CN115810016B - Method, system, storage medium and terminal for automatic recognition of CXR image of lung infection - Google Patents

Abstract

The invention discloses a method, system, storage medium and terminal for automatic recognition of CXR images of lung infection, and relates to the automatic recognition of pneumonia images. A technical problem with poor automatic image recognition. By building a densely connected network, each layer of dense blocks in the densely connected network is connected to the dense blocks of all previous layers and used as the input of the next layer of dense blocks; and the connection between dense blocks and dense blocks includes juxtaposed The left connection branch, the right connection branch, the left connection branch includes two convolutional layers, and the right connection branch includes four convolutional layers. The network structure of the present invention is changed to a bidirectional dense layer and a transition layer, which can perform non-compressed operations and extract multi-scale features of pneumonia, and the deepening of the densely connected network will not cause the gradient to disappear, and deepening the network can obtain better pneumonia recognition effect .

Description

Method, system, storage medium and terminal for automatic recognition of CXR image of lung infection

technical field

The invention relates to the technical field of image segmentation in artificial intelligence medicine, and relates to automatic recognition of pneumonia images, in particular to a method, system, storage medium and terminal for automatic recognition of CXR images of lung infection.

Background technique

The rapid spread of the novel coronavirus infection (COVID-19) disease has placed unprecedented pressure on healthcare systems in many countries. For viral infections, usually manifested as upper respiratory tract infection or pneumonia, can also progress to acute respiratory failure, multi-organ failure and death. In the existing clinical diagnosis and treatment, chest X-ray (CXR) photography is still the main means of screening, classification and diagnosis of various types of pneumonia, including bacterial, viral and other types of pneumonia. Recent developments in artificial intelligence (AI) provide new potential opportunities for the rapid growth of diagnostic radiology applications. For the diagnosis of common lung and heart diseases based on CXRs, AI using weakly supervised classification or attention-based convolutional neural networks has also been investigated. Model. Therefore, through the method of machine learning, the relevant pathological image features in the image are automatically extracted, and then pneumonia is judged based on the image features, which meets the actual needs of clinicians.

Pneumonia automatic identification methods have been extensively studied in recent years. These methods can be divided into traditional pneumonia identification methods based on statistical learning and pneumonia identification methods based on deep learning. But deep learning methods outperform traditional machine learning methods in many medical image analysis tasks, including detection, classification, and segmentation.

Deep learning has made great progress in various tasks of natural image recognition. The emergence of AlexNet (2012) expanded the depth of LeNet and opened the first of its kind in deep learning for image processing. Deep neural networks have been widely used. VGGNet (2014) further improved the performance of deep learning by increasing the depth of the network and reducing the size of the convolution kernel. ResNe (2015) uses residual modules to form more complex networks, trained using standard stochastic gradient descent. This structure can reduce gradient disappearance during model training. In the test, the recognition rate of the ImageNet dataset was reduced to 3.57%. In the same year, Inception increased the depth of the network while also increasing the width of the network, using convolution kernels of different sizes to extract features, and calculating convolutions of different sizes at the same time, and then connecting their results respectively, and then outputting the outputs of these filters along the The channel dimensions are stacked and passed to the next layer to obtain different information about the input image, parallelizing these operations and combining all the results will lead to a better image representation. In a further development, researchers are interested in 3D images, which have more spatial features. The researchers propose to use a 3D convolutional neural network to apply it to 3D data to achieve progress in 3D object recognition tasks.

In the existing technology, deep learning methods are also applied to pneumonia recognition tasks: CheXNet is a 121-layer dense convolutional network trained on the Huast-ray14 dataset, which effectively improves the effect of pneumonia recognition; there is also the use of convolutional neural network Network (CNN) analysis postulates earlier diagnosis and better patient management based on signs of pneumonia.

In addition, the invention patent application with the application number 202011184333 discloses a novel coronavirus infection CT image processing method based on lung non-uniform pooling, which includes the following steps: S1, FPN-based fully automatic lung segmentation algorithm; S2, Identification of the centerline of both lungs; S3, lung pooling operation; S4, convolutional neural network based on lung pooling. Among them, in step S1, a feature pyramid full convolutional neural network based on DenseNet121 is constructed to automatically segment the lung region from the CT image. The feature pyramid full convolutional neural network FPN uses the DenseNet network with pre-trained weights in ImageNet as the basic network, and then extracts the output of the last convolutional layer from each Dense block in DenseNet in the form of a feature pyramid as a multiple Then, the features of different scales are up-sampled step by step and stitched together. Finally, the segmented lung region is obtained through the full convolutional network, that is, the lung target region ROI. The segmentation algorithm builds a feature pyramid full convolutional neural network based on DenseNet121, amplifies the features of the lung area when the convolutional neural network is pooled, compresses the area outside the lung, eliminates redundant features, and strengthens the lung area. The image information improves the accuracy of the new coronavirus infection CT image processing method, and at the same time does not rely on any manually labeled images, which improves the practicability of the algorithm.

The invention patent application with the application number 202010666696 also discloses a method and system for predicting the recovery time of patients with new coronavirus infection based on deep learning. Sequence images are preprocessed; input to the lung lobe segmentation model and pneumonia segmentation model respectively, and the multi-day lung lobe area and lesion area are respectively extracted; the multi-day lesion area is obtained by calculating the ratio of the multi-day lesion area to the lung lobe area Area ratio value; use the multi-day lesion area ratio value R to fit the Gaussian process model to predict the recovery time of patients infected with the new coronavirus. The present invention uses Densenet as the backbone DeepLabV3+ framework and 3D UNet framework to segment lung lobes and pneumonia regions. The segmentation is fast and effective, and the Gaussian process can accurately predict the patient's recovery time, which serves as a reference for medical resource allocation.

Like the above two patent documents, there is also a pneumonia image recognition network based on the DenseNet network in the prior art. However, in the prior art, the DenseNet network usually has multiple layers of dense blocks, and each layer of dense blocks will be connected with the dense blocks of all previous layers as the input of the next layer of dense blocks. And when the dense block is connected with the dense block, the connection is a single-branch structure, as shown in Figure 1, which includes two convolutional layers with a convolution kernel of 1*1 and a convolution kernel of 3*3. The step sizes of the two convolution kernels are 4k and k respectively. Although the DenseNet network with this structure can realize the automatic recognition of pneumonia images, but due to the densely connected network, the features extracted by the data compression operation are not comprehensive, so the extracted multi-scale features of pneumonia are limited, and the automatic recognition effect of pneumonia images is relatively poor. Difference.

Contents of the invention

The purpose of the present invention is: in order to solve the technical problem in the above-mentioned prior art that the densely connected network has a poor automatic recognition effect of pneumonia images due to data compression, the present invention provides a method, system and storage medium for automatic recognition of CXR images of lung infection and terminal.

The present invention specifically adopts the following technical solutions in order to achieve the above object:

A method for automatic recognition of lung infection CXR images, comprising the following steps:

Step S1, get sample data

Obtain CXR image samples and corresponding chest pathological labels;

Step S2, build a densely connected DenseNet network

Build a densely connected DenseNet network. The densely connected DenseNet network includes multiple layers of dense blocks. Each layer of dense blocks is connected to the dense blocks of all previous layers through dense connections, and the next layer of dense blocks is connected to all previous layers of dense blocks through feedback connections. And each dense block extracts the output of the last layer of convolutional layer and upsamples step by step along the direction close to the input, and input to the first layer of dense block after convolution operation;

The dense connection and feedback connection between the dense block and the dense block include a left connection branch and a right connection branch arranged in parallel, the left connection branch includes two convolutional layers, and the right connection branch includes four convolutional layers;

Step S3, training densely connected DenseNet network

Using the CXR image samples and chest pathology labels acquired in step S1 to train the densely connected DenseNet network to obtain a mature densely connected DenseNet network;

Step S4, real-time recognition

Input the CXR image acquired in real time into the mature densely connected DenseNet network obtained in step S3, and output the chest pathological recognition result.

Further, in step S1, the CXR image samples and corresponding chest pathology labels come from the ChestX-ray14 dataset, which includes 112,120 frontal X-ray images of 30,805 patients, and each X-ray image is marked with up to 14 different chest pathology labels.

Further, in step S2, the left connection branch includes two convolution layers with a convolution kernel of 1*1 and a step size of 2k, a convolution kernel of 3*3 and a step size of k/2, and the right connection branch The branch includes sequentially setting the convolution kernel to 1*1 with a step size of 2k, the convolution kernel to 3*3 with a step size of k/2, the convolution kernel to 3*3 with a step size of k/2, and the convolution Four convolutional layers with a 3*3 kernel and a stride of k/2.

Further, in step S3, when training the densely connected DenseNet network, the weighted binary cross-entropy loss function used for:

in, samples for the network input, is the binary label output by the network, ; is the probability that the network assigns to label 1, is the probability that the network assigns to label 0, , Represents the number of positive cases and negative cases of pneumonia in the sample, respectively.

Further, in step S3, when training the densely connected DenseNet network, an initial learning rate of 0.001 is used, and each time the validation loss stabilizes after an epoch, it decays by 10%, and the model with the lowest validation loss is selected; After each iteration of learning, the network judges the total error of the pneumonia recognition result. If the current error is smaller than the error of the previous iteration, the current model is saved and training continues until the maximum number of iterations is reached.

A system for automatic recognition of lung infection CXR images, comprising:

A sample data acquisition module, configured to acquire a CXR image sample and a corresponding chest pathological label;

The densely connected DenseNet network building module is used to build a densely connected DenseNet network. The densely connected DenseNet network includes multiple layers of dense blocks. Each layer of dense blocks is connected to the dense blocks of all previous layers through dense connections, and the next layer of dense blocks is connected through feedback. Connect with all previous layers of dense blocks, and each dense block extracts the output of the last layer of convolutional layer and upsamples step by step along the direction close to the input, and input to the first layer of dense block after convolution operation;

The dense connection and feedback connection between the dense block and the dense block include a left connection branch and a right connection branch arranged in parallel, the left connection branch includes two convolutional layers, and the right connection branch includes four convolutional layers;

The densely connected DenseNet network training module is used to train the densely connected DenseNet network using the CXR image samples and chest pathological labels obtained in the sample data acquisition module to obtain a mature densely connected DenseNet network;

The real-time recognition module is used to input the CXR image acquired in real time into the mature densely connected DenseNet network obtained by the densely connected DenseNet network training module, and output the chest pathological recognition result.

A storage medium stores a computer program, and when the computer program is executed by a processor, the processor executes the steps of the above method.

A terminal includes a memory and a processor, the memory stores a computer program, and when the computer program is executed by the processor, the processor executes the steps of the above method.

The beneficial effects of the present invention are as follows:

1. In the present invention, there is not only a dense connection connecting to the output direction, but also a feedback connection connecting along the input direction between the dense block and the dense block. Through the dense connection and the feedback connection, it can accept its own information to form an effective The loop is more in line with the biological neural network structure combined with the context to train the model to share parameter information; both the dense connection and the feedback connection are set to adopt a double-branch structure, which can perform uncompressed operations on the data, thereby extracting the multi-scale features of pneumonia and deepening the network A better pneumonia recognition effect can be obtained; in addition, the output of the last convolutional layer is extracted for each dense block and upsampled step by step along the direction close to the input, and the convolution operation is input to the first layer of dense block, which It can improve the resolution of the image and connect it during upsampling to fuse the underlying position information with the deep semantics, and finally input it to the first dense block, so that further training can extract more dimensional features, which is helpful for model performance. promote.

2. In the present invention, the structure based on the densely connected network is adopted, and the feature map learned by any layer can be used by other subsequent layers through the operation of the link, which makes the feature reusable, and the model becomes more compact, achieving Small models with fewer parameters can significantly save bandwidth, reduce storage overhead, perform faster calculations, and have strong anti-fitting capabilities.

3. In the present invention, the optimized weighted binary cross-entropy loss function is used to train the network model, and the target evaluation index is directly optimized, so that the model can be trained more fully.

4. In the present invention, the trained network can be used for rapid detection and identification of pneumonia, which greatly saves manpower and material resources for primary screening, and provides more accurate detection conclusions for subsequent medical staff.

Description of drawings

Figure 1 is a schematic diagram of the connection between dense blocks and dense blocks in the existing DenseNet network;

Fig. 2 is a schematic flow sheet of the present invention;

Fig. 3 is a schematic structural diagram of a densely connected DenseNet network in the present invention;

Fig. 4 is a schematic structural diagram of connections between dense blocks and dense blocks in the DenseNet network of the present invention.

Detailed ways

In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments It is a part of embodiments of the present invention, but not all embodiments. The components of the embodiments of the invention generally described and illustrated in the figures herein may be arranged and designed in a variety of different configurations.

Accordingly, the following detailed description of the embodiments of the invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely represents selected embodiments of the invention. Based on the embodiments of the present invention, all other embodiments obtained by persons of ordinary skill in the art without creative efforts fall within the protection scope of the present invention.

Example 1

This embodiment provides a method for automatic recognition of CXR images of lung infection. By building a densely connected DenseNet network and training the network model, the trained network model can realize automatic recognition of pneumonia CXR images. The identification method is shown in Figure 2, and specifically includes the following steps:

Step S1, get sample data

Obtain CXR image samples and corresponding chest pathology labels.

In step S1, the CXR image samples and corresponding chest pathology labels come from the ChestX-ray14 dataset, which includes 112,120 frontal X-ray images of 30,805 patients, and each X-ray image is marked with up to 14 different chest pathologies Label.

The image with pneumonia was labeled as a positive example as an annotated pathology, and all other images were labeled as negative examples. In the pneumonia detection task, the dataset of CXR image samples was randomly divided into training training (28744 patients, 98637 images), validation (1672 patients, 6351 images) and testing (389 patients, 420 images), There is no overlap between adjacent two groups. Before feeding the CXR image samples into the network, the CXR image samples are downscaled to 224×224 and normalized according to the mean and standard deviation of the CXR image samples in the ImageNet training set.

Step S2, build a densely connected DenseNet network

Build a densely connected DenseNet network as shown in Figure 3. The densely connected DenseNet network includes multiple layers of dense blocks. Each layer of dense blocks is connected to the dense blocks of all previous layers through dense connections, and the next layer of dense blocks is connected to all previous layers through feedback connections. Layer dense blocks are connected, and each dense block extracts the output of the last convolutional layer and upsamples step by step along the direction close to the input, and input to the first dense block after convolution operation.

And the dense connection and feedback connection between the dense block and the dense block are shown in Figure 4, which includes a left connection branch and a right connection branch arranged in parallel, the left connection branch includes two convolutional layers, and the right connection branch includes four convolutional layers. laminated.

The left connection branch includes two convolution layers with a convolution kernel of 1*1 and a step size of 2k, a convolution kernel of 3*3 and a step size of k/2, and the right connection branch includes convolution layers set in sequence The kernel is 1*1 and the step size is 2k, the convolution kernel is 3*3 and the step size is k/2, the convolution kernel is 3*3 and the step size is k/2, the convolution kernel is 3*3 and the step Four convolutional layers of length k/2.

The left path of the previous layer is connected to the filter connection through 1*1 convolutional layer and 3*3 convolutional layer in turn, and the right path of the previous layer is sequentially passed through 1*1 convolutional layer and three 3*3 convolutional layers Connect to filter connection. In addition to halving the filter of the original main branch (that is, the left connection branch), the densely connected DenseNet network also adds a new branch (that is, the right connection branch), and uses three 3x3 convolutions in the new branch , the receptive field of this branch will become the same as the 7x7 convolution; in this way, the extracted features are not only of a single scale, but can take into account both small targets and large targets; in the densely connected DenseNet network, the transition layer is used to feature The spatial resolution of the map is reduced, and the number of channels in the transition layer will be smaller than the number of channels in the previous layer: it is found that this operation will weaken the expressive ability of network features, so in this network, the number of channels in the transition layer and the previous layer is set to The same value, so that the expressiveness can be better obtained.

Step S3, training densely connected DenseNet network

The densely connected DenseNet network is trained by using the CXR image samples and chest pathology labels obtained in step S1 to obtain a mature densely connected DenseNet network.

In step S3, when training the densely connected DenseNet network, the weighted binary cross-entropy loss function used is for:

in, samples for the network input, is the binary label output by the network, ; is the probability that the network assigns to label 1, is the probability that the network assigns to label 0, , Represents the number of positive cases and negative cases of pneumonia in the sample, respectively.

When training a densely connected DenseNet network, use an initial learning rate of 0.001, and each time the verification loss stabilizes after an epoch, it will decay by 10%, and the model with the lowest verification loss will be selected; after each iteration of learning, the network judges The total error of the pneumonia recognition result. If the current error is smaller than the error of the previous iteration, save the current model and continue training until the maximum number of iterations is reached.

Step S4, real-time recognition

Input the CXR image acquired in real time into the mature densely connected DenseNet network obtained in step S3, and output the chest pathological recognition result.

Example 2

This embodiment provides an automatic recognition system for CXR images of lung infection, which specifically includes:

The sample data acquisition module is used to acquire CXR image samples and corresponding chest pathological labels.

In the sample data acquisition module, the CXR image samples and the corresponding chest pathological labels come from the ChestX-ray14 dataset, which includes 112,120 frontal X-ray images of 30,805 patients, and each X-ray image is marked with up to 14 different Chest pathology label.

The image with pneumonia was labeled as a positive example as an annotated pathology, and all other images were labeled as negative examples. In the pneumonia detection task, the dataset of CXR image samples was randomly divided into training training (28744 patients, 98637 images), validation (1672 patients, 6351 images) and testing (389 patients, 420 images), There is no overlap between adjacent two groups. Before feeding the CXR image samples into the network, the CXR image samples are downscaled to 224×224 and normalized according to the mean and standard deviation of the CXR image samples in the ImageNet training set.

The densely connected DenseNet network building module is used to build the densely connected DenseNet network shown in Figure 3. The densely connected DenseNet network includes multiple layers of dense blocks, and each layer of dense blocks is connected to the dense blocks of all previous layers through dense connections. The next The dense block of the layer is connected with all the dense blocks of the previous layer through the feedback connection, and each dense block extracts the output of the last layer of convolutional layer and upsamples step by step along the direction close to the input, and input to the first layer of dense block after the convolution operation .

And the dense connection and feedback connection between the dense block and the dense block are shown in Figure 4, which includes a left connection branch and a right connection branch arranged in parallel, the left connection branch includes two convolutional layers, and the right connection branch includes four convolutional layers. layered;

The left connection branch includes two convolution layers with a convolution kernel of 1*1 and a step size of 2k, a convolution kernel of 3*3 and a step size of k/2, and the right connection branch includes convolution layers set in sequence The kernel is 1*1 and the step size is 2k, the convolution kernel is 3*3 and the step size is k/2, the convolution kernel is 3*3 and the step size is k/2, the convolution kernel is 3*3 and the step Four convolutional layers of length k/2.

The left path of the previous layer is connected to the filter connection through 1*1 convolutional layer and 3*3 convolutional layer in turn, and the right path of the previous layer is sequentially passed through 1*1 convolutional layer and three 3*3 convolutional layers Connect to filter connection. In addition to halving the filter of the original main branch (that is, the left connection branch), the densely connected DenseNet network also adds a new branch (that is, the right connection branch), and uses three 3x3 convolutions in the new branch , the receptive field of this branch will become the same as the 7x7 convolution; in this way, the extracted features are not only of a single scale, but can take into account both small targets and large targets; in the densely connected DenseNet network, the transition layer is used to feature The spatial resolution of the map is reduced, and the number of channels in the transition layer will be smaller than the number of channels in the previous layer: it is found that this operation will weaken the expressive ability of network features, so in this network, the number of channels in the transition layer and the previous layer is set to The same value, so that the expressiveness can be better obtained.

The densely connected DenseNet network training module is used to train the densely connected DenseNet network using the CXR image samples and chest pathological labels obtained in the sample data acquisition module to obtain a mature densely connected DenseNet network.

In the densely connected DenseNet network training module, the weighted binary cross-entropy loss function used when training the densely connected DenseNet network for:

in, samples for the network input, is the binary label output by the network, ; is the probability that the network assigns to label 1, is the probability that the network assigns to label 0, , Represents the number of positive cases and negative cases of pneumonia in the sample, respectively.

When training a densely connected DenseNet network, use an initial learning rate of 0.001, and each time the verification loss stabilizes after an epoch, it will decay by 10%, and the model with the lowest verification loss will be selected; after each iteration of learning, the network judges The total error of the pneumonia recognition result. If the current error is smaller than the error of the previous iteration, save the current model and continue training until the maximum number of iterations is reached.

The real-time recognition module is used to input the CXR image acquired in real time into the mature densely connected DenseNet network obtained by the densely connected DenseNet network training module, and output the chest pathological recognition result.

Example 3

This embodiment provides a storage medium. The storage medium is a computer-readable storage medium, and a computer program is stored in the storage medium. When the computer program is executed by a processor, the processor is made to perform the above-mentioned method for automatic identification of lung infection CXR images. step.

Wherein, the computer-readable storage medium stores an interface display program, and the interface display program can be executed by at least one processor, so that the at least one processor executes the steps of the above-mentioned method for automatically identifying lung infection CXR images.

Through the description of the above embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus a necessary general-purpose hardware platform, and of course also by hardware, but in many cases the former is better implementation. Based on such an understanding, the technical solution of the present application can be embodied in the form of a software product in essence or the part that contributes to the prior art, and the computer software product is stored in a storage medium (such as ROM/RAM, disk, CD-ROM) contains several instructions to enable a terminal device (which may be a mobile phone, computer, server or network device, etc.) to execute the method for automatic recognition of lung infection CXR images described in the embodiment of the present application.

Example 4

This embodiment provides a terminal. The terminal may be a computer device. The terminal includes a memory and a processor. The memory stores a computer program. When the computer program is executed by the processor, the processor executes the above method for automatic recognition of lung infection CXR images. A step of.

Wherein, the computer equipment may be computing equipment such as a desktop computer, a notebook, a palmtop computer, and a cloud server. The computer device can perform human-computer interaction with the user through keyboard, mouse, remote controller, touch panel or voice control device.

The memory includes at least one type of readable storage medium, and the readable storage medium includes flash memory, hard disk, multimedia card, card type memory (for example, SD or D interface display memory, etc.), random access memory (RAM), Static Random Access Memory (SRAM), Read Only Memory (ROM), Electrically Erasable Programmable Read Only Memory (EEPROM), Programmable Read Only Memory (PROM), Magnetic Memory, Magnetic Disk, Optical Disk, etc. In some embodiments, the memory may be an internal storage unit of the computer device, such as a hard disk or a memory of the computer device. In some other embodiments, the memory may also be an external storage device of the computer device, such as a plug-in hard disk equipped on the computer device, a smart memory card (Smart Media Card, SMC), a secure digital (Secure Digital , SD) card, flash memory card (Flash Card), etc. Of course, the storage may also include both the internal storage unit of the computer device and its external storage device. In this embodiment, the memory is often used to store the operating system and various application software installed in the computer device, such as the program code of the method for automatic recognition of CXR images of lung infection. In addition, the memory can also be used to temporarily store various types of data that have been output or will be output.

In some embodiments, the processor may be a central processing unit (Central Processing Unit, CPU), a controller, a microcontroller, a microprocessor, or other data processing chips. The processor is typically used to control the overall operation of the computer device. In this embodiment, the processor is configured to run program codes stored in the memory or process data, for example, run program codes of the method for automatic recognition of lung infection CXR images.

Claims (7)

1. An automatic identification method for CXR images of lung infection is characterized by comprising the following steps:

step S1, obtaining sample data

Acquiring CXR image samples and corresponding chest pathology labels;

step S2, setting up a dense connection DenseNet network

Constructing a dense connection DenseNet network, wherein the dense connection DenseNet network comprises a plurality of layers of dense blocks, each layer of dense block is connected with the dense blocks of all the previous layers through dense connection, the next layer of dense block is connected with the dense blocks of all the previous layers through feedback connection, and each dense block extracts the output of the last layer of convolution layer, gradually up-samples along the direction close to the input direction, convolves the output and inputs the output to the first layer of dense block;

the dense connection and feedback connection between the dense blocks comprise a left connection branch and a right connection branch which are arranged in parallel, wherein the left connection branch comprises two convolution layers, and the right connection branch comprises four convolution layers;

step S3, training dense connection DenseNet network

Training the densely connected DenseNet network by adopting the CXR image sample and the chest pathology label obtained in the step S1 to obtain a mature densely connected DenseNet network;

step S4, identifying in real time

Inputting the CXR image obtained in real time into the mature dense connection DenseNet network obtained in the step S3, and outputting a chest pathology recognition result;

in step S3, when training the dense connection DenseNet network, the weighted binary cross entropy loss function L (x, y) used is:

wherein x is a sample input by the network, y is a binary label output by the network, and y is {0,1}; p (y=1/x) is the probability that the network is allocated to the label 1, P (y=0/x) is the probability that the network is allocated to the label 0, and |p| and |n| respectively represent the number of positive cases and negative cases of pneumonia in the sample.

2. An automatic identification method for a CXR image for pulmonary infection as defined in claim 1, wherein: in step S1, the CXR image samples and corresponding chest pathology labels are from the ChestX-ray14 dataset, comprising 112120 frontal X-ray images of 30805 patients, each labeled with up to 14 different chest pathology labels.

3. An automatic identification method for a CXR image for pulmonary infection as defined in claim 1, wherein: in step S2, the left connecting branch comprises two convolution layers with a convolution kernel of 1*1 and a step length of 2k, a convolution kernel of 3*3 and a step length of k/2, which are sequentially arranged, and the right connecting branch comprises four convolution layers with a convolution kernel of 1*1 and a step length of 2k, a convolution kernel of 3*3 and a step length of k/2, a convolution kernel of 3*3 and a step length of k/2, a convolution kernel of 3*3 and a step length of k/2, which are sequentially arranged.

4. An automatic identification method for a CXR image for pulmonary infection as defined in claim 1, wherein: in step S3, when training a dense connection DenseNet network, using an initial learning rate of 0.001, each time the verification loss becomes stable after an epoch, it decays by 10%, and selecting the model with the lowest verification loss; after each iteration learning, the network judges the total error of the pneumonia recognition result, if the current error is smaller than the error of the previous iteration, the current model is saved, and then training is continued until the maximum iteration times are reached.

5. An automated lung infection, CXR, image recognition system comprising:

the sample data acquisition module is used for acquiring CXR image samples and corresponding chest pathology labels;

the dense connection DenseNet network building module is used for building a dense connection DenseNet network, the dense connection DenseNet network comprises a plurality of layers of dense blocks, each layer of dense block is connected with the dense blocks of all the previous layers through dense connection, the next layer of dense block is connected with the dense blocks of all the previous layers through feedback connection, and each dense block extracts the output of the last layer of convolution layer, gradually upsamples along the direction close to the input direction, and inputs the output to the first layer of dense block after convolution operation;

the dense connection and feedback connection between the dense blocks comprise a left connection branch and a right connection branch which are arranged in parallel, wherein the left connection branch comprises two convolution layers, and the right connection branch comprises four convolution layers;

the dense connection DenseNet network training module is used for training the dense connection DenseNet network by adopting the CXR image sample and the chest pathology label acquired by the sample data acquisition module to obtain a mature dense connection DenseNet network;

the real-time identification module is used for inputting CXR images acquired in real time into the mature dense connection DenseNet network obtained by the dense connection DenseNet network training module and outputting chest pathology identification results;

the DenseNet network comprises a plurality of layers of dense blocks, and when training dense connection of the DenseNet network, a weighted binary cross entropy loss function L (x, y) is adopted as follows:

wherein x is a sample input by the network, y is a binary label output by the network, and y is {0,1}; p (y=1/x) is the probability that the network is allocated to the label 1, P (y=0/x) is the probability that the network is allocated to the label 0, and |p| and |n| respectively represent the number of positive cases and negative cases of pneumonia in the sample.

6. A storage medium, characterized by: a computer program is stored which, when executed by a processor, causes the processor to perform the steps of the method according to any one of claims 1 to 4.

7. A terminal, characterized by: comprising a memory and a processor, the memory storing a computer program which, when executed by the processor, causes the processor to perform the steps of the method according to any one of claims 1 to 4.

Images