Children pneumonia pathogenesis typing system based on deep convolutional neural network

Abstract

The invention relates to the technical field of training models, in particular to a children pneumonia pathogenesis typing system based on a deep convolutional neural network. The system comprises a data acquisition module used for acquiring images and features in a training model; the adult comparative analysis module obtains comparative importance according to the expression deviation of the corresponding features of the children and adult pneumonia images on different causes and the expression deviation associated with the features; the clinical influence analysis module is used for obtaining clinical references through association of disease causes and clinical indexes in combination with characteristic related conditions; and the model adjustment training module comprehensively compares contribution degrees of importance and clinical reference characterization, and adjusts weight of training classification according to significant conditions of the contribution degrees in different causes to obtain a model. According to the invention, through feature comparison and extraction of children and adults and in combination with auxiliary classification features of clinical association in pathogenesis classification, classification errors caused by different growth stages are reduced, and the accuracy and robustness of a children pneumonia pathogenesis classification model are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of training models, and particularly to a pediatric pneumonia etiology classification system based on a deep convolutional neural network. Background Art

[0002] Pediatric pneumonia is an important disease leading to child mortality and health burden. Its etiology is complex, mainly including types such as bacterial, viral, and mycoplasma infections. Accurately and efficiently classifying the etiology of pneumonia is crucial for reasonably formulating treatment plans. In recent years, deep convolutional neural networks can achieve functions such as feature extraction, lesion area recognition, and automatic classification in the field of medical image analysis. The pediatric pneumonia etiology classification system based on deep convolutional neural networks can utilize pulmonary medical images to achieve efficient and accurate automatic classification, improve the recognition ability of different pathogen types of pediatric pneumonia, and provide intelligent assistance for clinical decision-making.

[0003] In the prior art, the medical image data of pediatric pneumonia is relatively scarce. Since high-quality pneumonia etiology classification datasets annotated by experts are relatively rare, deep learning models need large-scale data for training. Usually, a model pre-trained with large-scale medical image data is used for transfer learning to classify the etiology of pediatric pneumonia.

[0004] However, the immune system of children is in the developmental stage, and its physiological response mechanism to pneumonia pathogens is different from that of adults, resulting in different imaging features in terms of lesion manifestation, inflammation diffusion pattern, and tissue damage degree of lung imaging features. Directly migrating a deep learning model trained on general pneumonia data may be difficult to accurately capture the key lesion areas and microscopic structural changes of pediatric pneumonia due to the bias of its feature extraction layer towards adult pathological features, resulting in a decline in the generalization ability of the model and affecting the accuracy of classification. Summary of the Invention

[0005] In order to solve the technical problem that a deep learning model directly migrated and trained on general pneumonia data may be difficult to accurately capture the key lesion areas and microscopic structural changes of pediatric pneumonia due to the bias of its feature extraction layer towards adult pathological features, resulting in a decline in the generalization ability of the model and affecting the accuracy of classification, the purpose of the present invention is to provide a pediatric pneumonia etiology classification system based on a deep convolutional neural network. The specific technical solution adopted is as follows: The present invention provides a pediatric pneumonia etiology classification system based on a deep convolutional neural network, and the system includes: A data acquisition module, configured to respectively acquire pneumonia images and various clinical indicators of adults and children with pneumonia under different etiologies; and obtain features in the pneumonia images through the convolutional layer of a pre-trained model; An adult comparison analysis module is used to analyze, for each cause of disease, the deviation between the pneumonia images of adults and children in terms of the manifestation degree of each feature in each cause of disease, so as to obtain the age discrimination degree of each feature in each cause of disease; and analyze the deviation between the pneumonia images of adults and children in terms of the correlation between each feature and the correlation of other features in terms of synchronous distribution and distribution distance, so as to obtain the distribution specificity of each feature in each cause of disease; combining the age discrimination degree and the distribution specificity to obtain the comparison importance of each feature in each cause of disease. A clinical impact analysis module is used to screen out the reference clinical indicators for each cause of disease according to the correlation between different causes of disease and various clinical indicators in children's pneumonia images; for each cause of disease, through the correlation between each feature and the reference clinical indicators, combined with the correlation between the cause of disease and the reference clinical indicators, to obtain the clinical reference of each feature in each cause of disease. A model adjustment and training module is used to combine the clinical reference and the comparison importance to obtain the contribution degree of each feature in each cause of disease; through the significant distribution of the contribution degree of each feature in different causes of disease, to obtain the classification contribution degree of each feature; and adjust the weights of the features in the training model according to the classification contribution degree and train to obtain a cause-of-disease classification network model.

[0006] Further, the method for obtaining the age discrimination degree includes: For any feature of any cause of disease, according to the manifestation of the feature in children's pneumonia images of this cause of disease, obtain the manifestation typicality of the feature in this cause of disease in children's pneumonia images; similarly, obtain the manifestation typicality of the feature in this cause of disease in adult pneumonia images. Calculate the difference in the manifestation typicality of the feature in this cause of disease between children's pneumonia images and adult pneumonia images and perform normalization processing to obtain the manifestation deviation degree. Multiply the manifestation deviation degree by the manifestation typicality of the feature in this cause of disease in children's pneumonia images as the age discrimination degree of the feature in this cause of disease.

[0007] Further, the method for obtaining the manifestation typicality includes: Take the number of occurrences of the feature in children's pneumonia images of this cause of disease as the occurrence frequency of the feature in this cause of disease; take the average value of the occurrence frequencies of all features in children's pneumonia images of this cause of disease as the distribution feature frequency of this cause of disease; calculate the average value of the occurrence frequencies of the feature in children's pneumonia images of all causes of disease to obtain the distribution cause frequency of the feature. Take the ratio of the occurrence frequency of the feature in this cause of disease to the distribution feature frequency of this cause of disease as the first characterization index of the feature in this cause of disease; take the ratio of the occurrence frequency of the feature in this cause of disease to the distribution cause frequency of the feature as the second characterization index of the feature in this cause of disease. Combining the first characterization index and the second characterization index, the typicality of the manifestation of this feature in this etiology under the imaging of childhood pneumonia is obtained.

[0008] Further, the method for obtaining the distribution specificity includes: Under any one etiology, each feature is sequentially used as the analysis feature; for any other feature of the analysis feature, in the childhood pneumonia images of this etiology, by analyzing the distribution correlation between the analysis feature and this other feature in terms of distribution frequency and distribution position, the association pattern index between the analysis feature and this other feature is obtained. When the association pattern index between the analysis feature and other features is greater than the preset association threshold, the corresponding other feature is taken as the associated feature; similarly, in the adult pneumonia images of this etiology, the association pattern index between the analysis feature and the associated feature is obtained. After calculating the difference in the association pattern index between the analysis feature and the associated feature between the childhood pneumonia images and the adult pneumonia images, 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 this etiology.

[0009] Further, the method for obtaining the association pattern index includes: The sum value of the number of occurrences of this other feature in the childhood pneumonia images of this etiology and the number of occurrences of the analysis feature in the childhood pneumonia images of this etiology is taken as the frequency sum value of the analysis feature and this other feature; the number of occurrences of the analysis feature and this other feature simultaneously in the childhood pneumonia images of this etiology 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 this other feature. Under this etiology, when the analysis feature and this other feature appear simultaneously in the childhood pneumonia images, after calculating the distance between the center point of the area corresponding to the analysis feature and the center point of the area corresponding to this other feature in a single pneumonia image, the mean value of all distances is obtained and negatively correlated and mapped to obtain the distribution distance feature index of the analysis feature and this other feature. Under the childhood pneumonia images, the product of the synchronous distribution feature index and the distribution distance feature index of the analysis feature and this other feature is taken as the association pattern index of the analysis feature and this other feature.

[0010] Further, the method for obtaining the comparative importance includes: The product of the age discrimination and the distribution specificity of each feature in each etiology is taken as the comparative importance of each feature in each etiology.

[0011] Further, the method for obtaining the reference clinical index includes: Perform a chi-square test on each clinical index and different etiologies to obtain the p-value of each clinical index and the etiology; the clinical index with a p-value less than or equal to the preset significance level is taken as the reference clinical index of the etiology.

[0012] Further, the method for obtaining the clinical reference includes: For any feature of any cause, calculate the correlation between the feature and each reference clinical index of the cause, and the product of the correlation between the cause and the reference clinical index, as the relevant connection index of the feature between the cause and each reference clinical index; Take the sum value of the relevant connection indexes of the feature between the cause and all reference clinical indexes as the numerator, and the sum value of the correlations between the cause and all reference clinical indexes as the denominator to obtain the clinical reference of the feature for the cause.

[0013] Further, the method for obtaining the contribution degree includes: Take the product of the clinical reference and the comparative importance of each feature for each cause as the contribution degree of each feature for each cause.

[0014] Further, the method for obtaining the contribution degree of classification includes: For any feature, successively take the cause where the feature is located as the analysis cause; take the average value of the contribution degrees of the feature in all causes except the analysis cause as the separation significance of the feature for the analysis cause; After calculating the difference between the contribution degree of the feature for the analysis cause and the separation significance, take the product of the difference and the contribution degree as the significant contribution degree of the feature and the analysis cause; Take the maximum value of the significant contribution degrees of the feature and all causes as the contribution degree of classification of the feature.

[0015] The present invention has the following beneficial effects: The present invention first conducts a comparative analysis on the pneumonia image data of children and adults, and in the training model, reflects the degree of more obvious contrast of features under each cause through the performance deviation of features on different causes and the performance deviation of feature associations, optimizes the classification model of deep learning, makes it more adaptable to the expression of children's impacts, and improves the robustness of the cause classification of children's pneumonia. Further, in order to reduce the interference of irrelevant features and combine clinical indicators, analyze the reference clinical indicators that are more closely related to the cause, and make the important representation degree of features in children's pneumonia more clear through the combination of the cause and clinical indicators, thereby assisting in improving the accuracy of classification. By comprehensively comparing the importance and clinical reference, adjust the weight of feature classification analysis through the contribution of features, so that the final training model is more suitable for the cause classification of children's pneumonia. The present invention extracts features by comparing children with adults, combines clinical associations in the auxiliary classification features of cause classification, reduces the classification error caused by different growth stages, and improves the accuracy and robustness of the cause classification model of children's pneumonia. Description of the Drawings

[0016] To more clearly illustrate the technical solutions and advantages in the embodiments of the present invention or the prior art, the following will briefly introduce the accompanying drawings required in the description of the embodiments or the prior art. Obviously, the accompanying drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can also be obtained based on these drawings.

[0017] Figure 1 The structural diagram of a children's pneumonia etiology classification system based on a deep convolutional neural network provided by an embodiment of the present invention; Figure 2 A schematic diagram of pneumonia imaging provided by an embodiment of the present invention; Figure 3 A schematic diagram for differentiating mycoplasma pneumonia provided by an embodiment of the present invention; Figure 4 A schematic diagram for differentiating children's mycoplasma pneumonia provided by an embodiment of the present invention; Figure 5 A schematic diagram of partial clinical indicators for different etiologies provided by an embodiment of the present invention. Detailed implementation manners

[0018] To further elaborate on the technical means and effects adopted by the present invention to achieve the predetermined invention purpose, the following, in conjunction with the accompanying drawings and preferred embodiments, details the specific implementation manners, structures, features, and effects of a children's pneumonia etiology classification system based on a deep convolutional neural network proposed according to the present invention. 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.

[0019] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by those skilled in the technical field to which the present invention belongs.

[0020] The following specifically describes the specific solution of a children's pneumonia etiology classification system based on a deep convolutional neural network provided by the present invention with reference to the accompanying drawings.

[0021] Please refer to Figure 1 , which shows the structural diagram of a children's pneumonia etiology classification system based on a deep convolutional neural network provided by an embodiment of the present invention. The system includes: a data acquisition module 101, an adult comparison and analysis module 102, a clinical impact analysis module 103, and a model adjustment and training module 104.

[0022] The data acquisition module 101 is used to obtain pneumonia images and various clinical indicators of adults and children under different etiologies of pneumonia respectively; and obtain the features in the pneumonia images through the convolutional layer of the pre-trained model.

[0023] In the embodiment of the present invention, the image data of children with pneumonia is collected from hospitals, public databases (such as Kaggle, PhysioNet, etc.) and multi-center clinical collaborations, such as chest X-rays and CTs, to ensure that the image data covers pneumonia cases with different etiologies (such as bacterial, viral, mycoplasma, fungal, etc.) and different stages, as well as the corresponding clinical indicator data, such as the white blood cell count of blood indicators, etc.

[0024] It is worth mentioning that in the specific implementation manner of this application, during the data collection process, the data privacy protection regulations are strictly complied with, and the patient information is desensitized and anonymized to ensure data security, and the collection, use and processing of relevant data need to comply with the relevant laws, regulations and standards of relevant countries and regions.

[0025] It should be noted that the collected images and clinical data are subjected to quality inspection, and the data with blurring, excessive noise or serious information loss are excluded. The data format and data labels are unified to ensure accurate and standardized data annotation.

[0026] In the embodiment of the present invention, image preprocessing is performed: size normalization, all images are cropped and scaled to adjust to a unified size, such as 224×224 pixels, to meet the input requirements of the deep convolutional neural network. Gray level normalization, the pixel values are normalized between 0 and 1 to reduce the influence caused by illumination and equipment differences. Denoising and enhancement, methods such as filtering, contrast enhancement or autoencoders are used to remove noise and improve the clarity of image details. Please refer to Figure 2 , which shows a schematic diagram of a pneumonia image provided by an embodiment of the present invention.

[0027] Deep learning models usually require large-scale data for training. The high-quality pneumonia etiology classification datasets of children with pneumonia that are expert-annotated are relatively few. It is necessary to use the model pre-trained on large-scale medical image data for transfer learning to perform pneumonia etiology classification for children. The data source of the pre-trained model is dominated by adult pneumonia or adult pneumonia data. However, the lungs of children are not fully developed, the airways are narrow, and the alveolar function is not yet fully mature. The inflammatory response and infection spread patterns are different from those of adults, resulting in different pneumonia image manifestations of different etiologies in adults and children. Since the dataset mainly comes from adults, the true key features may be ignored in children's image data, leading to misjudgment of the pneumonia etiology type of child patients. Therefore, it is necessary to find the key features suitable for child pneumonia classification by comparing the image features of different types of pneumonia in adults and children, and fine-tune the pre-trained model to make the model more accurately adapted to child cases.

[0028] To improve the feature extraction ability of the deep learning model and optimize the classification effect of childhood pneumonia, it is first necessary to analyze the characteristics of pneumonia images with different etiologies in children and adults, and conduct statistical analysis to select the key features for classifying childhood pneumonia. In the embodiments of the present invention, the ResNet50 pre-trained model is selected, which has rich feature extraction ability. Through the convolutional layer of ResNet50, feature maps are obtained from the pre-processed childhood pneumonia images, and local information such as edges, textures, and densities is extracted to reflect the high-order features in the images, such as lesion morphology (e.g., patchy, reticular opacities, nodules, consolidation areas), density changes (e.g., ground-glass opacities, consolidation, bronchial wall thickening), and lesion distribution (e.g., focal, multifocal, diffuse), etc.

[0029] The feature maps of the input childhood pneumonia images are obtained through 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. For example, the features of channel 50 may focus on the thickening of the bronchial wall. Due to the need for comparative analysis, the feature conditions of adult pneumonia images can be obtained through a trained convolutional network model, which will not be elaborated here.

[0030] The adult comparison analysis module 102 is used to analyze, under each etiology, the deviation between the pneumonia images of adults and children in terms of the manifestation degree of each feature in each etiology, so as to obtain the age discrimination degree of each feature in each etiology; and analyze the deviation between the pneumonia images of adults and children in terms of the correlation between each feature and other features in terms of synchronous distribution and distribution distance, so as to obtain the distribution specificity of each feature in each etiology; combining the age discrimination degree and the distribution specificity, the comparative importance of each feature in each etiology is obtained.

[0031] Pneumonia with different etiologies has different manifestation characteristics in images. For example, childhood bacterial pneumonia may present more diffuse or multifocal consolidation, with blurred lesion boundaries, and bilateral involvement may be present in some cases. To determine the key features for classifying childhood pneumonia, it is necessary to select the manifestation features that are present in a certain or most cases of childhood bacterial pneumonia from a large number of possible image features as the basis for classification.

[0032] The immune response of children is different from that of adults, resulting in different disease development patterns, and the typical manifestation characteristics of pneumonia with different etiologies are also different. For example, the bronchial wall thickening in childhood mycoplasma pneumonia is significant, while the bronchial wall thickening in adult mycoplasma pneumonia is mild, and ground-glass opacity is the main manifestation. Bronchial wall thickening can be seen in adult fungal pneumonia. Therefore, if a directly transferred trained model dominated by adult pneumonia data is used for transfer learning, it will lead to misjudgment of the etiology classification of childhood pneumonia. To enable the model to focus on features that are more discriminative for the etiology classification of childhood pneumonia, it is necessary to extract the key features for the etiology classification of childhood pneumonia.

[0033] By comparing the typical manifestation characteristics of pneumonia in children and adults with different etiologies, key characteristics that contribute significantly to the etiological classification of childhood pneumonia are extracted. The same manifestation characteristic may belong to different etiologies in the pneumonia images of children and adults. When a certain characteristic appears in the chest images of children, if it can determine the etiology of childhood pneumonia and there are differences in the typicality of this characteristic in the same etiological pneumonia between children and adults, then this characteristic plays a key role in the etiological classification of childhood pneumonia. Please refer to Figure 3 , which shows a schematic diagram for differentiating mycoplasma pneumonia provided by an embodiment of the present invention. In the figure, the left picture is mycoplasma pneumonia in children, and the right picture is mycoplasma pneumonia in adults. The characteristic of bronchial wall thickening is significantly present in childhood pneumonia.

[0034] Therefore, first, comparative difference analysis is carried out on the typical manifestations of characteristics in terms of etiology to reflect the manifestation differences caused by age. Preferably, in the embodiment of the present invention, the method for obtaining the age discrimination degree includes: First, for any characteristic of any etiology, according to the manifestation of this characteristic in the pneumonia images of children with this etiology, the manifestation typicality of this characteristic in the pneumonia of this etiology in children's pneumonia images is obtained, which reflects the degree of typical manifestation of the characteristic. In the embodiment of the present invention, the method for obtaining the manifestation typicality includes: The number of times this characteristic appears in the pneumonia images of children with this etiology is used as the appearance frequency of this characteristic in this etiology, which reflects the degree of existence of the characteristic in the etiology. Furthermore, the mean value of the appearance frequencies of all characteristics in the pneumonia images of children with this etiology is used as the distribution characteristic frequency of this etiology, which reflects the average degree of appearance of each characteristic under this etiology.

[0035] Calculate the mean value of the appearance frequencies of this characteristic in the pneumonia images of children with all etiologies to obtain the distribution etiology frequency of this characteristic. Through the distribution etiology frequency, the possibility that this characteristic is a common characteristic is characterized.

[0036] Furthermore, the ratio of the appearance frequency of this characteristic in this etiology to the distribution characteristic frequency of this etiology is used as the first characterization index of this characteristic in this etiology, which reflects the representativeness of this characteristic in this etiology. When the first characterization index is larger, it indicates that this characteristic can better represent the pneumonia situation of this etiology and can be used as a reference basis for judging the etiological classification.

[0037] Furthermore, the ratio of the appearance frequency of this characteristic in this etiology to the distribution etiology frequency of this characteristic is used as the second characterization index of this characteristic in this etiology, which represents the uniqueness of this characteristic for the judgment of pneumonia of this etiology. The larger the second characterization index, the more credible the judgment ability.

[0038] Finally, by combining the first characterization index and the second characterization index, the manifestation typicality of this characteristic in the pneumonia images of children with this etiology is obtained. As an example, the expression of the manifestation typicality is: , where Indicates the typicality of the th feature in the th etiology; Indicates the th feature in the frequency of occurrence in the imaging data of the th etiology; Indicates the frequency of distribution characteristics of the th etiology; Indicates the frequency of distribution etiology of the th feature, Indicates the first characterization index.

[0039] Similarly, in the imaging of adult pneumonia, obtain the typicality of the manifestation of this feature in this etiology, and analyze according to the situation in the imaging of adult pneumonia when obtaining the frequency. In the embodiment of the present invention, the number of occurrences of this feature in the imaging of adult pneumonia of this etiology is used as the frequency of occurrence of this feature in this etiology. The average value of the frequencies of occurrence of all features in the imaging of adult pneumonia of this etiology is used as the frequency of distribution characteristics of this etiology. Calculate the average value of the frequencies of occurrence of this feature in the imaging of adult pneumonia of all etiologies to obtain the frequency of distribution etiology of this feature. The ratio of the frequency of occurrence of this feature in this etiology to the frequency of distribution characteristics of this etiology is used as the first characterization index of this feature in this etiology. The ratio of the frequency of occurrence of this feature in this etiology to the frequency of distribution etiology of this feature is used as the second characterization index of this feature in this etiology. Combine the first characterization index and the second characterization index to obtain the typicality of the manifestation of this feature in this etiology under the imaging of child pneumonia, and the specific meaning will not be elaborated here.

[0040] After obtaining the characterization typicality, conduct a comparative analysis, calculate the difference in the typicality of the manifestation of this feature in this etiology between the imaging of child pneumonia and the imaging of adult pneumonia and perform normalization processing to obtain the manifestation deviation degree. The greater the manifestation difference degree, the greater the importance for classifying child pneumonia.

[0041] Finally, the product of the manifestation deviation degree and the typicality of the manifestation of this feature in this etiology under the imaging of child pneumonia is used as the age discrimination degree of this feature in this etiology. The greater the age discrimination degree, the more obvious the manifestation of this feature in the etiology of child pneumonia, and the greater the contrast and distinction from adults, and the greater the possibility of being emphasized in the future. As an example, the expression of the age discrimination degree is: , where Indicates the th feature in the th etiology; In the imaging of child pneumonia ; In the imaging of adult pneumonia, the th feature in the The typicality of manifestations among various etiologies; Indicates the manifestation deviation degree. Expressed as a normalization function. It should be noted that normalization is a well-known technical means in the art. The choice of the normalization function can be linear normalization, standard normalization, etc. The specific normalization method is not limited herein.

[0042] The differences in the manifestations of the same feature between children and adults mainly stem from the differences in their lung anatomy, developmental stage, and immune response. If only the presence or absence of the feature is concerned, the detailed differences in the etiological classification of the feature between children and adults may be overlooked. It is also necessary to pay special attention to the distribution pattern of the feature, and the distribution pattern of the feature varies among different populations and different etiological pneumonias. For example, the thickening of a certain bronchial wall may be more concentrated and diffuse in children, while it may be limited to a single lung lobe in adults, determining the different contributions and weights of this feature to etiological classification 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, its typicality may be different in adults. Only by combining the distribution pattern can the accuracy of the model in discriminating the etiology of children's pneumonia be improved.

[0043] Children's pneumonia often consists of not a single feature, but multiple imaging manifestations that jointly reflect the inflammatory process. Pneumonias with different etiologies often present a pattern of feature combinations in imaging. By the association between features, the distribution pattern of the lesion in the lungs can be determined. Please refer to Figure 4 , which shows a schematic diagram for differentiating mycoplasma pneumonia in children provided by an embodiment of the present invention. Inside the circle are the phenomena of bronchial wall thickening and interstitial changes, and the lesions are mainly distributed along the bronchovascular bundles.

[0044] Therefore, further comparative analysis is carried out by combining the associated situations of the distributions to reflect the differences in the feature distribution patterns. Preferably, in the embodiment of the present invention, the method for obtaining the distribution specificity includes: First, under any one etiology, each feature is sequentially used as the analysis feature. For any other feature of the analysis feature, in the images of children's pneumonia with this etiology, by analyzing the distribution association situation between the analysis feature and this other feature in terms of distribution frequency and distribution position, the association pattern index between the analysis feature and this other feature is obtained, and the association existence distribution pattern is obtained through the positional relationship and co-occurrence probability of the feature on the image. In the embodiment of the present invention, the method for obtaining the association pattern index includes: First, take the sum of the number of occurrences of the other feature in the pediatric pneumonia images of this etiology and the number of occurrences of the analysis feature in the pediatric pneumonia images of this etiology as the frequency sum value of the analysis feature and the other feature. Take the number of times the analysis feature and the other feature simultaneously appear in the pediatric pneumonia images of this etiology as the numerator, and the frequency sum value as the denominator to obtain the synchronous distribution feature index of the analysis feature and the other feature, so as to reflect the probability when the two features need to appear simultaneously under the etiology.

[0045] Further, under this etiology, when the analysis feature and the other feature simultaneously appear in the pediatric pneumonia images, calculate the distance between the center point of the region corresponding to the analysis feature and the center point of the region corresponding to the other feature in a single pneumonia image, then find the average value of all distances and perform a negative correlation mapping to obtain the distribution distance feature index of the analysis feature and the other feature. Calculate the distance only when the two features simultaneously appear under the etiology. Since the regions with lesions caused by the distribution are relatively adjacent, the smaller the distance, the greater the possible distribution correlation of the features. Therefore, by performing a negative correlation mapping on the average value of the distances in several images where the features simultaneously appear, the distribution distance feature index is obtained.

[0046] It should be noted that both the distance calculation and the negative correlation mapping are technical means well-known to those skilled in the art. The negative correlation calculation can adopt forms such as inverse proportion or negative exponential power, etc., which will not be elaborated and restricted here.

[0047] Finally, combining the two aspects of position and simultaneous appearance, under the pediatric pneumonia images, take the product of the synchronous distribution feature index and the distribution distance feature index of the analysis feature and the other feature as the association pattern index of the analysis feature and the other feature. The larger the association pattern index, the higher the possibility that the two features have an associated distribution under this etiology. As an example, the expression of the association pattern index is: , where represents the associated lesion pattern of the th feature and the th feature under the th etiology; The th feature and the th feature simultaneously appear in the pediatric pneumonia images of the th etiology; represents the number of occurrences of the th feature in the pediatric pneumonia image data of the th etiology; represents the number of occurrences of the th feature in the pediatric pneumonia image data of the th etiology; represents the th feature and the th feature in the The distribution distance characteristic index of the cause of disease is expressed as the average distance between the center points and is expressed as an exponential function with the natural constant as the base.

[0048] Only the characteristic distribution with a strong correlation pattern is compared with that of adults. Therefore, when the correlation pattern index of the analysis feature and other features is greater than the preset correlation threshold, the corresponding other features are used as the associated features, and the cases with high correlation are selected for comparative analysis. In the embodiment of the present invention, the preset correlation threshold is set to 0.7, and the specific value can be adjusted by the implementer according to the specific implementation scenario, which is not limited here.

[0049] Similarly, in the adult pneumonia image of this cause of disease, the correlation pattern index of the analysis feature and the associated feature is obtained, and the correlation distribution of the features is analyzed based on the adult pneumonia image. In the embodiment of the present invention, for any associated feature, the sum of the number of times the associated feature appears in the adult pneumonia image of this cause of disease and the number of times the analysis feature appears in the adult pneumonia image of this cause of disease is used as the frequency sum value of the analysis feature and the associated feature. The number of times the analysis feature and the associated feature appear simultaneously in the adult pneumonia image of this cause of disease is used as the numerator, and the frequency sum value is used as the denominator to obtain the synchronous distribution characteristic index of the analysis feature and the associated feature. Under this cause of disease, when the analysis feature and the associated feature appear simultaneously in the adult pneumonia image, after calculating the distance between the center points of the corresponding regions of the analysis feature and the associated feature in a single pneumonia image, the average value of all distances is obtained and negatively correlated mapped to obtain the distribution distance characteristic index of the analysis feature and the associated feature. Under the adult pneumonia image, the product of the synchronous distribution characteristic index and the distribution distance characteristic index of the analysis feature and the associated feature is used as the correlation pattern index of the analysis feature and the associated feature, and the specific meaning is not elaborated here.

[0050] Furthermore, for the comparative analysis of the correlation distribution pattern, after calculating the difference in the correlation pattern index between the analysis feature and the associated feature in the child pneumonia image and the adult pneumonia image, the average value of the corresponding differences of the analysis feature and all associated features is obtained to obtain the distribution specificity of the analysis feature for this cause of disease. The greater the distribution specificity, the greater the contrast difference in the correlation distribution pattern of the analysis feature, and more attention is required. As an example, the expression of the distribution specificity is: , where in the formula represents the th distribution specificity of the th feature for the th cause of disease; represents the number of associated features associated with the ; Indicates the th feature and the th associated feature in the associated lesion pattern of the th etiology. Indicated as an absolute value extraction function.

[0051] Finally, through the age discrimination and distribution specificity, the importance of features under comparison is reflected. In the embodiments of the present invention, the product of the age discrimination and distribution specificity of each feature for each etiology is used as the comparative importance of each feature for each etiology. When the comparative importance is higher, it indicates that the corresponding feature has a higher reference value in the etiology classification, and the feature is relatively more critical.

[0052] The clinical impact analysis module 103 is used to screen out the reference clinical indicators for each etiology according to the correlation between different etiologies and various clinical indicators in the pediatric pneumonia image; under each etiology, through the correlation between each feature and the reference clinical indicators, combined with the correlation between the etiology and the reference clinical indicators, the clinical reference of each feature for each etiology is obtained.

[0053] Although key features with higher reference value for pediatric pneumonia etiology classification can be screened out by comparing the pneumonia image features of children and adults, there may still be important image features unique to children that may not be typical in adult pneumonia or have completely different manifestation patterns. Therefore, relying solely on the comparative analysis between children and adults is not enough. It is also necessary to combine the relationship between image features and clinical key indicators to further increase the contribution weight of the features and make them play a more important role in etiology classification. Please refer to Figure 5 , which shows a schematic diagram of partial clinical indicators for different etiologies provided by an embodiment of the present invention. Among them, WBC is the white blood cell count, and NEU is the neutrophil ratio.

[0054] To retain the significant clinical indicators related to pneumonia etiology, the reference clinical indicators are first screened through the correlation. In the embodiments of the present invention, the chi-square test is used for each clinical indicator and different etiologies to obtain the p-value of each clinical indicator and the etiology. The clinical indicators with p-values less than or equal to the preset significance level are used as the reference clinical indicators for the etiology. The reference clinical indicators are the clinical indicators that are more significant for the etiology, and the preset significance level is 0.05.

[0055] The chi-square test is mainly used to analyze the correlation and goodness of fit of categorical data, and evaluates whether the data conforms to the expected distribution or model by comparing the differences between observed values and expected values. 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, the less likely the result is caused by chance. By comparing with the significance level, when the p-value is less than or equal to the significance level, the null hypothesis is rejected, and it is considered that there is a significant association between variables. It should be noted that the chi-square test is a well-known technical means familiar to those skilled in the art and will not be elaborated here.

[0056] The greater the relationship between a certain feature and multiple clinical indicators that are of reference significance for the etiological classification of pneumonia, the greater the clinical reference value of this feature for the etiological classification of childhood pneumonia and the higher the attention. Therefore, preferably, in the embodiments of the present invention, the method for obtaining clinical reference includes: First, for any feature of any etiology, calculate the correlation between the feature and each reference clinical indicator of the etiology, as well as the product of the correlation between the etiology and the reference clinical indicators, as the relevant connection index of the feature between the etiology and each reference clinical indicator. The greater the relevant connection index, the higher the degree of correlation between the feature and the clinical indicator under the higher connection between the etiology and the clinical indicator. In the embodiments of the present invention, the correlation can be calculated by the point-biserial correlation coefficient. The point-biserial correlation coefficient is a statistical method for measuring the correlation between a binary variable and a continuous variable. The greater the correlation, the stronger the relevant association. It should be noted that the calculation of the correlation is a well-known technical means familiar to those skilled in the art and will not be elaborated here.

[0057] Further, use the sum value of the relevant connection indicators of the feature between the etiology and all reference clinical indicators as the numerator, and the sum value of the correlations between the etiology and all reference clinical indicators as the denominator to obtain the clinical reference value of the feature for the etiology. Through the relevant situation between the more reference-significant clinical indicators and the feature, the greater the clinical reference value, the more important the role played by the feature in the etiological classification.

[0058] The model adjustment training module 104 is used to obtain the contribution degree of each feature for each etiology by combining clinical reference and comparative importance; obtain the contribution degree of each feature for classification through the significant distribution of the contribution degrees of each feature in different etiologies; adjust the weights of the features in the training model according to the contribution degree of classification and train to obtain the etiological classification network model.

[0059] According to the comparative importance of features in the etiological classification of childhood pneumonia and their correlation with clinical key indicators, the contribution of each feature to different etiologies is obtained. In the embodiments of the present invention, the product of the clinical reference and comparative importance of each feature in each etiology is used as the contribution of each feature to each etiology. The higher the contribution, the higher the attention required in the etiological classification of childhood pneumonia.

[0060] The same feature may exist in pneumonias with different etiologies. When a single feature has a relatively high contribution in one type of etiology but a relatively low contribution in other types of pneumonia, this feature is more beneficial for the judgment of the etiological classification of childhood pneumonia. Therefore, by adjusting according to the distribution significance of the contribution of each feature to different etiologies, the classification contribution of each feature is obtained.

[0061] Preferably, in the embodiments of the present invention, the method for obtaining the classification contribution includes: For any feature, the etiology where the feature is located is sequentially used as the analysis etiology, and the average value of the contributions of the feature to all etiologies except the analysis etiology is used as the separation significance of the feature in the analysis etiology, reflecting the contribution of the feature to other etiologies as a whole.

[0062] After further calculating the difference between the contribution of the feature to the analysis etiology and the separation significance, the product of the difference and the contribution is used as the significant contribution of the feature to the analysis etiology. The higher the contribution of the feature to the analysis etiology and the greater the deviation from other etiologies, the more representative the feature is in the classification.

[0063] Therefore, the individual significant situations of all etiologies are analyzed in turn, and the maximum value of the significant contributions of the feature to all etiologies is used as the classification contribution of the feature, representing the degree to which the feature can be concerned in the classification.

[0064] Greater weights are assigned to important features with high classification contributions to optimize the childhood etiology classification model. In the embodiments of the present invention, in a deep convolutional neural network, the input and output data are usually multi-channel tensors, and each channel represents a specific feature in the image. Global averaging is performed on each channel of the output of the convolutional layer to compress each feature map into a numerical value, obtaining a vector describing the features of each channel. The classification contribution of each feature to the comprehensive etiology is used as the weight and multiplied by each channel of the feature map to achieve dynamic adjustment of the features of each channel. Channels with higher weights, that is, channels more important for the classification task, are amplified, while those with lower weights are suppressed, enabling the model to automatically focus on features more discriminative for the etiological classification of childhood pneumonia. Finally, the adjusted feature vectors of each channel are added to the fully connected layer and the Softmax classifier for the etiological classification of childhood pneumonia.

[0065] After obtaining the etiological classification network model that has been pre-trained on a large-scale general image dataset and fine-tuned on pediatric data, in the embodiments of the present invention, the image data of pediatric pneumonia, such as chest X-rays and CT images, are input. Using the etiological classification network model, high-level features in the images, such as lung consolidation, reticular opacities, thickening of the bronchial wall, or nodules, are extracted. The model will automatically learn the imaging patterns unique to the etiology of pediatric pneumonia, distinguish different manifestations such as bacterial, viral, mycoplasma, or fungal, capture the lesion details unique to children, reduce the interference of adult data features, and ensure that the extracted features are closely related to the etiology. The extracted features are used as input to a classifier to perform etiological classification on pediatric pneumonia, and classification results such as bacterial, viral, mycoplasma, or fungal pneumonia are output.

[0066] In summary, the present invention first conducts a comparative analysis of the pneumonia image data of children and adults. In the training model, through the performance deviation of features on different etiologies and the performance deviation of feature associations, it reflects the degree of more obvious feature contrast under each etiology, optimizes the deep learning classification model to make it more adaptable to the expression of children's images, and improves the robustness of etiological classification of pediatric pneumonia. Further, in order to reduce the interference of irrelevant features, combined with clinical indicators, reference clinical indicators that are more closely related to the etiology are analyzed, and the important characterization degree of features in pediatric pneumonia is made clearer through the combination of etiology and clinical indicators, assisting in improving the accuracy of classification. By comprehensively comparing the importance and clinical reference, the weights of feature classification analysis are adjusted according to the contribution of features, so that the final trained model is more suitable for etiological classification of pediatric pneumonia. The present invention extracts features by comparing children with adults, combines clinical associations in the auxiliary classification features of etiological classification, reduces the classification error caused by different growth stages, and improves the accuracy and robustness of the etiological classification model of pediatric pneumonia.

[0067] It should be noted that: the above sequence of the embodiments of the present invention is only for description and does not represent the advantages or 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 also possible or may be advantageous.

[0068] Each embodiment in this specification is described in a progressive manner. The same or similar parts among the embodiments can be referred to each other, and the key point of each embodiment is to illustrate the differences from other embodiments.

Claims

1. A children's pneumonia etiology classification system based on a deep convolutional neural network, characterized in that, The system includes: A data acquisition module, which is used to respectively acquire pneumonia images and various clinical indicators of adults and children under different etiologies of pneumonia; and obtain features in the pneumonia images through the convolutional layer of the pre-trained model. An adult comparison and analysis module, which is used to analyze, under each etiology, the deviation between the pneumonia images of adults and children and the deviation between the manifestation degrees of each feature in each etiology, so as to obtain the age discrimination degree of each feature in each etiology; and analyze the deviation between the pneumonia images of adults and children and the deviation between the correlation situations of each feature and other features in terms of synchronous distribution and distribution distance, so as to obtain the distribution specificity of each feature in each etiology; and combine the age discrimination degree and the distribution specificity to obtain the comparative importance of each feature in each etiology. A clinical impact analysis module, which is used to screen out the reference clinical indicators of each etiology according to the correlation situations between different etiologies and various clinical indicators in children's pneumonia images; and under each etiology, obtain the clinical reference of each feature in each etiology by combining the correlation situation between each feature and the reference clinical indicators and the correlation situation between the etiology and the reference clinical indicators. A model adjustment and training module, which is used to combine the clinical reference and the comparative importance to obtain the contribution degree of each feature in each etiology; obtain the typing contribution degree of each feature through the significant distribution situation of the contribution degrees of each feature in different etiologies; and adjust the weights of the features in the training model according to the typing contribution degree and train to obtain an etiology typing network model.

2. The children's pneumonia etiology classification system based on a deep convolutional neural network according to claim 1, characterized in that, The method for obtaining the age discrimination degree includes: For any feature of any etiology, obtain the manifestation typicality of the feature in the pneumonia images of children with this etiology according to the manifestation situation of the feature in the pneumonia images of children with this etiology; similarly, obtain the manifestation typicality of the feature in the pneumonia images of adults with this etiology. Calculate the difference in the manifestation typicality of the feature in the pneumonia images of children and adults with this etiology and perform normalization processing to obtain the manifestation deviation degree. Take the product of the manifestation deviation degree and the manifestation typicality of the feature in the pneumonia images of children with this etiology as the age discrimination degree of the feature in this etiology.

3. The children's pneumonia etiology classification system based on a deep convolutional neural network according to claim 2, characterized in that, The method for obtaining the manifestation typicality includes: Take the number of times the feature appears in the pneumonia images of children with this etiology as the appearance frequency of the feature in this etiology; take the average value of the appearance frequencies of all features in the pneumonia images of children with this etiology as the distribution feature frequency of this etiology; calculate the average value of the appearance frequencies of the feature in the pneumonia images of children with all etiologies to obtain the distribution etiology frequency of the feature. Take the ratio of the appearance frequency of the feature in this etiology to the distribution feature frequency of this etiology as the first characterization index of the feature in this etiology; take the ratio of the appearance frequency of the feature in this etiology to the distribution etiology frequency of the feature as the second characterization index of the feature in this etiology. Combine the first characterization index and the second characterization index to obtain the manifestation typicality of the feature in the pneumonia images of children with this etiology.

4. The children's pneumonia etiology classification system based on a deep convolutional neural network according to claim 1, characterized in that, The method for obtaining the distribution specificity includes: Under any one cause of disease, each feature is sequentially used as the analysis feature; for any other feature of the analysis feature, in the pediatric pneumonia images of this cause of disease, by analyzing the distribution correlation between the analysis feature and this other feature in terms of distribution frequency and distribution position, the association pattern index between the analysis feature and this other feature is obtained; When the association pattern index between the analysis feature and the other feature is greater than the preset association threshold, the corresponding other feature is taken as the associated feature; similarly, in the adult pneumonia images of this cause of disease, the association pattern index between the analysis feature and the associated feature is obtained; After calculating the difference in the association pattern index between the analysis feature and the associated feature between the pediatric pneumonia images and the adult pneumonia images, 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 for this cause of disease.

5. The children's pneumonia etiology classification system based on a deep convolutional neural network according to claim 4, characterized in that, The method for obtaining the association pattern index includes: Taking the sum of the number of occurrences of this other feature in the pediatric pneumonia images of this cause of disease and the number of occurrences of the analysis feature in the pediatric pneumonia images of this cause of disease as the frequency sum value of the analysis feature and this other feature; taking the number of occurrences of the analysis feature and this other feature simultaneously in the pediatric pneumonia images of this cause of disease as the numerator and the frequency sum value as the denominator to obtain the synchronous distribution feature index of the analysis feature and this other feature; Under this cause of disease, when the analysis feature and this other feature appear simultaneously in the pediatric pneumonia images, after calculating the distance between the center point of the region corresponding to the analysis feature and the center point of the region corresponding to this other feature in a single pneumonia image, the mean value of all distances is obtained and negatively correlated mapped to obtain the distribution distance feature index of the analysis feature and this other feature; Under the pediatric pneumonia images, taking the product of the synchronous distribution feature index and the distribution distance feature index of the analysis feature and this other feature as the association pattern index between the analysis feature and this other feature.

6. The children's pneumonia etiology classification system based on a deep convolutional neural network according to claim 1, characterized in that, The method for obtaining the comparison importance includes: Taking the product of the age discrimination degree and the distribution specificity of each feature for each cause of disease as the comparison importance of each feature for each cause of disease.

7. The children's pneumonia etiology classification system based on a deep convolutional neural network according to claim 1, characterized in that, The method for obtaining the reference clinical index includes: Using the chi-square test for each clinical index and different causes of disease to obtain the p-value of each clinical index and the cause of disease; taking the clinical index with the p-value less than or equal to the preset significance level as the reference clinical index of the cause of disease.

8. The children's pneumonia etiology classification system based on a deep convolutional neural network according to claim 1, characterized in that, The method for obtaining the clinical reference includes: For any feature of any cause of disease, calculating the product of the correlation degree between this feature and each reference clinical index of this cause of disease and the correlation degree between this cause of disease and the reference clinical index as the correlation link index between this feature and each reference clinical index of this cause of disease; Taking the sum value of the correlation link indexes between this feature and all reference clinical indexes of this cause of disease as the numerator and the sum value of the correlation degrees between this cause of disease and all reference clinical indexes as the denominator to obtain the clinical reference of this feature for this cause of disease.

9. The pediatric pneumonia etiology classification system based on a deep convolutional neural network according to claim 1, wherein, The method for obtaining the contribution degree includes: Taking the product of the clinical reference and the comparison importance of each feature for each cause of disease as the contribution degree of each feature for each cause of disease.

10. The pediatric pneumonia etiology classification system based on a deep convolutional neural network according to claim 1, wherein, The method for obtaining the classification contribution degree includes: For any one feature, sequentially use the cause of the disease where the feature is located as the analyzed cause of the disease; take the average contribution degree of the feature in all causes of the disease except the analyzed cause of the disease as the separation significance degree of the feature in the analyzed cause of the disease. After calculating the difference between the contribution degree of the feature in the analyzed cause of the disease and the separation significance degree, take the product of the difference and the contribution degree as the significant contribution degree of the feature and the analyzed cause of the disease. Take the maximum value of the significant contribution degrees of the feature and all causes of the disease as the classification contribution degree of the feature.

Images