1. The new coronary pneumonia data processing system based on artificial intelligence technology is characterized by comprising

The lung database is provided with a plurality of lung databases of pneumonia original images and normal original images, and pneumonia sample images and normal sample images obtained through training and screening are used as data sets to form a convolutional neural network model;

a convolutional neural network that extracts information layer by layer from pixel-level raw data to abstract semantic concepts using a stacked structure to obtain a wide input view;

the image preprocessing unit comprises an image amplification module and an image contrast enhancement module, wherein the image amplification module is used for increasing the number of training sample images, the image contrast enhancement module is used for enhancing the local contrast of the sample, and the sample images with the enhanced local contrast are reduced through a linear interpolation module to obtain sample enhanced images.

2. The system of claim 1, wherein the convolutional neural network model comprises a convolutional layer for extracting feature information of the input image, a pooling layer for removing the feature information for aggregation and purification, and a full-link layer for connecting neuron nodes in a current layer with neuron nodes in a previous layer.

3. The system for processing new coronary pneumonia according to claim 1, characterized in that a feature fusion module for enhancing the judgment ability of the model feature image is added between the fourth layer and the sixth layer of the convolution layer, and the feature fusion module is set as a convolution layer with convolution kernel of 1 x 1.

4. The new coronary pneumonia data processing system based on artificial intelligence technology as claimed in claim 1, wherein said convolutional neural network is further provided with an attention mechanism module, said attention mechanism module is used for weighting the feature channels according to global information learned by high level, and then feeding back to the output feature map to further realize the screening of important features.

5. The artificial intelligence technology-based new coronary pneumonia data processing system of claim 4, wherein the input end of the attention mechanism module is the eighth layer of the convolutional layer, and the output end of the attention mechanism module is the ninth layer of the convolutional layer.

6. The artificial intelligence technology-based new coronary pneumonia data processing system of claim 1, wherein the convolutional neural network sets and determines the weight parameters of model learning through a classification optimization module during training.

7. The artificial intelligence technology-based new crown pneumonia data processing system of claim 6, wherein said classification optimization module selects a mean square error loss function, a mean absolute error loss function or a log likelihood loss function.

8. The artificial intelligence technology-based new crown pneumonia data processing system of claim 2, wherein the pooling layer polymerization purification process employs maximum pooling or mean pooling.

9. The artificial intelligence technology-based new coronary pneumonia data processing system of claim 1, wherein the convolution layers are each set to 3 x 3 in convolution kernel size.

10. The system for processing new coronary artery disease data based on artificial intelligence technology as claimed in claim 2, wherein a Dropot layer is disposed between the fully connected layers, and the coefficient of the Dropot layer is set to 0.5.

Background

According to the study report of the world health organization, pneumonia remains the leading cause of death in many fatal diseases in children under five years of age, with at least 2400 children dying from pneumonia every day as a complication. In 2017, over 80 million children die worldwide from pneumonia, accounting for 15% of all deaths below 5 years of age, with deaths exceeding the sum of malaria, aids and measles. Of particular note, the vast majority of children diagnosed with clinical pneumonia are distributed in developing countries such as india, china, etc. The diagnosis of pneumonia in children is difficult. The data acquired by the infant patient is not high in fitting degree, and the imaging content and quality of the acquired pictures are remarkably different and diversified.

The principle of the chest X-ray is that when X-rays with uniform intensity penetrate tissue structures with different densities and equal thicknesses, X-ray images with black-white contrast and level difference are displayed on an X-ray film due to different absorption degrees, however, experts with knowledge and experience are required to check the X-ray images through careful reading, but image characteristics similar to pneumonia can be displayed on the images due to other diseases such as lung cancer, lung effusion and the like, so that the process of screening pneumonia by manually reading the X-ray images is very time-consuming on one hand, and the problem of low accuracy is caused by interference caused by unobvious image characteristics on the other hand.

Disclosure of Invention

In view of the above, the present invention provides a new coronary pneumonia data processing system based on artificial intelligence technology.

In order to solve the technical problems, the invention adopts the technical scheme that: the new coronary pneumonia data processing system based on the artificial intelligence technology comprises a lung database, wherein the lung database is provided with a plurality of lung databases of pneumonia original images and normal original images, and pneumonia sample images and normal sample images obtained through training and screening are used as data sets to form a convolutional neural network model;

a convolutional neural network that extracts information layer by layer from pixel-level raw data to abstract semantic concepts using a stacked structure to obtain a wide input view;

the image preprocessing unit comprises an image amplification module and an image contrast enhancement module, the image amplification module is used for increasing the number of training sample images, the image contrast enhancement module is used for enhancing the local contrast of the samples, and the sample images with the enhanced local contrast are reduced through a linear interpolation module to obtain sample enhanced images.

In the present invention, preferably, the convolutional neural network model includes a convolutional layer for extracting feature information of an input image, a pooling layer for removing the feature information and performing aggregation and purification, and a full-link layer for implementing connection between a current layer of neuron nodes and a previous layer of neuron nodes.

In the present invention, preferably, a feature fusion module for enhancing the judgment capability of the model feature image is added between the fourth layer and the sixth layer of the convolutional layer, and the feature fusion module is set as a convolutional layer with a convolution kernel of 1 × 1.

In the present invention, preferably, the convolutional neural network is further provided with an attention mechanism module, and the attention mechanism module is configured to weight the feature channels according to global information learned by a high layer, and then feed back the feature channels to the output feature map to further implement screening of important features.

In the present invention, preferably, the input end of the attention mechanism module is the eighth layer of the convolutional layer, and the output end of the attention mechanism module is the ninth layer of the convolutional layer.

In the invention, preferably, the convolutional neural network sets and determines the weight parameters of model learning through a classification optimization module in a training process.

In the present invention, preferably, the classification optimization module selects a mean square error loss function, a mean absolute error loss function, or a log-likelihood loss function.

In the present invention, preferably, the pooling layer polymerization purification process employs maximum pooling or uniform pooling.

In the present invention, it is preferable that the convolution kernel sizes of the convolution layers are each set to 3 × 3.

In the present invention, preferably, a Dropout layer is disposed between the full connection layers, and a coefficient of the Dropout layer is set to 0.5.

The invention has the advantages and positive effects that: the image preprocessing unit comprises an image amplification module and an image contrast enhancement module, the image amplification module can overcome the fitting transition problem caused by the condition of small data sample amount, the image contrast enhancement module enables the sample image after local contrast enhancement to greatly improve the contrast of an inflammation area and a surrounding tissue structure, the local contrast can be improved on the premise of not influencing the overall contrast, further the key features in the image are highlighted, and the reliability of key feature classification is promoted subsequently; the feature fusion module is arranged between the two convolution layers, feature channel amplification is carried out on the feature map of the front layer after feature dimension reduction, and then the feature channel amplification is carried out on the feature map of the front layer and the feature map of the rear layer to obtain a fused feature map, so that fine-grained feature information of an X-ray image is effectively extracted.

Drawings

The accompanying drawings, which are included to provide a further understanding of the invention and are incorporated in and constitute a part of this specification, illustrate embodiments of the invention and together with the description serve to explain the principles of the invention and not to limit the invention. In the drawings:

FIG. 1 is a schematic diagram of the structure of the convolutional neural network of the new coronary pneumonia data processing system based on artificial intelligence technology;

FIG. 2 is a schematic diagram of the structure of the attention mechanism module of the new coronary pneumonia data processing system based on artificial intelligence technology;

FIG. 3 is a schematic diagram of a sample image before local contrast enhancement for the new coronary pneumonia data processing system based on artificial intelligence technology of the present invention;

fig. 4 is a schematic diagram of a sample image after local contrast enhancement of the new coronary pneumonia data processing system based on artificial intelligence technology 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.

It will be understood that when an element is referred to as being "secured to" another element, it can be directly on the other element or intervening elements may also be present. When an element is referred to as being "connected" to another element, it can be directly connected to the other element or intervening elements may also be present. When an element is referred to as being "disposed on" another element, it can be directly on the other element or intervening elements may also be present. The terms "vertical," "horizontal," "left," "right," and the like as used herein are for illustrative purposes only.

Unless defined otherwise, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention belongs. The terminology used in the description of the invention herein is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention. As used herein, the term "and/or" includes any and all combinations of one or more of the associated listed items.

As shown in figure 1, the invention provides a new coronary pneumonia data processing system based on artificial intelligence technology, which comprises

The lung database is provided with a plurality of lung databases of pneumonia original images and normal original images, and pneumonia sample images and normal sample images obtained through training and screening are used as data sets to form a convolutional neural network model;

a convolutional neural network that extracts information layer by layer from pixel-level raw data to abstract semantic concepts using a stacked structure to obtain a wide input view;

the image preprocessing unit comprises an image amplification module and an image contrast enhancement module, the image amplification module is used for increasing the number of training sample images, the image contrast enhancement module is used for enhancing the local contrast of the samples, and the sample images with the enhanced local contrast are reduced through a linear interpolation module to obtain sample enhanced images. The image amplification module can overcome the fitting transition problem caused by the condition that the data sample amount is small, and particularly as shown in fig. 3 and 4, the image contrast enhancement module enables the sample image after enhancing the local contrast to greatly improve the contrast of an inflammation area and a surrounding tissue structure, and can improve the local contrast on the premise of not influencing the overall contrast, so that the key features in the image are highlighted, and the subsequent reliability of key feature classification is facilitated to be improved.

In this embodiment, the convolutional neural network model further includes a convolutional layer, a pooling layer, and a full-link layer, where the convolutional layer is used to extract feature information of the input image, the pooling layer is used to remove the feature information for aggregation and purification, and the full-link layer is used to implement connection between a current layer of neuron nodes and a previous layer of neuron nodes. The convolutional layer and the pooling layer are mainly used for executing image feature extraction operation, the convolutional neural network obtains abstract high-level semantic features through stacking of the convolutional layer and the pooling layer, and finally the semantic features are input into the full-connection layer, so that mapping from original input data to a final task target is realized. The pooling layer can select the maximum value from a specific area as an output value by gradually reducing the space size represented by the features and further reducing the quantity of parameters and model calculation in the grid, the input feature graph can be pooled through the maximum value, the pooling layer can remove feature redundancy extracted by the convolutional layer, aggregate image features and reduce the dimensionality of the features, and therefore the calculation quantity of the network is greatly reduced.

In this embodiment, a feature fusion module for enhancing the determination capability of the model feature image is further added between the fourth layer and the sixth layer of the convolutional layer, and the feature fusion module is set as a convolutional layer with a convolution kernel of 1 × 1. The number of parameters of the model is not obviously increased, and the problem that the model occupies too much memory resources is effectively avoided.

Since the feature fusion module introduces noise included in the underlying features of the model, an attention mechanism module is provided to solve the above-mentioned problem.

As shown in fig. 2, in this embodiment, the convolutional neural network is further provided with an attention mechanism module, where the attention mechanism module is configured to weight the feature channels according to global information learned by a high layer, further screen out information related to a current task from a large amount of feature information, further reduce interference of irrelevant noise, and then feed back to the output feature map to further screen out important features, and since the soft attention mechanism is differentiable, it can implement updating of weight parameters in the depth model forward propagation and backward propagation processes, so the soft attention mechanism is adopted in the present scheme. By arranging the attention mechanism module, importance measurement is carried out on global features learned by high layers, so that the learning capability and the discrimination capability of the convolutional neural network on key features are improved.

In this embodiment, further, the input end of the attention mechanism module is the eighth layer of the convolutional layer, and the output end of the attention mechanism module is the ninth layer of the convolutional layer. As shown in fig. 1, in operation, the global mean value pooling layer of the output feature map tree map of the eighth layer of the convolutional layer is used to obtain a feature vector, the dimension of the feature vector is consistent with the number of input feature map channels, by this way, global spatial information can be well preserved, and the number of parameters is reduced, then the obtained feature vector is input into two fully-connected layers to display the corresponding relationship between learning feature map channels, the feature vector output by the second fully-connected layer is equivalent to the importance measure for the input features from the convolutional layer, and thus the number of neuron nodes of the fully-connected layers is consistent with the number of channels. After the maximum pooling operation is performed on the output characteristic diagram of the eighth layer of the convolutional layer, the output of the second fully-connected layer is multiplied by the output characteristic diagram after the pooling operation in a channel-by-channel weighting mode, so that the original characteristic is recalibrated in the characteristic channel dimension, and the complexity of calculation is reduced. Starting from the global characteristics of a convolutional neural network, the input characteristics firstly pass through a global mean pooling layer and a full connection layer to be used for calculator attention distribution, then the input characteristics are weighted according to the distribution, so that the model can better pay attention to the channel characteristics related to the pneumonia diagnosis task, the influence of the unrelated characteristics on the model performance is inhibited, the sensitivity of the model to the characteristic channels can be improved by adding a small amount of parameters, and the judgment capability of the characteristics is further enhanced.

In this embodiment, further, the convolutional neural network sets and determines the weight parameters learned by the model through a classification optimization module in the training process. A loss function is selected through a gradient descent algorithm, so that the optimal setting of weight parameters of model learning is determined, and the whole process of fitting data through a calculation method is realized through the change of the loss function.

In this embodiment, the classification optimization module further selects a mean square error loss function, a mean absolute error loss function, or a log-likelihood loss function. For the binary problem, since the convolutional neural network only needs to predict the results corresponding to two cases, when the model prediction probability corresponding to each class is assumed to be p and 1-p, respectively, the expression of the loss function is as follows:wherein the positive class is 1, the negative class is 0, and N representsTotal number of samples of training data, y_iLabel, p, representing sample i_iRepresenting the probability that sample i is predicted as positive, 1- ρ_iRepresenting the probability that sample i is predicted as a negative class.

In this embodiment, further, the pooling layer polymerization purification process employs maximum pooling or mean pooling. In order to reduce the size of the feature map, information is extracted layer by layer from raw data at the pixel level to abstract semantic concepts through stacked convolutional layers to achieve a wider input view without increasing the computational complexity of the model, which in turn enables the model to better aggregate image context information.

In this embodiment, further, the convolution kernel sizes of the convolutional layers are all set to 3 × 3, and setting the convolution kernel size to a smaller size can make each neuron node sufficiently small relative to the total receptive field, so as to capture local detail features related to the target task, and meanwhile, the situation that the size of the local receptive field is limited due to the fact that adverse effects caused by a large increase in parameter scale are avoided, and the sensitivity of the convolutional kernel in comparison with remote information can be improved by setting an extended convolutional layer instead of a common convolutional layer on the basis of not increasing model parameters.

In this embodiment, a Dropout layer is further disposed between the full connection layers, and a coefficient of the Dropout layer is set to 0.5. Therefore, when the training is updated every time, the training of half of the neuron nodes can be stopped randomly, so that the neuron nodes of the hidden layer are prevented from being dependent on specific input, and the model performance is improved; on the other hand, the network is subjected to iterative training in a mode of randomly extracting small batches of data from training data in the training process, a model prediction result is obtained through calculation in the forward propagation process, then a random gradient descent algorithm is combined in the backward propagation process to minimize a loss function, weight parameters are updated until the model converges, and the occurrence of overfitting of the model in the training process can be avoided.

Generally speaking, in order to obtain better classification performance, the convolutional neural network needs to extract features of an input image to obtain enough receptive fields so as to obtain higher-level semantic information, however, conventionally, increasing the convolutional layer and the convolutional kernel size would result in increasing complexity of a model space and higher requirement on a memory space, so that problems of low efficiency and overfitting are generated in a model training process.

The pooling layer is used for performing a redundancy removal process on the characteristic information.

The system has a large demand on the number of original images of a sample, but in the actual situation, enough original images of pneumonia are difficult to obtain, because the structure of the convolutional neural network is different from the structure of a common pre-training model, it is difficult to directly perform parameter migration, and on the other hand, because the traditional ImageNet data set mainly consists of color images in nature, however, the data set of the system is composed of a gray scale image, the domain correlation degree of the gray scale image and the gray scale image is greatly different, the natural image and the medical image have the condition that the apparent distribution cannot be matched, the knowledge transfer learning method adopting the ImageNet data set can not only obtain the optimal solution, but also can cause the classification capability of the model in the medical image data set to be reduced to a great extent, so that the classification capability of the convolutional neural network is required to be optimized.

The embodiments of the present invention have been described in detail, but the description is only for the preferred embodiments of the present invention and should not be construed as limiting the scope of the present invention. All such changes and modifications which come within the scope of the invention are desired to be protected by the following claims.