CN116230172A

CN116230172A - System for acquiring electrocardio ST segment information through shooting electrocardiogram

Info

Publication number: CN116230172A
Application number: CN202310006125.4A
Authority: CN
Inventors: 刘澄玉; 孙佳乐; 张铄; 赵莉娜; 李建清
Original assignee: Southeast University
Current assignee: Southeast University
Priority date: 2023-01-04
Filing date: 2023-01-04
Publication date: 2023-06-06

Abstract

The invention discloses a system for acquiring electrocardio ST segment information through photographing an electrocardiogram, which comprises an electrocardio image reading module, an ST segment abnormality diagnosis module and an android mobile phone APP function module; the electrocardiographic image reading module is used for rapidly positioning and extracting electrocardiographic areas from electrocardiographic reports shot by the mobile phone by utilizing the YOLOv5 target detection model to obtain electrocardiographic images without redundant information; the ST abnormal diagnosis module judges whether the ST is depressed or raised by utilizing a convolutional neural network and an ST intelligent diagnosis algorithm; the android mobile phone APP functional module is provided with an image selection module and a result feedback module, wherein the image selection module is used for shooting through a mobile phone camera or selecting pictures through a mobile phone built-in photo album, so that a user can upload standard 12-lead electrocardiogram report images to a platform; the result feedback module outputs an ST segment diagnosis result; the invention can judge whether the ST segment is abnormal or not according to the standard 12-lead electrocardiogram report uploaded by the user, and effectively assist the patient to comprehensively consider the next intelligent analysis and medical measures.

Description

System for acquiring electrocardio ST segment information through shooting electrocardiogram

Technical Field

The invention relates to the field of mobile medical treatment and medical monitoring, in particular to a system for acquiring electrocardio ST segment information by taking a picture of an electrocardiogram.

Background

Electrocardiography is a non-invasive medical examination method for detecting heart conditions by tracking electrocardiographic activity, and is widely used. In a conventional electrocardiographic examination, only 4 limb lead electrodes and V1-V66 chest lead electrodes are usually placed, and a standard 12-lead electrocardiogram is recorded. Because the electrocardiographic signals are generated by electrocardiographic and structural changes, the electrocardiograph contains a large amount of information directly reflecting the physiological functions of the heart, and is used for helping to diagnose electrocardiographic diseases such as ST depression or ST elevation.

With the widespread popularity of wearable electrocardiography, electrocardiography is an extremely widespread and relatively easy medical examination, and diagnosis of electrocardiography is often based on the abundant medical knowledge and clinical experience of cardiologists. If a patient gets electrocardiographic data through a household wearable electrocardiograph device and goes to a hospital for consultation, the complicated medical treatment process and high time cost are unfavorable for the real-time diagnosis of electrocardiographic diseases of the patient. Therefore, the occurrence of the APP for diagnosing the electrocardiographic abnormality can help patients to know possible diseases of the heart, and can also help society to save medical resources.

At present, the electrocardio data stored in images with high circulation in society is less, the quality level of the images is uneven, a practical and sufficient large-scale data set can not be provided for further research of an electrocardio abnormality diagnosis neural network model, and research of the electrocardio abnormality diagnosis neural network is blocked. Therefore, the bottleneck can be well broken through by converting one-dimensional signals into two-dimensional images and creating an electrocardiogram data set by using image enhancement methods such as rotation, background superposition and the like.

Target detection is a popular direction of computer vision and digital image processing, and is a research hotspot of theory and application in recent years. The system which uses the target detection technology to extract the electrocardio area in the 12-lead electrocardiogram report and uses the electrocardio abnormality diagnosis algorithm to provide the electrocardio ST segment information acquisition system which is simple in algorithm and convenient to use and acquires the electrocardio ST segment information through the photographed electrocardiogram is an important improvement and breakthrough in the electrocardio abnormality diagnosis field.

Disclosure of Invention

Aiming at the problems in the prior art, the invention provides a system for acquiring the information of the ST segment of the electrocardio by photographing an electrocardiogram, which photographs a standard 12-lead electrocardiogram report through an application program constructed on android equipment, rapidly locates and extracts an electrocardio region by using a YOLOv5 target detection model, judges whether the ST segment is abnormal or not by using an electrocardio abnormality diagnosis algorithm, effectively assists a patient to comprehensively consider the next intelligent analysis and medical measures, and is suitable for the fields of mobile medical treatment, medical monitoring and the like.

In order to achieve the aim of the invention, a system for acquiring the information of an electrocardio ST segment by taking a picture of an electrocardiogram is provided, which is characterized by comprising an electrocardio image reading module, an ST segment abnormality diagnosis module and an android mobile phone APP functional module, wherein the android mobile phone APP functional module comprises an image selection module and a result feedback module; the output end of the electrocardiographic image reading module is connected with the input end of the ST-segment abnormality diagnosis module, the output end of the image selecting module is connected with the input end of the electrocardiographic image reading module, and the output end of the ST-segment abnormality diagnosis module is connected with the input end of the result feedback module.

Specifically, the electrocardiographic image reading module is used for processing standard 12-lead electrocardiographic image data uploaded by a patient user, carrying out rapid identification, positioning and extraction on electrocardiographic image parts through a pre-trained YOLOv5 target detection model on the acquired data, and transmitting the data to the ST-segment abnormality diagnosis module after processing.

Specifically, the ST segment abnormality diagnosis module predicts the electrocardiograph part extracted by the electrocardiograph image reading module through a pretrained convolutional neural network, and uses an ST segment intelligent diagnosis algorithm to diagnose whether the electrocardiograph ST segment is raised or depressed.

Specifically, the android mobile phone APP functional module comprises an image selection module and a result feedback module; the image selecting module is used for shooting through a mobile phone camera or selecting pictures through a mobile phone built-in photo album, so that a user can upload standard 12-lead electrocardiogram report images to a platform and transmit the standard 12-lead electrocardiogram report images to the electrocardiogram image reading module; the result feedback module can feed back corresponding diagnosis information to the user according to the result of the ST segment abnormality diagnosis module.

In particular, the electrocardiographic reading module includes data set processing, model training, and data extraction.

Specifically, in the data set processing, an image training set is derived from two parts, and the label position of an electric part of an image center is obtained through LabelImg software and Pycharm software; the first partial image is an electrocardiogram image which converts an electrocardiosignal of a dataset in a 2018 physiological signal challenge race into a grid-based electrocardiogram image, and the electrocardiogram image is superimposed in random positions of a plurality of real background pictures after data enhancement operations such as rotation, scaling and the like are carried out, so that the authenticity of image data is similar to that of an electrocardiogram shot by a mobile phone; the second partial image is an electrocardiogram report of real shooting, and the actual sample size of the data set is increased to improve the generalization capability of the model.

Specifically, in the model training, the model is divided into three parts of feature extraction, feature fusion and prediction.

Specifically, the feature extraction comprises three parts of a downsampling structure, a cross-stage local network and a spatial pyramid pooling module.

Specifically, in the downsampling structure, the electrocardiograph image is sliced before entering the backbone network, and information of key points is extracted to obtain a downsampling feature map.

Specifically, in the cross-stage local network, the cross-stage local network is formed by splicing a convolution layer and a residual error network, so that the network learning capacity is enhanced, the calculation bottleneck is reduced, and the learning accuracy is maintained while the weight is reduced.

Specifically, in the spatial pyramid pooling module, the spatial pyramid pooling module is composed of a plurality of multi-scale maximum pooling layers, so that fusion of different feature scales is realized.

Specifically, in the feature fusion part, the feature fusion network performs further feature extraction on a feature map input by a main network in feature extraction through a cross-stage local network and a convolution layer module, processes the feature map through operations such as spatial pyramid pooling and the like, and transmits tensor data obtained through processing to a prediction layer.

Specifically, in the prediction part, a method of introducing an intersection ratio is used for defining a loss function, and a GIoU loss function is used as a loss function of an object recognition prediction frame.

Specifically, a training model algorithm is mainly used for training a YOLOv5 target detection model according to model parameters adjusted according to processing effects, key information of an electrocardiogram image is thinned through layer-by-layer neural network learning, and the key information is compared with label information obtained in a data processing step, and is continuously optimized and corrected to obtain a trained weight file.

Specifically, in the data extraction, according to an electrocardiographic image and a trained model which are transmitted by a user, an electrocardiographic region of the electrocardiographic image is predicted and extracted by using a prediction function, and the electrocardiographic region is transmitted to an ST-segment abnormality diagnosis module.

Specifically, the ST segment abnormality diagnostic module includes model training and outcome diagnostics.

Specifically, in the model training, the electrocardiographic data set is divided into a training set, a verification set and a test set by using a random classification method, and then is trained by a network.

Specifically, in the network training, the convolutional neural network model is divided into three parts of an input end, a backbone network and an output end.

Specifically, in the backbone network, the backbone network has eight layers in total, the first seven layers are composed of a two-dimensional convolution layer, a batch processing layer, a random inactivation layer and a correction linear unit layer, and the eighth layer is composed of 2 full connection layers and 1 correction linear unit layer.

Specifically, in the first seven layers of network, tensors which are output by an electrocardiographic image reading module and are preprocessed by an ST-segment abnormality diagnosis model are input, and sequentially pass through a two-dimensional convolution layer, a batch processing layer and a random inactivation layer respectively, and finally, the characteristics of electrocardiographic images are extracted for multiple times through the activation processing of a correction linear unit layer, so that the characteristics of the acquired electrocardiographic images are more accurate.

Specifically, in the eighth layer network, the data obtained by the output of the seventh layer is used as the input of the full-connection layer, and then the predicted result of the ST segment information is obtained through the RelU activation function processing and the full-connection layer processing.

Specifically, the training model algorithm is mainly used for testing the ST-segment abnormality diagnosis model performance under different learning rates and convolution kernels according to the prediction accuracy, and continuously optimizing and correcting through layer-by-layer neural network learning to obtain a trained weight file.

Specifically, in the result diagnosis, the diagnosis prediction function is used for carrying out abnormality diagnosis on the ST segment according to the electrocardiographic region image and the trained model obtained by the electrocardiographic image reading module.

Specifically, the ST segment diagnostic results include normal, ST segment elevation, and ST segment depression.

Specifically, the image selecting module of the android mobile phone APP functional module is used for selecting 12-lead electrocardiogram report pictures through mobile phone cameras or mobile phone built-in photo albums, helping users upload the pictures to a platform and transmitting the pictures to an electrocardiogram image reading module.

Specifically, the result feedback module of the android mobile phone APP function module outputs the ST segment diagnosis result.

Compared with the prior art, the system for acquiring the electrocardio ST segment information by photographing the electrocardiogram has the beneficial effects that:

(1) By adopting the electrocardiographic image reading module, the image of the electrocardiograph under the real background is simulated by using the data enhancement and image migration technology, the bottleneck of insufficient electrocardiographic image report data is broken through, and the positioning and the extraction of an electrocardiographic region are completed by using a lightweight target detection network structure, so that the registration accuracy can reach more than 95.3 percent.

(2) The adopted ST segment abnormality diagnosis module realizes the electrocardio ST abnormality diagnosis function by using the convolutional neural network, can identify whether the ST segment has a lifting or pressing condition, has the accuracy of 86.9 percent and 96.7 percent, and can allow any patient with a standard 12-lead electrocardiogram report to be screened for electrocardio abnormality in an auxiliary way.

(3) The invention is constructed on the android mobile phone APP, is closer to a user, and can complete all functions of shooting an electrocardio report and reasoning a deep learning model without a remote server by a personal mobile phone.

Drawings

Fig. 1 is a system configuration diagram.

Fig. 2 is a flowchart of an electrocardiographic image reading module.

Fig. 3 is a diagram of the YOLOv5 object detection network.

Fig. 4 is a diagram of a ST segment abnormality diagnosis network.

Fig. 5 is an APP page diagram of an android phone.

Detailed description of the preferred embodiments

The present invention will be described in further detail with reference to the drawings and examples, in order to make the objects, technical solutions and advantages of the present invention more apparent. It should be understood that the specific embodiments described herein are for purposes of illustration only and are not intended to limit the scope of the invention.

Reference herein to "an embodiment" means that a particular feature, structure, or characteristic described in connection with the embodiment may be included in at least one embodiment of the invention. The appearances of such phrases in various places in the specification are not necessarily all referring to the same embodiment, nor are separate or alternative embodiments mutually exclusive of other embodiments.

Referring to fig. 1, fig. 1 is a schematic structural diagram of a system for acquiring information of an ST segment of an electrocardiograph by taking a photograph of the electrocardiograph according to the first embodiment, which includes an electrocardiograph image reading module, an ST segment abnormality diagnosis module, and an android mobile phone APP function module, wherein the android mobile phone APP function module includes an image selecting module and a result feedback module; the output end of the electrocardiographic image reading module is connected with the input end of the ST-segment abnormality diagnosis module, the output end of the image selecting module is connected with the input end of the electrocardiographic image reading module, and the output end of the ST-segment abnormality diagnosis module is connected with the input end of the result feedback module.

As shown in fig. 1 to 3, in the data set processing, the image training set is derived from two parts, the first part of image is a data set for challenging the race by using physiological signals in 2018, electrocardiosignals of known ST disease category labels are randomly selected, converted into electrocardiogram images with grids as the background, and after data enhancement operations such as rotation, scaling and the like are carried out, the electrocardiogram images are superimposed in random positions of a plurality of real background images, so that the authenticity of image data is similar to that of an electrocardiogram shot by a mobile phone; the second partial image is an electrocardiogram report of real shooting, and the actual sample size of the data set is increased to improve the generalization capability of the model. And positioning and acquiring the label positions of the electric parts of the center of the image by using LabelImg software and Pycharm software, and naming the image and the label thereof correspondingly as a training set.

As shown in fig. 1 to 3, the input end processes the input image by using a mosaics data enhancement method, adaptive anchor calculation, adaptive image scaling and the like, wherein the mosaics data enhancement method adds robustness by splicing different images to form a new image.

As shown in fig. 1 to 3, in the model training, the electrocardiographic data set is divided into a training set, a verification set and a test set by using a random classification method, which specifically includes: (1) Setting the file name of the label to which the image belongs to be the same as the file name of the image, and respectively placing the file name of the label in the image folder and the label folder; (2) obtaining a file name list in a folder; (3) Calculating the quantity of the electrocardiographic images in the training set, the verification set and the test set according to the set proportion and the file name list length; (4) scrambling the list of file names by a function; (5) Distributing the electrocardio images and corresponding labels into a training set, a verification set and a test set according to the different numbers calculated in the step (3) from left to right, wherein the training set, the verification set and the test set respectively comprise an image folder and a label folder; (6) Repeating the operation (2) and (5) until the data under all feature classification are completely divided. The model adopts a YOLOv5s model, has higher running speed, smaller parameter number and better detection precision, and can reduce the requirement on computer computing resources.

The YOLOv5s model is divided into three parts, namely feature extraction, feature fusion and prediction.

The feature extraction part comprises a downsampling structure, a cross-stage local network and a spatial pyramid pooling module.

The downsampling structure performs slicing processing on the electrocardiographic image before the electrocardiographic image enters the backbone network, namely, the image data are cut into 4 parts, each part of data is obtained by downsampling by 2 times, then splicing is performed in the channel dimension, and information of key points is extracted to obtain a downsampling characteristic diagram. The expression is as follows: the input of 3×640×640 defaults is duplicated 4 copies, 4 pictures are cut into 4 slices of 3×320×320 by slicing operation, the slices are connected in depth and output as output of 12×320×320, finally, output of 32×320×320 is generated through convolution layer with convolution kernel of 32, and finally, the result is transmitted to the next convolution layer through batch processing operation and activation function.

The cross-stage local network is formed by splicing a convolution layer and a residual network, so that the network learning capacity is enhanced, the calculation bottleneck is reduced, and the learning accuracy is maintained while the weight is reduced. The backbone network has 4 cross-stage local networks, and the image with the feature map 320 is input from the upper layer, and the feature map is changed from 320 to 160, 80, 40 and 20 through the 4 cross-stage local networks.

The spatial pyramid pooling module is composed of a plurality of multi-scale maximum pooling layers, so that fusion of different feature scales is realized, and the feature map size is 20×20×256.

In the feature fusion part, the feature fusion network further extracts features of the feature map input by the main network in the feature extraction through the cross-stage local network and the convolution layer module, and then processes the feature map through operations such as spatial pyramid pooling, so that an image feature matrix output by the main network can be better utilized, and finally tensor data obtained through processing is transmitted to a prediction layer.

In the prediction part, a method of introducing an intersection ratio is used for defining a loss function, which is an important parameter for measuring the positioning accuracy of a target detection algorithm, a GIoU loss function is used as a loss function of an object identification prediction frame,

wherein B is the area of the prediction frame, B _gt The area of the real frame is C, and the minimum frame area containing both the predicted frame and the real frame is C.

The model algorithm mainly trains a Yolov5 target detection model according to the processing effect adjustment model parameters, and the model algorithm needs to be explained that the weight parameters of the Yolov5 pre-trained model are obtained as initial weights of an electrocardiographic image recognition model to learn and update. And updating parameters in the YOLOv5 neural network by using an optimization algorithm until the YOLOv5 neural network converges.

After the model test trained by the YOLOv5, the obtained test result is related to the confidence coefficient set by the test, and the confidence coefficient parameter is adjusted, so that the recognition of the electrocardiograph data area by the YOLOv5 achieves the maximum effect; the key information of the electrocardiogram image is thinned through layer-by-layer neural network learning, and is compared with the label information obtained in the data processing step, and the weight file after training is obtained.

As shown in fig. 1 to 3, in the data extraction, according to an electrocardiographic image and a trained model transmitted by a user, an electrocardiographic region of the electrocardiographic image is predicted and extracted by using a prediction function to obtain a pure electrocardiographic image without interference of redundant objects such as personal information of a patient or shooting background, the image is matched with a corresponding disease label of a 2018 physiological signal challenge match dataset, and a label value is converted into one-hot code to obtain a sample target value y= { Y ₁ ,Y ₂ ,Y ₃ ,Y _4, …,Y _i ,…,Y _n -wherein Y is _i The coding value of the ith sample label is 0 or 1,0 represents ST elevation, 1 represents ST depression, and the obtained label and image are transmitted to an ST abnormality diagnosis module to be used as a training set.

As shown in fig. 1 and 4, in the network construction and model training, the electrocardiographic data set is divided into a training set, a verification set and a test set by using a random classification method, which specifically includes: (1) Placing electrocardiographic images corresponding to the elevation of the ST segment and the depression of the ST segment in corresponding folders; (2) obtaining a file name list in a folder; (3) Calculating the quantity of the electrocardiographic images in the training set, the verification set and the test set according to the set proportion and the file name list length; (4) scrambling the list of file names by a function; (5) Assigning electrocardiographic images to the training set, the validation set and the test set according to the different numbers calculated in step (3) in the left-to-right order; (6) Repeating the operation (2) and (5) until the data under all feature classification are completely divided.

The convolutional neural network model is divided into an input end, a backbone network and an output end.

The backbone network is composed of eight layers, wherein the first seven layers are composed of a two-dimensional convolution layer, a batch processing layer, a random inactivation layer and a correction linear unit layer, and the eighth layer is composed of 2 full-connection layers and 1 correction linear unit layer.

In the first seven layers, tensor 675 x 1450 x 3 which is output by an electrocardiographic image reading module and is preprocessed by an ST-segment abnormality diagnosis model is input, and the tensor 675 x 1450 x 3 sequentially passes through a two-dimensional convolution layer with a convolution kernel size of 5*5, a batch processing layer and a random inactivation layer respectively, wherein the random inactivation layer randomly deletes part of network edges with a probability of 0.4, even if the number of neurons is kept at 60% to improve the robustness, and finally the network edges are subjected to the activation processing of a corrected linear unit layer. Through the first seven layers, the characteristics of the electrocardiograph images are extracted for multiple times, so that the characteristics of the acquired electrocardiograph images are more accurate.

The convolution layer is utilized to capture the remarkable characteristics, and the convolution layer has the characteristics of translational invariance, size invariance, rotation invariance and the like. The convolution-based distribution property mapping layer handles forward and backward propagation changes, and the following layer is named batch normalization to play a role in normalization, which can keep a stable gradient size and continuously adopt new changes.

Correcting the linear cell layer plays an important role in improving the performance of the convolutional neural network model based on nonlinear transformation, and the Relu function can be expressed as: f (x) =max (x; 0), x being the value input to the Relu activation function, introduces sparsity, reducing the occurrence of overfitting by reducing the interdependence between the parameters.

In the eighth layer, 2×8×512 data obtained by the output of the seventh layer is used as the input of the full-connection layer, and then is processed by the RelU activation function, and finally passes through the full-connection layer again to obtain the prediction result of the ST segment information. The prediction results may be classified as ST elevation or ST depression; the probability of a two-class outcome may also be output, e.g., (p, q) for the outcome, where p represents the probability of an ST elevation class, q represents the probability of an ST depression class, and if (0.2, 0.8) for the outcome, the outcome is interpreted as indicating that the outcome is predictive of ST depression.

The training model algorithm is mainly used for testing the performance of ST segment abnormal diagnosis models under different learning rates and convolution kernels according to the prediction accuracy, and continuously optimizing and correcting the model by adjusting the batch size, the initial learning rate and the like through layer-by-layer neural network learning to obtain a trained weight file.

As shown in fig. 1 and 4, the trained model is deployed at the android mobile phone end and is embedded in the APP in the mobile phone end.

As shown in fig. 1 and 4, in the result verification, the diagnosis prediction function is used to verify the abnormal diagnosis of the ST segment according to the electrocardiographic region image obtained by the electrocardiographic image reading module and the trained model.

As shown in fig. 1 and 5, the image selection module of the android mobile phone APP function module is divided into a "photo" and a "photo album", and clicking the "photo" button can open the mobile phone camera, and the mobile phone camera is used for shooting a 12-lead electrocardiogram report picture, so that a simple background with few impurities is selected as much as possible during shooting, and shooting is performed in an environment with sufficient light and no shadow as much as possible, and a parallel angle is formed between the camera and the picture as much as possible; clicking the "album" button can select the electrocardiograph picture stored locally in the mobile phone, and through the function, the user is assisted in uploading the picture to the platform, and the uploading electrocardiograph sources are not limited, and can be reports obtained by electrocardiographs of different models, but are necessarily standard 12-lead electrocardiograph reports.

As shown in fig. 1 and 5, on the APP of the mobile phone, the original image is subjected to an electrocardiographic image reading module to obtain an image with enhanced characteristics, which is done to highlight the key information of the image, remove the influence of redundant information such as name, gender, age, report completion time or background sundries, and enable the ST segment abnormality diagnosis network to better identify the symptoms.

As shown in fig. 1 and 5, the result feedback module of the android phone APP function module presents the diagnosis result of the ST segment to the front end.

The technical features of the above embodiments may be arbitrarily combined, and all possible combinations of the technical features in the above embodiments are not described for brevity of description, however, as long as there is no contradiction between the combinations of the technical features, they should be considered as the scope of the description.

It should be noted that, the term "first\second\third" related to the embodiment of the present invention is merely to distinguish similar objects, and does not represent a specific order for the objects, it is to be understood that "first\second\third" may interchange a specific order or sequence where allowed. It is to be understood that the "first\second\third" distinguishing aspects may be interchanged where appropriate to enable embodiments of the invention described herein to be implemented in sequences other than those illustrated or described.

The terms "comprising" and "having" and any variations thereof, are intended to cover a non-exclusive inclusion. For example, a process, method, apparatus, article, or device that comprises a list of steps or modules is not limited to the particular steps or modules listed and may optionally include additional steps or modules not listed or inherent to such process, method, article, or device.

The above examples illustrate only a few embodiments of the invention, which are described in detail and are not to be construed as limiting the scope of the invention. It should be noted that it will be apparent to those skilled in the art that several variations and modifications can be made without departing from the spirit of the invention, which are all within the scope of the invention. Accordingly, the scope of protection of the present invention is to be determined by the appended claims.

Claims

1. The system for acquiring the electrocardio ST segment information through photographing an electrocardiogram is characterized by comprising an electrocardio image reading module, an ST segment abnormality diagnosis module and an android mobile phone APP functional module, wherein the android mobile phone APP functional module comprises an image selection module and a result feedback module; the output end of the image selecting module is connected with the input end of the electrocardiographic image reading module, and the output end of the ST-segment abnormality diagnosis module is connected with the input end of the result feedback module;

the electrocardiographic image reading module is used for processing standard 12-lead electrocardiographic image data uploaded by a patient user, carrying out quick identification, positioning and extraction on electrocardiographic image parts through a pre-trained YOLOv5 target detection model on the acquired data, and transmitting the data to the ST-segment abnormality diagnosis module after processing;

the ST-segment abnormality diagnosis module predicts the electrocardio part extracted by the electrocardiograph image reading module through a pretrained convolutional neural network, and uses an ST-segment intelligent diagnosis algorithm to diagnose whether the electrocardio ST segment is raised or depressed;

the android mobile phone APP functional module comprises an image selection module and a result feedback module; the image selecting module is used for shooting through a mobile phone camera or selecting pictures through a mobile phone built-in photo album, so that a user can upload standard 12-lead electrocardiogram report images to a platform and transmit the standard 12-lead electrocardiogram report images to the electrocardiogram image reading module; the result feedback module can feed back corresponding diagnosis information to the user according to the result of the ST segment abnormality diagnosis module.

2. A system for acquiring cardiac ST segment information by taking a picture of an electrocardiogram according to claim 1, wherein the electrocardiograph image reading module comprises:

(1) The data set is processed, wherein the image training set is derived from two parts, and the label position of the electric part of the image center is obtained through LabelImg software and Pycharm software;

(2) Model training, wherein the model training is divided into three parts of feature extraction, feature fusion and prediction;

(3) And (3) extracting data, predicting and extracting an electrocardio region of the electrocardio image by using a prediction function according to an electrocardio image transmitted by a user and a trained model, and transmitting the electrocardio region to an ST-segment abnormality diagnosis module.

3. The system for acquiring the information of the electrocardio ST segment by taking a picture of an electrocardiogram according to claim 2, wherein in the data set processing, a first part of images in an image training set are electrocardiogram images taking grids as the background, and the images are superimposed in random positions of a plurality of real background pictures after data enhancement operations such as rotation, scaling and the like are carried out, so that the authenticity of image data is similar to that of the electrocardiogram taken by a mobile phone; the second partial image is an electrocardiogram report of real shooting, and the actual sample size of the data set is increased to improve the generalization capability of the model.

4. The system for acquiring the information of the ST segment of the heart by taking a picture of an electrocardiogram according to claim 2, wherein the feature extraction of the model training comprises three parts of a downsampling structure, a cross-stage local network and a spatial pyramid pooling module:

(1) In the downsampling structure, slicing the electrocardiograph before the electrocardiograph enters a backbone network, extracting information of key points, and obtaining a downsampling feature map;

(2) In the cross-stage local network, the cross-stage local network is formed by splicing a convolution layer and a residual error network, so that the network learning capacity is enhanced, the calculation bottleneck is reduced, and the learning accuracy is maintained while the weight is reduced;

(3) In the space pyramid pooling module, the space pyramid pooling module is composed of a plurality of multi-scale maximum pooling layers, so that fusion of different feature scales is realized.

5. The system for acquiring the information of the electrocardio ST segment by taking a picture of an electrocardiogram according to claim 2, wherein in a feature fusion part of model training, a feature fusion network performs further feature extraction on a feature map input by a main network in feature extraction through a cross-stage local network and a convolution layer module, processes the feature map through operations such as spatial pyramid pooling and the like, and transmits tensor data obtained by processing to a prediction layer.

6. The system for acquiring the information of the ST segment of the heart by taking a picture of an electrocardiogram according to claim 2, wherein the loss function is defined by introducing an intersection ratio in a prediction part of the model training, and the GIoU loss function is used as a loss function of an object recognition prediction frame.

7. The system for acquiring the electrocardio ST segment information by taking a picture of an electrocardiogram according to claim 2, wherein a training model algorithm trains a YOLOv5 target detection model according to the processing effect adjustment model parameters, key information of an electrocardiogram image is thinned through layer-by-layer neural network learning, and the key information is compared with label information obtained in the data processing step, and is continuously optimized and corrected to obtain a trained weight file.

8. The system for acquiring cardiac ST segment information from a photographed electrocardiogram of claim 1, wherein the ST segment abnormality diagnosis module comprises model training and result diagnosis:

(1) In model training, dividing the electrocardiogram data set into a training set, a verification set and a test set by using a random classification method, and training by using a network, wherein a training model algorithm is mainly used for testing ST segment abnormality diagnosis model performances under different learning rates and convolution kernels according to prediction accuracy, and continuously optimizing and correcting by using layer-by-layer neural network learning to obtain a trained weight file;

(2) In the result diagnosis, in the network training, a convolutional neural network model is divided into an input end, a main network and an output end, and according to an electrocardiographic region image obtained by an electrocardiographic image reading module and a trained model, the diagnosis prediction function is utilized to carry out abnormal diagnosis on the ST segment, including ST segment elevation and ST segment depression.

9. The system for acquiring cardiac ST segment information from a photographed electrocardiogram of claim 8, wherein the convolutional neural network comprises:

s1, in a main network, the main network has eight layers, wherein the first seven layers consist of a two-dimensional convolution layer, a batch processing layer, a random inactivation layer and a correction linear unit layer, and the eighth layer consists of 2 full-connection layers and 1 correction linear unit layer;

s2, in the first seven layers of networks, tensors which are output by an electrocardiographic image reading module and are preprocessed by an ST-segment abnormality diagnosis model are input, and sequentially pass through a two-dimensional convolution layer, a batch processing layer and a random inactivation layer respectively, and finally, the characteristics of electrocardiographic images are extracted for multiple times through the activation processing of a correction linear unit layer, so that the characteristics of the acquired electrocardiographic images are more accurate;

s3, in the eighth layer network, taking the data obtained by the output of the seventh layer as the input of the full-connection layer, and then obtaining the prediction result of the ST segment information through the RelU activation function processing and the full-connection layer processing.

10. The system for acquiring the electrocardio ST segment information by taking a picture of an electrocardiogram according to claim 1, wherein the image selecting module of the android mobile phone APP functional module is used for selecting 12-lead electrocardiogram report pictures by taking a picture through a mobile phone camera or a mobile phone built-in photo album, helping a user to upload the pictures to a platform and transmitting the pictures to an electrocardiograph image reading module; and a result feedback module of the android mobile phone APP functional module outputs an ST segment diagnosis result.