System for classifying causes of pneumonia in children based on deep convolutional neural networks

Abstract

The present application relates to the technical field of training model, in particular to a children pneumonia cause classification system based on deep convolutional neural network, which comprises a data acquisition module for acquiring images and features in the training model; an adult contrast analysis module obtains contrast importance by the performance deviation of corresponding features of children and adult pneumonia images on different causes, and the performance deviation of feature correlation; a clinical influence analysis module obtains clinical reference by the association of cause and clinical index combined with the feature correlation; a model adjustment training module adjusts the weight of training classification according to the contribution degree of contrast importance and clinical reference representation in the significant situation of different causes to obtain the model. The present application extracts the features of children and adults by contrast, combines the auxiliary classification features of cause classification with clinical association, reduces the classification error caused by different growth stages, and improves the accuracy and robustness of the children pneumonia cause classification model.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of training models, in particular to a child pneumonia etiology typing system based on a deep convolutional neural network. BACKGROUND

[0002] Child pneumonia is an important disease that causes child death and health burden, and its etiology is complex, mainly including bacterial, viral and mycoplasma infection types, and accurate and efficient pneumonia etiology typing is crucial for developing a reasonable treatment plan. In recent years, deep convolutional neural networks can realize feature extraction, lesion area identification and automatic classification in medical image analysis, and a child pneumonia etiology typing system based on a deep convolutional neural network can use lung medical images to realize efficient and accurate automatic classification and improve the recognition ability of different pathogen types of child pneumonia, providing intelligent assistance for clinical decision-making.

[0003] In the prior art, the medical image data of child pneumonia is relatively small, and since high-quality pneumonia etiology typing data sets annotated by experts are rare, deep learning models need large-scale data for training, and usually use models trained on large-scale medical image data for transfer learning to realize pneumonia etiology typing.

[0004] However, the immune system of children is in the development stage, and its physiological response mechanism to pneumonia pathogens is different from that of adults, resulting in different imaging features of lung imaging in lesion manifestation, inflammatory diffusion pattern and tissue damage degree. Directly migrating a deep learning model trained on general pneumonia data may not accurately capture the key lesion area and microstructure changes of child pneumonia due to the bias of the feature extraction layer to adult pathological features, resulting in decreased model generalization ability and affecting the typing accuracy. SUMMARY

[0005] To solve the technical problem that directly migrating a deep learning model trained on general pneumonia data may not accurately capture the key lesion area and microstructure changes of child pneumonia due to the bias of the feature extraction layer to adult pathological features, resulting in decreased model generalization ability and affecting the typing accuracy in the prior art, the purpose of the present application is to provide a child pneumonia etiology typing system based on a deep convolutional neural network, and the technical solution adopted is as follows:

[0006] The present application provides a child pneumonia etiology typing system based on a deep convolutional neural network, which comprises:

[0007] A data acquisition module is used to acquire pneumonia images and various clinical indicators of adults and children under different etiologies of pneumonia, respectively; and a convolutional layer of a pre-training model is used to acquire features in the pneumonia images;

[0008] The adult contrast analysis module is used for analyzing the deviation between each feature in each cause and the performance degree of each feature in the pneumonia image of children and adults under each cause, obtaining the age discrimination degree of each feature in each cause, and analyzing the deviation between each feature and the correlation between other features in the pneumonia image of children and adults under each cause, obtaining the distribution specificity of each feature in each cause; combining the age discrimination degree and the distribution specificity, the contrast importance of each feature in each cause is obtained;

[0009] The clinical influence analysis module is used for screening reference clinical indicators of each cause according to the correlation between different causes and various clinical indicators under the pneumonia image of children, and obtaining the clinical reference of each feature in each cause through the correlation between each feature and the reference clinical indicators, combining the correlation between the cause and the reference clinical indicators.

[0010] The model adjustment training module is used for obtaining the contribution degree of each feature in each cause by combining the clinical reference and the contrast importance, obtaining the typing contribution degree of each feature through the distribution significance of the contribution degree of each feature in different causes, and adjusting the weight of the feature in the training model according to the typing contribution degree and training to obtain the cause typing network model.

[0011] Further, the age discrimination degree acquisition method comprises:

[0012] For any feature of any cause, the performance typicality of the feature in the cause in the pneumonia image of children is obtained according to the performance of the feature in the pneumonia image of children in the cause; similarly, the performance typicality of the feature in the cause in the pneumonia image of adults is obtained.

[0013] The difference between the performance typicality of the feature in the cause in the pneumonia image of children and the pneumonia image of adults is calculated and normalized to obtain the performance deviation degree.

[0014] The product of the performance deviation degree and the performance typicality of the feature in the cause in the pneumonia image of children is taken as the age discrimination degree of the feature in the cause.

[0015] Further, the performance typicality acquisition method comprises:

[0016] The number of occurrences of the feature in the pneumonia image of children in the cause is taken as the occurrence frequency of the feature in the cause; the mean of the occurrence frequencies of all features in the pneumonia image of children in the cause is taken as the distribution feature frequency of the cause; and the mean of the occurrence frequencies of the feature in the pneumonia image of children in all causes is calculated to obtain the distribution cause frequency of the feature.

[0017] The ratio of the frequency of occurrence of the feature to the frequency of distribution of the cause is taken as the first representation index of the feature to the cause, and the ratio of the frequency of occurrence of the feature to the frequency of distribution of the cause is taken as the second representation index of the feature to the cause.

[0018] The first representation index and the second representation index are combined to obtain the typicality of the feature in the cause in the image of child pneumonia.

[0019] Further, the distribution specificity acquisition method comprises:

[0020] Under any cause, each feature is sequentially taken as an analysis feature; for any other feature of the analysis feature, the association mode index of the analysis feature and the other feature is obtained by analyzing the distribution association between the analysis feature and the other feature in the distribution frequency and the distribution position in the child pneumonia image of the cause;

[0021] When the association mode index of the analysis feature and the other feature is greater than a preset association threshold, the corresponding other feature is taken as an associated feature; similarly, the association mode index of the analysis feature and the associated feature is obtained in the adult pneumonia image of the cause.

[0022] After calculating the difference between the association mode index of the analysis feature and the associated feature in the child pneumonia image and the adult pneumonia image, the mean value of the corresponding differences between the analysis feature and all associated features is obtained to obtain the distribution specificity of the analysis feature in the cause.

[0023] Further, the association mode index acquisition method comprises:

[0024] The sum of the number of occurrences of the other feature in the child pneumonia image of the cause and the number of occurrences of the analysis feature in the child pneumonia image of the cause is taken as the frequency sum of the analysis feature and the other feature; the number of occurrences of the analysis feature and the other feature in the child pneumonia image of the cause is taken as the numerator, and the frequency sum is taken as the denominator to obtain the synchronous distribution feature index of the analysis feature and the other feature.

[0025] Under the cause, when the analysis feature and the other feature appear simultaneously in the child pneumonia image, the distance between the center points of the corresponding regions of the analysis feature and the other feature in the single pneumonia image is calculated, the mean value of all distances is obtained and is negatively correlated to obtain the distribution distance feature index of the analysis feature and the other feature.

[0026] In the child pneumonia image, the product of the synchronous distribution feature index and the distribution distance feature index of the analysis feature and the other feature is taken as the association mode index of the analysis feature and the other feature.

[0027] Further, the method for obtaining the contrast importance comprises:

[0028] The product of the age discrimination and the distribution specificity of each feature in each cause is taken as the contrast importance of each feature in each cause.

[0029] Further, the method for obtaining the reference clinical indicator comprises:

[0030] The p value of each clinical indicator and cause is obtained by using the chi-square test on the clinical indicators and different causes; the clinical indicator with a p value less than or equal to a preset significance level is taken as the reference clinical indicator of the cause.

[0031] Further, the method for obtaining the clinical reference comprises:

[0032] For any feature and any cause, the product of the correlation degree of the feature and each reference clinical indicator of the cause and the correlation degree of the cause and the reference clinical indicator is taken as the correlation connection indicator of the feature and each reference clinical indicator of the cause.

[0033] The sum of the correlation connection indicators of the feature and all reference clinical indicators of the cause is taken as the numerator, and the sum of the correlation degrees of the cause and all reference clinical indicators is taken as the denominator to obtain the clinical reference of the feature in the cause.

[0034] Further, the method for obtaining the contribution degree comprises:

[0035] The product of the clinical reference and the contrast importance of each feature in each cause is taken as the contribution degree of each feature in each cause.

[0036] Further, the method for obtaining the typing contribution degree comprises:

[0037] For any feature, the cause in which the feature is located is sequentially taken as an analysis cause; the average of the contribution degrees of the feature in all causes except the analysis cause is taken as the separation significance of the feature in the analysis cause.

[0038] After the difference between the contribution degree and the separation significance of the feature in the analysis cause is calculated, the product of the difference and the contribution degree is taken as the significant contribution degree of the feature and the analysis cause.

[0039] The maximum of the significant contribution degrees of the feature and all causes is taken as the typing contribution degree of the feature.

[0040] The present application has the following beneficial effects:

[0041] The application first compares and analyzes the image data of pneumonia of children and adults, reflects the degree of contrast of features under each cause by the performance deviation of features on different causes and the performance deviation of feature association in the training model, optimizes the classification model of deep learning, makes it more suitable for the expression of children's influence, and improves the robustness of children's pneumonia cause typing. Further, in order to reduce the interference of irrelevant features, the reference clinical indicators closely related to the cause are analyzed, and the important representation degree of features in children's pneumonia is more clearly combined with the cause and clinical indicators, which helps to improve the accuracy of typing. By comparing the importance and clinical reference, the weight of feature classification analysis is adjusted by the contribution of features, so that the final training model is more suitable for children's pneumonia cause typing. The application extracts the features of children and adults, combines the auxiliary classification features of clinical association in cause typing, reduces the classification error caused by different growth stages, and improves the accuracy and robustness of the children's pneumonia cause typing model. BRIEF DESCRIPTION OF DRAWINGS

[0042] In order to more clearly illustrate the technical solutions and advantages of the embodiments of the present application or the prior art, the following will briefly introduce the drawings needed to be used in the embodiments or prior art description. Obviously, the drawings in the following description are only some embodiments of the present application, and those skilled in the art can obtain other drawings according to these drawings without creative labor.

[0043] Figure 1 The structural diagram of a children's pneumonia cause typing system based on a deep convolutional neural network provided by an embodiment of the present application;

[0044] Figure 2 The pneumonia image schematic diagram provided by an embodiment of the present application;

[0045] Figure 3 The mycoplasma pneumonia differentiation schematic diagram provided by an embodiment of the present application;

[0046] Figure 4 The children's mycoplasma pneumonia differentiation schematic diagram provided by an embodiment of the present application;

[0047] Figure 5 The partial clinical indicator schematic diagram of different causes provided by an embodiment of the present application. DETAILED DESCRIPTION

[0048] In order to further illustrate the technical means and effects taken by the present application to achieve the predetermined inventive objectives, the specific implementation, structure, features and effects of a child pneumonia etiology typing system based on a deep convolutional neural network according to the present application are described in detail as follows in combination with the drawings and preferred embodiments. In the following description, different "one embodiment" or "another embodiment" do not necessarily refer to the same embodiment. In addition, the specific features, structures or characteristics in one or more embodiments can be combined in any suitable form.

[0049] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which the present application belongs.

[0050] The specific scheme of the child pneumonia etiology typing system based on a deep convolutional neural network provided by the present application is specifically described below in combination with the drawings.

[0051] Please refer to Figure 1 which shows the structural diagram of the child pneumonia etiology typing system based on a deep convolutional neural network provided by one embodiment of the present application. The system comprises a data acquisition module 101, an adult contrast analysis module 102, a clinical impact analysis module 103 and a model adjustment training module 104.

[0052] The data acquisition module 101 is used to acquire pneumonia images and various clinical indicators of adults and children under different etiological pneumonias respectively; and the features in the pneumonia images are acquired through the pre-trained convolutional layer of the model.

[0053] In the embodiment of the present application, image data of child pneumonia, such as chest X-ray and CT, is collected from hospitals, public databases (such as Kaggle, PhysioNet, etc.) and multi-center clinical cooperation, to ensure that the image data covers pneumonia cases of different etiologies (such as bacterial, viral, mycoplasma and fungal) and different stages, and corresponding clinical indicator data, such as white blood cell count of blood indicators.

[0054] It is worth mentioning that in the specific embodiments of the present application, during the data acquisition process, the data privacy protection regulations are strictly followed, the patient information is desensitized and anonymized, the data security is ensured, and the collection, use and processing of relevant data need to comply with relevant laws, regulations and standards of relevant countries and regions.

[0055] It should be noted that the collected images and clinical data are subjected to quality inspection, and data with blur, excessive noise or serious information loss are excluded. The data format and data labels are unified to ensure accurate and standardized data labeling.

[0056] In the embodiment of the present application, the image is preprocessed: size normalization, all images are cropped and scaled, and adjusted to a uniform size, such as 224x224 pixels, to meet the input requirements of the deep convolutional neural network. Gray scale normalization, the pixel value is normalized to 0 to 1, to reduce the influence caused by light, equipment difference. Denoising and enhancement, using filtering, contrast enhancement or autoencoder method to remove noise and improve the clarity of image details. Please refer to Figure 2 Fig. 3 shows a schematic diagram of a pneumonia image provided by an embodiment of the present application.

[0057] Deep learning models usually need large-scale data for training, and high-quality, expert-labeled pneumonia etiology classification datasets for children are relatively scarce. It is necessary to use a model trained on large-scale medical image data for transfer learning to classify the etiology of children's pneumonia. The data source of the trained model is dominated by adult pneumonia or adult pneumonia data. However, the child's lung is not fully mature, the airway is narrow, and the alveolar function is not fully mature. The inflammatory response and infection spread pattern is different from that of adults, resulting in different image manifestations of pneumonia of different etiologies in adults and children. Since the dataset is mainly derived from adults, the true key features may be ignored on the child image data, leading to misjudgment of the etiology type of the child patient's pneumonia. Therefore, it is necessary to compare the image features of different types of pneumonia in adults and children, find the key features suitable for the classification of children's pneumonia, and fine-tune the trained model to make the model more accurate and suitable for child cases.

[0058] To improve the feature extraction capability of the deep learning model and optimize the classification effect of children's pneumonia, it is necessary to first analyze the features of children's pneumonia images of different etiologies and select the key features of children's pneumonia classification through statistical analysis. In the embodiment of the present application, a ResNet50 pre-trained model is selected, which has rich feature extraction capability. The convolutional layer of ResNet50 is used to obtain feature maps from preprocessed child pneumonia images, extract local information such as edges, textures, and densities, and reflect high-order features in the image, such as lesion morphology (such as patchy, grid shadow, nodules, and consolidation area), density changes (such as ground glass shadow, consolidation, and bronchial wall thickening), and lesion distribution (such as focal, multifocal, and diffuse).

[0059] The feature maps of the input child pneumonia images are obtained by the output results of the convolutional layer of the pre-trained model. The channels of the feature maps represent different dimensions of the feature maps, and each channel corresponds to different features, such as the feature of channel 50, which may focus on the thickening of the bronchial wall. Due to the need for comparative analysis, the feature information of adult pneumonia images can be obtained through a trained convolutional neural network model, which is not described here.

[0060] The adult contrast analysis module 102 is used for analyzing the deviation between the images of pneumonia of adults and children, the appearance degree of each feature in each cause, and obtaining the age discrimination degree of each feature in each cause; and analyzing the deviation between the images of pneumonia of adults and children, the correlation between each feature and other features in synchronous distribution and distribution distance, and obtaining the distribution specificity of each feature in each cause; and combining the age discrimination degree and the distribution specificity, obtaining the contrast importance of each feature in each cause.

[0061] The performance features of pneumonia of different causes are different in images. For example, bacterial pneumonia in children may present more diffuse or multifocal consolidation, and the lesion boundary is relatively blurred. Some cases may have bilateral involvement at the same time. In order to determine the key features of the classification of children's pneumonia, it is necessary to select the performance features that may appear in children with bacterial pneumonia in a certain or most cases as the basis for typing from a large number of possible image features.

[0062] The immune response of children is different from that of adults, which leads to different disease development patterns, and the typical performance features of pneumonia of different causes are also different. For example, the bronchial wall thickening of mycoplasma pneumonia in children is significant, while the bronchial wall thickening of mycoplasma pneumonia in adults is mild, and ground glass shadow is the main performance. In the adult fungal pneumonia, bronchial wall thickening can be seen. Therefore, if the trained model dominated by adult pneumonia data is directly migrated for learning, it will lead to misjudgment of the cause typing of children's pneumonia. In order to make the model pay attention to the features that are more discriminative for the cause typing of children's pneumonia, it is necessary to extract the key features of the cause typing of children's pneumonia.

[0063] By comparing the typical performance features of children and adults with pneumonia of different causes, the key features that contribute more to the cause typing of children's pneumonia are extracted. The same performance feature may belong to different causes in children and adults. When a certain feature appears in the chest image of children, it can determine the cause of children's pneumonia, and there is a difference in the typicality of the feature in the same cause of pneumonia between adults and children. Therefore, the feature can play a key role in the cause typing of children's pneumonia. Please refer to Figure 3 which shows a mycoplasma pneumonia differentiation diagram provided by an embodiment of the present application. In the diagram, the left diagram is mycoplasma pneumonia in children, and the right diagram is mycoplasma pneumonia in adults. The feature of bronchial wall thickening occurs obviously in children's pneumonia.

[0064] Therefore, first, the feature is compared and analyzed for the contrast difference in the typical performance of the cause, reflecting the performance discrimination caused by age. Preferably, in the embodiment of the present application, the method for obtaining the age discrimination degree comprises:

[0065] First, for any feature of any etiology, based on the manifestation of that feature in the imaging of children with pneumonia of that etiology, the typicality of the manifestation of that feature in the imaging of children with pneumonia of that etiology is obtained, reflecting the degree of typical manifestation of the feature. In the embodiments of the present invention, the method for obtaining the typicality of manifestation includes:

[0066] The frequency of occurrence of this feature in images of children with pneumonia of this etiology is taken as the frequency of occurrence of this feature in that etiology, reflecting the degree of presence of the feature in the etiology. Furthermore, the mean frequency of occurrence of all features in images of children with pneumonia of this etiology is taken as the distribution feature frequency of that etiology, reflecting the average degree of occurrence of each feature under that etiology.

[0067] Calculate the mean frequency of this feature in images of children with pneumonia of all etiologies to obtain the distributed etiological frequency of this feature. By analyzing the distributed etiological frequency, we can characterize the feature as a common feature.

[0068] Furthermore, the ratio of the frequency of occurrence of this feature in this etiology to the frequency of the distribution characteristics of this etiology is used as the first characterization index of this feature in this etiology, reflecting the representativeness of this feature under this etiology. The larger the first characterization index, the more representative the feature is of pneumonia under this etiology, and the more it can be used as a reference for judging etiological classification.

[0069] Furthermore, the ratio of the frequency of occurrence of this feature in this etiology to the frequency of distribution of this feature in other etiologies is used as a second characterization index of this feature in this etiology, representing the uniqueness of this feature in the diagnosis of pneumonia of this etiology. The larger the second characterization index, the more reliable the judgment ability.

[0070] Finally, by combining the first and second characterization indicators, the typicality of this feature in pediatric pneumonia imaging for this etiology is obtained. As an example, the expression for typicality is: In the formula, Indicates the first Type of feature in the first Typical manifestations in various etiologies; Indicates the first Type of feature in the first Frequency of occurrence in imaging data of various etiologies; Indicates the first Distribution characteristics and frequencies of various etiologies; Indicates the first Distribution frequency of etiological characteristics Indicates the first characterization index. This represents the second characterization index.

[0071] Similarly, the feature is acquired in adult pneumonia images, and the typicality of the feature in the etiology is analyzed according to the adult pneumonia images. In the embodiment of the present application, the frequency of the feature in the adult pneumonia images of the etiology is taken as the frequency of the feature in the etiology. The average of the frequencies of all features in the adult pneumonia images of the etiology is taken as the distribution feature frequency of the etiology. The average of the frequencies of the feature in the adult pneumonia images of all etiologies is calculated to obtain the distribution etiology frequency of the feature. The ratio of the frequency of the feature in the etiology to the distribution feature frequency of the etiology is taken as the first representation index of the feature in the etiology. The ratio of the frequency of the feature in the etiology to the distribution etiology frequency of the feature is taken as the second representation index of the feature in the etiology. The typicality of the feature in the etiology in the pediatric pneumonia images is obtained by combining the first representation index and the second representation index, and the specific meaning is not described here.

[0072] After the representation typicality is acquired, comparative analysis is performed, the difference of the representation typicality of the feature in the etiology between the pediatric pneumonia images and the adult pneumonia images is calculated and normalized to obtain the representation deviation degree. The greater the representation difference degree is, the greater the importance of the pediatric pneumonia typing is.

[0073] Finally, the product of the representation deviation degree and the representation typicality of the feature in the etiology in the pediatric pneumonia images is taken as the age discrimination degree of the feature in the etiology. The greater the age discrimination degree is, the more obvious the feature in the pediatric pneumonia etiology is, and the greater the comparison with adults is, and the greater the possibility of being emphasized in the subsequent is. As an example, the expression of the age discrimination degree is: , in the formula, the age discrimination degree of the i-th feature in the j-th etiology is represented; the representation typicality of the i-th feature in the j-th etiology in the pediatric pneumonia images is represented; the representation typicality of the i-th feature in the j-th etiology in the adult pneumonia images is represented; the representation deviation degree is represented. The representation deviation degree is represented by a normalization function. It should be noted that normalization is a technique familiar to those skilled in the art, and the normalization function can be linear normalization or standard normalization. The specific normalization method is not limited here.

[0074] ​​​​​The difference in the performance of the same feature in children and adults is mainly due to the difference in lung anatomy, development stage and immune response between the two, and if only the presence or absence of the feature is concerned, the detailed difference of the feature in the etiological typing of children and adults may be ignored, and special attention should be paid to the distribution pattern of the feature, which is different in different populations and different etiological pneumonias. For example, a certain bronchial wall thickening may be more concentrated in children and diffuse, while in adults it may be limited to a single lung lobe, which determines the contribution and weight of the feature to the etiological typing in different populations. By analyzing the distribution pattern of the feature, different etiologies can be more accurately distinguished. Even if a feature can indicate a certain etiology in children, it may have different typicality in adults, and only by combining the distribution pattern can the model improve the accuracy of the etiology of children's pneumonia.

[0075] Children's pneumonia is often not composed of a single feature, but a variety of imaging manifestations reflect the inflammatory process together, and different etiologies of pneumonia often present a pattern of feature combination on imaging. By correlating between features, the distribution pattern of the lesion in the lung can be determined. Please refer to Figure 4 , which shows a schematic diagram for distinguishing mycoplasma pneumonia in children according to an embodiment of the present application, the circle is bronchial wall thickening and interstitial change, and the lesion is mainly distributed along the bronchial vascular bundle.

[0076] Therefore, further comparative analysis is carried out by combining the distribution correlation, reflecting the difference in the distribution pattern of the feature, and preferably, in the embodiment of the present application, the acquisition method of the distribution specificity comprises:

[0077] First, under any etiology, each feature is sequentially taken as an analysis feature, and for any other feature of the analysis feature, the correlation pattern index of the analysis feature and the other feature is obtained by analyzing the distribution correlation between the analysis feature and the other feature in the distribution frequency and distribution position of the children's pneumonia imaging of the etiology, and the correlation exists in the distribution pattern by the positional relationship of the features on the image and the common appearance probability. In the embodiment of the present application, the acquisition method of the correlation pattern index comprises:

[0078] First, the sum of the number of occurrences of the other feature in the children's pneumonia imaging of the etiology and the number of occurrences of the analysis feature in the children's pneumonia imaging of the etiology is taken as the frequency sum of the analysis feature and the other feature, and the number of occurrences of the analysis feature and the other feature in the children's pneumonia imaging of the etiology is taken as the numerator, and the frequency sum is taken as the denominator to obtain the synchronous distribution feature index of the analysis feature and the other feature, to reflect the probability that the two features need to appear at the same time under the etiology.

[0079] Further in this etiology, when the analysis feature and the other feature appear simultaneously in the child pneumonia image, the distance between the center points of the corresponding regions of the analysis feature and the other feature in the single pneumonia image is calculated, the mean of all distances is calculated and negatively correlated, and the distribution distance feature index of the analysis feature and the other feature is obtained. The distance is calculated only when the two features appear simultaneously in the etiology. Since the region where the distribution produces the lesion is relatively close, the smaller the distance, the greater the distribution correlation of the feature. Therefore, by negatively correlating the mean of the distances in several simultaneous images, the distribution distance feature index is obtained.

[0080] It should be noted that distance calculation and negative correlation mapping are well-known technical means to those skilled in the art. Negative correlation calculation can adopt inverse proportion or negative exponential power form, etc., which will not be repeated and limited here.

[0081] Finally, in combination with the position and synchronous appearance, the product of the synchronous distribution feature index and the distribution distance feature index of the analysis feature and the other feature in the child pneumonia image is taken as the correlation mode index of the analysis feature and the other feature. The greater the correlation mode index, the higher the possibility of the correlation distribution of the two features in the etiology. As an example, the expression of the correlation mode index is: , in which, indicates the correlation lesion mode of the first feature and the second feature in the first etiology; the number of times that the first feature and the second feature appear simultaneously in the first etiology of child pneumonia image; indicates the number of times that the first feature appears in the first etiology of child pneumonia image data; indicates the number of times that the second feature appears in the first etiology of child pneumonia image data; indicates the distribution distance feature index of the first feature and the second feature in the first etiology, indicates the mean distance between the center points, indicates the exponential function with natural constant as base.

[0082] Only the feature distribution with strong correlation mode is compared with the adult, so when the correlation mode index of the feature and other features is greater than the preset correlation threshold, the corresponding other features are taken as the correlation features, and the high correlation situation is screened out for comparison and analysis. In the embodiment of the present application, the preset correlation threshold is set to 0.7, and the specific value can be adjusted by the implementer according to the specific implementation scene, which is not limited here.

[0083] Similarly, in the adult pneumonia image of the cause, the correlation mode index of the analysis feature and the correlation feature is obtained, and the correlation distribution between the features is analyzed according to the adult pneumonia image. In the embodiment of the present application, for any one correlation feature, the sum of the number of times of the correlation feature in the adult pneumonia image of the cause and the number of times of the analysis feature in the adult pneumonia image of the cause is taken as the frequency sum value of the analysis feature and the correlation feature. The number of times of the analysis feature and the correlation feature appearing in the adult pneumonia image of the cause is taken as the numerator, and the frequency sum value is taken as the denominator to obtain the synchronous distribution feature index of the analysis feature and the correlation feature. When the analysis feature and the correlation feature appear in the adult pneumonia image at the same time, the distance between the center points of the corresponding regions of the analysis feature and the correlation feature in the single pneumonia image is calculated, and then the mean value of all distances is calculated and negatively correlated to obtain the distribution distance feature index of the analysis feature and the correlation feature. In the adult pneumonia image, the product of the synchronous distribution feature index and the distribution distance feature index of the analysis feature and the correlation feature is taken as the correlation mode index of the analysis feature and the correlation feature, and the specific meaning is not described here.

[0084] Further comparison and analysis of the correlation distribution mode are performed, the difference between the correlation mode index of the analysis feature and the correlation feature in the child pneumonia image and the adult pneumonia image is calculated, and then the mean value of the corresponding difference of the analysis feature and all correlation features is obtained to obtain the distribution specificity of the analysis feature in the cause. The greater the distribution specificity, the greater the contrast difference of the analysis feature in the correlation distribution mode, which needs more attention. As an example, the expression of the distribution specificity is: , in which, represents the distribution specificity of the first feature in the first cause; represents the number of correlation features of the first feature; represents the correlation of the first feature in the child pneumonia image; represents the correlation of the first feature and the first correlation feature in the adult pneumonia image in the first cause. is expressed as an absolute value extraction function.

[0085] Finally, by age discrimination and distribution specificity, the feature importance under the contrast is reflected, and in the embodiment of the present application, the product of each feature and the age discrimination and distribution specificity of each cause is taken as the contrast importance of each feature in each cause. The higher the contrast importance is, the higher the reference value of the corresponding feature under the cause typing is, and the feature is more critical.

[0086] The clinical impact analysis module 103 is used to screen reference clinical indicators of each cause under the image of child pneumonia according to the correlation between different causes and various clinical indicators, and obtain the clinical reference of each feature in each cause by the correlation between each feature and the reference clinical indicator and the correlation between the cause and the reference clinical indicator.

[0087] Although the key features with higher reference value for child pneumonia cause typing can be screened by comparing the image features of child and adult pneumonia, there may still be important image features specific to children, which may not be typical in adult pneumonia or have completely different performance modes. Therefore, relying only on the comparative analysis of children and adults is not enough, and the relationship between image features and clinical key indicators also needs to be combined to further improve the contribution weight of the features so that they play a more important role in cause typing. Please refer to Figure 5 , which shows a partial clinical indicator diagram of different causes provided by an embodiment of the present application, wherein WBC is white blood cell count and NEU is neutrophil ratio.

[0088] To retain significant clinical indicators related to pneumonia causes, reference clinical indicators are obtained by screening through correlation. In the embodiment of the present application, chi-square test is used for each clinical indicator and different causes to obtain the p value of each clinical indicator and cause. The clinical indicators with p value less than or equal to the preset significance level are taken as the reference clinical indicators of the cause. The reference clinical indicators are the clinical indicators with higher reference significance for the cause, and the preset significance level is 0.05.

[0089] Chi-square test is mainly used to analyze the correlation and goodness of fit of classification data. By comparing the difference between the observed value and the expected value, it is evaluated whether the data conforms to the expected distribution or model. The p value is the probability of observing the current or more extreme chi-square statistic under the premise that the null hypothesis is true. The smaller the p value is, the less likely the result is caused by chance. By comparison with the significance level, when the p value is less than or equal to the significance level, the original hypothesis is rejected, and it is considered that there is a significant correlation between variables. It should be noted that the chi-square test is a well-known technical means familiar to those skilled in the art, which will not be described here.

[0090] The greater the relationship between a certain feature and a plurality of clinical indicators having reference significance for the etiological typing of pneumonia, the greater the clinical reference of the feature for the etiological typing of pneumonia in children, and the higher the attention, and therefore preferably, in the embodiments of the present application, the method for obtaining clinical reference includes:

[0091] First, for any feature of any etiology, the correlation between the feature and each reference clinical indicator of the etiology is calculated, and the product of the correlation between the etiology and the reference clinical indicator is taken as the correlation indicator of the feature with the etiology and each reference clinical indicator. The greater the correlation indicator, the higher the correlation between the feature and the clinical indicator. In the embodiments of the present application, the correlation can be calculated by point two column correlation coefficient, which is a statistical method for measuring the correlation between binary variables and continuous variables. The greater the correlation, the stronger the correlation.

[0092] Further, the sum of the correlation indicators of the feature with all reference clinical indicators of the etiology is taken as the numerator, and the sum of the correlations of the etiology with all reference clinical indicators is taken as the denominator to obtain the clinical reference of the feature for the etiology. The greater the clinical reference, the more important the feature plays in the etiological typing.

[0093] The model adjustment training module 104 is used to obtain the contribution degree of each feature to each etiology by combining the clinical reference and the comparative importance; the typing contribution degree of each feature is obtained by the distribution of the contribution degree of each feature in different etiologies; the features in the training model are adjusted in weight and trained according to the typing contribution degree, and the etiological typing network model is obtained.

[0094] According to the comparative importance of the feature in the etiological typing of pneumonia in children and the correlation with the clinical key indicators, the contribution degree of the feature in different etiologies is obtained. In the embodiments of the present application, the product of the clinical reference and the comparative importance of each feature in each etiology is taken as the contribution degree of each feature in each etiology. The higher the contribution degree, the higher the attention required in the etiological typing of pneumonia in children.

[0095] The same feature may exist in different etiologies of pneumonia. When a single feature has a higher contribution degree in one type of etiology and a relatively lower contribution degree in other types of pneumonia, the feature is more beneficial to the judgment of the etiological typing of pneumonia in children. Therefore, the typing contribution degree of each feature is obtained by adjusting the distribution of the contribution degree of each feature in different etiologies.

[0096] Preferably, in the embodiments of the present application, the method for obtaining the typing contribution degree comprises:

[0097] For any one feature, the cause of the feature is sequentially taken as the analysis cause, and the average contribution degree of the feature in all causes except the analysis cause is taken as the separation significance of the feature in the analysis cause, reflecting the contribution of the feature in the overall other causes.

[0098] Further, after calculating the difference between the contribution degree and the separation significance of the feature in the analysis cause, the product of the difference and the contribution degree is taken as the significant contribution degree of the feature and the analysis cause. The higher the contribution degree of the feature in the analysis cause and the greater the deviation from other causes, the more representative the feature is in typing.

[0099] Therefore, the separate significance of all causes is sequentially analyzed, the maximum value of the significant contribution degree of the feature and all causes is taken as the typing contribution degree of the feature, and the degree of attention that can be paid to the feature in typing is represented.

[0100] Important features with high typing contribution degrees are given greater weights, thereby optimizing the child cause typing model. In the embodiments of the present application, in the deep convolutional neural network, the input and output data are usually multi-channel tensors, each channel represents a certain specific feature in the image, the global average of each channel of the convolutional layer output is performed, each feature map is compressed into a numerical value, and a vector describing each channel feature is obtained. The typing contribution degree of each feature comprehensive cause is taken as a weight, multiplied by each feature map channel, dynamic adjustment of each channel feature is realized, the channel with a higher weight, that is, the channel more important to the classification task is amplified, while the channel with a lower weight is suppressed, which can make the model automatically pay attention to the features more discriminative to the typing of the child pneumonia cause. Finally, the adjusted each channel feature vector is added to the fully connected layer and the Softmax classifier for the typing of the child pneumonia cause.

[0101] After obtaining the cause typing network model pre-trained on a large-scale general image data set and fine-tuned on child data, in the embodiments of the present application, the image data of child pneumonia, such as chest X-ray and CT images, is input. Using the cause typing network model, high-level features in the image, such as pulmonary consolidation, grid shadow, bronchial wall thickening or nodule, etc. are extracted. The model automatically learns the image patterns specific to the cause of child pneumonia, distinguishes different manifestations such as bacterial, viral, mycoplasma or fungal pneumonia, captures the details of the child-specific lesions, reduces the interference of adult data features, and ensures that the extracted features are closely related to the cause. The extracted features are input into the classifier to perform cause typing on the child pneumonia, and the classification results such as bacterial, viral, mycoplasma or fungal pneumonia are output.

[0102] To sum up, the application first compares and analyzes the image data of pneumonia of children and adults, reflects the degree of contrast of features under each cause by the performance deviation of features on different causes and the performance deviation of feature association in the training model, optimizes the classification model of deep learning, makes it more suitable for the expression of children's influence, and improves the robustness of children's pneumonia cause typing. Further, in order to reduce the interference of irrelevant features, the reference clinical indicators more closely related to the cause are analyzed, the important expression degree of the features in children's pneumonia is more clearly reflected by combining the cause and the clinical indicators, and the accuracy of typing is improved. By comparing the importance and clinical reference, the weight of feature classification analysis is adjusted according to the contribution of the features, so that the final training model is more suitable for children's pneumonia cause typing. The application extracts the features of children and adults by comparison, combines the auxiliary classification features of cause typing with clinical association, reduces the classification error caused by different growth stages, and improves the accuracy and robustness of the children's pneumonia cause typing model.

[0103] It should be noted that the above-mentioned order of the embodiments of the application is only for description, and does not represent the advantages and disadvantages of the embodiments. The processes depicted in the drawings do not necessarily require the specific order or continuous order shown to achieve the desired results. In some embodiments, multitasking and parallel processing are possible or may be advantageous.

[0104] Each embodiment in the specification is described in a progressive manner, and the same or similar parts between each embodiment can be referred to each other. Each embodiment mainly describes the difference from other embodiments.

Claims

1. A deep convolutional neural network-based pediatric pneumonia etiology typing system, characterized in that, The system comprises: a data acquisition module, configured to acquire pneumonia images and clinical indicators of adults and children under different causes of pneumonia respectively, and acquire features in the pneumonia images through a convolution layer of a pre-trained model; an adult contrast analysis module, configured to analyze, under each cause, a deviation between the pneumonia images of adults and children, a performance degree of each feature in each cause, and obtain age discrimination of each feature in each cause; and analyze a deviation between a correlation between each feature and other features in synchronous distribution and distribution distance between the pneumonia images of adults and children, and obtain distribution specificity of each feature in each cause; and obtain contrast importance of each feature in each cause by combining the age discrimination and the distribution specificity; a clinical influence analysis module, configured to filter reference clinical indicators of each cause according to a correlation between different causes and the clinical indicators under the pneumonia images of children; and obtain clinical reference of each feature in each cause by combining a correlation between each feature and the reference clinical indicators and a correlation between the cause and the reference clinical indicators under each cause; a model adjustment and training module, configured to obtain contribution of each feature in each cause by combining the clinical reference and the contrast importance; obtain typing contribution of each feature by a distribution significant situation of the contribution of each feature in different causes; and obtain a cause typing network model by adjusting weights of the features in the training model according to the typing contribution and training. The contrast importance comprises: multiplying the age discrimination and the distribution specificity of each feature in each cause to obtain the contrast importance of each feature in each cause; The clinical reference comprises: for any feature in any cause, calculating a product between a correlation between the feature and each reference clinical indicator of the cause and a correlation between the cause and the reference clinical indicators to obtain a correlation index between the feature and each reference clinical indicator of the cause; calculating a sum of the correlation indexes between the feature and all reference clinical indicators of the cause as a numerator, and a sum of the correlations between the cause and all reference clinical indicators as a denominator to obtain the clinical reference of the feature in the cause; The contribution comprises: multiplying the clinical reference and the contrast importance of each feature in each cause to obtain the contribution of each feature in each cause; The typing contribution comprises: for any feature, sequentially taking a cause where the feature is located as an analysis cause; taking an average of the contribution of the feature in all causes except the analysis cause as a separation significance of the feature in the analysis cause; calculating a difference between the contribution of the feature in the analysis cause and the separation significance, and taking a product of the difference and the contribution as a significant contribution of the feature and the analysis cause; taking a maximum of the significant contributions of the feature and all causes as the typing contribution of the feature. 2.The child pneumonia etiological typing system based on deep convolutional neural network according to claim 1, wherein, The age discrimination comprises: For any one feature of any one cause, according to the performance of the feature in the child pneumonia image of the cause, the performance typicality of the feature in the child pneumonia image of the cause is obtained; similarly, the performance typicality of the feature in the adult pneumonia image of the cause is obtained; The difference between the performance typicality of the feature in the child pneumonia image and the performance typicality of the feature in the adult pneumonia image of the cause is calculated and normalized to obtain the performance deviation degree; The product of the performance deviation degree and the performance typicality of the feature in the child pneumonia image of the cause is taken as the age discrimination degree of the feature in the cause. 3.The system of claim 2, wherein, The performance typicality obtaining method comprises: The frequency of the feature in the child pneumonia image of the cause is taken as the appearance frequency of the feature in the cause; the mean value of the appearance frequencies of all features in the child pneumonia image of the cause is taken as the distribution feature frequency of the cause; the mean value of the appearance frequencies of the feature in the child pneumonia images of all causes is calculated to obtain the distribution cause frequency of the feature; The ratio of the appearance frequency of the feature in the cause to the distribution feature frequency of the cause is taken as the first representation index of the feature in the cause; the ratio of the appearance frequency of the feature in the cause to the distribution cause frequency of the feature is taken as the second representation index of the feature in the cause; The product of the first representation index and the second representation index is taken as the performance typicality of the feature in the child pneumonia image of the cause. 4.The system of classifying the causes of pneumonia in children based on deep convolutional neural network according to claim 1, wherein, The distribution specificity obtaining method comprises: Under any one cause, each feature is sequentially taken as an analysis feature; for any one other feature of the analysis feature, in the child pneumonia image of the cause, the association mode index of the analysis feature and the other feature is obtained by analyzing the distribution association between the analysis feature and the other feature in the distribution frequency and the distribution position; When the association mode index of the analysis feature and the other feature is greater than a preset association threshold, the corresponding other feature is taken as an associated feature; similarly, the association mode index of the analysis feature and the associated feature in the adult pneumonia image of the cause is obtained; After calculating the difference between the association mode indexes of the analysis feature and the associated features in the child pneumonia image and the adult pneumonia image, the mean value of the corresponding differences of the analysis feature and all associated features is obtained to obtain the distribution specificity of the analysis feature in the cause.

5. The deep convolutional neural network-based system for children pneumonia etiology typing according to claim 4, wherein, The association mode index obtaining method comprises: The sum of the appearance frequency of the other feature in the child pneumonia image of the cause and the appearance frequency of the analysis feature in the child pneumonia image of the cause is taken as the frequency sum of the analysis feature and the other feature; the distance between the center points of the corresponding regions of the analysis feature and the other feature in a single pneumonia image is calculated when the analysis feature and the other feature appear simultaneously in the child pneumonia image of the cause, and the mean value of all distances is obtained and negatively correlated to obtain the distribution distance feature index of the analysis feature and the other feature; ​ Under the image of the child pneumonia, the product of the analysis feature and the synchronous distribution feature index and the distribution distance feature index of the other feature is analyzed as the association mode index of the analysis feature and the other feature. 6.The system of classifying the causes of pneumonia in children based on deep convolutional neural network according to claim 1, wherein, The method for obtaining the reference clinical index comprises: The p value of each clinical index and the cause is obtained by using the chi-square test on each clinical index and different causes; the clinical index with a p value less than or equal to a preset significance level is taken as the reference clinical index of the cause.

Images