Lung disease classification model construction method and device

Abstract

The invention discloses a lung disease classification model construction method and device, and relates to the field of computer-aided diagnosis and treatment and image recognition, and the model construction method comprises the steps: constructing a sample data set with labels; the sample data set with the labels comprises a plurality of chest X-ray images and lung disease categories corresponding to the chest X-ray images; a classification model based on Mobile-Former is constructed; the classification model based on Mobile-Former comprises an enhanced Mobile-Former module group, a feature fusion module and a classification module which are connected in sequence; and training the classification model based on the Mobile-Former by adopting the sample data set with the label to obtain a trained classification model based on the Mobile-Former, and taking the trained classification model as the lung disease classification model. According to the invention, the accuracy of lung disease classification by using the chest X-ray image is improved.

Description

Technical Field

[0001] This application relates to the fields of computer-aided diagnosis and image recognition, and in particular to a method and apparatus for constructing a classification model for lung diseases. Background Technology

[0002] Lung infections caused by viral and bacterial pathogens, such as pneumonia and COVID-19, have become a major challenge in global public health due to their high morbidity and mortality rates. Notably, early and accurate diagnosis of lung diseases plays a crucial role in reducing mortality and promoting effective clinical intervention.

[0003] Among existing neural network-based methods for diagnosing lung diseases, the Transformer model, while effectively capturing global dependency patterns, suffers from high computational complexity. Convolutional neural networks (CNNs) have inherent limitations in modeling long-range dependencies and global context, impacting their performance in lung disease classification. Therefore, a solution capable of accurate lung disease classification is urgently needed. Summary of the Invention

[0004] The purpose of this application is to provide a model construction method and apparatus capable of accurate classification of lung diseases. To achieve the above objective, this application provides the following solution: In a first aspect, this application provides a method for constructing a lung disease classification model, characterized in that the method includes: Construct a labeled sample dataset; the labeled sample dataset includes multiple chest X-ray images and the corresponding lung disease categories for each chest X-ray image; A classification model based on Mobile-Former is constructed; the classification model based on Mobile-Former includes, in sequence: an enhanced Mobile-Former module group, a feature fusion module, and a classification module; The Mobile-Former-based classification model is trained using the labeled sample dataset to obtain the trained Mobile-Former-based classification model, which serves as the lung disease classification model.

[0005] The enhanced Mobile-Former module group includes N sequentially connected enhanced Mobile-Former modules; each of the N sequentially connected enhanced Mobile-Former modules includes: a first summation sub-block, a second summation sub-block, a first bidirectional cross-attention sub-block, a second bidirectional cross-attention sub-block, a Mobile sub-block, and a Former sub-block; The output of the first bidirectional cross-attention sub-block is connected to the input of the first summation sub-block; the output of the first summation sub-block is connected to the input of the Former sub-block, the output of the Former sub-block is connected to the input of the second bidirectional cross-attention sub-block, the output of the second bidirectional cross-attention sub-block is connected to the input of the second summation sub-block, and the output of the Mobile sub-block is connected to another input of the second summation sub-block; the output of the Former sub-block serves as the first output, and the output of the second summation sub-block serves as the second output. The first enhanced Mobile-Former module receives the labeled sample dataset as its input, and its input is connected to the inputs of the Mobile sub-block and the first bidirectional cross-attention sub-block. The second to Nth enhanced Mobile-Former modules each have two inputs: a first input and a second input. The first input is connected to the first output of the previous enhanced Mobile-Former module, and the second input is connected to the second output of the previous enhanced Mobile-Former module.

[0006] The feature fusion module includes two input terminals, namely a third input terminal and a fourth input terminal; The third input terminal is connected to the first output terminal of the Nth enhanced Mobile-Former module, and the fourth input terminal is connected to the second output terminal of the Nth enhanced Mobile-Former module; The third input terminal is connected to the input terminal of the third summation sub-block, and the fourth input terminal is connected to the input terminal of the improved parallel block perceptual attention module; the output terminal of the improved parallel block perceptual attention module is connected to the input terminal of the third bidirectional cross-attention module and the input terminal of the improved wavelet-guided energy recalibration module; the output terminal of the third bidirectional cross-attention module is connected to another input terminal of the third summation sub-block, and the output terminal of the third summation sub-block is connected to the input terminal of the fusion sub-block; the output terminal of the improved wavelet-guided energy recalibration module is connected to another input terminal of the fusion sub-block, and the output terminal of the fusion sub-block serves as the output terminal of the feature fusion module.

[0007] The classification module comprises, in sequence: a first linear layer, a Swish activation function, a second linear layer, a ReLU activation function, a Dropout layer, and a third linear layer; the first linear layer is connected to the output of the feature fusion module.

[0008] The Mobile sub-block includes multiple sets of sequentially connected first 1×1 pointwise convolutional layers, first dynamic ReLU, 3×3 depthwise convolutional layers, second dynamic ReLU, second 1×1 pointwise convolutional layers, and multilayer perceptron sub-blocks; The Former sub-block includes a multi-head attention layer and a feedforward network connected in sequence; The input of the multilayer sensing sub-block is connected to the output of the second bidirectional cross-attention sub-block, and the output of the multilayer sensing sub-block is connected to the first dynamic ReLU and the second dynamic ReLU.

[0009] The improved parallel block-aware attention module includes a multi-branch feature extraction module, a multi-statistic high-efficiency channel attention module, a depth-separable attention collaboration module, and a random path discarding module connected in sequence. The multi-statistic efficient channel attention module includes a first adaptive average pooling sub-block, a first max pooling sub-block, a first standard deviation pooling sub-block, a one-dimensional convolutional block, and an activation function; The inputs of the first adaptive average pooling sub-block, the first max pooling sub-block, and the first standard deviation pooling sub-block are all connected to the output of the multi-branch feature extraction module; the outputs of the first adaptive average pooling sub-block, the first max pooling sub-block, and the first standard deviation pooling sub-block are added together and then connected to the input of the one-dimensional convolution sub-block. The output of the one-dimensional convolutional sub-block is connected to the input of the activation function; the output of the activation function and the input of the one-dimensional convolutional sub-block are added together, and then multiplied by the output of the multi-branch feature extraction module, which is then used as the output of the multi-statistic efficient channel attention module.

[0010] The depthwise separable convolutional module includes: a channel partitioning sub-block, a first depthwise convolutional sub-block, and a context-aware feature space attention module; The input of the channel partitioning submodule is connected to the output of the multi-statistic efficient channel attention module; the channel partitioning submodule is used to divide the output of the multi-statistic efficient channel attention module into four branches along the channel dimension, namely the identity branch, the square kernel branch, the vertical strip branch, and the horizontal strip branch; The input of the first deep convolutional sub-block is connected to the output of the channel partitioning sub-block. The first deep convolutional sub-block is used to capture the features of each of the four branches. After concatenating the features of each of the four branches in the channel dimension, a spatial attention mask is generated by using the first Sigmoid activation function. The spatial attention mask is multiplied by the output of the multi-statistic efficient channel attention module to obtain the concatenated features. The context-aware feature space attention module includes a second adaptive average pooling sub-block, a second max pooling sub-block, a second standard deviation pooling sub-block, a second depthwise convolution sub-block, a second sigmoid activation function, and a feature optimization sub-block; the feature optimization sub-block includes a spatial dimension mean sub-block, a fully connected sub-block, and a third sigmoid activation function connected in sequence; the second depthwise convolution sub-block includes a 7×7 convolution sub-block, a batch normalization sub-block, a ReLU activation function, and a 1×1 convolution sub-block connected in sequence; The inputs to the second adaptive average pooling sub-block, the second max pooling sub-block, and the second standard deviation pooling sub-block are all the concatenated features; the outputs of the second adaptive average pooling sub-block, the second max pooling sub-block, and the second standard deviation pooling sub-block are concatenated along the channel dimension and then input to the second depthwise convolution sub-block and the second sigmoid activation function, which are then output; the output of the second sigmoid activation function is multiplied by the concatenated features and then input to the feature optimization sub-block.

[0011] The improved wavelet-guided energy recalibration module includes: a 3×3 depth convolution sub-block, a first convolution sub-block, a second convolution sub-block, a first center shear sub-block, a second center shear sub-block, and a SimAM simple parameterless attention sub-block; the first convolution sub-block and the second convolution sub-block each include a fixed wavelet convolution sub-block and a channel grouping convolution sub-block connected in sequence; The input of the 3×3 depth convolutional sub-block is connected to the output of the depth-separable attention collaboration module, and the output of the 3×3 depth convolutional sub-block is connected to the input of the first convolutional sub-block and the second convolutional sub-block; the first convolutional module is used to decompose the output features of the 3×3 depth convolutional sub-block into low-frequency components, and the second convolutional module is used to decompose the output features of the 3×3 depth convolutional sub-block into high-frequency components; The output of the first convolutional sub-block is connected to the input of the first central shearing sub-block; the output of the second convolutional sub-block is connected to the input of the second central shearing sub-block, and the output of the second central shearing sub-block is connected to the input of the SimAM simple parameterless attention sub-block. The SimAM simple parameterless attention sub-block is used to calculate the weight of each spatial location in the high-frequency feature, and multiply the weight of each spatial location of the high-frequency feature by the feature value of the corresponding spatial location to obtain the enhanced high-frequency feature. The low-frequency features and the enhanced high-frequency features are added together, and then multiplied by the output of the improved parallel block perception attention module, which is then used as the output of the improved wavelet-guided energy recalibration module.

[0012] The lung disease classification model also includes a lightweight bottleneck module. The output of the lightweight bottleneck module is connected to the input of the Mobile sub-block. The input of the lightweight bottleneck module is the chest X-ray image. The lightweight bottleneck module includes a 3×3 depth convolution sub-block and a 1×1 pointwise convolution sub-block connected in sequence.

[0013] Secondly, this application provides a lung disease classification device, characterized in that the lung disease classification device includes: an acquisition device and a host computer, the acquisition device being connected to the host computer; the host computer is provided with a lung disease classification model, and the host computer is used to classify lung diseases based on chest X-ray images acquired by the acquisition device.

[0014] According to the specific embodiments provided in this application, the following technical effects are disclosed: This application provides a method and apparatus for constructing a lung disease classification model. The lung disease classification model achieves efficient information exchange by introducing a bidirectional cross-attention module in the Mobile-Former, thereby fusing local and global features. Adaptive average pooling, max pooling, and standard deviation pooling are introduced into the multi-statistic efficient channel attention module to solve the information loss problem caused by single pooling. Channel partitioning and feature extraction in the deep separable attention collaboration module enable the model to selectively capture spatial and contextual features of different dimensions in the input. Channel fusion is achieved through two steps: depthwise convolution and pointwise convolution in the depthwise convolution sub-block. Batch normalization and ReLU activation functions are introduced during this process to preserve channel feature interaction capabilities while further reducing computational overhead. The SimAM simple parameterless attention mechanism is introduced into the improved wavelet-guided energy recalibration module, which further enhances high-frequency feature information through adaptive weighting, suppresses noise interference in the high-frequency feature map, and retains more complete features of the input image. In summary, this application can achieve accurate classification of lung diseases based on chest X-ray images, providing scientific guidance for the early diagnosis of lung diseases. Attached Figure Description

[0015] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0016] Figure 1 This is a schematic diagram of the improved parallel block-aware attention module of this application; Figure 2 This is a schematic diagram of the improved wavelet-guided energy recalibration module of this application; Figure 3 This is a framework diagram of the enhanced Mobile-Former module group in this application; Figure 4 This is a test image according to an embodiment of this application; wherein, Figure 4 (a) shows test samples in the COVID dataset categorized as normal, COVID-19, and pneumonia. Figure 4 (b) Test samples for the normal, COVID-19, and pneumonia categories in the Curated dataset; Figure 5 This is a confusion matrix generated on a COVID dataset according to an embodiment of this application; Figure 6 This is a confusion matrix generated on the Curated dataset according to an embodiment of this application; Figure 7 This is a schematic diagram of the structure of a computer device provided in an embodiment of this application. Detailed Implementation

[0017] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0018] To make the above-mentioned objectives, features and advantages of this application more apparent and understandable, the application will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0019] 1. Parallel Block-Aware Attention Mechanism Based on Inception Architecture Parallelized patch-aware attention (PPA) is a typical multi-branch feature extraction method capable of efficiently capturing multi-scale and multi-dimensional feature details. This study, based on PPA, innovatively proposes an Inception-style parallelized patch-aware attention (IPPA) mechanism based on the Inception architecture to enhance the representation ability of spatial and multi-scale features. This mechanism guides the model to focus on key image regions, strengthening its ability to capture complex features and local patterns, thereby enhancing pathological representation and optimizing model performance.

[0020] Specifically, this study introduces depthwise separable convolution into the Inception architecture and combines it with context-aware feature spatial attention (CFSA) to design the Inception depthwise attention synergy (IDAS) module.

[0021] Furthermore, this study constructed a Multi-statistic Efficient Channel Attention (MS-ECA) module to provide the model with more discriminative channel attention features. The IPPA mechanism mainly consists of three basic modules: multi-branch feature extraction, IDAS, and MS-ECA, as illustrated in the diagram below. Figure 1 As shown, the specific implementation steps are as follows: Step 1: Since the multi-branch feature extraction module involves a series of convolutional, local, and global block-aware attention branches, this study first inputs the feature map into the local and global branches of this module to capture feature information at different scales. Assume the input feature map is represented as... ,in Indicates the number of input channels. and Corresponding to The height and width of the input feature map. First pass Convolution (i.e.) Processing, to obtain the transformed features In block perception, After being unfolded and reshaped, it is decomposed into a spatially continuous block structure with dimensions of... , This is the block size parameter. Then, a channel averaging operation is performed on it, resulting in a dimension of... The feature maps are then further input into a feed-forward network (FFN) consisting of two multilayer perceptrons (MLPs) and layer normalization (LayerNorm) to perform linear transformation. The spatial probability distribution is then calculated using the Softmax activation function to achieve feature selection. Finally, based on task relevance, the weights of tokens (i.e., labels, a set of learnable latent representation vectors initialized randomly and not directly related to the input image itself; through attention interaction with convolutional features, these labels can dynamically encode the global context information of the current image and guide and modulate the convolutional network during feature extraction and representation learning) and channels are adaptively adjusted. This study sets segmentation size parameters. ,in Corresponding to the local block-aware branch, while This is a global block-aware branch. The transformed features. Perception through two blocks Branch generation corresponds to local features and global features In mathematics, it can be represented as (1) Step 2: A series of convolutional branches use three... convolution( The sequential structure composed of layers expands the receptive field, enabling the capture of mesoscale pathological features. The output features of this branch are represented as follows: ,in This corresponds to each convolutional layer within the branch. The specific operation of this process is described below: (2) in, express convolution, and These are the ReLU activation function and batch normalization, respectively.

[0022] Step 3: Initial features The features extracted from the three branches are then aggregated, as shown in the following expression: (3) in, This represents the fused feature representation. By employing residual connections and multi-scale fusion, the model effectively reduces the single-branch limitation, thereby further optimizing the attention modulation effect.

[0023] Step 4: To refine the response characteristics of fused features in the channel dimension, this study introduces an MS-ECA module based on the ECA mechanism. This module will fuse the features... Adaptive average pooling, max pooling, and standard deviation pooling are applied sequentially to mitigate information loss caused by a single pooling method. The outputs of the different pooling methods are then summed along the channel dimension to obtain the features. The calculation process can be expressed as follows: (4) in, , and These correspond to adaptive average pooling, max pooling, and standard deviation pooling operations, respectively. The generated features... After one-dimensional convolution ( Further processing is performed to obtain the output features. This study introduces an activation function to maintain the continuity of channel features through residual connections, thereby enhancing the ability to model channel dependencies. Specifically, this can be expressed mathematically as follows: (5) in, This represents the features after activation and residual integration. This is the Sigmoid activation function. Subsequently... With fusion feature map Element-wise multiplication yields the channel-weighted features used for adaptive enhancement. This serves as the final output feature representation of the MS-ECA module.

[0024] Step 5: To improve feature representation capabilities while optimizing computational efficiency and simplifying model parameters, this study uses the IDAS module to weight channel features. The input is based on an Inception structure with depthwise separable convolutions. Channel fusion is achieved through two steps: depthwise convolution and pointwise convolution. Batch normalization (BN) and ReLU activation functions are introduced during the process to preserve the interaction capabilities of channel features while further reducing computational overhead. It is worth noting that the features... The system is divided into four distinct branches along the channel dimension: the identity branch, the square kernel branch, the vertical stripe branch, and the horizontal stripe branch. The square kernel branch and the stripe branch each have a set number of channels. ( (This is the number of output channels), and the remaining channels are allocated to the identity branch. The specific operation is as follows: (6) in, and These correspond to the characteristics of the identity branch, square kernel branch, vertical strip branch, and horizontal strip branch outputs, respectively. Represented as Depthwise separable convolution ( (), which aims to capture the two-dimensional local spatial features of the input. Corresponding to Depthwise separable convolution ( ), used to capture long-range dependencies in the horizontal axis direction. The table is as follows Depthwise separable convolution ( This extracts long-range dependency features in the vertical direction. Based on the branched convolution design described above, the model is able to selectively capture spatial and contextual features of different dimensions in the input.

[0025] Step 6: Fuse features from separable convolutions of different depths in the IDAS module along the channel dimension, then use Sigmoid activation to generate the corresponding spatial attention mask, and combine it with the features. Modulation characteristics are obtained by element-wise multiplication. : (7) in, This indicates a feature splicing operation.

[0026] Step 7: Inspired by spatial attention mechanisms, this study designed a CFSA module to further optimize lesion region discrimination by enhancing relevant spatial features. Specifically, in the CFSA module, adaptive average pooling, max pooling, and standard deviation pooling are applied to the features. The outputs of each pooling operation are then concatenated along the channel dimension to obtain statistical features. This step can be represented as (8) Statistical characteristic graph implement Convolution (i.e.) ), to extract spatial features, the results of which are expressed as Refined spatial features Batch normalization and ReLU activation are performed sequentially: (9) Subsequently, the activated features implement Pointwise convolution (denoted as) ) and Sigmoid activation to generate a spatial attention map : (10) This attention map is used to adaptively weight channels, enhancing feature representations by selectively highlighting important spatial correlations. Ultimately, the features... With attention map Perform element-wise multiplication to generate the output features of the CFSA module. .

[0027] Step 8: Spatial weighted features Dynamic feature selection is performed to generate dynamic channel weights, achieving effective feature optimization. First, the mean of the spatially weighted features along the spatial dimension is calculated to evaluate the global importance of each channel. Then, dynamic weights are obtained through a fully connected layer and a sigmoid activation function. : (11) in, This indicates a global average pooling operation, while This is for fully connected layer operations. This study uses spatially weighted features. With dynamic weights Element-wise multiplication is performed to refine the contribution of each channel, and then residual fusion features are used to formally represent the process. (12) in, It is the final transformed feature output by the IDAS module.

[0028] Step 9: Features generated by the IDAS module DropPath (random path discarding) regularization is performed, and grouping normalization (GN) is introduced to ensure the stability of feature scaling. A nonlinear transformation is introduced through the GELU activation function to further enhance feature representation capabilities. Finally, the enhanced features obtained by the IPPA mechanism are... Can be represented as (13) in, Represents the GELU activation function. and These represent DropPath regularization and group normalization operations, respectively.

[0029] In summary, the IPPA mechanism proposed in this study can enhance the representation ability of spatial and multi-scale features, enabling the model to more effectively focus on key diagnostic regions, thereby further improving the performance of lung disease detection tasks.

[0030] 2. Wavelet-guided energy recalibration module Wavelet transform, by constructing a spatial-frequency localized representation, enables the model to decompose features at multiple scales, effectively distinguishing low-frequency structural information from high-frequency detail components, thus compensating for the shortcomings of traditional spatial convolution in frequency domain representation. Based on this, this study designs a wavelet-guided energy recalibration (WGER) module based on wavelet transform and attention enhancement to achieve multi-scale modeling of complex structures and lesion features. Figure 2 The structural diagram of this module is shown, and the specific implementation is as follows: Step 1: First, perform the following steps on the input feature map: Depthwise convolution ( This operation enhances the local spatial representation capability. Assume the input feature map is... ,in , and They represent The number of channels, height, and width. (Through) Feature maps generated by depthwise convolution The specific process can be expressed as (14) in, express Depthwise convolution operation.

[0031] Step 2: This study introduces a fixed wavelet convolution module to process the features. It is decomposed into low-frequency structural components and high-frequency detail components. To maintain the continuity of horizontal texture and efficiently decompose the high-frequency and low-frequency components in the horizontal direction, features... First, a horizontally symmetrical mirror padding process is performed, followed by a channel grouping convolution (GConv) operation. The feature maps of the low-frequency and high-frequency components generated in this process are denoted as follows: and Its formal representation is as (15) in, The horizontal symmetrical mirror fill operation is designed to mitigate edge distortion and ensure texture continuity. and These represent low-pass and high-pass grouped convolutions in the horizontal direction, respectively.

[0032] Step 3: To achieve separation of low-frequency and high-frequency components, this study analyzes the feature map obtained after horizontal transformation. and Further, symmetrical mirror padding is performed in the vertical direction, followed by channel-grouped convolution to complete the feature transformation in the vertical direction. The corresponding output feature map is denoted as... and , can be expressed as (16) in, This indicates a symmetrical mirror fill operation in the vertical direction to mitigate edge distortion. and The low-pass and high-pass groups are convolutionally applied vertically. After the above two-dimensional decomposition, and The primary encoding is coarse-grained anatomical structural information, while and It focuses on capturing fine-grained pathological details such as lesion boundaries and local textures.

[0033] Step 4: Crop the vertically transformed feature map to... The size of the central target region is used to obtain the low-frequency and high-frequency feature maps of the final decomposition (i.e., and The specific calculation formula is as follows: (17) in, This is a cropping operation targeting the center area. It's worth noting that after cropping... and It preserves the channel dimension of the original feature map and satisfies .

[0034] Step 5: This study introduces SimAM (Simple parameter-free attention module) attention into the WGER module, further enhancing high-frequency feature information through adaptive weighting to suppress noise interference in the high-frequency feature map. Assumption Represents the given channel in the feature map. line, number The characteristic values ​​of the column, where In the WGER module, SimAM calculates the spatial mean for each channel. With sample variance And calculate each spatial location based on the energy function. The corresponding minimum energy value To quantify the distinguishability of this location feature from the surrounding area, its mathematical expression is: (18) in, This is a very small regularization constant. This study will... Importance weight is defined as The high-frequency enhanced feature map is then multiplied element-wise with the original high-frequency features. This selectively enhances diagnostic-relevant details while suppressing noise. The specific process can be defined as follows: (19) Step 6: Enhance the high-frequency feature map Low-frequency feature map that preserves the entire thoracic anatomy The feature maps are fused to obtain enhanced feature maps. This feature map Compared with the original input features Element-wise multiplication yields the fine features output by the WGER module. : (20) Based on the above steps, the WGER module constructed in this study organically combines wavelet decomposition with attention mechanism, effectively maintaining anatomical structural information while enhancing key lesion features related to diagnosis, thereby improving the model's ability to identify lung lesions.

[0035] 3. A lung disease diagnosis method based on an enhanced Mobile-Former network Lung infections caused by viral and bacterial pathogens have become a major global public health challenge. To address this issue, this study proposes a distributed attention-driven enhanced Mobile-Former network architecture for accurate diagnosis of lung diseases via X-ray. The core objective of this research is to accurately identify pneumonia, COVID-19, and normal cases, providing reliable decision support for clinical diagnostic assistance. The enhanced Mobile-Former network proposed in this study integrates a distributed attention paradigm composed of innovative IPPA and WGER modules on top of the original model, aiming to effectively improve the model's predictive performance for lung diseases. Figure 3 As shown, this study provides an enhanced Mobile-Former network framework diagram for the diagnosis of lung diseases, which is described in detail below: Step 1: The enhanced Mobile-Former network integrates MobileNet and Transformer modules through bidirectional cross-attention (CSA) to capture both local and global features. First, the image is input to... The backbone formed by convolution, through Depth convolution and A lightweight bottleneck module constructed using pointwise convolution is used to adjust the channel dimension.

[0036] Step 2: This study utilizes a series of Mobile-Former modules to extract local features and model global information. Each Mobile-Former module involves MobileNet (… ) and Transformer ) sub-blocks, and bidirectional CSA mechanism ( and ). The sub-blocks use an inverted bottleneck structure of Dynamic ReLU (DY-ReLU) to enhance feature learning capabilities. Specifically, the first... layer Sub-blocks of input features proceed in sequence Pointwise convolution and Depthwise convolutions are used to extract local features, and DY-ReLU is inserted between two convolutional processes to achieve dynamic feature scaling. Sub-blocks then process learnable tokens. ,in and These represent the number and dimension of tokens, respectively, and global dependencies are captured through Multi-head Attention (MHA) and FFN. Furthermore, the above sub-modules interact using a bidirectional CSA mechanism to achieve the fusion of local and global features. From the local feature map... To global token The calculation process for the forward CSA and the corresponding reverse CSA is as follows: (twenty one) in, and Classified as One point of attention, namely and . This represents the standard MHA mechanism, which executes queries in parallel on attention heads. ),key( ) and value ( Projection operation. For The mechanism involves concatenating the MHA outputs of each attention head before outputting them through an output matrix. This process is used to transfer information from local features to the global token. Conversely, in the opposite direction... Mechanism projection and The key and value are transformed and concatenated to obtain the final output. Both attention operations achieve the fusion of local and global features, thus realizing efficient information exchange and feature refinement. It is worth noting that the relevant parameters are obtained by using two MLP layers ( ) Applied to The global token output of the submodule is obtained, thus avoiding reliance on average pooling operations. After completing multi-head attention calculation and parameter generation, local features are updated. and global token , and enter the next Iterative optimization is performed within the module.

[0037] Step 3: Input features are iteratively stacked After obtaining the Mobile-Former module, local features are obtained. With global features First, local features. Enhancement is performed using the IPPA mechanism to generate features. The corresponding mapping relationship is as follows (twenty two) in, This represents the IPPA module. Optimized local features. Through CSA mechanism and global features By merging, a more refined global representation can be obtained. To further enhance the detail of the lesions, the above-mentioned enhanced local features... The WGER module is then used for processing to generate high-frequency refined local features. The process can be described as follows: (twenty three) in, This represents the WGER module. Ultimately, the refined local features... With global features By concatenating along the feature dimensions, aggregated features are obtained. The mathematical formula is expressed as (twenty four) Step 4: In the final stage of the enhanced Mobile-Former network, aggregate features. The input consists of a classification module with three linear layers to achieve lung disease detection based on X-ray images. Specifically, firstly, the input feature vector... The first linear layer maps the features to a high-dimensional feature space, while the Swish activation function introduces non-linearity, enriching the feature representation and adapting it to complex distributions. Subsequently, a second linear layer further transforms the high-dimensional features linearly and combines it with the ReLU activation function to enhance non-linear expressiveness, while Dropout regularization mitigates the risk of overfitting. Finally, a third linear layer maps the processed features to the target class space, achieving multi-class classification. The overall design, through the combination of layer-by-layer linear transformations and non-linear activations, provides high-quality feature representations for subsequent classification, thereby effectively improving the overall performance of lung disease prediction.

[0038] Following the steps above, the application of the distributed attention-driven enhanced Mobile-Former model proposed in this study in the diagnosis of lung diseases is expected to accurately distinguish between pneumonia, COVID-19, and normal cases, providing a reliable and effective technical reference for clinical auxiliary diagnosis.

[0039] To verify the effectiveness of the proposed scheme in the three-class classification of lung diseases, this study conducted a series of experiments on two publicly available X-ray datasets and quantified its performance through numerical analysis. Different evaluation metrics and confusion matrix analysis were used as the core evaluation methods in the experimental phase to systematically demonstrate the accuracy and robustness of the scheme in lung disease classification tasks.

[0040] 1. Dataset Description and Environment Setup For a three-class classification task of lung diseases, this study conducted experiments on two publicly available datasets: the COVID-19 Radiography Dataset and the Curated X-Ray Dataset. The COVID-19 Radiography Dataset (COVID) contains 3616 chest X-ray images of COVID-19, 10192 normal images, 6012 images of lung opacities, and 1345 images of pneumonia. This study focuses on the three-class classification task of lung diseases, therefore only using the 3616 COVID-19, 10192 normal, and 1345 pneumonia images. This dataset is from the Kaggle database. In addition, the Curated X-ray Dataset (Curated) contains 1281 chest X-ray images of COVID-19, 3270 normal images, 1656 images of viral pneumonia, and 3001 images of bacterial pneumonia. This study merged the viral and bacterial pneumonia categories into a single pneumonia category, resulting in 4657 pneumonia images. Specifically, the Curated dataset can be obtained from the Kaggle database.

[0041] like Figure 4As shown, this study presents test image samples with clear annotations and a uniform size of 224×224 pixels selected from the two datasets mentioned above. Among them, Figure 4 (a) shows test samples in the COVID dataset categorized as normal, COVID-19, and pneumonia. Figure 4 (b) are test samples of the above three categories in the Curated dataset.

[0042] To conduct a comprehensive and systematic evaluation of the model, this study randomly divided the dataset into three subsets—training, validation, and testing—in a 7:1:2 ratio. Table 1 shows the distribution statistics of the training, validation, and testing data using the COVID and Curated datasets.

[0043] Table 1. Distribution statistics of training, validation, and test sets for different datasets used in this study.

[0044] During model training, the Adam optimizer was used with a batch size of 32 and an initial learning rate of 1e-4, for a total of 250 training epochs. To enhance the model's robustness and generalization ability, the training data underwent consistency processing. and Rotation, color jittering, and Gaussian noise enhancement were applied to systematically enlarge the training data sample. Furthermore, online data augmentation operations were performed on the augmented training set, including a 50% probability random horizontal flip and a random affine transformation (translation amplitude of 5%, scaling range of 0.8–1.1, and shearing angle of ). All X-ray images were uniformly adjusted to 224×224 pixels to meet the input requirements of the designed Mobile-Former architecture. Furthermore, image pixel values ​​were normalized according to the channel mean [0.485, 0.456, 0.406] and standard deviation [0.229, 0.224, 0.225] to ensure input consistency and improve computational stability.

[0045] To validate the diagnostic performance of the enhanced Mobile-Former network for lung diseases under different clinical conditions, this study performed three independent random partitions on two publicly available datasets. Notably, the model used the same parameter settings and experimental procedures on all datasets to ensure the consistency and reliability of the results. This study used evaluation metrics including accuracy (Acc), precision (Pre), recall (Rec), and F1 score (F1) to quantify the model's performance. To mitigate the impact of random fluctuations, the mean and standard deviation were calculated based on the experimental results for each dataset. This systematic experimental design and performance evaluation effectively validated the predictive ability of the proposed network framework in classifying pneumonia, COVID-19, and normal cases, thus providing reliable and valuable technical support for the automated diagnosis of lung diseases.

[0046] 2. Experimental Results and Analysis This study conducted a series of experiments on the COVID and Curated datasets to verify the effectiveness of the proposed method. Table 2 shows the experimental results of the proposed method on the two public datasets, COVID and Curated, for classifying pneumonia, COVID-19, and normal samples. Specifically, the proposed method performs exceptionally well on the COVID dataset, achieving an average accuracy of 98.67%, demonstrating excellent diagnostic performance for lung diseases. The third random split (Split-3) achieved the highest accuracy at 98.71%. Furthermore, the precision of the third random split reached 98.47%, indicating that the model also has high recognition accuracy for positive lung disease samples. Therefore, based on the above experimental results, the proposed method possesses stable and excellent diagnostic capabilities on the COVID dataset. Notably, the proposed method also exhibits excellent performance on the Curated dataset, with an average accuracy of 98.64% for three random splits, an even higher average precision of 98.88%, and a precision of 98.91% for the third random split. Furthermore, the recall and F1 score of the second random split (Split-2) were 98.78% and 98.83%, respectively. Overall, the proposed method demonstrates significant advantages in classification accuracy on the Curated dataset, and also exhibits good consistency across key evaluation metrics. Its performance stability is essentially on par with its performance on the COVID dataset. Therefore, the proposed method not only possesses superior classification accuracy for the three-class classification task of lung diseases, but also demonstrates good reliability in terms of performance stability and result consistency, fully validating its feasibility, effectiveness, and robustness in the auxiliary diagnosis of lung diseases.

[0047] Table 2. Experimental data of the proposed method for three-class classification of lung diseases on the COVID and Curated datasets.

[0048] Confusion matrices are an important tool for measuring the classification performance of a model, visually presenting the model's ability to distinguish between different categories and its misclassification patterns. To comprehensively validate the performance of the proposed model in the identification tasks of three types of lung diseases—pneumonia, COVID-19, and normal lungs—this study conducted confusion matrix analysis on classification results obtained from three independently and randomly partitioned datasets of COVID and Curated.

[0049] Figure 5 The confusion matrix of the model in three experiments on the COVID dataset is shown, where Figure 5 (a)-(c) correspond to the experimental results under three different data partitions. It can be observed that the model consistently exhibits relatively better true positive identification performance in the normal category. In the 2039 normal category samples in this dataset, the number of true positives obtained from the three different data partition experiments were 2022, 2026, and 2030, respectively, with corresponding false negatives of only 17, 13, and 9, indicating that the model's false negative rate in this category remains at a low level. Furthermore, the model also shows stable and good detection capabilities for samples in the COVID-19 and pneumonia categories. In 724 COVID-19 category samples, the highest number of true positives was 708; in 270 pneumonia category samples, the highest number of true positives was 263. Overall, the model's classification performance on the COVID dataset demonstrates high accuracy and stability, especially showing strong diagnostic capabilities in the normal and pneumonia categories.

[0050] Figure 6 The confusion matrix for the model's three-class classification of lung diseases on the Curated dataset is provided, where Figure 6 (a)-(c) show the experimental results for three different data partitions. In the 257 COVID-19 cases, the true positive numbers from the three random partitions were 253, 256, and 249, respectively, all corresponding to low false negative levels (4, 1, and 8, respectively), indicating a low risk of misclassification for this category. Furthermore, in the 654 normal category samples, the model achieved a maximum of 650 true positives; in the 933 pneumonia category samples, the highest true positive number was 921, demonstrating the excellent predictive performance of the proposed scheme across different categories. Therefore, this study also exhibits high classification accuracy on the Curated dataset, effectively reducing misclassification while maintaining the reliability of overall predictive performance. Based on the above experimental results and analysis, the proposed model demonstrates significant effectiveness in the three-class classification task of lung diseases, laying a solid foundation for its application in practical clinical lung disease auxiliary diagnosis scenarios.

[0051] Table 3 shows the evaluation results of the proposed scheme on different datasets for classifying pneumonia, COVID-19, and normal cases.

[0052] To rigorously verify the superiority of this study in the multi-class classification task of medical images of lung diseases, Table 3 shows the predictive performance of the model in classifying pneumonia, COVID-19 and normal samples on the COVID and Curated datasets, including the comprehensive evaluation results of different evaluation metrics Acc, Pre, Rec and F1.

[0053] On the COVID dataset, the proposed method maintained high classification accuracy across all three classes. Specifically, the average accuracy, precision, recall, and F1 score for the COVID category were 99.13%, 98.84%, 97.51%, and 98.17%, respectively. The normal category demonstrated even greater robustness, with an average accuracy of 98.76%, and average recall and F1 score of 99.36% and 99.08%, respectively, indicating a low risk of missed diagnoses. The average accuracy, precision, recall, and F1 score for the pneumonia category were 99.45%, 97.27%, 96.54%, and 96.90%, respectively, also demonstrating high predictive performance. Furthermore, the standard deviation of the recall for the normal category was only 0.20, and the standard deviation of the F1 score was 0.03, indicating that the model maintained a highly consistent recognition ability under different random partitions. The standard deviations for the COVID and pneumonia categories were also within a reasonable range, demonstrating the model's strong stability on the COVID dataset.

[0054] On the other hand, the model's performance on the Curated dataset was further improved, demonstrating more significant stability across all categories. Specifically, the average values ​​of the four indicators for the COVID-19 category were 99.77%, 100%, 98.31%, and 99.15%, respectively. Precision reached 100% in all three random data partitions, achieving zero false positives, fully validating the model's excellent ability to identify this category of samples. The performance for the normal and pneumonia categories was also stable, with average precision, recall, and F1 score all remaining above 98.25%, and standard deviations remaining within a low range, indicating that the method can reliably adapt to different data distribution conditions.

[0055] The above analysis shows that the experimental results from three random partitions based on these two publicly available X-ray datasets demonstrate that the proposed scheme exhibits significant classification performance in image classification tasks for pneumonia, COVID-19, and normal categories. It not only demonstrates high stability within the same dataset but also maintains good generalization ability in cross-dataset evaluations. This result fully demonstrates the application potential and reliability of the proposed model in practical clinical scenarios for auxiliary diagnosis of lung diseases, providing solid support for the subsequent promotion and application of the model.

[0056] In one exemplary embodiment, this application also provides a lung disease classification device comprising: an acquisition device and a host computer, wherein the acquisition device is connected to the host computer; the host computer is configured with a lung disease classification model for classifying lung diseases based on chest X-ray images acquired by the acquisition device.

[0057] In one exemplary embodiment, this application provides a computer device, which may be a server or a terminal, and its internal structure diagram may be as follows. Figure 7 As shown, this computer device includes a processor, memory, input / output (I / O) interfaces, and a communication interface. The processor, memory, and I / O interfaces are connected via a system bus, and the communication interface is also connected to the system bus via the I / O interfaces. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and databases. The internal memory provides the environment for the operating system and computer programs stored in the non-volatile storage media to run. The I / O interfaces are used for exchanging information between the processor and external devices. The communication interface is used for communicating with external terminals via a network connection.

[0058] Those skilled in the art will understand that Figure 7 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0059] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of the relevant data must comply with relevant regulations.

[0060] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments described above. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM).

[0061] The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, etc., and are not limited to these.

[0062] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0063] This document uses specific examples to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only for the purpose of helping to understand the methods and core ideas of this application. Furthermore, those skilled in the art will recognize that, based on the ideas of this application, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of this application.

Claims

1. A method for constructing a classification model for lung diseases, characterized in that, The method for constructing the lung disease classification model includes: Construct a labeled sample dataset; the labeled sample dataset includes multiple chest X-ray images and the corresponding lung disease categories for each chest X-ray image; A classification model based on Mobile-Former is constructed; the classification model based on Mobile-Former includes, in sequence: an enhanced Mobile-Former module group, a feature fusion module, and a classification module; The Mobile-Former-based classification model is trained using the labeled sample dataset to obtain the trained Mobile-Former-based classification model, which serves as the lung disease classification model.

2. The method for constructing a lung disease classification model according to claim 1, characterized in that, The enhanced Mobile-Former module group includes N sequentially connected enhanced Mobile-Former modules; each of the N sequentially connected enhanced Mobile-Former modules includes: a first summation sub-block, a second summation sub-block, a first bidirectional cross-attention sub-block, a second bidirectional cross-attention sub-block, a Mobile sub-block, and a Former sub-block; The output of the first bidirectional cross-attention sub-block is connected to the input of the first summation sub-block; the output of the first summation sub-block is connected to the input of the Former sub-block, the output of the Former sub-block is connected to the input of the second bidirectional cross-attention sub-block, the output of the second bidirectional cross-attention sub-block is connected to the input of the second summation sub-block, and the output of the Mobile sub-block is connected to another input of the second summation sub-block; the output of the Former sub-block serves as the first output, and the output of the second summation sub-block serves as the second output. The first enhanced Mobile-Former module receives the labeled sample dataset as its input, and its input is connected to the inputs of the Mobile sub-block and the first bidirectional cross-attention sub-block. The second to Nth enhanced Mobile-Former modules each have two inputs: a first input and a second input. The first input is connected to the first output of the previous enhanced Mobile-Former module, and the second input is connected to the second output of the previous enhanced Mobile-Former module.

3. The method for constructing a lung disease classification model according to claim 2, characterized in that, The feature fusion module includes: a third summation sub-block, a third bidirectional cross-attention sub-block, an improved parallel block perceptual attention module, an improved wavelet-guided energy recalibration module, and a fusion sub-block; the feature fusion module includes two input terminals, namely the third input terminal and the fourth input terminal; The third input terminal is connected to the first output terminal of the Nth enhanced Mobile-Former module, and the fourth input terminal is connected to the second output terminal of the Nth enhanced Mobile-Former module; The third input terminal is connected to the input terminal of the third summing sub-block, and the fourth input terminal is connected to the input terminal of the improved parallel block perception attention module; The output of the improved parallel block perceptual attention module is connected to the input of the third bidirectional cross-attention sub-block and the input of the improved wavelet-guided energy recalibration module; the output of the third bidirectional cross-attention module is connected to the other input of the third summation sub-block, and the output of the third summation sub-block is connected to the input of the fusion sub-block; the output of the improved wavelet-guided energy recalibration module is connected to the other input of the fusion sub-block, and the output of the fusion sub-block serves as the output of the feature fusion module.

4. The method for constructing a lung disease classification model according to claim 3, characterized in that, The classification module comprises, in sequence: a first linear layer, a Swish activation function, a second linear layer, a ReLU activation function, a Dropout layer, and a third linear layer; the input of the first linear layer is connected to the output of the feature fusion module.

5. The method for constructing a lung disease classification model according to claim 2, characterized in that, The Mobile sub-block includes multiple sets of sequentially connected first 1×1 pointwise convolutional layers, first dynamic ReLU, 3×3 depthwise convolutional layers, second dynamic ReLU, second 1×1 pointwise convolutional layers, and multilayer perceptron sub-blocks; The Former sub-block includes a multi-head attention layer and a feedforward network connected in sequence; The input of the multilayer sensing sub-block is connected to the output of the second bidirectional cross-attention sub-block, and the output of the multilayer sensing sub-block is connected to the first dynamic ReLU and the second dynamic ReLU.

6. The method for constructing a lung disease classification model according to claim 3, characterized in that, The improved parallel block-aware attention module includes a multi-branch feature extraction module, a multi-statistic high-efficiency channel attention module, a depth-separable attention collaboration module, and a random path discarding module connected in sequence. The multi-statistic efficient channel attention module includes a first adaptive average pooling sub-block, a first max pooling sub-block, a first standard deviation pooling sub-block, a one-dimensional convolutional block, and an activation function; The inputs of the first adaptive average pooling sub-block, the first max pooling sub-block, and the first standard deviation pooling sub-block are all connected to the output of the multi-branch feature extraction module; the outputs of the first adaptive average pooling sub-block, the first max pooling sub-block, and the first standard deviation pooling sub-block are added together and then connected to the input of the one-dimensional convolution sub-block. The output of the one-dimensional convolutional sub-block is connected to the input of the activation function; the output of the activation function and the input of the one-dimensional convolutional sub-block are added together, and then multiplied by the output of the multi-branch feature extraction module, which is then used as the output of the multi-statistic efficient channel attention module.

7. The method for constructing a lung disease classification model according to claim 6, characterized in that, The depthwise separable convolutional module includes: a channel partitioning sub-block, a first depthwise convolutional sub-block, and a context-aware feature space attention module; The input of the channel partitioning submodule is connected to the output of the multi-statistic efficient channel attention module; the channel partitioning submodule is used to divide the output of the multi-statistic efficient channel attention module into four branches along the channel dimension, namely the identity branch, the square kernel branch, the vertical strip branch, and the horizontal strip branch; The input of the first deep convolutional sub-block is connected to the output of the channel partitioning sub-block. The first deep convolutional sub-block is used to capture the features of each of the four branches. After concatenating the features of each of the four branches in the channel dimension, a spatial attention mask is generated by using the first Sigmoid activation function. The spatial attention mask is multiplied by the output of the multi-statistic efficient channel attention module to obtain the concatenated features. The context-aware feature space attention module includes a second adaptive average pooling sub-block, a second max pooling sub-block, a second standard deviation pooling sub-block, a second depthwise convolution sub-block, a second sigmoid activation function, and a feature optimization sub-block; the feature optimization sub-block includes a spatial dimension mean sub-block, a fully connected sub-block, and a third sigmoid activation function connected in sequence; the second depthwise convolution sub-block includes a 7×7 convolution sub-block, a batch normalization sub-block, a ReLU activation function, and a 1×1 convolution sub-block connected in sequence; The inputs to the second adaptive average pooling sub-block, the second max pooling sub-block, and the second standard deviation pooling sub-block are all the concatenated features; the outputs of the second adaptive average pooling sub-block, the second max pooling sub-block, and the second standard deviation pooling sub-block are concatenated along the channel dimension and then input to the second depthwise convolution sub-block and the second sigmoid activation function, which are then output; the output of the second sigmoid activation function is multiplied by the concatenated features and then input to the feature optimization sub-block.

8. The method for constructing a lung disease classification model according to claim 3, characterized in that, The improved wavelet-guided energy recalibration module includes: a 3×3 depth convolution sub-block, a first convolution sub-block, a second convolution sub-block, a first center shear sub-block, a second center shear sub-block, and a SimAM simple parameterless attention sub-block; the first convolution sub-block and the second convolution sub-block each include a fixed wavelet convolution sub-block and a channel grouping convolution sub-block connected in sequence; The input of the 3×3 depth convolutional sub-block is connected to the output of the depth-separable attention collaboration module, and the output of the 3×3 depth convolutional sub-block is connected to the input of the first convolutional sub-block and the second convolutional sub-block; the first convolutional module is used to decompose the output features of the 3×3 depth convolutional sub-block into low-frequency components, and the second convolutional module is used to decompose the output features of the 3×3 depth convolutional sub-block into high-frequency components; The output of the first convolutional sub-block is connected to the input of the first central shearing sub-block; the output of the second convolutional sub-block is connected to the input of the second central shearing sub-block, and the output of the second central shearing sub-block is connected to the input of the SimAM simple parameterless attention sub-block. The SimAM simple parameterless attention sub-block is used to calculate the weight of each spatial location in the high-frequency feature, and multiply the weight of each spatial location of the high-frequency feature by the feature value of the corresponding spatial location to obtain the enhanced high-frequency feature. The low-frequency features and the enhanced high-frequency features are added together, and then multiplied by the output of the improved parallel block perception attention module, which is then used as the output of the improved wavelet-guided energy recalibration module.

9. The method for constructing a lung disease classification model according to claim 2, characterized in that, The lung disease classification model also includes a lightweight bottleneck module. The output of the lightweight bottleneck module is connected to the input of the Mobile sub-block. The input of the lightweight bottleneck module is the chest X-ray image. The lightweight bottleneck module includes a 3×3 depth convolution sub-block and a 1×1 pointwise convolution sub-block connected in sequence.

10. A lung disease classification device, characterized in that, The lung disease classification device includes: an acquisition device and a host computer, wherein the acquisition device is connected to the host computer; the host computer is equipped with a lung disease classification model according to any one of claims 1-9, used to classify lung diseases based on chest X-ray images acquired by the acquisition device.

Images