Chronic obstructive pulmonary disease grading system and method based on X-ray chest radiography

Abstract

According to the chronic obstructive pulmonary disease grading system and method based on the X-ray chest radiograph, the original X-ray chest radiograph is obtained, and the original X-ray chest radiograph is preprocessed; a conditional generative adversarial network is adopted to separate the preprocessed original X-ray chest radiograph, and a bone suppression image and a bone tissue feature image are obtained; inputting the bone tissue feature image into the constructed double-branch deep neural network to obtain a pulmonary parenchyma feature map and a rib feature map; obtaining a clinical variable, preprocessing the clinical variable, performing cross-modal fusion on the preprocessed clinical variable, the lung parenchyma feature map and the rib feature map, and generating a fusion feature vector; processing the fusion feature vector by combining a regression-classification output header model to obtain an FEV1% predicted value and a GOLD staging result; a conditional generative adversarial network, a double-branch deep neural network, cross-modal fusion and a joint regression-classification output head model are introduced, and low-cost, low-radiation and high-universality GOLD automatic grading can be realized.

Description

Technical Field

[0001] The present invention relates to the technical field of clinical diagnosis, and in particular to a chronic obstructive pulmonary disease grading system and method based on chest X-rays. Background Art

[0002] Chronic obstructive pulmonary disease (COPD) is the third leading cause of death worldwide. Early screening and accurate staging are key to improving prognosis. Existing screening methods primarily include pulmonary function tests (PFTs), single-slice computed tomography (CT), and chest X-rays. PFTs are the gold standard for diagnosis, but they rely on active patient cooperation (such as forced expiratory maneuvers), resulting in a false-negative rate exceeding 30% in elderly and frail patients. While CT can quantitatively assess emphysema and small airway disease, the radiation dose of a single chest CT scan (approximately 8 mSv) is 400 times that of a conventional chest X-ray (0.02 mSv), making it unsuitable for routine screening. Therefore, chest X-rays remain the imaging tool of choice for COPD screening due to their low cost, low radiation exposure, and widespread availability. However, conventional chest radiographs have a sensitivity of less than 40% for early COPD signs (such as subtle lung marking disturbances) and are unable to accurately classify patients according to the GOLD (Global Initiative for Chronic Obstructive Lung Disease) classification system (1-4).

[0003] This technical bottleneck has given rise to new solutions driven by artificial intelligence (AI). Deep learning (DL) models have demonstrated accuracy that surpasses that of radiologists in tasks such as pneumonia detection and pneumothorax segmentation on chest radiographs. However, existing AI technologies mostly focus on extracting single lung parenchymal features, ignoring the value of rib structural changes (such as widening of the intercostal spaces and changes in chest morphology) in indicating abnormal chest dynamics. The lack of cross-modal fusion of clinical variables (such as inflammatory markers and blood gas analysis) and imaging features results in insufficient clinical interpretability of the models. Furthermore, the masking effect of high-density rib shadows on lung parenchymal texture features in chest radiographs is not addressed, affecting the accuracy of detecting microstructural abnormalities. Summary of the Invention

[0004] In view of this, the present invention proposes a chronic obstructive pulmonary disease grading system and method based on chest X-ray. Through dual-tissue feature analysis and multimodal data fusion, combined with clinical variables, low-cost, low-radiation, and highly universal GOLD automatic staging can be achieved.

[0005] The technical solution of the present invention is achieved as follows:

[0006] A method for grading chronic obstructive pulmonary disease based on chest X-rays comprises the following steps:

[0007] Step S1: obtaining an original chest X-ray film and preprocessing the original chest X-ray film;

[0008] Step S2: Using a conditional generative adversarial network to separate the preprocessed original chest X-ray to obtain a bone inhibition image and a bone tissue feature image;

[0009] Step S3: input the bone suppression image and the bone tissue feature image into the constructed two-branch deep neural network to obtain the lung parenchyma feature map and the rib feature map;

[0010] Step S4: Acquire clinical variables, preprocess the clinical variables, perform cross-modal fusion on the preprocessed clinical variables, lung parenchyma feature maps, and rib feature maps, and generate a fusion feature vector;

[0011] Step S5: The combined regression-classification output head model processes the fused feature vector to obtain the FEV1% predicted value and the GOLD staging result.

[0012] Preferably, the specific steps of step S1 are:

[0013] Step S11: receiving an original chest X-ray film, adjusting the window width of the original chest X-ray film, performing grayscale processing after resampling;

[0014] Step S12: normalize, enhance contrast, and perform Gaussian filtering noise suppression on the original chest X-ray film after grayscale processing.

[0015] Preferably, the specific steps of step S2 are:

[0016] Step S21: construct a conditional generative adversarial network using a U-Net generator and a PatchGAN discriminator;

[0017] Step S22: The pre-processed original chest X-ray film is input into the generator, and the generator is forward propagated and obtains a bone inhibition image and a bone tissue feature image through an encoding-decoding path;

[0018] Step S23: input the generated bone suppression image into the discriminator for authenticity determination and obtain a authenticity score;

[0019] Step S24: Calculate pixel-level L1 loss, adversarial loss, and anatomical constraint loss, and jointly optimize the generator and discriminator by backpropagation based on the authenticity score, pixel-level L1 loss, adversarial loss, and anatomical constraint loss;

[0020] Step S25 , iteratively optimize until the PSNR of the generated bone suppression image is greater than 32 dB and the Dice coefficient of the rib mask is greater than 0.85, and output the bone suppression image and bone tissue feature image generated by the generator.

[0021] Preferably, when the discriminator performs authenticity judgment, its input also includes a real bone suppression image obtained by CT projection. By stacking convolutional layers, a authenticity score in the range of [0,1] is generated for each 70*70 pixel receptive field area of the two input images, where the authenticity target of the real bone suppression image is 1, and the authenticity target of the generated bone suppression image is 0. The optimization goal of the discriminator is to maximize the ability to distinguish between the real bone suppression image and the generated bone suppression image, and the optimization goal of the generator is to minimize the discriminator's ability to distinguish the generated bone suppression image.

[0022] Preferably, the dual-branch deep neural network includes a lung parenchyma branch, and the lung parenchyma branch adopts an improved ResNet-50 network structure. The ResNet-50 network structure includes: Conv1, Conv2_x, Conv3_x, Conv4_x, Conv5_x, a spatial-channel attention module, a pyramid pooling module, a global average pooling layer, and a fully connected layer. The improvement of the ResNet-50 network structure includes replacing the standard 3×3 convolution with a dilated convolution with a dilated rate of 2 in the Conv4_x stage and introducing a spatial-channel attention module and a pyramid pooling module. The specific steps of obtaining the lung parenchyma feature map in step S3 are:

[0023] Step S31: normalize the bone suppression image and input it into the improved ResNet-50 network structure;

[0024] Step S32: The multi-scale features are sequentially passed through Conv1-Conv5_x, and the feature maps of each stage are applied with the attention mechanism of the spatial-channel attention module;

[0025] In step S33, the output of Conv5_x is fused with multi-scale information through the pyramid pooling module, and then a lung parenchyma feature map is generated through a global average pooling layer and a fully connected layer.

[0026] Preferably, the dual-branch deep neural network further includes a rib parsing branch, and the rib parsing branch adopts a 3D PointNet++ network structure. The 3D PointNet++ network structure includes a sampling layer consisting of SA layer 1, SA layer 2, and SA layer 3, and a feature extraction layer consisting of a fully connected layer 1, a Dropout layer, and a fully connected layer 2. The steps for obtaining the rib feature map are:

[0027] Step S34: input the bone tissue feature image into the high-resolution network HRNet-W32 to detect rib key points;

[0028] Step S35: After extracting high-confidence rib edge points from the key point coordinates through non-maximum suppression, the key point coordinates are converted into 3D point cloud data;

[0029] Step S36: Input the 3D point cloud data into the 3D PointNet++ network structure. After the hierarchical features are extracted by the three SA layers, the rib feature map is extracted by the feature extraction layer.

[0030] Preferably, the specific steps of obtaining clinical variables and preprocessing the clinical variables in step S4 are:

[0031] Step S41: obtaining continuous variables and performing standardization on the continuous variables, wherein the continuous variables include age and BMI;

[0032] Step S42: Obtain discrete variables and encode the discrete variables, wherein the discrete variables include gender;

[0033] Step S43: Obtain ordinal variables, perform logarithmic transformation on the ordinal variables, and then perform linear standardization, wherein the ordinal variables include smoking pack-years.

[0034] Preferably, the specific steps of performing cross-modal fusion on the pre-processed clinical variables, lung parenchyma feature map and rib feature map and generating a fusion feature vector in step S4 are:

[0035] Step S44: Continuous variables are directly entered into the low-dimensional space, and discrete variables and ordinal variables are mapped into the low-dimensional space through the embedding layer and concatenated into clinical feature vectors;

[0036] Step S45: using a cross-attention module to calculate the correlation between the lung parenchyma feature map and the rib feature map, and obtaining a rib-lung parenchyma collaborative lesion index, and fusing the rib-lung parenchyma collaborative lesion index with the lung parenchyma feature map and the rib feature map respectively through a gating mechanism to obtain an enhanced lung parenchyma feature map and an enhanced rib feature map;

[0037] Step S46: Splicing the enhanced lung parenchyma feature map and the enhanced rib feature map to obtain a spliced feature, using a multi-head attention mechanism to calculate the cross-attention weights of the clinical feature variables and the spliced feature, and applying the splicing results of multiple attention heads to obtain a weighted image feature;

[0038] Step S47: Integrate the clinical feature variables and the weighted image features through residual connection and gating mechanism to obtain a fused feature vector.

[0039] Preferably, the specific steps of step S5 are:

[0040] Step S51: The regression branch in the joint regression-classification output head model processes the fused feature vector through a three-layer fully connected network and outputs the FEV1% predicted value;

[0041] Step S52: The classification branch in the joint regression-classification output head model outputs each GOLD level threshold through the ordinal regression layer;

[0042] Step S53: Calculate the cumulative probability, and determine the GOLD staging result based on the cumulative probability.

[0043] A chest X-ray-based grading system for chronic obstructive pulmonary disease (COPD) consisting of:

[0044] A chest X-ray acquisition module is used to acquire and pre-process the original chest X-ray film;

[0045] Generative bone suppression module, which uses a conditional generative adversarial network to separate the preprocessed original chest X-ray to obtain bone suppression images and bone tissue feature images;

[0046] A dual-branch parsing module is used to input the bone suppression image and the bone tissue feature image into the constructed dual-branch deep neural network to obtain the lung parenchyma feature map and the rib feature map;

[0047] The cross-modal fusion module is used to obtain clinical variables, pre-process the clinical variables, perform cross-modal fusion on the pre-processed clinical variables, lung parenchyma feature maps, and rib feature maps, and generate a fusion feature vector;

[0048] The joint regression classification output module is used to process the fused feature vector to obtain the predicted FEV1% value and GOLD staging results;

[0049] The chest X-ray acquisition module, the generative bone suppression module, the dual-branch parsing module, the cross-modal fusion module and the joint regression classification output module are sequentially connected.

[0050] Compared with the prior art, the present invention has the following beneficial effects:

[0051] The present invention provides a chronic obstructive pulmonary disease grading system and method based on chest X-rays. After preprocessing the original chest X-rays, a conditional generative adversarial network is used to separate the original chest X-rays, eliminating the rib masking effect and improving the visualization of subtle changes in lung texture. Then, by constructing a two-branch deep neural network, bone suppression images and bone tissue feature images can be processed, lung texture features can be extracted, and rib morphological parameters can be quantified. Clinical variables are then introduced and cross-modally fused with lung parenchyma feature maps and rib feature maps, and image features are associated with clinical variables to quantify the contribution of clinical variables. Finally, a regression-classification output head model is used to process the fused feature vector to achieve continuous and precise GOLD staging. At the same time, the predicted FEV1% estimated value can also provide a theoretical basis for treatment. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only preferred embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0053] Figure 1 This is a flow chart of a method for grading chronic obstructive pulmonary disease based on chest X-rays of the present invention;

[0054] Figure 2 This is a flow chart of step S1 of a method for grading chronic obstructive pulmonary disease based on chest X-rays of the present invention;

[0055] Figure 3 This is a flow chart of step S2 of a method for grading chronic obstructive pulmonary disease based on chest X-rays of the present invention;

[0056] Figure 4 This is a flow chart of step S3 of a method for grading chronic obstructive pulmonary disease based on chest X-rays of the present invention;

[0057] Figure 5 This is a flow chart of step S4 of a method for grading chronic obstructive pulmonary disease based on chest X-rays of the present invention;

[0058] Figure 6 This is a flow chart of step S5 of a method for grading chronic obstructive pulmonary disease based on chest X-rays of the present invention;

[0059] Figure 7 This is a schematic diagram of a chronic obstructive pulmonary disease grading system based on chest X-rays according to the present invention;

[0060] Figure 8 This is a structural diagram of a generative bone inhibition module of a chronic obstructive pulmonary disease grading system based on chest X-rays of the present invention;

[0061] Figure 9 This is a structural diagram of an improved ResNet-50 network structure for a chronic obstructive pulmonary disease grading system based on chest X-rays of the present invention;

[0062] Figure 10 This is a structural diagram of a joint regression classification output module of a chronic obstructive pulmonary disease grading system based on chest X-rays of the present invention;

[0063] In the figure, 1. Chest X-ray acquisition module; 2. Generative bone suppression module; 3. Dual-branch parsing module; 4. Cross-modal fusion module; 5. Joint regression and classification output module. DETAILED DESCRIPTION

[0064] In order to better understand the technical content of the present invention, a specific embodiment is provided below, and the present invention is further described in conjunction with the accompanying drawings.

[0065] See also Figures 1 to 10 The present invention provides a method for grading chronic obstructive pulmonary disease based on chest X-ray, comprising the following steps:

[0066] Step S1: Obtain an original chest X-ray film and pre-process the original chest X-ray film. The specific steps are as follows:

[0067] Step S11: receiving an original chest X-ray film, adjusting the window width of the original chest X-ray film, performing grayscale processing after resampling;

[0068] Step S12: normalize, enhance contrast, and perform Gaussian filtering noise suppression on the original chest X-ray film after grayscale processing.

[0069] Methods: Raw chest radiographs (512 × 512 pixels, DICOM format) and structured clinical data tables were received. The DICOM files were loaded and parsed, and pixel data and metadata were extracted. The window width and position were adjusted to 1500 HU and -600 HU, and the images were resampled to a resolution of 1024 × 1024 pixels. Grayscale values were normalized to the range [-1, 1]. U-Net was then used to pre-segment the lung regions for subsequent processing. Contrast-limited adaptive histogram equalization (CLAHE) was applied to enhance lung field details, and Gaussian filtering (σ = 0.5) was applied to suppress noise.

[0070] Step S2: Use a conditional generative adversarial network to separate the preprocessed original chest X-ray to obtain a bone inhibition image and a bone tissue feature image. The specific steps are as follows:

[0071] Step S21: construct a conditional generative adversarial network using a U-Net generator and a PatchGAN discriminator;

[0072] Step S22: The pre-processed original chest X-ray film is input into the generator, and the generator is forward propagated and obtains a bone inhibition image and a bone tissue feature image through an encoding-decoding path;

[0073] Step S23: input the generated bone suppression image into the discriminator for authenticity determination and obtain a authenticity score;

[0074] Step S24: Calculate pixel-level L1 loss, adversarial loss, and anatomical constraint loss, and jointly optimize the generator and discriminator by backpropagation based on the authenticity score, pixel-level L1 loss, adversarial loss, and anatomical constraint loss, using the AdamW optimizer;

[0075] Step S25 , iteratively optimize until the PSNR of the generated bone suppression image is greater than 32 dB and the Dice coefficient of the rib mask is greater than 0.85, and output the bone suppression image and bone tissue feature image generated by the generator.

[0076] The role of the conditional generative adversarial network is to eliminate the masking effect of the ribs on the lung parenchyma while retaining the morphological characteristics of the ribs. It includes a U-Net generator and a PatchGAN discriminator. The U-Net generator includes an input layer, an encoding path (downsampling stage), and a decoding path (upsampling stage, including residual jump connections), while the PatchGAN discriminator only includes an input layer and an output layer. After the preprocessed original X-ray chest film is input into the U-Net generator, forward propagation is performed in the generator, and after the encoding-decoding path, the generated bone suppression image and the generated bone tissue can be obtained. The generated bone suppression image is input into the PatchGAN discriminator for authenticity determination. When the PatchGAN discriminator makes authenticity determination, its input also includes the real bone suppression image obtained by CT projection. By stacking convolutional layers, a authenticity score in the range of [0,1] is generated for each 70*70 pixel receptive field area of the two input images. The authenticity target of the real bone suppression image is 1, and the authenticity target of the generated bone suppression image is 0. Then, three loss functions are calculated, namely pixel-level L1 loss, adversarial loss, and anatomical constraint loss, where:

[0077] The expression of pixel-level L1 loss is: L_L1 = E[‖Iboneless-I_target‖_1], where Iboneless is the bone-suppressed image and I_target is the boneless standard image generated by practical CT image avatar. The pixel-level L1 loss is used to constrain the structural similarity between Iboneless and the gold standard CT registration image.

[0078] Adversarial loss is used to determine whether it can distinguish generated images from real bone suppression images. Its expression is: L_GAN = E[log(D(I_target))]+E[log(1-D(G(I_input)))], where D represents the discriminator, G represents the generator, and I_input is the input original chest X-ray.

[0079] The expression of anatomical constraint loss is: L_anatomy = DiceLoss(M_pred, M_GT), where M_pred is the binarized rib mask and M_GT is the rib mask annotated by the radiologist.

[0080] The total loss function is L_total = λ1*L_L1+λ2*L_GAN+λ3*L_anatomy, where λ1 = 100, λ2 = 1, and λ3 = 10 are weight hyperparameters.

[0081] Based on the calculated three-part loss function and the authenticity score, the generator and discriminator can be jointly optimized through back-propagation, and iterative optimization is performed until the PSNR of the generated bone suppression image is greater than 32dB and the Dice coefficient of the rib mask is greater than 0.85. Optimization is then stopped. At this point, the bone suppression image and bone tissue feature image can be output. The optimization goal of the discriminator is to maximize the ability to distinguish between the real bone suppression image and the generated bone suppression image, and the optimization goal of the generator is to minimize the discriminator's ability to distinguish the generated bone suppression image.

[0082] Step S3: input the bone suppression image and the bone tissue feature image into the constructed two-branch deep neural network to obtain the lung parenchyma feature map and the rib feature map;

[0083] Preferably, the dual-branch deep neural network includes a lung parenchyma branch, and the lung parenchyma branch adopts an improved ResNet-50 network structure. The ResNet-50 network structure includes: Conv1, Conv2_x, Conv3_x, Conv4_x, Conv5_x, a spatial-channel attention module, a pyramid pooling module, a global average pooling layer, and a fully connected layer. The improvement of the ResNet-50 network structure includes replacing the standard 3×3 convolution with a dilated convolution with a dilated rate of 2 in the Conv4_x stage and introducing a spatial-channel attention module and a pyramid pooling module. The specific steps of obtaining the lung parenchyma feature map in step S3 are:

[0084] Step S31: normalize the bone suppression image and input it into the improved ResNet-50 network structure;

[0085] Step S32: Multi-scale features are sequentially passed through Conv1-Conv5_x, and the dilated convolution increases the receptive field. The feature maps at each stage are applied with the attention mechanism of the spatial-channel attention module to suppress interference from non-lung areas.

[0086] In step S33, the output of Conv5_x is fused with multi-scale information through the pyramid pooling module, and then a lung parenchyma feature map is generated through the global average pooling layer and the fully connected layer. At the same time, emphysema quantitative indicators (% LAA-950, heterogeneity index, etc.) can also be calculated as auxiliary features.

[0087] The functional goal of the lung parenchyma branch is to quantify COPD-specific signs such as emphysema and bronchial wall thickening. It improves the original ResNet-50, which has the following limitations: a limited receptive field that makes it difficult to capture large-scale emphysema patterns; insufficient ability to distinguish subtle texture structures in the lung field; and background features (such as the mediastinum and shoulders) that interfere with lung feature extraction.

[0088] Therefore, the present invention improves ResNet-50, and the specific improvement measures are as follows: improvement of dilated convolution: in the conv4_x stage, the standard 3×3 convolution is replaced by a dilated convolution with a dilated rate of 2, which expands the effective receptive field by about 48% without increasing the number of parameters. The advantage is that it can better capture the global distribution pattern of large-area emphysema areas; the spatial-channel attention module is introduced, and the spatial weight is multiplied by the channel weight and then applied to the feature map, which can suppress the interference of the background area and highlight the texture features inside the lung field; the pyramid pooling module is introduced, and 4 levels of pooling of different scales (1×1, 2×2, 3×3, 6×6) are added after the final feature layer. The pooling features at each level are upsampled to the original size and then spliced. The advantage is that it captures multi-scale emphysema patterns and enhances the local-global feature association.

[0089] The feature extraction strategy of the improved ResNet-50 network is as follows: low-level features (conv1-2): capture local details of lung texture (such as bronchial wall edges) with high feature resolution; mid-level features (conv3): identify medium-scale lung structures (vascular distribution, interlobular septa); high-level features (conv4-5): identify emphysema areas (low-density areas) and global lung field morphology.

[0090] Preferably, the dual-branch deep neural network further includes a rib parsing branch, and the rib parsing branch adopts a 3D PointNet++ network structure. The 3D PointNet++ network structure includes a sampling layer consisting of SA layer 1, SA layer 2, and SA layer 3, and a feature extraction layer consisting of a fully connected layer 1, a Dropout layer, and a fully connected layer 2. The steps for obtaining the rib feature map are:

[0091] Step S34: input the bone tissue feature image into the high-resolution network HRNet-W32 to detect rib key points;

[0092] Step S35: After extracting high-confidence rib edge points from the key point coordinates through non-maximum suppression, the key point coordinates are converted into 3D point cloud data, and the rib number is added as the third dimension;

[0093] Step S36: Input the 3D point cloud data into the 3D PointNet++ network structure. After the hierarchical features are extracted by the three SA layers, the rib feature map is extracted by the feature extraction layer.

[0094] The functional goal of the rib parsing branch is to quantify thoracic cage dynamics parameters and capture COPD-related rib morphological changes. When performing rib key point detection, the high-resolution network HRNet-W32 uses the following network structure: 4 high-resolution branches (decreasing resolution); multi-resolution feature fusion (each stage); output of key point heat map. After obtaining the rib key points, the 2D key points (x, y) are converted into 3D point clouds (x, y, z), where z is the rib ordinal number (1-12). The 3D point cloud data is centered and scaled to the [-1, 1]^3 space, and 10% Gaussian noise is added to improve generalization ability. Finally, the 3D point cloud data is input into the 3D PointNet++ network structure. After the three SA layers extract hierarchical features, the feature extraction layer extracts the rib feature map. At the same time, clinically relevant rib morphological parameters (intercostal space, thoracic cage proportion, etc.) can also be calculated as auxiliary features.

[0095] Step S4: Acquire clinical variables, preprocess the clinical variables, perform cross-modal fusion on the preprocessed clinical variables, lung parenchyma feature maps, and rib feature maps, and generate a fusion feature vector. The specific steps are as follows:

[0096] Step S41: obtaining continuous variables and performing standardization on the continuous variables, wherein the continuous variables include age and BMI;

[0097] Step S42: Obtain discrete variables and encode the discrete variables, wherein the discrete variables include gender;

[0098] Step S43: Obtain ordinal variables, perform logarithmic transformation on the ordinal variables, and then perform linear standardization, wherein the ordinal variables include smoking pack-years.

[0099] Step S44: Continuous variables are directly entered into the low-dimensional space, and discrete variables and ordinal variables are mapped into the low-dimensional space through the embedding layer and concatenated into clinical feature vectors;

[0100] Step S45: using a cross-attention module to calculate the correlation between the lung parenchyma feature map and the rib feature map, and obtaining a rib-lung parenchyma collaborative lesion index, and fusing the rib-lung parenchyma collaborative lesion index with the lung parenchyma feature map and the rib feature map respectively through a gating mechanism to obtain an enhanced lung parenchyma feature map and an enhanced rib feature map;

[0101] Step S46: The enhanced lung parenchyma feature map and the enhanced rib feature map are spliced together to obtain a spliced feature. A multi-head attention mechanism is used to calculate the cross-attention weights of the clinical feature variables and the spliced feature. The splicing results are then applied using eight attention heads to obtain a weighted image feature.

[0102] Step S47: Integrate the clinical feature variables and the weighted image features through residual connection and gating mechanism to obtain a fused feature vector.

[0103] For continuous variables such as age and BMI, Z-score standardization is performed, while discrete variables such as gender can be encoded using one-hot encoding. Ordinal variables such as smoking pack-years are log-transformed and divided by 5 for standardization. Then, clinical variables need to be embedded. Continuous variables do not need to be embedded and their original values are retained. Discrete and ordinal variables need to be mapped to a low-dimensional space after passing through the embedding layer. Each discrete variable is mapped to an 8-dimensional vector, and ordinal variables are mapped to a 16-dimensional variable. After embedding and splicing all types of variables, a 32-dimensional clinical feature vector is formed.

[0104] For the lung parenchyma feature map and the rib feature map, a lung-bone tissue association model can be established through the cross-attention mechanism to capture the coordinated evolution of alveolar destruction and chest compensatory expansion in the progression of COPD, and obtain the rib-lung parenchyma coordinated lesion index (ERR). The ERR can be fused with the lung parenchyma feature map and the rib feature map respectively to obtain an enhanced lung parenchyma feature map and an enhanced rib feature map, thereby strengthening the structural-functional association between image features and ERR and avoiding premature interference of clinical variables in the learning of image features. The enhanced lung parenchyma feature map and the enhanced rib feature map can be spliced to obtain spliced features. The multi-head attention mechanism can be used to calculate the cross-attention weights of the clinical feature variables and the spliced features, so that weighted splicing can be performed, and 8 attention heads are applied and the splicing results are spliced to obtain weighted image features. The weighted image features and clinical feature variables can be fused through residual connection and gating mechanism integration to obtain the final fusion feature vector. The residual connection and gating mechanism integration can be used for the joint regression-classification output head model.

[0105] Step S5: The combined regression-classification output head model processes the fused feature vector to obtain the predicted FEV1% value and GOLD staging result. The specific steps are as follows:

[0106] Step S51: The regression branch in the joint regression-classification output head model processes the fused feature vector through a three-layer fully connected network and outputs the FEV1% predicted value;

[0107] Step S52: The classification branch in the joint regression-classification output head model outputs each GOLD level threshold through the ordinal regression layer;

[0108] Step S53: Calculate the cumulative probability, and determine the GOLD staging result based on the cumulative probability.

[0109] The functional goal of the joint regression-classification output head model is to simultaneously achieve the prediction of FEV1% continuous values and GOLD staging. Its network structure includes a feature input layer, a regression branch, and a classification branch. The regression branch includes fully connected layer 1, fully connected layer 2, fully connected layer 3, and an activation function. The classification branch includes a feature extraction layer, an ordinal regression layer, and cumulative probability calculation. The ordinal regression layer can output three threshold parameters w_k. The expression for cumulative probability calculation is:

[0110] P(y>k)=σ(φ(x)-w_k-b_k), where φ(x) is the output of the feature extraction layer, w_k and b_k are learnable threshold parameters, and the staging decision rule of the joint regression-classification output head model is:

[0111] y_pred=0if P(y>0)<0.5else

[0112] 1if P(y>1)<0.5else

[0113] 2if P(y>2)<0.5else

[0114] 3if P(y>3)<0.5else 4

[0115] The corresponding GOLD stages are: 0-normal, 1-mild, 2-moderate, 3-severe, 4-very severe.

[0116] A chest X-ray-based grading system for chronic obstructive pulmonary disease (COPD) consisting of:

[0117] Chest X-ray acquisition module 1, used to acquire original chest X-ray films and pre-process the original chest X-ray films;

[0118] Generative bone inhibition module 2, used to separate the pre-processed original chest X-ray using a conditional generative adversarial network to obtain a bone inhibition image and a bone tissue feature image;

[0119] The dual-branch parsing module 3 is used to input the bone suppression image and the bone tissue feature image into the constructed dual-branch deep neural network to obtain the lung parenchyma feature map and the rib feature map;

[0120] Cross-modal fusion module 4 is used to obtain clinical variables, pre-process the clinical variables, perform cross-modal fusion on the pre-processed clinical variables, lung parenchyma feature maps, and rib feature maps, and generate a fusion feature vector;

[0121] The combined regression classification output module 5 is used to process the fused feature vector to obtain the FEV1% predicted value and the GOLD staging result;

[0122] The chest X-ray acquisition module 1, the generative bone suppression module 2, the dual-branch analysis module 3, the cross-modal fusion module 4 and the joint regression classification output module 5 are sequentially connected.

[0123] The chest film acquisition module 1 can acquire and preprocess the original X-ray chest film. The preprocessed original X-ray chest film can be transmitted to the generative bone suppression module 2. The generative bone suppression module 2 uses a conditional generative adversarial network for processing and can obtain bone suppression images and bone tissue feature images, which can eliminate the rib masking effect and improve the visualization of subtle changes in lung texture. The dual-branch analysis module 3 can extract lung parenchyma texture features and rib structure dynamic features respectively through the lung parenchyma branch and the rib analysis branch, and output lung parenchyma feature maps and rib feature maps respectively. The cross-modal fusion module 4 can cross-modally fuse the acquired clinical variables with the lung parenchyma feature map and the rib feature map, overcoming the shallow fusion defects of the traditional multimodal model. Finally, the joint regression classification output module 5 can obtain the FEV1% predicted value and GOLD staging results, which can be used for clinical evaluation and improve the diagnostic effect of COPD.

[0124] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A method for grading chronic obstructive pulmonary disease based on chest X-ray, characterized in that: The following steps are involved: Step S1: obtaining an original chest X-ray film and preprocessing the original chest X-ray film; Step S2: Using a conditional generative adversarial network to separate the preprocessed original chest X-ray to obtain a bone inhibition image and a bone tissue feature image; Step S3: input the bone tissue feature image into the constructed two-branch deep neural network to obtain the lung parenchyma feature map and the rib feature map; Step S4: Acquire clinical variables, preprocess the clinical variables, perform cross-modal fusion on the preprocessed clinical variables, lung parenchyma feature maps, and rib feature maps, and generate a fusion feature vector; Step S5: The combined regression-classification output head model processes the fused feature vector to obtain the FEV1% predicted value and the GOLD staging result.

2. The method for grading chronic obstructive pulmonary disease based on chest X-ray according to claim 1, characterized in that: The specific steps of step S1 are: Step S11: receiving an original chest X-ray film, adjusting the window width of the original chest X-ray film, performing grayscale processing after resampling; Step S12: normalize, enhance contrast, and perform Gaussian filtering noise suppression on the original chest X-ray film after grayscale processing.

3. The method for grading chronic obstructive pulmonary disease based on chest X-ray according to claim 1, characterized in that: The specific steps of step S2 are: Step S21: construct a conditional generative adversarial network using a U-Net generator and a PatchGAN discriminator; Step S22: The pre-processed original chest X-ray film is input into the generator, and the generator is forward propagated and obtains a bone inhibition image and a bone tissue feature image through an encoding-decoding path; Step S23: input the generated bone suppression image into the discriminator for authenticity determination and obtain a authenticity score; Step S24: Calculate pixel-level L1 loss, adversarial loss, and anatomical constraint loss, and jointly optimize the generator and discriminator by backpropagation based on the authenticity score, pixel-level L1 loss, adversarial loss, and anatomical constraint loss; Step S25 , iteratively optimize until the PSNR of the generated bone suppression image is greater than 32 dB and the Dice coefficient of the rib mask is greater than 0.85, and output the bone suppression image and bone tissue feature image generated by the generator.

4. The method for grading chronic obstructive pulmonary disease based on chest X-ray according to claim 3, characterized in that: When the discriminator performs authenticity determination, its input also includes a real bone suppression image obtained by CT projection. By stacking convolutional layers, a authenticity score in the range of [0, 1] is generated for each 70*70 pixel receptive field area of the two input images, where the authenticity target of the real bone suppression image is 1, and the authenticity target of the generated bone suppression image is 0. The optimization goal of the discriminator is to maximize the ability to distinguish between the real bone suppression image and the generated bone suppression image, and the optimization goal of the generator is to minimize the discriminator's ability to distinguish the generated bone suppression image.

5. The method for grading chronic obstructive pulmonary disease based on chest X-ray according to claim 1, characterized in that: The dual-branch deep neural network includes a lung parenchyma branch, which adopts an improved ResNet-50 network structure. The ResNet-50 network structure includes: Conv1, Conv2_x, Conv3_x, Conv4_x, Conv5_x, a spatial-channel attention module, a pyramid pooling module, a global average pooling layer, and a fully connected layer. The improvement to the ResNet-50 network structure includes replacing the standard 3×3 convolution with a dilated convolution with a dilated rate of 2 in the Conv4_x stage and introducing a spatial-channel attention module and a pyramid pooling module. The specific steps of obtaining the lung parenchyma feature map in step S3 are as follows: Step S31: normalize the bone suppression image and input it into the improved ResNet-50 network structure; Step S32: The multi-scale features are sequentially passed through Conv1-Conv5_x, and the feature maps of each stage are applied with the attention mechanism of the spatial-channel attention module; In step S33, the output of Conv5_x is fused with multi-scale information through the pyramid pooling module, and then a lung parenchyma feature map is generated through a global average pooling layer and a fully connected layer.

6. The method for grading chronic obstructive pulmonary disease based on chest X-ray according to claim 1, characterized in that: The dual-branch deep neural network also includes a rib parsing branch, which adopts a 3D PointNet++ network structure. The 3D PointNet++ network structure includes a sampling layer consisting of SA layer 1, SA layer 2, and SA layer 3, and a feature extraction layer consisting of a fully connected layer 1, a Dropout layer, and a fully connected layer 2. The steps for obtaining the rib feature map are as follows: Step S34: input the bone tissue feature image into the high-resolution network HRNet-W32 to detect rib key points; Step S35: After extracting high-confidence rib edge points from the key point coordinates through non-maximum suppression, the key point coordinates are converted into 3D point cloud data; Step S36: Input the 3D point cloud data into the 3D PointNet++ network structure. After the hierarchical features are extracted by the three SA layers, the rib feature map is extracted by the feature extraction layer.

7. The method for grading chronic obstructive pulmonary disease based on chest X-ray according to claim 1, characterized in that: The specific steps of obtaining clinical variables and preprocessing the clinical variables in step S4 are as follows: Step S41: obtaining continuous variables and performing standardization on the continuous variables, wherein the continuous variables include age and BMI; Step S42: Obtain discrete variables and encode the discrete variables, wherein the discrete variables include gender; Step S43: Obtain ordinal variables, perform logarithmic transformation on the ordinal variables, and then perform linear standardization, wherein the ordinal variables include smoking pack-years.

8. The method for grading chronic obstructive pulmonary disease based on chest X-ray according to claim 1, characterized in that: The specific steps of step S4 for performing cross-modal fusion on the pre-processed clinical variables, lung parenchyma feature map, and rib feature map and generating a fusion feature vector are as follows: Step S44: Continuous variables are directly entered into the low-dimensional space, and discrete variables and ordinal variables are mapped into the low-dimensional space through the embedding layer and concatenated into clinical feature vectors; Step S45: using a cross-attention module to calculate the correlation between the lung parenchyma feature map and the rib feature map, and obtaining a rib-lung parenchyma collaborative lesion index, and fusing the rib-lung parenchyma collaborative lesion index with the lung parenchyma feature map and the rib feature map respectively through a gating mechanism to obtain an enhanced lung parenchyma feature map and an enhanced rib feature map; Step S46: Splicing the enhanced lung parenchyma feature map and the enhanced rib feature map to obtain a spliced feature, using a multi-head attention mechanism to calculate the cross-attention weights of the clinical feature variables and the spliced feature, and applying the splicing results of multiple attention heads to obtain a weighted image feature; Step S47: Integrate the clinical feature variables and the weighted image features through residual connection and gating mechanism to obtain a fused feature vector.

9. The method for grading chronic obstructive pulmonary disease based on chest X-ray according to claim 1, characterized in that: The specific steps of step S5 are: Step S51: The regression branch in the joint regression-classification output head model processes the fused feature vector through a three-layer fully connected network and outputs the FEV1% predicted value; Step S52: The classification branch in the joint regression-classification output head model outputs each GOLD level threshold through the ordinal regression layer; Step S53: Calculate the cumulative probability, and determine the GOLD staging result based on the cumulative probability.

10. A system using the method for grading chronic obstructive pulmonary disease based on chest X-ray according to any one of claims 1 to 9, characterized in that: include: A chest X-ray acquisition module is used to acquire and pre-process the original chest X-ray film; Generative bone suppression module, which uses a conditional generative adversarial network to separate the preprocessed original chest X-ray to obtain bone suppression images and bone tissue feature images; A dual-branch parsing module is used to input the bone suppression image and the bone tissue feature image into the constructed dual-branch deep neural network to obtain the lung parenchyma feature map and the rib feature map; The cross-modal fusion module is used to obtain clinical variables, pre-process the clinical variables, perform cross-modal fusion on the pre-processed clinical variables, lung parenchyma feature maps, and rib feature maps, and generate a fusion feature vector; The joint regression classification output module is used to process the fused feature vector to obtain the predicted FEV1% value and GOLD staging results; The chest X-ray acquisition module, the generative bone suppression module, the dual-branch parsing module, the cross-modal fusion module and the joint regression classification output module are sequentially connected.

Images