operation, the calibration data to be measured is complex and there are many problems such as easy to make mistakes, heavy workload, complex manual operation, long training time, high cost, and low efficiency, etc. Therefore, how to realize the automation and intelligentization of digital electrocardiograph machine calibration has become a problem to be solved.

[0003] The digital electrocardiograph machine calibration method is mainly an electrical signal processing method and an image recognition method based on machine vision. Invention CN104840194B and CN104720790A input the electrocardiograph data output by the digital electrocardiograph machine into the recognition module for electrical signal processing analysis. Patent CN115512130A and CN101978931B use image recognition method to extract characteristic parameters of electrocardiograph signal.

[0004] The above specific patent comparison files are:

[0005] 1) "A paper electrocardiograph waveform parameter automatic measurement method", patent number CN115512130A, this invention belongs to the field of digital image processing technology, and specifically refers to a paper electrocardiograph waveform parameter automatic measurement method, which includes electrocardiograph image acquisition, electrocardiograph image tilt detection and correction, electrocardiograph grid feature removal, electrocardiograph waveform extraction, and electrocardiograph data calculation. This invention uses a high-resolution industrial camera combined with an AI algorithm to complete automatic identification, detection, calculation of electrocardiograph waveform in the specified area, extraction of wave peak, wave trough, and horizontal line in the field of view, automatic marking of measurement points, identification of isopleth position, identification of heartbeat waveform, identification of voltage waveform, obtaining of data, high detection efficiency, high measurement data precision, and improvement of robustness and reliability of electrocardiograph machine index calibration. The key points and feature parameter extraction method of the present application are different from the above-mentioned patents, and the DeepLabCut deep learning combined with SIFT corner point detection method is used to simplify the construction process of the deep learning training set, simplify the calibration process, and reduce the workload of manual marking.

[0006] 2), "A digital electrocardiograph calibration method and system", patent number CN104840194B, the invention discloses a kind of digital electrocardiograph calibration method and system, for realizing the intelligent calibration of digital electrocardiograph.The present application receives the electrocardio data output by the digital electrocardiograph to be tested by access module, and the received electrocardio data is transmitted to identification module, identification module identifies the calibration item to which the received electrocardio data belongs, and then the judgment module judges whether the received electrocardio data is qualified according to the identification result.The present application can realize the intelligent calibration of digital electrocardiograph, improve the calibration efficiency, reduce human error, improve the calibration accuracy, and at the same time, it is suitable for existing calibration instrument.The feature parameter extraction method of the present application is different from the above-mentioned patent.The electrocardio signal processed by the above-mentioned patent is transmitted to the calibration system through data line, different analog electrocardio templates are selected for analog electrocardio waveform, and the present application adopts DeepLabCut deep learning combined with SIFT corner detection method for paper electrocardio simulation electrocardio image, which has higher detection robustness and stronger generalization ability.

[0007] 3), "Multi-channel synchronous electrocardiograph calibrator", patent number CN104720790A.The present application provides a kind of multi-channel synchronous electrocardiograph calibrator, belongs to the field of measuring biological electric signal of human body or each part of human body.The calibrator has human electrocardio signal simulation module for storing database in electrocardio waveform data and converting it into potential waveform that electrocardiograph can detect and electrocardiograph calibration module for comparing the waveform of human electrocardio signal simulation module with the waveform actually collected by electrocardiograph, to judge whether the detection function and analysis function of the electrocardiograph are qualified.The calibrator can simulate the real electrocardiograph working state, realize the synchronous output of 12 lead standard and abnormal electrocardio waveform signal, and is used for the measurement and evaluation of automatic analysis function of electrocardiograph.The detection method of the present application is different from the above-mentioned patent, and the above-mentioned application compares whether the electrocardio module stored in acquisition module and human electrocardio simulation module is same or whether it satisfies minimum deviation by waveform comparison module, and the present application describes the extraction method of feature point and the calculation method of feature parameter in detail, to achieve the purpose of calibrating electrocardiograph by feature point extraction and calibration parameter calculation.

[0008] 4) "A kind of calibrating device and method of electrocardiogram diagnostic analysis system", patent number CN101978931B, the application discloses a kind of calibrating device and method of electrocardiogram diagnostic analysis system, including: central processing module, power module, man-machine communication module, calibration module, weak signal module and communication module;The power module, man-machine communication module and the calibration module are all connected with the central processing module;The central processing module is connected with external electrocardiogram diagnostic analysis system by the weak signal module;The central processing module is connected with external intelligent terminal equipment or electrocardiogram diagnostic analysis system respectively by the communication module.This application realizes the automation of verification process by program design, eliminates the problems such as low efficiency and artificial error caused by artificial measurement in the past verification, greatly improves the verification efficiency, accuracy of corresponding project;And have anti-cheating algorithm module, can prevent unscrupulous electrocardiogram diagnostic analysis equipment manufacturer through technical cheating to respond to the verification requirements.The image processing method of the application is different from the above, the above application converts the waveform number signal into image data back to the intelligent terminal equipment by image acquisition module, the application directly detects feature point and extracts characteristic parameter to the paper quality simulation electrocardiogram image printed by digital electrocardiograph, retains the image characteristics of paper quality electrocardiogram feature point. SUMMARY

[0009] To solve the above technical problems, the purpose of the present application is to provide a kind of digital electrocardiograph verification method based on deep learning combined with corner detection.

[0010] The purpose of the present application is realized by the following technical solutions:

[0011] A kind of digital electrocardiograph verification method based on deep learning combined with corner detection, comprising:

[0012] A, paper quality simulation verification electrocardiogram data set is made, and SIFT corner feature is extracted by digital image processing to paper quality simulation verification electrocardiogram data set;

[0013] B, the paper quality simulation verification electrocardiogram data set after digital image processing is detected by SIFT corner detection algorithm, and multiple SIFT corner key points are selected as the label of DeepLabCut paper quality simulation verification electrocardiogram training data set;

[0014] C, DeepLabCut is used to test paper quality simulation verification electrocardiogram data set, and the positions of multiple key points are predicted;

[0015] D, according to the predicted key point position, the amplitude parameter and the time parameter are calculated, the amplitude parameter and the time parameter are used as the amplitude-time parameter required for digital electrocardiograph verification, and the physical quantity is converted.

[0016] Compared with the prior art, one or more embodiments of the present application can have the following advantages:

[0017] By introducing SIFT corner points, artificial selection, improving key point prediction accuracy and robustness, simplifying the DeepLabCut training set construction process and verification process, reducing the workload of artificial marking, improving the verification efficiency, obtaining 21 amplitude-time parameters required for verifying electrocardiograph; directly detecting feature points and extracting feature parameters from the paper simulation electrocardiogram image printed by the digital electrocardiograph, retaining the image features of the paper electrocardiogram feature points, and meeting the requirements of JJG 1041-2008 Digital Electrocardiograph Verification Regulation. The method provides a new solution for automatic intelligent verification of digital electrocardiographs, has the characteristics of high automation, high speed, low labor cost and high accuracy, and has practical significance and popularization value. BRIEF DESCRIPTION OF DRAWINGS

[0018] Figure 1 is a flow chart of a digital electrocardiograph verification method based on deep learning combined with corner point detection;

[0019] Figure 2 is a schematic diagram of 17 key points of a paper simulation verification electrocardiogram in a digital electrocardiograph verification method based on deep learning combined with corner point detection;

[0020] Figure 3 is a schematic diagram of SIFT corner key points and non-SIFT corner key points in a digital electrocardiograph verification method based on deep learning combined with corner point detection;

[0021] In the figure: 31-SIFT corner key point; 32-non-SIFT corner key point. DETAILED DESCRIPTION

[0022] In order to make the purpose, technical scheme and advantages of the present application clearer, the present application will be further described in detail below in combination with examples and drawings.

[0023] As shown in Figure 1 , it is a digital electrocardiograph verification method based on deep learning combined with corner point detection, which comprises:

[0024] Step 10: Make a paper simulation verification electrocardiogram data set, perform digital image processing on the paper simulation verification electrocardiogram data set, and extract SIFT corner features;

[0025] Step 20: Perform corner point detection on the paper simulation verification electrocardiogram data set processed by digital image processing through the SIFT corner point detection algorithm, and select a plurality of SIFT corner key points as the labels of the DeepLabCut paper simulation verification electrocardiograph training data set;

[0026] Step 30 tests the paper simulation calibration electrocardiogram dataset by DeepLabCut to predict the positions of multiple key points;

[0027] Step 40 calculates the amplitude parameter and the time parameter according to the predicted key point positions, takes the amplitude parameter and the time parameter as the amplitude-time parameters required for digital electrocardiograph calibration, and converts the physical quantity.

[0028] The above step 10 specifically comprises: performing digital image processing on the paper simulation calibration electrocardiogram dataset, including grayscale, binarization and median filtering; the paper simulation calibration electrocardiogram dataset is I origin , the image dataset is {F origin ,F origin-1 ,F origin-2 ,F origin-3 ,……,F origin-N}, where N is the number of images in I origin ; an image pixel coordinate system is established, with the coordinate origin O being the top left vertex of the image, the positive direction of the x-axis being horizontal to the right, and the positive direction of the y-axis being vertical downward; I origin is subjected to grayscale processing to obtain I grey , the image dataset is {F grey ,F grey-1 ,F grey-2 ,F grey-3 ,……,F grey-N}, where N is the number of images in I grey ; I grey is subjected to binarization processing to obtain I binary , the image dataset is {F binary ,F binary-1 ,F binary-2 ,F binary-3 ,……,F binary-N}, where N is the number of images in I binary ; I binary is subjected to median filtering processing to obtain I median filtering , the image dataset is {F median filtering ,F median filtering-1 ,F median filtering-2 ,F median filtering-3 ,……,F median filtering-N}, where N is the number of images in I median filtering .

[0029] The above step 20 specifically comprises: the object detected by the SIFT corner point detection algorithm is I median filtering , after being processed by the SIFT corner point detection algorithm, the image dataset is {F SIFT ,F SIFT-1 ,F SIFT-2 ,F SIFT-3 ,……,F SIFT-N}, where N is the number of images in ISIFT The number of images in the middle, for each F SIFT-i , i∈N, contains SIFT corner images and coordinates, where:

[0030] {P SIFT-1 ,P SIFT-2 ,P SIFT-3 ,……,P SIFT-n} constitutes the SIFT corner subset i SIFT , in i SIFT There are several P SIFT-i These are the key points required for labeling the training dataset of paper-based simulated electrocardiogram verification, (x click ,y click The coordinates of manually labeled sampling points are used. The following formula, SIFT corner points, can be used to assist in manually selecting high-precision key points and filtering out redundant SIFT corner points. The sampling point tolerance distance is α:

[0031]

[0032] Obtain SIFT corner keypoints 31 subset i SIFT-key i SIFT-key For {P SIFT-key-1 ,P SIFT-key-2 ,P SIFT-key-3 ,……,P SIFT-key-n}, where n is the number of SIFT corner keypoints, n<17, and this embodiment selects 17 keypoints, p1-p 17 (like Figure 2 (As shown).

[0033] 32 non-SIFT corner key points were manually selected, {P non-SIFT-key-1 ,P non-SIFT-key-2 ,P non-SIFT-key-3 ,……,P non-SIFT-key-n}Construct a subset i of non-SIFT corner keypoints non-SIFT From SIFT corner keypoint subset i SIFT Non-SIFT corner keypoint subset i non-SIFT Merged into a keypoint annotation subset i key-label :

[0034] i SIFT ∪i non-SIFT =i key-label (2)

[0035] i key-label Contains {P key-label-1 ,P key-label-2 ,P key-label-3 ,……,P key-label-17}, for each I origin F in the image dataset origin-iEach has a corresponding subset of key points i key-label (like Figure 3 (As shown).

[0036] Step 30 above specifically includes: I origin With the corresponding key point subset i key-label After being converted to a training set format, the data is input into DeepLabCut for training. The trained DeepLabCut model then performs key point detection on paper-based simulated electrocardiograms, detecting {P}. key-test-1 ,P key-test-2 ,P key-test-3 ,……,P key-test-17}, for P key-test-i Including x and y coordinates, it can be represented as {(x key-test-1 ,y key-test-1 ),(x key-test-2 ,y key-test-2 ),(x key-test-3 ,y key-test-3 ),……,(x key-test-17 ,y key-test-17 )}.

[0037] Step 40 above specifically includes: calculating 10 amplitude parameters (A1-A10) and 11 time parameters (T1-T11), which serve as the 21 amplitude-time parameters required for digital electrocardiograph calibration. According to {(x key-test-1 ,y key-test-1 ),(x key-test-2 ,y key-test-2 ),(x key-test-3 ,y key-test-3 ),……,(x key-test-17 ,y key-test-17 The calculated amplitude parameter is h. Ai (pix), amplitude parameter is w Ti (pix), meaning in pixels.

[0038]

[0039]

[0040]

[0041]

[0042]

[0043]

[0044]

[0045]

[0046]

[0047]

[0048]

[0049]

[0050]

[0051]

[0052]

[0053]

[0054]

[0055]

[0056]

[0057]

[0058]

[0059] In the standard calibration conditions (the recording sensitivity S of the electrocardiograph is 10 mm / mV, the recording speed of the electrocardiograph calibration pattern is v is 25 mm / s), the voltage V of each small grid of the calibration pattern grid longitudinal direction is 0.1 mV, the time t of each small grid of the horizontal direction is 0.04 s, and the resolution DPI of the sampling pattern equipment. The number of pixel points P contained in each small grid of the calibration pattern is calculated by using formulas (24), (25) and (26) respectively. n The amplitude parameter H of the calibrated ECG signal Ai (mm), the calibrated ECG time parameter W Ti (mm), that is, in millimeters (mm).

[0060]

[0061]

[0062]

[0063] Under the standard calibration conditions, according to the regulation file JJG 1041-2008, the standard value of the calibration parameter is determined. The physical quantity H Ai (mm) is converted to i.e. in millivolts (mV); converted to physical quantity W Ti (mm) converted to i.e. in milliseconds (ms).

[0064]

[0065]

[0066] While the present application has been disclosed with reference to the embodiments described above, it is to be understood that the disclosure is intended in an illustrative rather than a limiting sense, as it is contemplated that modifications and variations will occur to those skilled in the art, which modifications and variations will be within the spirit and scope of the present application. Accordingly, it will be appreciated that the scope of the present application is not intended to be limited to the embodiments described herein but is intended to be accorded the full scope inherent to the principles described herein including modifications and variations thereof.

Claims

1. A digital electrocardiograph verification method based on deep learning combined with corner detection, characterized in that, The method comprises: A, making a paper simulation calibration electrocardiogram data set, carrying out digital image processing on the paper simulation calibration electrocardiogram data set, and extracting SIFT corner point features; B, detecting the corner points of the paper simulation calibration electrocardiogram data set after digital image processing through a SIFT corner point detection algorithm, and selecting multiple SIFT corner point key points as labels of a DeepLabCut paper simulation calibration electrocardiogram training data set; C, testing the paper simulation calibration electrocardiogram data set by DeepLabCut, and predicting the positions of the multiple key points; D, calculating amplitude parameters and time parameters according to the predicted key point positions, taking the amplitude parameters and time parameters as amplitude-time parameters required for digital electrocardiograph calibration, and converting physical quantities; The object detected by the SIFT corner point detection algorithm in the B is I medianfiltering , which is processed by the SIFT corner point detection algorithm, I SIFT , and the image data set is {F SIFT-1 , F SIFT-2 , F SIFT-3 , …, F SIFT-N}, wherein N is the number of images in I SIFT , and for each F SIFT-i , i∈N, the SIFT corner point image and coordinates are included, wherein: {P SIFT-1 ,P SIFT-2 ,P SIFT -3,……,P SIFT-n} constitute a SIFT corner subset i SIFT , i SIFT There are several P SIFT-i is the key point required for labeling paper simulation verification electrocardiogram training data set, (x click , y click ) is the artificial labeling click sampling point coordinates, according to formula (1) and (2) SIFT corner auxiliary selection high-precision key point, and screen out redundant SIFT corner, the sampling fault tolerance distance is α: Get SIFT corner key point subset i SIFT-key , i SIFT-key is {P SIFT-key-1 , P SIFT-key-2 , P SIFT-key-3 , …, P SIFT-key-n}, wherein n is the number of SIFT corner key points, n < 17; Select non-SIFT corner key points, {P non-SIFT-key-1 ,P non-SIFT-key-2 ,P non-SIFT-key-3 ,……,P non-SIFT-key-n} constitute a non-SIFT corner key point subset i non-SIFT , the SIFT corner key point subset i SIFT and the non-SIFT corner key point subset i non-SIFT are merged into a key point label subset i key-label : i SIFT ∪i non-SIFT =i key-label (2) i key-label comprising {P key-label-1 i key-label-2 i key-label-3 i key-label-17}, for each I origin F origin-i i key-label in the image dataset.

2. The digital electrocardiograph verification method based on deep learning combined with corner detection according to claim 1, wherein, The digital image processing on the paper simulation calibration electrocardiogram data set in A comprises grayscale processing, binarization and median filtering.

3. The digital electrocardiograph verification method based on deep learning combined with corner detection according to claim 1, wherein, The paper quality simulation test ECG data set in A is I origin , I origin The image data set is {F origin-1 ,F origin-2 ,F origin-3 ,……,F origin-N}, wherein N is the number of images in I origin .

4. The digital electrocardiograph verification method based on deep learning combined with corner detection according to claim 2, characterized in that, The grayscale processing comprises establishing an image pixel coordinate system, and the coordinate origin O is the top left vertex of the image, the positive direction of the x-axis is horizontal right, and the positive direction of the y-axis is vertical downward; to I origin Grayscale processing yields I grey ,I grey The image dataset is {F grey-1 ,F grey-2 ,F grey-3 ,……,F grey-N }, where N is I grey Number of images in the middle; for I grey Binarization yields I binary ,I binary The dataset is {F binary-1 ,F binary-2 ,F binary-3 ,……,F binary-N }, where N is I binary Number of images in the middle; for I binary I is obtained by median filtering. medianfiltering ,I medianfiltering The dataset is {F medianfiltering-1 ,F medianfiltering-2 ,F medianfiltering-3 ,……,F medianfiltering-N }, where N is I medianfiltering Number of images in the text.

5. The digital electrocardiograph verification method based on deep learning combined with corner detection according to claim 3, characterized in that, In the C, I origin With the corresponding key point subset i key-label After conversion to the training set format, input to DeepLabCut for training, and after training, the DeepLabCut model detects key points of the paper simulation test ECG, and detects {P key-test-1 ,P key-test-2 ,P key-test-3 ,……,P key-test-17} for P key-test-i Contains x coordinates and y coordinates, represented as {(x key-test-1 ,y key-test-1 ),(x key-test-2 ,y key-test-2 ),(x key-test-3 ,y key-test-3 ),……,(x key-test-17 ,y key-test-17 )}.

6. The digital electrocardiograph verification method based on deep learning combined with corner detection according to claim 1, wherein, The amplitude parameters A1-A10 and time parameters T1-T11 in the D are calculated as the amplitude-time parameters required for digital electrocardiogram testing, according to {(x key-test-1 ,y key-test-1 ),(x key-test-2 ,y key-test-2 ),(x key-test-3 ,y key-test-3 ),……,(x key-test-17 ,y key-test-17} to obtain the amplitude parameter h Ai (pix) and the time parameter w Ti (pix), namely in units of pixels (pix); In the electrocardiograph, the recording sensitivity S is 10 mm / mV, the recording speed of the electrocardiograph verification pattern is v is 25 mm / s, the voltage V of each small grid in the longitudinal direction of the verification pattern grid is 0.1 mV, the time t of each small grid in the transverse direction is 0.04 s, and the resolution DPI of the acquisition pattern device; the number of pixel points P contained in each small grid of the verification pattern is calculated by formulas (24), (25) and (26) respectively n , the amplitude parameter H of the verified ECG signal Ai (mm), the verified ECG time parameter W Ti (mm), that is, in millimeters, determining the standard value of the assay parameter; converting the physical quantity H Ai (mm) into i.e. in millivolts; converting the physical quantity W Ti (mm) into i.e. in milliseconds;

Images