Deep learning-based children chest radiography evaluation method, apparatus and device, and medium

Abstract

The invention discloses a child chest radiograph evaluation method, device and equipment based on deep learning and a medium, and the method comprises the steps: dividing a pre-collected child chest X-ray image into a training set, an internal verification set and an external verification set, and constructing a U-Net segmentation model, the method comprises the steps that a training set is established, region segmentation is conducted on the training set through a U-Net segmentation model, a DenseNetECA model is established, a core architecture of the DenseNetECA model is composed of four ECABlock modules, dense connection and an efficient channel attention mechanism are integrated in the ECABlock modules, feature dimensionality reduction and space down-sampling are conducted between the two ECABlock modules through a Transformation module, and the two ECABlock modules are connected through a Decision module; and inputting the processed training set, the internal verification set and the external verification set into a DenseNetECA model to complete model training, inputting a to-be-detected X-ray image into a U-Net segmentation model to perform region segmentation, and inputting the segmented to-be-detected X-ray image into the DenseNetECA model to perform evaluation to obtain an image evaluation result. The accuracy of the chest image of the child can be improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of child chest radiograph quality control, and in particular to a child chest radiograph detection method and device based on deep learning, equipment and a medium. BACKGROUND

[0002] In the process of child chest X-ray photography, most children cannot complete the breath-holding instruction due to their young age, and the shooting is usually done in a supine position, making it difficult to fully open the shoulder blades, further increasing the difficulty of shooting.

[0003] The current quality control method based on adult chest frontal position as the standard is difficult to adapt to the particularity of child chest radiographs, resulting in a lack of comprehensiveness and accuracy in quality control.

[0004] Currently, the quality control of child radiological images mainly relies on artificial subjective identification. Although this method can achieve image quality control to a certain extent, due to differences in individual understanding of quality control standards, artificial identification method has significant limitations in quality control efficiency, accuracy and comprehensiveness, and it is difficult to meet the precise needs of high-quality images in clinical practice. SUMMARY

[0005] The present application aims to at least solve one of the technical problems existing in the prior art. To this end, the present application proposes a child chest radiograph evaluation method based on deep learning, which can improve the accuracy of child chest images.

[0006] The present application also proposes a child chest radiograph evaluation device based on deep learning.

[0007] The present application also proposes a child chest radiograph evaluation device based on deep learning.

[0008] The present application also proposes a computer-readable storage medium.

[0009] In a first aspect, an embodiment of the present application provides a child chest radiograph evaluation method based on deep learning, comprising:

[0010] The pre-collected child chest X-ray images are divided into a training set, an internal validation set and an external validation set;

[0011] A U-Net segmentation model is constructed, and the training set is regionally segmented by the U-Net segmentation model;

[0012] A DenseNet_ECA model is constructed, the core architecture of the DenseNet_ECA model is composed of four ECA_Block modules, the ECA_Block module integrates dense connection and efficient channel attention mechanism, and the two ECA_Block modules are connected by a Transition module for feature dimension reduction and spatial down-sampling;

[0013] inputting the processed training set, the internal validation set and the external validation set into the DenseNet_ECA model to complete model training;

[0014] inputting the to-be-detected X-ray image into the U-Net segmentation model for region segmentation, and inputting the segmented to-be-detected X-ray image into the DenseNet_ECA model for evaluation to obtain an image evaluation result.

[0015] The child chest radiograph evaluation method has at least the following beneficial effects: the division of the training set, the internal validation set and the external validation set establishes a multi-level verification system, ensuring the stability of the model under different data distributions. The U-Net segmentation model effectively identifies the overlapping region of the scapula caused by the child's lying position through the multi-scale feature fusion capability, providing an accurate region of interest for subsequent evaluation. The DenseNet_ECA model promotes the reuse of features at different levels through dense connection, and the channel attention mechanism dynamically adjusts the feature channel weight, so that the model focuses on the key anatomical structure in the child's chest radiograph. The processed training set, internal validation set and external validation set are input into the DenseNet_ECA model for model training, and the original DenseNet_ECA model is optimized. The Transition module reduces the feature dimension while retaining spatial information, preventing feature loss in deep networks. When the segmented to-be-detected X-ray image is input into the DenseNet_ECA model, the interference of irrelevant regions is excluded, so that the attention mechanism can more effectively act on the target region, improving the recognition accuracy of the special anatomical structure of the child.

[0016] According to the child chest radiograph evaluation method of some embodiments of the application, the U-Net segmentation model is constructed by:

[0017] constructing a multi-task learning framework to train the corresponding U-Net segmentation model for each target region;

[0018] using an adaptive learning rate adjustment strategy combined with a Dice coefficient and a cross-entropy loss function to perform multi-objective optimization on the U-Net segmentation model;

[0019] adjusting the key parameters of the U-Net segmentation model through hyperparameter grid search and cross-validation methods to obtain the optimized U-Net segmentation model for each region.

[0020] According to the chest radiograph evaluation method of the child according to another embodiment of the application, the ECA_Block module is formed by multi-layer network dense connection, the network layers contained in the four ECA_Block modules are 6 layers, 12 layers, 24 layers and 16 layers in turn, and each layer of network is formed by a Dense_Layer feature extraction layer and an ECA feature recalibration module; wherein the Dense_Layer feature extraction layer reduces the dimension of the channel number of the input image through 1*1 convolution, and performs feature extraction operation through 3*3 convolution, and a PReLU activation function is used after each convolution layer of the Dense_Layer to provide the network with nonlinear modeling capability.

[0021] According to the chest radiograph evaluation method of the child according to another embodiment of the application, the specific processing steps of the ECA feature recalibration module on the input chest radiograph X-ray image of the child are as follows:

[0022] Performing global average pooling operation on the chest radiograph X-ray image of the child to obtain a global average value;

[0023] Through one-dimensional convolution with a size of K, cross-channel information interaction is realized between the target channel and the K adjacent channels; wherein the convolution kernel size K is calculated by the following formula: , C is the input channel number, b and is used to change the ratio between the channel number and the convolution kernel size, and odd represents the nearest odd number selected upward;

[0024] Inputting the global average value into a Sigmoid function for mapping to obtain a group of scalars between 0 and 1 as the target weights of each channel;

[0025] According to the target weights, weighted calculation is performed on the original input chest radiograph X-ray image of the child to realize recalibration of the feature channel, and an attention feature map is obtained.

[0026] According to the chest radiograph evaluation method of the child according to another embodiment of the application, the PReLU activation function includes the following formula: ; wherein x is the function input value, a is the modified linear unit parameter, and the modified linear unit parameter is learned adaptively so that the function value is not directly 0 when the function input value is negative.

[0027] According to the chest radiograph evaluation method of the child according to another embodiment of the application, the input of the processed training set, internal validation set and external validation set into the DenseNet_ECA model to complete model training includes:

[0028] optimizing parameters according to the processed training set to fit data rules of the DenseNet_ECA model;

[0029] evaluating performance of the DenseNet_ECA model under the same data distribution according to the internal validation set;

[0030] verifying applicability and generalization ability of the DenseNet_ECA model under different data distribution according to the external validation set.

[0031] According to the chest radiograph evaluation method of some embodiments of the present application, after the X-ray image to be detected is input into the U-Net segmentation model for region segmentation, and the segmented X-ray image to be detected is input into the DenseNet_ECA model for evaluation, an image evaluation result is obtained, the method further comprises:

[0032] performing consistency test on the image evaluation result according to artificial evaluation result based on Kappa analysis;

[0033] performing multi-dimensional performance evaluation on the DenseNet_ECA model according to accuracy, sensitivity, precision and F1 score indicators to obtain model performance evaluation result.

[0034] In a second aspect, an embodiment of the present application provides a deep learning-based child chest radiograph evaluation device, comprising:

[0035] An image set classification module is configured to randomly divide pre-collected child chest X-ray images into a training set, an internal validation set and an external validation set;

[0036] A segmentation model construction module is configured to construct a U-Net segmentation model and perform region segmentation on the training set through the U-Net segmentation model;

[0037] An evaluation model construction module is configured to construct a DenseNet_ECA model, the core architecture of the DenseNet_ECA model is composed of four ECA_Block modules, the ECA_Block module integrates dense connection and efficient channel attention mechanism, and the two ECA_Block modules are connected through a Transition module for feature dimension reduction and spatial down-sampling;

[0038] An evaluation model training module is configured to input the processed training set, the internal validation set and the external validation set into the DenseNet_ECA model to complete model training;

[0039] An image evaluation module is configured to input an X-ray image to be detected into the U-Net segmentation model for region segmentation, and input the segmented X-ray image to be detected into the DenseNet_ECA model for evaluation, to obtain an image evaluation result.

[0040] In a third aspect, an embodiment of the present application provides a deep learning-based child chest radiograph evaluation device, comprising:

[0041] at least one processor, and

[0042] a memory in communication connection with the at least one processor; wherein

[0043] The memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to perform the deep learning-based child chest radiograph evaluation method according to the first aspect.

[0044] In a fourth aspect, an embodiment of the present application provides a computer readable storage medium, which stores computer executable instructions for causing a computer to perform the deep learning-based child chest radiograph evaluation method according to the first aspect.

[0045] Other features and advantages of the present application will be set forth in the following description, and in part will become apparent from the description, or can be learned by practice of the present application. The objects and other advantages of the present application will be realized and attained by the structure particularly pointed out in the written description and claims thereof. BRIEF DESCRIPTION OF DRAWINGS

[0046] Figure 1 is a specific embodiment flowchart of the child chest radiograph evaluation method in the embodiment of the present application;

[0047] Figure 2 is Figure 1 a specific embodiment flowchart of step 102 in the embodiment of the present application;

[0048] Figure 3 is another specific embodiment flowchart of the child chest radiograph evaluation method in the embodiment of the present application;

[0049] Figure 4 is Figure 1 a specific embodiment flowchart of step 104 in the embodiment of the present application;

[0050] Figure 5 is another specific embodiment flowchart of the child chest radiograph evaluation method in the embodiment of the present application;

[0051] Figure 6is a specific embodiment module block diagram of the child chest radiograph evaluation device in the embodiment of the present application;

[0052] Figure 7 is a specific embodiment effect diagram of image delineation in the embodiment of the present application;

[0053] Figure 8 is a specific embodiment neural network architecture diagram of the U-Net segmentation model in the embodiment of the present application;

[0054] Figure 9 is a specific embodiment effect diagram of image segmentation in the embodiment of the present application;

[0055] Figure 10 is a specific embodiment schematic diagram of model construction based on radiographic range in the embodiment of the present application;

[0056] Figure 11 is a specific embodiment schematic diagram of model construction based on body position specification in the embodiment of the present application;

[0057] Figure 12 is a specific embodiment schematic diagram of model construction based on foreign matter in the embodiment of the present application;

[0058] Figure 13 is a specific embodiment schematic diagram of model construction based on shoulder bone positioning specification in the embodiment of the present application;

[0059] Figure 14 is a specific embodiment structure diagram of the Dense_Layer feature extraction layer in the embodiment of the present application;

[0060] Figure 15 is a specific embodiment structure diagram of the ECA feature re-calibration module and the ECA_Block module in the embodiment of the present application.

[0061] Explanation of reference signs:

[0062] Image set classification module 601, segmentation model construction module 602, evaluation model construction module 603, evaluation model training module 604, image evaluation module 605. DETAILED DESCRIPTION

[0063] The concept and the generated technical effects of the present application will be described below in conjunction with embodiments to make the purpose, features and effects of the present application clear. Obviously, the described embodiments are only a part of the embodiments of the present application, but not all the embodiments. Based on the embodiments of the present application, other embodiments obtained by those skilled in the art without creative labor are within the protection scope of the present application.

[0064] In the description of the present application, if the orientation description such as "upper", "lower", "front", "back", "left", "right" and the like indicates the orientation or positional relationship shown in the drawings, it is only for the convenience of describing the present application and simplifying the description, and does not indicate or imply that the device or element referred to must have a particular orientation, be constructed and operated in a particular orientation, and therefore cannot be understood as a limitation on the present application. If a certain feature is referred to as "provided", "fixed", "connected", "mounted" on another feature, it can be directly provided, fixed, connected or mounted on the other feature, or indirectly provided, fixed, connected or mounted on the other feature.

[0065] In the description of the embodiments of the present application, if "several" is referred to, it means more than one, if "multiple" is referred to, it means more than two, and if "greater than", "less than", "exceeding" is referred to, it should be understood as not including the number itself, if "and above", "and below", "and within" are referred to, it should be understood as including the number itself. If "first", "second" are referred to, it should be understood as being used to distinguish technical features, and cannot be understood as indicating or implying relative importance or implicitly indicating the number of indicated technical features or implicitly indicating the sequence of indicated technical features.

[0066] In the process of X-ray photography of children's chest, most of the children cannot cooperate to complete the breath-holding instruction due to their young age, and usually take a supine position for shooting, which makes it difficult to fully open the scapula, further increasing the difficulty of shooting.

[0067] The current quality control method with adult chest as the standard is difficult to adapt to the particularity of children's chest X-ray, resulting in lack of comprehensiveness and accuracy of quality control effect.

[0068] At present, the quality control of children's radiological images mainly relies on artificial subjective identification, although this method can achieve image quality control to a certain extent, but due to the differences in understanding of individual quality control standards, artificial identification method has significant limitations in quality control efficiency, accuracy and comprehensiveness, and it is difficult to meet the precise needs of high-quality images in clinical practice.

[0069] The present application aims to at least solve one of the technical problems existing in the prior art. To this end, the present application proposes a children's chest X-ray evaluation method based on deep learning, which can improve the accuracy of children's chest X-ray images.

[0070] Reference Figure 1 , Figure 1 The flowchart of the children's chest X-ray evaluation method based on deep learning in the embodiments of the present application is shown. In some embodiments, the children's chest X-ray evaluation method can include but is not limited to steps 101 to 105:

[0071] Step 101, the pre-collected chest X-ray images of children are divided into a training set, an internal validation set and an external validation set.

[0072] In step 101, specifically, based on a total of 6500 chest X-ray images of children in two years in the hospital, 5000 chest X-ray images of children are randomly divided into a training set of 3500 and an internal validation set of 1500, and the remaining 1500 are set as an external validation set.

[0073] Step 102, constructing a U-Net segmentation model, and performing region segmentation on the training set through the U-Net segmentation model.

[0074] In step 102, the U-Net segmentation model refers to a medical image segmentation network based on an encoder-decoder structure, which is used to accurately locate the target anatomical region in the chest X-ray of children.

[0075] Step 103, constructing a DenseNet_ECA model, the core architecture of the DenseNet_ECA model is composed of four ECA_Block modules, the ECA_Block module integrates dense connection and efficient channel attention mechanism, and two ECA_Block modules are connected through Transition module for feature dimension reduction and spatial down sampling.

[0076] In step 103, the DenseNet_ECA model refers to a deep network that integrates dense connection and channel attention mechanism, which combines attention mechanism to strengthen key channel feature extraction. The ECA_Block module refers to a basic unit containing multi-layer dense connection and channel attention mechanism, which realizes feature abstraction through hierarchical progression. The Transition module refers to a transition structure connecting adjacent ECA_Block, balancing the calculation efficiency and feature reservation.

[0077] Step 104, inputting the processed training set, internal validation set and external validation set into the DenseNet_ECA model to complete the model training.

[0078] Step 105, inputting the to-be-detected X-ray image into the U-Net segmentation model for region segmentation, and inputting the segmented to-be-detected X-ray image into the DenseNet_ECA model for evaluation to obtain the image evaluation result.

[0079] The steps 101 to 105 shown in the embodiments of the present application, the division of the training set, the internal validation set and the external validation set establishes a multi-level verification system to ensure the stability of the model under different data distribution. The U-Net segmentation model effectively identifies the overlapping region of the scapula caused by the child's lying position through the multi-scale feature fusion capability, and provides an accurate region of interest for subsequent evaluation. The dense connection in the DenseNet_ECA model promotes the reuse of features at different levels, and the channel attention mechanism dynamically adjusts the feature channel weight, so that the model focuses on the key anatomical structure in the child's chest X-ray. The processed training set, internal validation set and external validation set are input into the DenseNet_ECA model for model training, and the original DenseNet_ECA model is optimized. The Transition module reduces the feature dimension while retaining spatial information, preventing feature loss in deep networks. When the segmented X-ray image to be detected is input into the DenseNet_ECA model, the interference of irrelevant areas is excluded, so that the attention mechanism can more effectively act on the target area, improving the recognition accuracy of the special anatomical structure of the child.

[0080] Reference Figure 2 , Figure 2 A flowchart of a deep learning-based child chest X-ray evaluation method in embodiments of the present application is shown. In some embodiments, constructing a U-Net segmentation model includes but is not limited to steps 201 to 203:

[0081] Step 201, constructing a multi-task learning framework to train corresponding U-Net segmentation models for each target region.

[0082] In step 201, the multi-task learning framework refers to a parallel training architecture for constructing independent segmentation models for different anatomical regions, which solves the problem that a single model cannot adapt to multi-region segmentation requirements through differentiated feature extraction and region-specific optimization.

[0083] Step 202, using an adaptive learning rate adjustment strategy combined with Dice coefficient and cross-entropy loss function to perform multi-objective optimization on the U-Net segmentation model.

[0084] In step 202, the adaptive learning rate adjustment strategy refers to a mechanism for dynamically adjusting the weights of each optimization target according to the gradient changes of the loss function, which balances the optimization speed of the Dice coefficient and the cross-entropy loss to avoid the model falling into local optimum.

[0085] Step 203, adjusting the key parameters of the U-Net segmentation model through hyperparameter grid search and cross-validation method to obtain the optimized U-Net segmentation model for each region.

[0086] In step 203, the hyperparameter grid search refers to a tuning method of systematically combining and traversing key parameters of a model, and the optimal parameter combination is screened in combination with a cross-validation method to improve the generalization ability of the model.

[0087] The steps 201 to 203 shown in the embodiments of the application are aimed at the morphological differences of the lung field, clavicle, rib and other anatomical structures in the chest radiograph of children. The multi-task learning framework allocates a dedicated U-Net segmentation model for the six key anatomical target regions, and captures regional features through an independent decoder branch. In the training process, the Dice coefficient is used to measure the overlap degree of the predicted region and the real label, and the cross-entropy loss is used for pixel-level classification accuracy control. The two are dynamically weighted and fused through an adaptive learning rate strategy, so that the model retains details while improving overall segmentation consistency. The hyperparameter grid search performs full combination traversal on the learning rate, batch size and regularization coefficient, and screens the optimal parameter configuration of each regional model in combination with five-fold cross-validation to ensure the best performance under limited data. The problems of low training efficiency, insufficient segmentation accuracy and lack of systematicness in parameter optimization in the multi-region segmentation task of children's chest radiographs are solved, efficient learning of regional differentiated features is realized, the model convergence speed and segmentation consistency are optimized, and the objectivity and scientificity of parameter configuration are ensured.

[0088] Specifically, referring to Figure 7 , Figure 8 and Figure 9 , Figure 7 an effect diagram of image delineation in the embodiments of the application is shown, Figure 8 a neural network architecture diagram of the U-Net segmentation model in the embodiments of the application is shown, Figure 9 an effect diagram of image segmentation in the embodiments of the application. In some embodiments, the image delineation comprises: since chest radiograph data with relevant lung region, scapula region and clavicle region annotation information is needed to train the model, the image data is annotated by using an annotation tool based on software labelme under a Python framework, and the annotation numbers represent 1 left lung field, 2 right lung field, 3 left clavicle, 4 right clavicle, 5 left scapula and 6 right scapula in order, Figure 7 in which left and right refer to the left and right of the image display, which is opposite to the actual human body structure direction, but does not affect the final intelligent evaluation result.

[0089] The image segmentation includes: adopting a U-Net segmentation model to perform fine segmentation on six key anatomical regions in medical images, such as bilateral lung fields, clavicles, scapulae, etc. By constructing a multi-task learning framework, an independent U-Net model is trained for each target region. In the model optimization process, an adaptive learning rate adjustment strategy is adopted, and a multi-objective optimization is performed in combination with the Dice coefficient and the cross-entropy loss function. Through the hyperparameter grid search and cross-validation method, the network depth, convolution kernel size, dropout rate and other key parameters are systematically adjusted, and finally the optimal segmentation model of each region is obtained. The segmentation accuracy of the optimized model on the six target regions on the test set needs to reach more than 90%, and the Dice coefficient of the lung field region drawn needs to reach 0.95.

[0090] It should be noted that the U-Net segmentation model is a deep learning model specially designed for medical image segmentation, which realizes pixel-level accurate segmentation through the process of "feature extraction-feature restoration-feature fusion".

[0091] Further, in some embodiments, in the DenseNet_ECA model, the ECA_Block module is composed of multiple layers of network dense connection, and the network layers contained in the four ECA_Block modules are 6 layers, 12 layers, 24 layers and 16 layers in turn, and each layer of network is composed of a Dense_Layer feature extraction layer and an ECA feature re-calibration module. Among them, the Dense_Layer feature extraction layer reduces the channel number of the input image through 1x1 convolution, and performs feature extraction operation through 3x3 convolution, and a PReLU activation function is provided after each convolution layer of the Dense_Layer to provide the network with nonlinear modeling ability.

[0092] The Dense_Layer feature extraction layer refers to a composite structure that realizes feature compression and extraction through convolution operation. The ECA feature re-calibration module refers to a component that dynamically adjusts the feature channel weight through the channel attention mechanism. The PReLU activation function refers to an improved linear rectifier unit with learnable parameters.

[0093] Specifically, in some embodiments, the model selection includes: the DenseNet_ECA model adopts an innovative modular design, the core architecture of which is composed of four feature extraction ECA_Block modules, and efficient feature learning is achieved through a hierarchical feature fusion mechanism. Specifically, each ECA_Block module integrates a dense connection (Dense Connection) and an efficient channel attention (Efficient Channel Attention) mechanism, which can fully extract multi-scale features and adaptively enhance key feature channels. During feature extraction, a Transition module is used between every two ECA_Block modules for feature dimension reduction and spatial downsampling, which is composed of a 1x1 convolution layer and a 2x2 average pooling layer in sequence, which reduces the resolution of the feature map while maintaining the feature expression ability. This hierarchical structure design not only realizes the progressive extraction of features, but also ensures the effective fusion of shallow and deep features through the dense connection mechanism, thereby improving the model's ability to extract medical image features.

[0094] Referring to Figure 10 , Figure 11 , Figure 12 and Figure 13 , Figure 10 shows a schematic diagram of model construction based on the scope of the film in the embodiment of the application, Figure 11 shows a schematic diagram of model construction based on the body position specification in the embodiment of the application, Figure 12 shows a schematic diagram of model construction based on foreign matter in the embodiment of the application, Figure 13 shows a schematic diagram of model construction based on shoulder bone positioning specification in the embodiment of the application. In DenseNet_ECA, each feature extraction module ECA_Block is composed of multiple layers of network dense connection, and the number of network layers contained in the four ECA_Block is 6, 12, 24, and 16 in turn. Each layer of network is mainly composed of a feature extraction layer Dense_Layer and a feature re-calibration module ECA. The structure of Dense_Layer can refer to Figure 14 , Figure 14 shows a structure diagram of the Dense_Layer feature extraction layer in the embodiment of the application, which first uses a 1x1 convolution to reduce the channel number of the input image, and then uses a 3x3 convolution for feature extraction operation, thereby effectively reducing the feature number and improving the calculation efficiency. At the same time, in order to help the network better learn the complex patterns in the data, the Dense_Layer uses a PReLU activation function after each convolution layer to provide the network with nonlinear modeling ability.

[0095] It should be noted that the DenseNet_ECA model is a hybrid architecture that combines the dense connection mechanism of DenseNet with the ECA (Efficient Channel Attention) module, aiming to optimize the model performance through efficient feature reuse and channel attention, especially suitable for medical image analysis. The feature extraction layer Dense_Layer is the core component of DenseNet, and the input of each layer not only contains the output of the previous layer, but also contains the output of all previous layers. This design maximizes feature reuse, alleviates the problem of gradient disappearance, and is especially suitable for feature extraction of medical images.

[0096] Referring to Figure 3 , Figure 3 The flowchart of the deep learning-based pediatric chest radiograph evaluation method in the embodiments of the application is shown. In some embodiments, the specific processing steps of the ECA feature re-labeling module for the input pediatric chest X-ray image include but are not limited to steps 301 to 304:

[0097] Step 301, performing a global average pooling operation on the pediatric chest X-ray image to obtain a global average value.

[0098] In step 301, the global average pooling operation refers to an operation of compressing the spatial dimension of the feature map, and obtaining the global statistical features of the channel dimension by eliminating the spatial position difference.

[0099] Step 302, cross-channel information interaction is achieved between the target channel and the K adjacent channels through one-dimensional convolution with a size of K. Wherein, the convolution kernel size K is calculated by the following formula: , C is the input channel number, b and is used to change the ratio between the number of channels and the size of the convolution kernel, and odd represents the nearest odd number selected upwards.

[0100] In step 302, one-dimensional convolution refers to a linear transformation operation along the channel dimension, and cross-channel correlation is established by adaptively adjusting the interaction range of adjacent channels.

[0101] Step 303, input the global average value into the Sigmoid function to obtain a group of scalars between 0 and 1 as the target weight of each channel.

[0102] In step 303, the Sigmoid function refers to an activation function that maps real numbers to the 0-1 interval, and generates interpretable channel attention weights through nonlinear transformation.

[0103] Step 304, according to the group of target weights, the original input pediatric chest X-ray image is weighted and calculated to realize the re-labeling of the feature channel, and the feature map with attention is obtained.

[0104] The steps 301 to 304 shown in the embodiments of the present application eliminate the spatial position interference caused by the body position difference through global average pooling, and extract the global statistical features of the channel dimension as the basis for attention calculation. Then, the one-dimensional convolution kernel size is dynamically calculated according to the input channel number, so that each channel can interact with the adjacent two channels. This adaptive mechanism avoids the problem of limited local feature interaction caused by fixed size convolution kernel, while maintaining the parameter efficiency and enhancing the correlation between channels. The statistical features are converted into weight coefficients through the Sigmoid function, so that the model can autonomously strengthen the response intensity of the channels related to the pathological features. The problem of attention dispersion caused by insufficient channel feature interaction in the traditional manual identification method is effectively solved, and the capture accuracy of the key region features of the child chest radiograph is enhanced through the adaptive channel interaction mechanism. On the premise of maintaining the calculation efficiency, the recognition ability of the network for the subtle pathological features under the interference of the overlapping region of the scapula and the respiratory motion artifact is significantly improved, and the misjudgment risk caused by the subjectivity of manual identification is overcome.

[0105] Specifically, in order to enable the network to pay more attention to the image features related to the task, the feature re-calibration module ECA is added after each Dense_Layer in this paper. The specific structure of the module can refer to Figure 15 (a), Figure 15 The structure diagram of the ECA feature re-calibration module and the ECA_Block module in the embodiments of the present application is shown. The specific processing steps of the ECA feature re-calibration module for the child chest radiograph feature map include: 1, performing a global average pooling operation (Global Average Pooling, GAP) on the input feature map, and compressing the global information into a group of scalars with global receptive field. 2, using a one-dimensional convolution with a size of K to realize the local cross-channel information interaction of each channel and its adjacent K channels. 3, the output result of the last layer is sent into the Sigmoid function to map a group of 0 to 1 scalars as the weight of each channel, and the group of weights is the final obtained attention value. 4, using the group of weights to perform weighted calculation on the original input data, realizing the re-calibration of the feature channels, and obtaining a new feature map with attention.

[0106] Referring to Figure 15 (b), in each ECA_Block, Figure 15 The dashed box part in (b) is each network layer. The features extracted by the feature extraction layer Dense_Layer in each network layer are re-calibrated by the ECA feature re-calibration module before entering the next network layer, so that the network can distinguish the importance of different features, and learn the image features related to the task better. At the same time, through the use of dense connection mechanism, the features extracted by all network layers are fused in the channel dimension, effectively enhancing the propagation and reuse of features.

[0107] It should be noted that ECA_Block (Efficient Channel Attention Block) is a lightweight channel attention mechanism, and the core advantage is to greatly reduce the amount of calculation while ensuring the effect of channel attention, especially suitable for medical image segmentation. The ECA feature re-scaling module is an efficient feature re-scaling module, which simplifies the calculation process of channel attention, greatly reduces the calculation cost while ensuring the re-scaling effect, and is especially suitable for medical image tasks sensitive to model efficiency.

[0108] Further, in some embodiments, the PReLU activation function includes the following formula: ; wherein x is the function input value, and a is the rectified linear unit parameter. By adaptively learning the rectified linear unit parameter, the function value will not be directly 0 when the function input value is negative.

[0109] It should be noted that PReLU is an improved version of ReLU activation function, which can effectively avoid the problem of neuron "necrosis".

[0110] Referring to Figure 4 , Figure 4 The flowchart of the deep learning-based child chest radiograph evaluation method in the embodiment of the application is shown. In some embodiments, the processed training set, internal validation set and external validation set are input into the DenseNet_ECA model to complete the model training, which includes but is not limited to steps 401 to 404:

[0111] Step 401: According to the processed training set, the parameters are optimized to fit the data law of the DenseNet_ECA model.

[0112] In step 401, the training set refers to a data set used to adjust the model parameters, which is used to make the model learn the core feature distribution law of the child chest radiograph.

[0113] Step 402: According to the internal validation set, the performance of the DenseNet_ECA model under the same data distribution is evaluated.

[0114] In step 402, the internal validation set refers to a data subset that is homologous to the training set and consistent with the data distribution, which is used to verify the fitting degree of the model to the known data mode and prevent overfitting phenomenon.

[0115] Step 403: According to the external validation set, the applicability and generalization ability of the DenseNet_ECA model under different data distributions are tested.

[0116] In step 403, the external validation set refers to a data set collected independently and having a difference in data distribution, which is used to simulate heterogeneous data in actual application scenarios to test the model generalization ability.

[0117] The steps 401 to 403 shown in the embodiments of the application, in the model training phase, first, the weight parameters of the DenseNet_ECA model are iteratively optimized through the training set, so that the network gradually fits the correlation pattern between the pathological features and normal anatomical structure of the child chest film. Subsequently, the internal validation set is periodically called during the training process to evaluate the performance, and the learning rate and other hyperparameters are dynamically adjusted by monitoring the loss function changes on the validation set, to ensure that the model maintains the fitting ability to the training data while avoiding falling into a local optimal solution. Finally, after the model training is completed, the external validation set is used for cross-domain testing, and the classification accuracy and feature extraction stability of the model under unknown data distribution are calculated to verify its robustness in actual clinical environment.

[0118] Referring to Figure 5 , Figure 5 A flowchart of a child chest film evaluation method based on deep learning in the embodiments of the application is shown. In some embodiments, after the to-be-detected X-ray image is input into the U-Net segmentation model for region segmentation, and the segmented to-be-detected X-ray image is input into the DenseNet_ECA model for evaluation to obtain an image evaluation result, the child chest film evaluation method specifically further includes but is not limited to steps 501 to 502:

[0119] Step 501, performing consistency test on the image evaluation result according to the artificial evaluation result based on Kappa analysis.

[0120] In step 501, Kappa analysis refers to quantifying the consistency degree between the model evaluation result and the artificial evaluation result by statistical methods, which is used to verify the degree of agreement between the model output and the clinical diagnosis standard.

[0121] Step 502, performing multi-dimensional performance evaluation on the DenseNet_ECA model according to the accuracy, sensitivity, precision and F1 score indicators to obtain a model performance evaluation result.

[0122] In step 502, the meaning of accuracy is that the proportion of correct positive and negative samples identified by the model to the total number of samples. Sensitivity represents the proportion of correct positive samples identified by the model to the true positive samples, which is used to measure the ability of the model to find all positive samples, and the higher the sensitivity, the lower the corresponding missed diagnosis rate. Precision represents the proportion of correct positive samples identified by the model to all samples identified as positive samples, which is used to measure the accuracy of the model for positive samples, and the higher the precision, the lower the misdiagnosis rate. F1 score is the harmonic mean of precision and sensitivity, which is used to evaluate the comprehensive identification ability of the model.

[0123] The steps 501 to 502 shown in the embodiments of the present application, after the model outputs the image evaluation result, first compare the model prediction result with the manual evaluation result of at least two radiologists, and judge whether there is statistical significance of consistency between the two by calculating Kappa analysis, to ensure that the model evaluation result is highly consistent with the expert evaluation. Subsequently, in the model performance verification stage, the accuracy, sensitivity, precision and F1 score of the model are calculated to realize the comprehensive evaluation of the model performance, and to ensure that the model meets the requirements of clinical application in various indicators.

[0124] Specifically, the accuracy, sensitivity, precision and F1 score are calculated according to the following formula:

[0125]

[0126]

[0127]

[0128]

[0129] In addition, the embodiments of the present application also disclose a child chest radiograph evaluation device based on deep learning, please refer to Figure 6 , Figure 6 An embodiment of the present application discloses a module block diagram of a child chest radiograph evaluation device. The child chest radiograph evaluation device can realize the child chest radiograph evaluation method described above, and the child chest radiograph evaluation device comprises: an image set classification module 601, a segmentation model construction module 602, an evaluation model construction module 603, an evaluation model training module 604 and an image evaluation module 605, and the image set classification module 601, the segmentation model construction module 602, the evaluation model construction module 603, the evaluation model training module 604 and the image evaluation module 605 are all in communication connection.

[0130] The image set classification module 601 randomly divides the pre-collected child chest X-ray images into a training set, an internal validation set and an external validation set. The segmentation model construction module 602 constructs a U-Net segmentation model, and performs regional segmentation on the training set through the U-Net segmentation model. The evaluation model construction module 603 constructs a DenseNet_ECA model, the core architecture of which is composed of four ECA_Block modules. The ECA_Block module integrates dense connection and efficient channel attention mechanism, and the two ECA_Block modules are connected through a Transition module for feature dimension reduction and spatial down-sampling. The evaluation model training module 604 inputs the processed training set, internal validation set and external validation set into the DenseNet_ECA model to complete model training. The image evaluation module 605 inputs the to-be-detected X-ray image into the U-Net segmentation model for regional segmentation, and inputs the segmented to-be-detected X-ray image into the DenseNet_ECA model for evaluation to obtain an image evaluation result.

[0131] The division of the training set, internal validation set and external validation set of the image set classification module 601 establishes a multi-level verification system, ensuring the stability of the model under different data distributions. The segmentation model construction module 602 constructs a U-Net segmentation model, which effectively identifies the overlapping region of the scapula caused by the child's lying position through the multi-scale feature fusion capability, and provides an accurate region of interest for subsequent evaluation. The evaluation model construction module 603 constructs a DenseNet_ECA model, in which the dense connection promotes the reuse of features at different levels, and the channel attention mechanism dynamically adjusts the feature channel weight, so that the model focuses on the key anatomical structure in the child's chest film. The evaluation model training module 604 inputs the processed training set, internal validation set and external validation set into the DenseNet_ECA model for model training, and completes the optimization of the original DenseNet_ECA model. The Transition module reduces the feature dimension while preserving spatial information, preventing feature loss in deep networks. When the segmented to-be-detected X-ray image is input into the DenseNet_ECA model by the image evaluation module 605, the interference of irrelevant regions is excluded, so that the attention mechanism can more effectively act on the target region, improving the recognition accuracy of the special anatomical structure of the child.

[0132] In the operation process of the child chest film evaluation device of the embodiment, the image set classification module 601 randomly divides the pre-collected child chest X-ray images into a training set, an internal validation set and an external validation set. The segmentation model construction module 602 constructs a U-Net segmentation model, and performs regional segmentation on the training set through the U-Net segmentation model. The evaluation model construction module 603 constructs a DenseNet_ECA model, the core architecture of which is composed of four ECA_Block modules. The ECA_Block module integrates dense connection and efficient channel attention mechanism, and the two ECA_Block modules are connected through a Transition module for feature dimension reduction and spatial down-sampling. The evaluation model training module 604 inputs the processed training set, internal validation set and external validation set into the DenseNet_ECA model to complete model training. The image evaluation module 605 inputs the to-be-detected X-ray image into the U-Net segmentation model for regional segmentation, and inputs the segmented to-be-detected X-ray image into the DenseNet_ECA model for evaluation to obtain an image evaluation result. Figure 1 、 Figure 2 、 Figure 3 、 Figure 4 、 Figure 5The steps S101-S105, S201-S203, S301-S304, S401-S403 and S501-S502 of the child chest radiograph evaluation method in the foregoing embodiments are not repeated here.

[0133] Another embodiment of the present application discloses a child chest radiograph evaluation device based on deep learning, comprising: at least one processor, and a memory connected with the at least one processor in communication; wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to perform the child chest radiograph evaluation method as Figure 1 The steps S101-S105 of the control method in the foregoing embodiments, Figure 2 The steps S201-S203 of the control method in the foregoing embodiments, Figure 3 The steps S301-S304 of the control method in the foregoing embodiments, Figure 4 The steps S401-S403 of the control method in the foregoing embodiments, and Figure 5 The steps S501 and S502 of the control method in the foregoing embodiments.

[0134] Another embodiment of the present application discloses a computer readable storage medium, comprising: the storage medium stores computer executable instructions, and the computer executable instructions are used to enable a computer to perform the child chest radiograph evaluation method as Figure 1 The steps S101-S105 of the control method in the foregoing embodiments, Figure 2 The steps S201-S203 of the control method in the foregoing embodiments, Figure 3 The steps S301-S304 of the control method in the foregoing embodiments, Figure 4 The steps S401-S403 of the control method in the foregoing embodiments, and Figure 5 The steps S501 and S502 of the control method in the foregoing embodiments.

[0135] The device embodiments described above are only schematic, and units described as separate components can or can not be physically separate, i.e., can be located in one place, or can be distributed over multiple network units. Part or all of the modules can be selected according to actual needs to achieve the purpose of the present embodiment.

[0136] As will be appreciated by one of ordinary skill in the art, all or some steps, systems of the above-disclosed methods can be implemented as software, firmware, hardware, or suitable combinations thereof. Some or all of the physical components can be implemented as software executed by a processor, such as a central processing unit, a digital signal processor, or a microprocessor, or as hardware, or as an integrated circuit, such as an application-specific integrated circuit. Such software can be distributed on computer readable media, which can comprise computer storage media (or non-transitory media), and communication media (or transitory media). As is well known to those of ordinary skill in the art, the term computer storage media includes both volatile and non-volatile, removable and non-removable media implemented in any method or technology for storage of information such as computer readable instructions, data structures, program modules or other data. Computer storage media include, but is not limited to, RAM, ROM, EEPROM, flash memory or other memory technology, CD-ROM, digital versatile disks (DVD) or other optical disk storage, magnetic cassettes, magnetic tape, magnetic disk storage or other magnetic storage devices, or any other medium which can be used to store the desired information and which can be accessed by a computer. Further, as is well known to those of ordinary skill in the art, communication media typically embodies computer readable instructions, data structures, program modules, or other data in a modulated data signal, such as a carrier wave or other transport mechanism, and includes any information delivery media.

[0137] The embodiments of the present application are described in detail above with reference to the accompanying drawings, but the present application is not limited to the above-described embodiments, and various changes can be made within the scope of knowledge of those skilled in the art without departing from the gist of the present application. Furthermore, the embodiments of the present application and features in the embodiments can be combined with each other without conflict.

Claims

1. A deep learning-based method for evaluating pediatric chest X-rays, characterized in that, include: The pre-collected chest X-ray images of children were divided into a training set, an internal validation set, and an external validation set; Construct a U-Net segmentation model, and use the U-Net segmentation model to segment regions in the training set; A DenseNet_ECA model is constructed. The core architecture of the DenseNet_ECA model consists of four ECA_Block modules. The ECA_Block modules integrate dense connections and efficient channel attention mechanisms. The two ECA_Block modules are connected by a Transition module for feature dimensionality reduction and spatial downsampling. The processed training set, the internal validation set, and the external validation set are input into the DenseNet_ECA model to complete model training; The X-ray image to be detected is input into the U-Net segmentation model for region segmentation, and the segmented X-ray image to be detected is input into the DenseNet_ECA model for evaluation to obtain the image evaluation result.

2. The method for evaluating a child's chest X-ray according to claim 1, characterized in that, The construction of the U-Net segmentation model includes: A multi-task learning framework is constructed to train the corresponding U-Net segmentation model for each target region; An adaptive learning rate adjustment strategy combined with the Dice coefficient and cross-entropy loss function is used to perform multi-objective optimization on the U-Net segmentation model. The key parameters of the U-Net segmentation model are adjusted by hyperparameter grid search and cross-validation methods to obtain the optimized U-Net segmentation model for each region.

3. The method for evaluating a child's chest X-ray according to claim 1, characterized in that, In the DenseNet_ECA model, the ECA_Block module is composed of densely connected multi-layer networks. The four ECA_Block modules contain 6, 12, 24, and 16 layers respectively. Each network layer consists of a Dense_Layer feature extraction layer and an ECA feature recalibration module. The Dense_Layer feature extraction layer reduces the dimensionality of the input image by using 1×1 convolutions and performs feature extraction by using 3×3 convolutions. After each convolutional layer of the Dense_Layer, a PReLU activation function is used to provide non-linear modeling capability to the network.

4. The method for evaluating a child's chest X-ray according to claim 3, characterized in that, The specific processing steps of the ECA feature recalibration module for the input pediatric chest X-ray image are as follows: Global average pooling is performed on the child's chest X-ray image to obtain the global average value; A one-dimensional convolution of size K enables cross-channel information exchange between the target channel and its K neighboring channels; the kernel size K is calculated using the following formula: C is the number of input channels, b and To change the ratio between the number of channels and the kernel size, `odd` means to select the nearest odd number upwards; The global average value is input into the Sigmoid function for mapping, resulting in a set of scalars from 0 to 1 as the target weights for each channel; The original input pediatric chest X-ray image is weighted according to the target weights described in the group to achieve recalibration of the feature channels and obtain a feature map with attention.

5. The method for evaluating a child's chest X-ray according to claim 3, characterized in that, The PReLU activation function includes the following formula: Where x is the function input value and a is the corrected linear unit parameter. By adaptively learning the corrected linear unit parameter, the function value will not be 0 directly when the function input value is negative.

6. The method for evaluating a child's chest X-ray according to claim 1, characterized in that, The step of inputting the processed training set, the internal validation set, and the external validation set into the DenseNet_ECA model to complete model training includes: The parameters are optimized based on the processed training set to fit the data patterns of the DenseNet_ECA model. The performance of the DenseNet_ECA model under the same data distribution is evaluated based on the internal validation set. The applicability and generalization ability of the DenseNet_ECA model under different data distributions are tested using the external validation set.

7. The method for evaluating a child's chest X-ray according to claim 1, characterized in that, After inputting the X-ray image to be detected into the U-Net segmentation model for region segmentation, and inputting the segmented X-ray image to be detected into the DenseNet_ECA model for evaluation to obtain the image evaluation result, the method further includes: Based on Kappa analysis, the consistency of the image evaluation results is checked according to the results of human evaluation. The DenseNet_ECA model was evaluated in multiple dimensions based on accuracy, sensitivity, precision, and F1 score to obtain the model performance evaluation results.

8. A deep learning-based pediatric chest X-ray assessment device, characterized in that, include: The image set classification module is used to randomly divide pre-collected pediatric chest X-ray images into a training set, an internal validation set, and an external validation set; A segmentation model construction module is used to construct a U-Net segmentation model and perform region segmentation on the training set using the U-Net segmentation model. The evaluation model building module is used to build the DenseNet_ECA model. The core architecture of the DenseNet_ECA model consists of four ECA_Block modules. The ECA_Block modules integrate dense connections and efficient channel attention mechanisms. The two ECA_Block modules are connected by a Transition module for feature dimensionality reduction and spatial downsampling. The evaluation model training module is used to input the processed training set, the internal validation set, and the external validation set into the DenseNet_ECA model to complete model training; The image evaluation module is used to input the X-ray image to be detected into the U-Net segmentation model for region segmentation, and then input the segmented X-ray image to be detected into the DenseNet_ECA model for evaluation to obtain the image evaluation result.

9. A deep learning-based pediatric chest X-ray assessment device, characterized in that, include: At least one processor, and, A memory communicatively connected to the at least one processor; wherein, The memory stores instructions that can be executed by the at least one processor to enable the at least one processor to perform the deep learning-based pediatric chest X-ray assessment method as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer-executable instructions for causing a computer to perform the deep learning-based pediatric chest X-ray assessment method as described in any one of claims 1 to 7.

Images