CN113344022A - Chest radiography detection method based on deep learning - Google Patents

Abstract

The invention belongs to the technical field of image processing, and particularly relates to a chest radiography detection method based on deep learning, which comprises the following steps: the method comprises the following steps of data acquisition, data preprocessing, data set division, model construction, model training and evaluation, wherein five images of pulmonary opacity, fracture, pneumothorax, pleural effusion and normal X-ray chest radiograph are collected and labeled by the data acquisition; the data preprocessing is used for cutting, scaling and normalizing the data, and converting the label into an One-Hot form; the data set division divides the data set into 5 data sets by using a K-fold cross validation method; the model construction comprises a CNN layer, an X-D layer, a full convolution layer and a full connection layer 4 to form an X-ray chest film identification depth network model; the model training and evaluation uses a K-fold intersection method to train the model, an optimal parameter model is selected according to the model identification effect, the training is stopped when the model loss value is not reduced, and the model is stored and evaluated.

Description

Chest radiography detection method based on deep learning

Technical Field

The invention belongs to the technical field of image processing, and particularly relates to a chest radiography detection method based on deep learning.

Background

At present, X-ray chest radiographs account for 40% of all X-ray diagnostic images in the world, but due to the lack of medical personnel with diagnostic qualification, the X-ray chest radiographs often have overstock conditions, patients cannot be diagnosed at the first time, the optimal treatment time is missed, the diagnosis of the X-ray chest radiographs mainly depends on manual diagnosis of doctors, and due to limited manpower, the diagnosis time is usually long, and the patients cannot be treated at the fastest speed.

Cause of problems or defects: the current X-ray chest radiography intelligent identification technology mostly depends on manual feature selection, the feature selection is difficult, and because algorithm designers have limited understanding on medical images and professional doctors have insufficient understanding on computer identification algorithms, manual feature engineering often cannot contain all effective features, so that the identification effect is poor.

Disclosure of Invention

Aiming at the problems of poor model identification effect and the like of the chest radiography identification processing technology, the invention provides a chest radiography detection method based on deep learning.

In order to solve the technical problems, the invention adopts the technical scheme that:

a chest radiography detection method based on deep learning comprises the following steps:

s100, data acquisition: five images of turbid lung, fracture, pneumothorax, pleural effusion and normal X-ray chest radiograph are collected and marked by a professional doctor;

s200, data preprocessing: cutting and scaling the data, carrying out normalization processing, converting the label into a One-Hot form, and constructing a standard data set for training and identifying a deep learning network;

s300, data set division: dividing a data set into 5 data sets by using a K-fold cross validation method;

s400, model construction: constructing an X-ray chest radiography recognition depth network model, wherein the model consists of a CNN layer, an X-D layer, a full convolution layer and a full connection layer 4 part, the former three-layer network carries out feature extraction of different receptive fields and different modes on features, and the full connection layer outputs classification results;

s500, model training and evaluation: specifying training parameters, training the model by using a K-fold intersection method, selecting an optimal parameter model according to the model identification effect, stopping training when the model loss value is not reduced, and storing and evaluating the model.

In the S100 data acquisition, focus X-ray image data is acquired from a medical institution, and a professional physician labels the data, where the data set includes 1600 focus X-ray images, 400X-ray images for 4 types of diseases, 600 normal chest X-ray images, and the labels of pulmonary opacity, fracture, pneumothorax, pleural effusion, and normal X-ray images are 1-5 types respectively.

In the S200 data preprocessing, the data is all scaled to an image of 600 × 600, the normalization processing is performed by dividing all the pixels by 255, and normalizing all the data to be in the range of [0, 1], and the One-Hot labels are pulmonary turbidity [1,0,0,0,0], fracture [0,1,0,0,0, 0], pneumothorax [0,0,1,0,0], pleural effusion [0,0,0,1,0], and normal [0,0,0, 1], respectively.

In the S300 data set division, a K-fold cross validation method is adopted for data sets to be divided into a training set and a validation set, where K is 5, and all data is averagely divided into 5 data sets, and the number of the data sets is data set a/b/c/d/e.

In the S400 model construction, a model is constructed based on Xconcept and is improved, the model consists of 4 parts, namely a CNN layer, an X-D layer, a full convolution layer and a full connection layer, the CNN layer is used for reducing feature map, improving data dimension and mining data characteristics; the X-D layer is composed of an Xception network with an expansion convolution kernel, the data characteristics are learned by utilizing depth separable convolution, and the characteristics under different receptive fields are analyzed by utilizing the expansion convolution; the full convolution layer performs dimensionality reduction analysis on the features by using a full convolution network; the full connectivity layer is used to complete the classification task.

In the S400 model construction, the CNN layer, the X-D layer, the full convolution layer and the full connection layer respectively have the following working modes:

CNN layer: performing convolution operation for three times, performing maximal pooling after each convolution operation, and activating by using a ReLU, wherein the convolution kernel size of the first convolution is 5 x 5, the step size is 2, the pooling kernel size is 4 x 4, and the step size is 2; the convolution kernel size of the second convolution is.3 x 3, the step size is 2, the pooling kernel size is 3 x 3, and the step size is 1; the convolution kernel size of the third convolution is.3 x 3, the step size is 1, the pooling kernel size is 2 x 2, and the step size is 1; performing dropout operation with the coefficient of 0.3 after pooling is completed each time to prevent overfitting, performing Batch Normalization on the obtained feature map after three times of convolution is completed to process data and prevent gradient from disappearing, and respectively increasing the data dimensionality to 4-dimension, 8-dimension and 16-dimension by 3 times of convolution;

X-D layer: after the CNN layer outputs a feature extraction result, firstly performing 1 × 1 convolution on input features by X-D, then respectively performing 3 × 3 convolution operation on each channel, performing an expansion convolution mode on 50% convolution kernels of the 3 × 3 convolution, wherein the expansion convolution scale is 1, after the convolution operation is completed, performing concatement on all outputs and the output of the CNN layer to obtain high-dimensional features, and performing Batch Normalization on the features obtained by the concatement;

full rolling layers: performing scatter on feature maps output by the X-D layers, performing full convolution operation on data features by adopting a 1X 1 convolution kernel, performing dimensionality reduction on data, and performing 3-layer 1X 1 convolution to reduce dimensionality of the data to 1;

full connection layer: and performing full-connection operation on the output characteristics of the full convolution layer, and outputting the calculation result by using softmax to obtain a final classification result.

In the S400 model construction, after the network is built, training is carried out on network parameters by using training set data, BGD is used as an optimizer, the initial learning rate is 0.02, the learning rate of each 100 epochs is attenuated by 50%, the size of batch size is 32, the loss function uses a cross entropy loss function, namely J- [ y, log (p) + (1-y), log (1-p) ], wherein y is a sample label, p is the probability of a prediction sample, 500 epochs are set for training, the training is stopped when the loss values of 15 continuous epochs are not reduced, and the model is stored.

In the S500 model training and evaluation, an a/b/c/d/e data set is taken as a verification set, the rest four data sets are taken as training sets to train the model, 5 data models are obtained, the prediction results of the 5 models on the verification set are evaluated and compared, if the model performances are close, the model is proved to have no over-fitting or under-fitting phenomenon, the model is stored, the model building is completed, if the 5 models have larger performance difference, the K-fold cross-validation division data set is carried out again, and the model is trained again by adjusting the learning rate until the optimal model is obtained.

In the S500 model training and evaluation, the trained model is used for carrying out chest radiography classification prediction on the data of the test set, the prediction result is compared with the corresponding label, and the identification effect evaluation is carried out, wherein the evaluation mode is F1-Score, the higher the F1-Score value is, the better the identification effect is,

wherein F1 is F1-score, A is accuracy, R is recall, TP is positive class number, FP is negative class number, FN is positive class number, and TN is negative class number.

Compared with the prior art, the invention has the following beneficial effects:

according to the invention, the X-ray chest radiography is intelligently identified through a deep learning method, characteristics are not required to be manually selected in the whole process, and the X-ray chest radiography is independently learned through a network, so that subjective errors are avoided, the identification speed is extremely high, the problem of low X-ray picture identification speed is effectively solved, and the model has a good identification effect based on the excellent performance of deep learning in picture identification.

Drawings

FIG. 1 is a flow chart of the main steps of the present invention;

FIG. 2 is a diagram of a network model of the present invention;

Detailed Description

The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the drawings in the embodiments of the present invention, and it is obvious that the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. All other embodiments, which can be derived by a person skilled in the art from the embodiments given herein without making any creative effort, shall fall within the protection scope of the present invention.

A chest radiography detecting method based on deep learning is disclosed, as shown in fig. 1, and includes the following steps:

s100, data acquisition: five images of turbid lung, fracture, pneumothorax, pleural effusion and normal X-ray chest radiograph are collected and marked by a professional doctor;

s200, data preprocessing: cutting and scaling the data, carrying out normalization processing, converting the label into a One-Hot form, and constructing a standard data set for training and identifying a deep learning network;

s300, data set division: dividing a data set into 5 data sets by using a K-fold cross validation method;

s400, model construction: constructing an X-ray chest radiography recognition depth network model, wherein the model consists of a CNN layer, an X-D layer, a full convolution layer and a full connection layer 4 part, the former three-layer network carries out feature extraction of different receptive fields and different modes on features, and the full connection layer outputs classification results;

s500, model training and evaluation: specifying training parameters, training the model by using a K-fold intersection method, selecting an optimal parameter model according to the model identification effect, stopping training when the model loss value is not reduced, and storing and evaluating the model.

Further, in the step S100, data acquisition is performed, focus X-ray image data is acquired from a medical institution, and a professional physician labels the data, wherein the data set comprises 1600 focus X-ray images, 400X-ray images of 4 types of diseases and 600X-ray images of a normal chest, and labels of pulmonary opacity, fracture, pneumothorax, pleural effusion and normal X-ray images are respectively 1-5 types.

Further, in the data preprocessing of step S200, the data is all scaled to an image of 600 × 600 size, the normalization processing is performed by dividing all the pixels by 255, and normalizing all the data to the range of [0, 1], and the One-Hot labels are pulmonary opacity [1,0,0,0,0], fracture [0,1,0,0,0], pneumothorax [0,0,1,0,0, 0], pleural effusion [0,0,0, 1], and normal [0,0,0,0, 0,1], respectively.

Further, in the step S300, in the data set division, the data set is divided into a training set and a verification set by using a K-fold cross-validation method, where K is 5, and all data are averagely divided into 5 data sets, which are numbered as data sets a/b/c/d/e.

Further, in the step S400 of constructing the model, as shown in fig. 2, the model is constructed based on Xception, and is improved, where the model is composed of 4 parts, which are a CNN layer, an X-D layer, a full convolution layer, and a full connection layer, respectively, and the CNN layer is used to reduce feature map, improve data dimension, and mine data features; the X-D layer is composed of an Xception network with an expansion convolution kernel, the data characteristics are learned by utilizing depth separable convolution, and the characteristics under different receptive fields are analyzed by utilizing the expansion convolution; the full convolution layer performs dimensionality reduction analysis on the features by using a full convolution network; the full connectivity layer is used to complete the classification task.

Further, in the step S400 of constructing the model, the CNN layer, the X-D layer, the full convolutional layer, and the full link layer respectively have the following working modes:

CNN layer: performing convolution operation for three times, performing maximal pooling after each convolution operation, and activating by using a ReLU, wherein the convolution kernel size of the first convolution is 5 x 5, the step size is 2, the pooling kernel size is 4 x 4, and the step size is 2; the convolution kernel size of the second convolution is.3 x 3, the step size is 2, the pooling kernel size is 3 x 3, and the step size is 1; the convolution kernel size of the third convolution is.3 x 3, the step size is 1, the pooling kernel size is 2 x 2, and the step size is 1; performing dropout operation with the coefficient of 0.3 after pooling is completed each time to prevent overfitting, performing Batch Normalization on the obtained feature map after three times of convolution is completed to process data and prevent gradient from disappearing, and respectively increasing the data dimensionality to 4-dimension, 8-dimension and 16-dimension by 3 times of convolution;

X-D layer: after the CNN layer outputs a feature extraction result, firstly performing 1 × 1 convolution on input features by X-D, then respectively performing 3 × 3 convolution operation on each channel, performing an expansion convolution mode on 50% convolution kernels of the 3 × 3 convolution, wherein the expansion convolution scale is 1, after the convolution operation is completed, performing concatement on all outputs and the output of the CNN layer to obtain high-dimensional features, and performing Batch Normalization on the features obtained by the concatement;

full rolling layers: performing scatter on feature maps output by the X-D layers, performing full convolution operation on data features by adopting a 1X 1 convolution kernel, performing dimensionality reduction on data, and performing 3-layer 1X 1 convolution to reduce dimensionality of the data to 1;

full connection layer: and performing full-connection operation on the output characteristics of the full convolution layer, and outputting the calculation result by using softmax to obtain a final classification result.

Further, in the step S400 of model construction, after the network is built, training the network parameters by using training set data, using BGD as an optimizer, where the initial learning rate is 0.02, the learning rate of each 100 epochs attenuates by 50%, the size of the batchsize is 32, and the loss function uses a cross entropy loss function, i.e., J ═ y · log (p) + (1-y) · log (1-p) ], where y is a sample label and p is the probability of predicting samples, setting 500 epochs for training, stopping training if 15 continuous epoch model loss values are not decreased, and storing the model.

Further, in the step S500 of model training and evaluation, the a/b/c/d/e data set is taken as a verification set, the rest four data sets are taken as training sets to train the model, 5 data models are obtained, the prediction results of the 5 models on the verification set are evaluated and compared, if the model performances are close, the model is proved to have no over-fitting or under-fitting phenomenon, the model is stored, the model building is completed, if the 5 model performances are large in difference, the K-fold cross-validation division data set is carried out again, the learning rate is adjusted, and the model is trained again until the best model is obtained.

Further, in the step S500 of model training and evaluation, the trained model is used for chest radiography classification prediction of the test set data, the prediction result is compared with the corresponding label, and the identification effect evaluation is carried out, wherein the evaluation mode is F1-Score, the higher the F1-Score value is, the better the identification effect is,

wherein F1 is F1-score, A is accuracy, R is recall, TP is positive class number, FP is negative class number, FN is positive class number, and TN is negative class number.

Although only the preferred embodiments of the present invention have been described in detail, the present invention is not limited to the above embodiments, and various changes can be made without departing from the spirit of the present invention within the knowledge of those skilled in the art, and all changes are encompassed in the scope of the present invention.

Claims (9)

1. A chest radiography detection method based on deep learning is characterized in that: comprises the following steps:

s100, data acquisition: five images of turbid lung, fracture, pneumothorax, pleural effusion and normal X-ray chest radiograph are collected and marked by a professional doctor;

s200, data preprocessing: cutting and scaling the data, carrying out normalization processing, converting the label into a One-Hot form, and constructing a standard data set for training and identifying a deep learning network;

s300, data set division: dividing a data set into 5 data sets by using a K-fold cross validation method;

s400, model construction: constructing an X-ray chest radiography recognition depth network model, wherein the model consists of a CNN layer, an X-D layer, a full convolution layer and a full connection layer 4 part, the former three-layer network carries out feature extraction of different receptive fields and different modes on features, and the full connection layer outputs classification results;

s500, model training and evaluation: specifying training parameters, training the model by using a K-fold intersection method, selecting an optimal parameter model according to the model identification effect, stopping training when the model loss value is not reduced, and storing and evaluating the model.

2. The chest radiography detection method based on deep learning according to claim 1, wherein: in the S100 data acquisition, focus X-ray image data is acquired from a medical institution, and a professional physician labels the data, where the data set includes 1600 focus X-ray images, 400X-ray images for 4 types of diseases, 600 normal chest X-ray images, and the labels of pulmonary opacity, fracture, pneumothorax, pleural effusion, and normal X-ray images are 1-5 types respectively.

3. The chest radiography detection method based on deep learning according to claim 2, wherein: in the S200 data preprocessing, the data is all scaled to an image of 600 × 600, the normalization processing is performed by dividing all the pixels by 255, and normalizing all the data to be in the range of [0, 1], and the One-Hot labels are pulmonary turbidity [1,0,0,0,0], fracture [0,1,0,0,0, 0], pneumothorax [0,0,1,0,0], pleural effusion [0,0,0,1,0], and normal [0,0,0, 1], respectively.

4. The chest radiography detection method based on deep learning according to claim 3, wherein: in the S300 data set division, a K-fold cross validation method is adopted for data sets to be divided into a training set and a validation set, where K is 5, and all data is averagely divided into 5 data sets, and the number of the data sets is data set a/b/c/d/e.

5. The chest radiography detection method based on deep learning according to claim 4, wherein: in the S400 model construction, a model is constructed based on Xconcept and is improved, the model consists of 4 parts, namely a CNN layer, an X-D layer, a full convolution layer and a full connection layer, the CNN layer is used for reducing feature map, improving data dimension and mining data characteristics; the X-D layer is composed of an Xception network with an expansion convolution kernel, the data characteristics are learned by utilizing depth separable convolution, and the characteristics under different receptive fields are analyzed by utilizing the expansion convolution; the full convolution layer performs dimensionality reduction analysis on the features by using a full convolution network; the full connectivity layer is used to complete the classification task.

6. The chest radiography detection method based on deep learning according to claim 5, wherein: in the S400 model construction, the CNN layer, the X-D layer, the full convolution layer and the full connection layer respectively have the following working modes:

CNN layer: performing convolution operation for three times, performing maximal pooling after each convolution operation, and activating by using a ReLU, wherein the convolution kernel size of the first convolution is 5 x 5, the step size is 2, the pooling kernel size is 4 x 4, and the step size is 2; the convolution kernel size of the second convolution is.3 x 3, the step size is 2, the pooling kernel size is 3 x 3, and the step size is 1; the convolution kernel size of the third convolution is.3 x 3, the step size is 1, the pooling kernel size is 2 x 2, and the step size is 1; performing dropout operation with the coefficient of 0.3 after pooling is completed each time to prevent overfitting, performing Batch Normalization on the obtained feature map after three times of convolution is completed to process data and prevent gradient from disappearing, and respectively increasing the data dimensionality to 4-dimension, 8-dimension and 16-dimension by 3 times of convolution;

X-D layer: after the CNN layer outputs a feature extraction result, firstly performing 1 × 1 convolution on input features by X-D, then respectively performing 3 × 3 convolution operation on each channel, performing an expansion convolution mode on 50% convolution kernels of the 3 × 3 convolution, wherein the expansion convolution scale is 1, after the convolution operation is completed, performing concatement on all outputs and the output of the CNN layer to obtain high-dimensional features, and performing Batch Normalization on the features obtained by the concatement;

full rolling layers: performing scatter on feature maps output by the X-D layers, performing full convolution operation on data features by adopting a 1X 1 convolution kernel, performing dimensionality reduction on data, and performing 3-layer 1X 1 convolution to reduce dimensionality of the data to 1;

full connection layer: and performing full-connection operation on the output characteristics of the full convolution layer, and outputting the calculation result by using softmax to obtain a final classification result.

7. The chest radiography detection method based on deep learning according to claim 6, wherein: in the S400 model construction, after the network is built, training is carried out on network parameters by using training set data, BGD is used as an optimizer, the initial learning rate is 0.02, the learning rate of each 100 epochs is attenuated by 50%, the size of batch size is 32, the loss function uses a cross entropy loss function, namely J- [ y, log (p) + (1-y), log (1-p) ], wherein y is a sample label, p is the probability of a prediction sample, 500 epochs are set for training, the training is stopped when the loss values of 15 continuous epochs are not reduced, and the model is stored.

8. The chest radiography detection method based on deep learning according to claim 7, wherein: in the S500 model training and evaluation, an a/b/c/d/e data set is taken as a verification set, the rest four data sets are taken as training sets to train the model, 5 data models are obtained, the prediction results of the 5 models on the verification set are evaluated and compared, if the model performances are close, the model is proved to have no over-fitting or under-fitting phenomenon, the model is stored, the model building is completed, if the 5 models have larger performance difference, the K-fold cross-validation division data set is carried out again, and the model is trained again by adjusting the learning rate until the optimal model is obtained.

9. The chest radiography detection method based on deep learning of claim 8, wherein: in the S500 model training and evaluation, the trained model is used for carrying out chest radiography classification prediction on the data of the test set, the prediction result is compared with the corresponding label, and the identification effect evaluation is carried out, wherein the evaluation mode is F1-Score, the higher the F1-Score value is, the better the identification effect is,

wherein F1 is F1-score, A is accuracy, R is recall, TP is positive class number, FP is negative class number, FN is positive class number, and TN is negative class number.

Images