A dust lung disease staging recognition method and system based on multi-modal artificial iconography feature fusion

Abstract

The application discloses a pneumoconiosis staging recognition method and system based on multi-modal artificial imaging feature fusion, and relates to the technical field of medical image recognition. The method comprises the following steps: standardizing and pre-processing an input chest X-ray image and extracting a lung field region; in the lung field region, detecting and counting micro nodules based on clinical imaging prior knowledge, and extracting micro nodule quantity features reflecting the quantity and spatial distribution characteristics of nodular lesions; calculating the first-order entropy features of the lung field region image to quantify the complexity of lung parenchyma texture, and extracting high-dimensional features representing gray heterogeneity and structure statistical characteristics by using an imaging feature analysis method. The multi-class artificial imaging features are standardized and fused in a unified feature space, and the fused features are input into a learning classification model for pneumoconiosis period recognition. The application improves the stability and interpretability of pneumoconiosis staging recognition through multi-modal artificial imaging feature fusion modeling.

Description

Technical Field

[0001] This invention relates to the field of medical image recognition technology, and in particular to a method and system for staging pneumoconiosis based on multimodal artificial imaging feature fusion. Background Technology

[0002] Currently, the clinical diagnosis and staging of pneumoconiosis mainly rely on the manual interpretation of chest X-ray images by radiologists. The diagnostic results are easily affected by factors such as the radiologist's experience, subjective perception, and workload, making it difficult to guarantee consistency in staging and accuracy in early identification. To assist physicians in identifying the stages of pneumoconiosis, using deep learning methods to learn the characteristics of each stage is a promising approach.

[0003] Although deep learning-based automatic diagnostic methods have made some progress in pneumoconiosis image analysis in recent years, their internal feature expression and decision-making processes lack interpretability and are highly dependent on large-scale, high-quality labeled data. In practical application scenarios with limited sample size and uneven class distribution, their stability is insufficient, making it difficult to meet clinical requirements for reliability and interpretability.

[0004] Therefore, there is an urgent need for a pneumoconiosis staging identification method that can integrate prior knowledge of clinical imaging and has good interpretability under limited sample conditions. Summary of the Invention

[0005] The purpose of this invention is to provide a method and system for staging pneumoconiosis based on multimodal artificial imaging feature fusion, aiming to solve or improve at least one of the above-mentioned technical problems.

[0006] To achieve the above objectives, the present invention provides the following solution: A method for staging pneumoconiosis based on multimodal artificial imaging feature fusion includes: Acquire chest X-ray images, perform grayscale conversion, pixel normalization and contrast enhancement processing on the chest X-ray images, and segment the lung field regions to obtain standardized lung field images. In the standardized lung field images, microfocal lesions are detected and counted based on prior knowledge of clinical imaging, and the number of microfocal lesions and spatial distribution characteristics are extracted. First-order entropy features are calculated on the standardized lung field images to extract first-order entropy features that characterize the complexity and irregularity of lung parenchyma texture. Based on radiomics analysis, radiomics features characterizing gray-scale distribution, texture structure, and spatial heterogeneity are extracted from the standardized lung field images. The numerical characteristics of the microfoci, the first-order entropy characteristics, and the radiomics characteristics are standardized and fused in a unified feature space to construct a multimodal artificial imaging feature vector. The multimodal artificial imaging feature vectors are input into the classification model for feature training, and the preset pneumoconiosis classification results are output. The pneumoconiosis classification results include no pneumoconiosis, stage I pneumoconiosis, stage II pneumoconiosis, and stage III pneumoconiosis. The classification model adopts a gradient boosting decision tree-based classification model.

[0007] Optionally, in the standardized lung field image, microfocals are detected and counted based on prior clinical imaging knowledge, and microfocal quantity features reflecting the number and spatial distribution characteristics of microfocals are extracted, specifically including: Contrast-restricted adaptive histogram equalization and the Laplacian operator are used to highlight the microfocal signals in the standardized lung field image, resulting in a microfocal signal enhancement image. Based on the enhanced image of the micro foci signal, a superpixel segmentation algorithm is used to divide the region, morphological opening operation is introduced to remove the rib structure, and a multi-threshold strategy is used to perform binarization processing in the divided candidate regions to generate preliminary candidate micro foci regions. For the preliminary candidate microfocal regions, morphological opening operation is applied again to remove isolated noise points to obtain the final microfocal detection region. An adaptive traversal mechanism is then used to combine the contrast enhancement coefficient, binarization threshold, and morphological kernel size to search for the number and corresponding location of microfocals. Feature extraction is performed based on the number and corresponding location of the microfoci to obtain microfoci quantity features that reflect the number and spatial distribution characteristics of microfoci.

[0008] Optionally, first-order entropy features are calculated on the standardized lung field image to extract first-order entropy features that characterize the complexity and irregularity of lung parenchyma texture, specifically including: Spatial scanning mapping was performed on the standardized lung field images, and the two-dimensional images were converted into one-dimensional signal sequences using Hilbert-Turbo scan conversion; The one-dimensional signal sequence is subjected to distribution mapping and symbolization processing to generate a symbol sequence for entropy calculation; Based on the symbol sequence, Shannon entropy, sample entropy, and multi-scale entropy are calculated respectively to obtain first-order entropy features for characterizing the complexity and irregularity of lung parenchyma texture.

[0009] Optionally, based on radiomics analysis methods, radiomics features characterizing grayscale distribution, texture structure, and spatial heterogeneity are extracted from the standardized lung field images, specifically including: Based on radiomics analysis, high-throughput artificial imaging features are extracted from the standardized lung field images to construct an initial set of radiomics features. Among them, radiomics features include, but are not limited to, morphological features, first-order gray-level statistical features, gray-level co-occurrence matrix features, gray-level run matrix features, gray-level region size matrix features, gray-level dependency matrix features, and neighborhood gray-level tone difference matrix features. Robustness screening and correlation analysis were performed on the initial radiomics feature set to remove features with insufficient stability and redundant correlations, thereby obtaining a candidate radiomics feature set. Based on the feature importance evaluation method, the discriminative ability of the candidate radiomics feature set is evaluated, and the preset number of features with the highest discriminative ability are selected as the final radiomics features used for pneumoconiosis staging identification.

[0010] Optionally, the numerical characteristics of the microfoci, the first-order entropy characteristics, and the radiomics characteristics are standardized and fused in a unified feature space to construct a multimodal artificial imaging feature vector, specifically including: The numerical characteristics of the microfoci, the first-order entropy characteristics, and the radiomics characteristics are standardized. An early fusion strategy was adopted to construct a unified high-dimensional feature vector by concatenating three types of standardized features in columns, thus obtaining a multimodal artificial imaging feature vector.

[0011] This invention also provides a pneumoconiosis staging identification system based on multimodal artificial imaging feature fusion, comprising: The preprocessing module is used to acquire chest X-ray images, perform grayscale conversion, pixel normalization and contrast enhancement on the chest X-ray images, and segment the lung field regions to obtain standardized lung field images. The microfocal feature extraction module is used to detect and count microfocals in the standardized lung field image based on prior knowledge of clinical imaging, and extract the microfocal number feature reflecting the number and spatial distribution characteristics of microfocals. The entropy feature extraction module is used to perform first-order entropy feature calculation on the standardized lung field image and extract first-order entropy features to characterize the complexity and irregularity of lung parenchyma texture. The radiomics feature extraction module is used to extract radiomics features that characterize gray-scale distribution, texture structure, and spatial heterogeneity from the standardized lung field image based on radiomics analysis methods. The feature fusion module is used to standardize the feature of the number of microfoci, the first-order entropy feature and the radiomics feature, and fuse them in a unified feature space to construct a multimodal artificial imaging feature vector. The staging diagnosis module is used to input the multimodal artificial imaging feature vectors into the classification model for feature training and output the preset pneumoconiosis classification results; the pneumoconiosis classification results include no pneumoconiosis, stage I pneumoconiosis, stage II pneumoconiosis, and stage III pneumoconiosis; the classification model adopts a classification model based on gradient boosting decision tree.

[0012] Optionally, the microfocal feature extraction module specifically includes: The microfocal enhancement unit is used to highlight the microfocal signal in the standardized lung field image by employing contrast-limited adaptive histogram equalization and the Laplacian operator to obtain a microfocal signal enhancement image. The region segmentation unit is used to segment the region based on the enhanced image of the micro foci signal using a superpixel segmentation algorithm, introduce morphological opening operation to remove rib structures, and perform binarization processing in the segmented candidate regions using a multi-threshold strategy to generate preliminary candidate micro foci regions. The region traversal unit is used to apply morphological opening operation again to remove isolated noise points for the preliminary candidate microfocal regions to obtain the final microfocal detection region. An adaptive traversal mechanism is used to combine the contrast enhancement coefficient, binarization threshold and morphological kernel size to search and obtain the number and corresponding location of microfocals. The feature extraction unit is used to extract features based on the number and corresponding location of the microfocals to obtain microfocal quantity features that reflect the number and spatial distribution characteristics of the microfocals.

[0013] Optionally, the entropy feature extraction module specifically includes: The image dimensionality reduction unit is used to perform spatial scan mapping on the standardized lung field image and to convert the two-dimensional image into a one-dimensional signal sequence using Hilbert cross scan transformation; A symbolization processing unit is used to perform distribution mapping and symbolization processing on the one-dimensional signal sequence to generate a symbol sequence for entropy calculation; The entropy calculation unit is used to calculate Shannon entropy, sample entropy and multi-scale entropy based on the symbol sequence, respectively, to obtain first-order entropy features for characterizing the complexity and irregularity of lung parenchyma texture.

[0014] Optionally, the radiomics feature extraction module specifically includes: The initial radiomics feature set construction unit is used to extract high-throughput artificial imaging features from the standardized lung field images based on radiomics analysis methods to construct an initial radiomics feature set; wherein, the radiomics features include, but are not limited to, morphological features, first-order gray-level statistical features, gray-level co-occurrence matrix features, gray-level run matrix features, gray-level region size matrix features, gray-level dependency matrix features, and neighborhood gray-level tone difference matrix features. The candidate radiomics feature set construction unit is used to perform robust screening and correlation analysis on the initial radiomics feature set, remove features with insufficient stability and redundant correlation, and obtain the candidate radiomics feature set. The feature evaluation unit is used to evaluate the discriminative power of the candidate radiomics feature set based on the feature importance evaluation method, and select the preset number of features with the highest discriminative power as the final radiomics features used for pneumoconiosis staging identification.

[0015] Optionally, the feature fusion module specifically includes: A standardization processing unit is used to standardize the numerical characteristics of the microfoci, the first-order entropy characteristics, and the radiomics characteristics. The multi-feature stitching unit is used to construct a unified high-dimensional feature vector by stitching together three types of standardized features through column fusion using an early fusion strategy, thereby obtaining a multimodal artificial imaging feature vector.

[0016] According to specific embodiments provided by the present invention, the present invention discloses the following technical effects: This invention discloses a method and system for staging and identifying pneumoconiosis based on multimodal artificial imaging feature fusion. The method includes acquiring chest X-ray images, performing grayscale conversion, pixel normalization, and contrast enhancement on the chest X-ray images, and segmenting the lung field region to obtain a standardized lung field image; in the standardized lung field image, detecting and counting microfocals based on prior clinical imaging knowledge, and extracting microfocal quantity features reflecting the number and spatial distribution characteristics of microfocals; calculating first-order entropy features on the standardized lung field image, and extracting first-order entropy features to characterize the complexity and irregularity of lung parenchyma texture; and... Using radiomics analysis, radiomics features characterizing grayscale distribution, texture structure, and spatial heterogeneity are extracted from the standardized lung field images. The number of microfocal features, the first-order entropy features, and the radiomics features are standardized and fused in a unified feature space to construct a multimodal artificial imaging feature vector. This multimodal artificial imaging feature vector is then input into a classification model for feature training, outputting a pre-defined pneumoconiosis classification result. The pneumoconiosis classification result includes no pneumoconiosis, stage I pneumoconiosis, stage II pneumoconiosis, and stage III pneumoconiosis. The classification model employs a gradient boosting decision tree-based classification model.

[0017] This invention improves the stability and interpretability of pneumoconiosis staging by fusing multimodal artificial imaging features, providing technical support for clinical staging assessment and decision support. Specifically, the invention features highly complementary multidimensional features. By fusing the number of microfocals, entropy-based texture features, and radiomics features, it jointly models pneumoconiosis images from multiple levels of morphology, texture, and structure, significantly improving the accuracy and stability of staging. The interpretability of this invention is significantly enhanced; all constructed features have clear radiological or statistical physical significance, avoiding the "black box" problem of pure deep learning models and enhancing the interpretability and clinical credibility of staging results. This invention is applicable to small sample and imbalanced data scenarios. Based on artificial features and ensemble learning modeling, it still exhibits good generalization performance under limited sample conditions, making it suitable for practical applications such as pneumoconiosis where sample acquisition is difficult. This invention has strong clinical transferability, with standardized and highly automated methods that can be directly embedded into existing occupational disease screening and imaging diagnostic systems, showing promising prospects for widespread application. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of the present invention 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 the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 This is a flowchart illustrating the pneumoconiosis staging identification method of the present invention; Figure 2 This is a flowchart of the microfocal feature extraction process in step S2 of this embodiment. Detailed Implementation

[0020] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] The purpose of this invention is to provide a method and system for staging pneumoconiosis based on multimodal artificial imaging feature fusion, aiming to solve or improve at least one of the above-mentioned technical problems.

[0022] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0023] As a first aspect, such as Figures 1-2 As shown, this invention provides a method for staging pneumoconiosis based on multimodal artificial imaging feature fusion, comprising: S1. Acquire chest X-ray images, perform grayscale conversion, pixel normalization, and contrast enhancement processing on the chest X-ray images, and segment the lung field regions to obtain standardized lung field images.

[0024] S2. In the standardized lung field image, based on prior knowledge of clinical imaging, microfocal lesions are detected and counted, and the microfocal lesion quantity feature reflecting the number and spatial distribution characteristics of microfocal lesions is extracted.

[0025] This step specifically includes: Contrast-limited adaptive histogram equalization and the Laplacian operator are used to highlight the microfocal signals in the standardized lung field image, resulting in a microfocal signal enhancement image. Based on the microfocal signal enhancement image, a superpixel segmentation algorithm is used for region division, morphological opening operation is introduced to remove rib structures, and a multi-threshold strategy is used for binarization processing in the divided candidate regions to generate preliminary candidate microfocal regions. For the preliminary candidate microfocal regions, morphological opening operation is applied again to remove isolated noise points, resulting in the final microfocal detection region. An adaptive traversal mechanism is used to combine the contrast enhancement coefficient, binarization threshold, and morphological kernel size for search to obtain the number and corresponding location of microfocals. Based on the number and corresponding location of microfocals, feature extraction is performed to obtain microfocal quantity features reflecting the number and spatial distribution characteristics of microfocals.

[0026] S3. Calculate the first-order entropy features of the standardized lung field image and extract the first-order entropy features used to characterize the complexity and irregularity of lung parenchyma texture. This step specifically includes: The standardized lung field image is spatially scanned and mapped, and the two-dimensional image is converted into a one-dimensional signal sequence using Hilbert cross-scan transformation. The one-dimensional signal sequence is then distributed and symbolized to generate a symbol sequence for entropy calculation. Based on the symbol sequence, Shannon entropy, sample entropy, and multi-scale entropy are calculated to obtain first-order entropy features that characterize the complexity and irregularity of lung parenchyma texture.

[0027] S4. Based on radiomics analysis, extract radiomics features that characterize gray-scale distribution, texture structure, and spatial heterogeneity from the standardized lung field images; This step specifically includes: Based on radiomics analysis, high-throughput artificial imaging features are extracted from the standardized lung field images to construct an initial radiomics feature set. These features include, but are not limited to, morphological features, first-order gray-level statistical features, gray-level co-occurrence matrix features, gray-level run-length matrix features, gray-level region size matrix features, gray-level dependency matrix features, and neighborhood gray-level tone difference matrix features. Robustness screening and correlation analysis are performed on the initial radiomics feature set to remove features with insufficient stability and redundant correlations, resulting in a candidate radiomics feature set. Based on feature importance evaluation methods, the discriminative power of the candidate radiomics feature set is assessed, and a predetermined number of features with the highest discriminative power are selected as the final radiomics features used for pneumoconiosis staging identification.

[0028] S5. The numerical characteristics of the microfoci, the first-order entropy characteristics, and the radiomics characteristics are standardized and fused in a unified feature space to construct a multimodal artificial imaging feature vector. This step specifically includes: The numerical characteristics of the microfoci, the first-order entropy characteristics, and the radiomics characteristics are standardized; an early fusion strategy is used to construct a unified high-dimensional feature vector by concatenating the three types of standardized features in columns, thus obtaining a multimodal artificial imaging feature vector.

[0029] S6. Input the multimodal artificial imaging feature vector into the classification model for feature training, and output the preset pneumoconiosis classification result; the pneumoconiosis classification result includes no pneumoconiosis, stage I pneumoconiosis, stage II pneumoconiosis, and stage III pneumoconiosis; the classification model adopts a classification model based on gradient boosting decision tree.

[0030] As one specific implementation method, the processing steps described above include: S1. Chest X-ray image preprocessing and lung field segmentation: To acquire the chest X-ray (CXR) images to be analyzed, the raw images are first standardized and preprocessed to reduce the impact of differences in imaging conditions on subsequent analysis. This includes: (1) Grayscale conversion and pixel intensity normalization The original chest X-ray images were converted into single-channel grayscale images, and the pixel intensity was standardized using the min-max normalization method to unify the grayscale range between different images. The calculation formula is as follows: in, Represents the pixels in the original image grayscale value, and These represent the minimum and maximum grayscale values ​​in the image, respectively. These are the normalized pixel values.

[0031] (2) Imaging inhomogeneity correction and contrast enhancement To reduce the bias effect caused by uneven irradiation intensity or detector response differences during X-ray imaging, bias correction processing was performed on the normalized image. Subsequently, the Adaptive Histogram Equalization (AHE) method was used to enhance the local contrast of the image, thereby improving the discernibility of small nodules, texture structures and edge details in the lung field.

[0032] (3) Automatic segmentation of lung field regions After the above preprocessing, the chest X-ray image is automatically segmented using a lung field segmentation model to accurately extract the bilateral lung regions and remove non-lung regions such as the heart and background, resulting in a standardized lung field image. This lung field image serves as the unified input for subsequent multimodal image feature extraction and pneumoconiosis staging diagnosis.

[0033] S2. Extraction of the number of microfocal features: Based on lung field images, a small foci quantification detection algorithm is designed. The lung field images are converted to grayscale images, and contrast-limited adaptive histogram equalization is used to enhance local contrast, thereby highlighting potential small foci signals. Subsequently, the Laplacian operator is used for edge enhancement to strengthen high-frequency details and highlight the boundary features of small foci. Its two-dimensional form is as follows: in, Represents a two-dimensional image.

[0034] To further improve local region consistency, the lung field is divided using the Simple Linear Iterative Clustering (SLIC) algorithm to obtain superpixel units with clear boundaries and good local consistency, thereby enhancing the aggregation of candidate microfocal regions and reducing interference from vascular artifacts. To suppress linear artifacts, a morphological opening operation (kernel=25×3) is introduced to remove structures such as ribs, thus improving the saliency of microfocal regions. In the candidate region extraction stage, a multi-threshold strategy (thresholds set to 15, 25, and 35) is used for binarization to generate preliminary candidate microfocal regions. For the binarization results, a small-scale kernel morphological opening operation is applied again to remove isolated noise points, making the boundaries of the microfocal regions more regular.

[0035] For parameter optimization, an adaptive traversal mechanism is adopted, combining contrast enhancement coefficient, binarization threshold, and morphological kernel size for search, automatically selecting the parameter combination that yields the optimal number of microfocals as the final detection result. Finally, the system outputs the optimal number of microfocals and their spatial distribution characteristics for each image, and simultaneously saves the original image, the binarization results of the microfocals, and a statistical table of microfocal values, providing a reliable basis for subsequent feature analysis and modeling.

[0036] S3, First-order entropy feature extraction: To quantify the complexity and irregularity of lung parenchyma texture, first-order entropy features are extracted from the standardized lung field image.

[0037] First, spatial scanning mapping is performed on the two-dimensional chest X-ray images of the lung field region, and the two-dimensional images are converted into one-dimensional signal sequences using Hilbert-Turbo conversion: in, S Represents a one-dimensional signal sequence. s L Represents pixels within the lung field region. This represents the total number of pixels within the lung field region.

[0038] Subsequently, the one-dimensional signal is subjected to distribution mapping and symbolization processing to generate a symbol sequence for entropy calculation.

[0039] (1) Shannon entropy Shannon entropy, proposed by Shannon in information theory, is used to measure the uncertainty of a system. Its mathematical definition is as follows: in, Let the random variable X take values The probability of grayscale distribution. For image signals, Shannon entropy reflects the uniformity and uncertainty of grayscale distribution: the higher the value, the more complex the image texture.

[0040] (2) Sample entropy Sample entropy is used to characterize the self-similarity of a signal, and it is defined as: Where A represents the sequence length as... B represents the number of matches when the sequence length is m, excluding self-matches. r This indicates the similarity tolerance. NSampEn represents the length of the time series. A lower SampEn value indicates stronger self-similarity of the time series. Compared to ApEn1D, SampEn1D is less dependent on signal length and exhibits stronger consistency. SampEn is calculated for Hilbert scan signals using the antropy.sample_entropy library function.

[0041] (3) Multiscale information entropy To further characterize the texture complexity at different spatial scales, the one-dimensional signal is scaled by a factor of 1. The coarsening process generates coarse-grained sequences: in, i This represents the index of the sampling point in the original one-dimensional signal. j Indicates an index in the coarse-grained sequence. N Indicates the sequence length.

[0042] Each coarse-grained sequence The sample entropy is used to obtain the entropy value at the scale: in, and These represent sequence lengths of 1 and 2 respectively. m +1 and m The number of matches, excluding self-matches. For random signals, The entropy gradually decreases with increasing scale; however, for correlated signals, it remains at a high level across multiple scales. An improved multi-scale symbolic dynamic structure entropy is used to analyze one-dimensional chest radiograph signals obtained through Hilbert transform. The signal after Hilbert transform... Perform a normal cumulative distribution mapping and discretize it into a sequence of symbols. Secondly, improve coarsening by adjusting the scale factor. The following generates a coarse-grained symbol sequence: Calculate the Rényi entropy for each coarse-grained symbol pattern set: in, The probability of the symbol pattern occurring. This is the Rényi index. In this embodiment, the scaling factor... The IMDSSE is calculated to obtain a ten-dimensional multi-scale entropy vector.

[0043] S4. Radiomics Feature Extraction and Screening: Based on radiomics analysis, high-throughput artificial imaging features were extracted from the standardized lung field images obtained in step S1 to quantify the morphological features, grayscale distribution features, and texture heterogeneity features of lung tissue, and to construct an initial set of radiomics features: in, Indicates the first One radiomics characteristic, This serves as the initial feature dimension. In this embodiment, the PyRadiomics 3.0.1 toolkit, conforming to the IBSI (Image Biomarker Standardisation Initiative) standard, is used to extract a total of 386 hand-designed features for each lung field region of interest (ROI). These features include, but are not limited to: morphological features, first-order gray-level statistical features, gray-level co-occurrence matrix features (GLCM), gray-level run-length matrix features (GLRLM), gray-level region size matrix features (GLSZM), gray-level dependency matrix features (GLDM), and neighborhood gray-level tone difference matrix features (NGTDM).

[0044] To improve the robustness of the feature set, reduce redundancy, and enhance the model's generalization ability, the initial radiomics feature set was sequentially filtered as follows: (1) Robustness screening The stability of features under different partitioning or perturbation conditions is evaluated using the intra-class correlation coefficient (ICC), retaining features that satisfy the following conditions. Features that indicate insufficient stability are eliminated.

[0045] (2) Correlation analysis and redundancy elimination Spearman rank correlation analysis was performed on the features that passed the robustness screening. When the correlation coefficient between any two features satisfies... At the same time, retain the features with higher discriminative power to reduce feature redundancy.

[0046] (3) Evaluation and ranking of feature importance After completing the above screening, a feature mapping relationship for the training set is constructed based on the XGBoost classification model. The game theory-based SHAP (SHapley Additive exPlanations) method is then used to calculate the global contribution of each candidate feature to the model's classification decision. The mean absolute SHAP value (mean |SHAP|) is used as the feature importance evaluation metric to rank all candidate features.

[0047] (4) Construction of radiomics feature subsets Based on the ranking of feature importance, the top 20 radiomics features with the highest discriminative power were selected to construct a subset of radiomics features: The aforementioned feature subset encompasses various radiological information, including gray-level dependence, first-order gray-level distribution, tissue morphological complexity, and texture spatial heterogeneity. It can characterize the radiological features of pneumoconiosis lesions at multiple scales, such as uneven gray-level distribution, complex texture, and structural irregularities. Ultimately, this radiomics feature subset serves as the core input for subsequent multimodal feature fusion and the construction of a pneumoconiosis staging and classification model.

[0048] S5, Multimodal Artificial Imaging Feature Fusion: Characteristics of the number of microfocal lesions First-order entropy complexity characteristics and radiomics characteristics The three types of features are fused, based on the micro-spot count and local density extracted by image processing algorithms, primarily for quantitative assessment of early lesion stages. Only stage I and II samples are retained; white spot features for other stages are set to zero to avoid noise interference during model training. To eliminate differences in units and numerical ranges among the different features, the concatenated multimodal feature vector is... Standardization process: in, For the first The first sample One characteristic, and The first The mean and standard deviation of each feature, with 12 being the total feature dimension. For the real number space, d w Let be the dimension of the white point feature subspace, corresponding to the feature set composed of the counts of tiny spots. d To represent the total dimension of the multimodal feature vector. Standardization ensures that features from different sources are aligned on the same statistical scale, preventing large numerical features from dominating model training; their relative contributions are automatically learned through gradient boosting trees during model training. Subsequently, an early fusion strategy is used to construct a unified high-dimensional feature vector by concatenating the three types of standardized features column-wise. S6. Model Training and Optimization Strategies: The pneumoconiosis staging prediction system is a typical multi-classification task with discrete categories. These represent pneumoconiosis-free lung, stage I, stage II, and stage III, respectively. The model is trained using the XGBoost algorithm, whose core idea is to iteratively construct a decision tree to fit the residuals from the previous round, thereby minimizing the loss function and regularization term. in, To indicate the first i The true class label of each sample To represent the model for the th i Prediction results for each sample For the first k The regularization term for the decision trees, where K is the total number of decision trees in the ensemble model. For multi-class cross-entropy loss function, For the first A decision tree, The number of leaf nodes. The weights of the leaf nodes are... These are the leaf node regularization and L2 regularization coefficients, used to control model complexity and prevent overfitting. Five-fold cross-validation is employed during training to effectively utilize the limited sample size and reduce random bias in data partitioning. To improve model performance, a grid search is performed on key hyperparameters—tree depth, learning rate, subsampling rate, and column sampling rate—and the optimal parameter combination is selected based on the average accuracy of cross-validation.

[0049] XGBoost automatically learns the contribution of each feature to the stage prediction through a tree-splitting structure. To further enhance the model's interpretability, this embodiment combines XGBoost's built-in feature importance analysis with the SHAP method to quantify the marginal contribution of each feature in the prediction, thereby achieving a transformation from "black box prediction" to "interpretable inference".

[0050] As a second aspect, the present invention also provides a pneumoconiosis staging identification system based on multimodal artificial imaging feature fusion, which applies the method described above, including: The preprocessing module is used to acquire chest X-ray images, perform grayscale conversion, pixel normalization and contrast enhancement on the chest X-ray images, and segment the lung field regions to obtain standardized lung field images.

[0051] The microfocal feature extraction module is used to detect and count microfocals in the standardized lung field image based on prior knowledge of clinical imaging, and extract the microfocal number feature reflecting the number and spatial distribution characteristics of microfocals.

[0052] The entropy feature extraction module is used to perform first-order entropy feature calculation on the standardized lung field image and extract first-order entropy features to characterize the complexity and irregularity of lung parenchyma texture.

[0053] The radiomics feature extraction module is used to extract radiomics features that characterize gray-scale distribution, texture structure, and spatial heterogeneity from the standardized lung field image based on radiomics analysis methods.

[0054] The feature fusion module is used to standardize the feature of the number of microfoci, the first-order entropy feature and the radiomics feature, and fuse them in a unified feature space to construct a multimodal artificial imaging feature vector.

[0055] The staging diagnosis module is used to input the multimodal artificial imaging feature vectors into the classification model for feature training and output the preset pneumoconiosis classification results; the pneumoconiosis classification results include no pneumoconiosis, stage I pneumoconiosis, stage II pneumoconiosis, and stage III pneumoconiosis; the classification model adopts a classification model based on gradient boosting decision tree.

[0056] The microfocal feature extraction module specifically includes: The microfocal enhancement unit is used to highlight the microfocal signal in the standardized lung field image by employing contrast-limited adaptive histogram equalization and the Laplacian operator, thereby obtaining a microfocal signal enhancement image.

[0057] The region segmentation unit is used to segment regions based on the enhanced image of the micro foci signal using a superpixel segmentation algorithm, introduce morphological opening operations to remove rib structures, and perform binarization processing in the segmented candidate regions using a multi-threshold strategy to generate preliminary candidate micro foci regions.

[0058] The region traversal unit is used to apply morphological opening operation again to remove isolated noise in the preliminary candidate microfocal regions to obtain the final microfocal detection region. An adaptive traversal mechanism is used to combine the contrast enhancement coefficient, binarization threshold and morphological kernel size to search and obtain the number and corresponding location of microfocals.

[0059] The feature extraction unit is used to extract features based on the number and corresponding location of the microfocals to obtain microfocal quantity features that reflect the number and spatial distribution characteristics of the microfocals.

[0060] The entropy feature extraction module specifically includes: The image dimensionality reduction unit is used to perform spatial scan mapping on the standardized lung field image and to convert the two-dimensional image into a one-dimensional signal sequence using Hilbert cross scan transformation.

[0061] The symbolization processing unit is used to perform distribution mapping and symbolization processing on the one-dimensional signal sequence to generate a symbol sequence for entropy calculation.

[0062] The entropy calculation unit is used to calculate Shannon entropy, sample entropy and multi-scale entropy based on the symbol sequence, respectively, to obtain first-order entropy features for characterizing the complexity and irregularity of lung parenchyma texture.

[0063] The radiomics feature extraction module specifically includes: The initial radiomics feature set construction unit is used to extract high-throughput artificial imaging features from the standardized lung field images based on radiomics analysis methods, and construct the initial radiomics feature set; wherein, the radiomics features include, but are not limited to, morphological features, first-order gray-level statistical features, gray-level co-occurrence matrix features, gray-level run matrix features, gray-level region size matrix features, gray-level dependency matrix features, and neighborhood gray-level tone difference matrix features.

[0064] The candidate radiomics feature set construction unit is used to perform robust screening and correlation analysis on the initial radiomics feature set, and remove features with insufficient stability and redundant correlation to obtain the candidate radiomics feature set.

[0065] The feature evaluation unit is used to evaluate the discriminative power of the candidate radiomics feature set based on the feature importance evaluation method, and select the preset number of features with the highest discriminative power as the final radiomics features used for pneumoconiosis staging identification.

[0066] The feature fusion module specifically includes: The standardization processing unit is used to standardize the numerical characteristics of the microfoci, the first-order entropy characteristics, and the radiomics characteristics.

[0067] The multi-feature stitching unit is used to construct a unified high-dimensional feature vector by stitching together three types of standardized features through column fusion using an early fusion strategy, thereby obtaining a multimodal artificial imaging feature vector.

[0068] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other.

[0069] This document uses specific examples to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the core ideas of the present invention. Furthermore, those skilled in the art will recognize that, based on the ideas of the present invention, 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 the present invention.

Claims

1. A method for staging pneumoconiosis based on multimodal artificial imaging feature fusion, characterized in that, include: Acquire chest X-ray images, perform grayscale conversion, pixel normalization and contrast enhancement processing on the chest X-ray images, and segment the lung field regions to obtain standardized lung field images. In the standardized lung field images, microfocal lesions are detected and counted based on prior knowledge of clinical imaging, and the number of microfocal lesions and spatial distribution characteristics are extracted. First-order entropy features are calculated on the standardized lung field images to extract first-order entropy features that characterize the complexity and irregularity of lung parenchyma texture. Based on radiomics analysis, radiomics features characterizing gray-scale distribution, texture structure, and spatial heterogeneity are extracted from the standardized lung field images. The numerical characteristics of the microfoci, the first-order entropy characteristics, and the radiomics characteristics are standardized and fused in a unified feature space to construct a multimodal artificial imaging feature vector. The multimodal artificial imaging feature vectors are input into the classification model for feature training, and the preset pneumoconiosis classification results are output. The pneumoconiosis classification results include no pneumoconiosis, stage I pneumoconiosis, stage II pneumoconiosis, and stage III pneumoconiosis. The classification model adopts a gradient boosting decision tree-based classification model.

2. The method for staging pneumoconiosis based on multimodal artificial imaging feature fusion according to claim 1, characterized in that, In the standardized lung field images, microfocal lesions are detected and counted based on prior clinical imaging knowledge. Microfocal lesion quantity features reflecting the number and spatial distribution characteristics of microfocal lesions are extracted, specifically including: Contrast-restricted adaptive histogram equalization and the Laplacian operator are used to highlight the microfocal signals in the standardized lung field image, resulting in a microfocal signal enhancement image. Based on the enhanced image of the micro foci signal, a superpixel segmentation algorithm is used to divide the region, morphological opening operation is introduced to remove the rib structure, and a multi-threshold strategy is used to perform binarization processing in the divided candidate regions to generate preliminary candidate micro foci regions. For the preliminary candidate microfocal regions, morphological opening operation is applied again to remove isolated noise points to obtain the final microfocal detection region. An adaptive traversal mechanism is then used to combine the contrast enhancement coefficient, binarization threshold, and morphological kernel size to search for the number and corresponding location of microfocals. Feature extraction is performed based on the number and corresponding location of the microfoci to obtain microfoci quantity features that reflect the number and spatial distribution characteristics of microfoci.

3. The method for staging pneumoconiosis based on multimodal artificial imaging feature fusion according to claim 1, characterized in that, First-order entropy features are calculated on the standardized lung field images to extract first-order entropy features that characterize the complexity and irregularity of lung parenchyma texture, specifically including: Spatial scanning mapping was performed on the standardized lung field images, and the two-dimensional images were converted into one-dimensional signal sequences using Hilbert-Turbo scan conversion; The one-dimensional signal sequence is subjected to distribution mapping and symbolization processing to generate a symbol sequence for entropy calculation; Based on the symbol sequence, Shannon entropy, sample entropy, and multi-scale entropy are calculated respectively to obtain first-order entropy features for characterizing the complexity and irregularity of lung parenchyma texture.

4. The method for staging pneumoconiosis based on multimodal artificial imaging feature fusion according to claim 1, characterized in that, Based on radiomics analysis methods, radiomics features characterizing grayscale distribution, texture structure, and spatial heterogeneity are extracted from the standardized lung field images, specifically including: Based on radiomics analysis, high-throughput artificial imaging features are extracted from the standardized lung field images to construct an initial set of radiomics features. Among them, radiomics features include, but are not limited to, morphological features, first-order gray-level statistical features, gray-level co-occurrence matrix features, gray-level run matrix features, gray-level region size matrix features, gray-level dependency matrix features, and neighborhood gray-level tone difference matrix features. Robustness screening and correlation analysis were performed on the initial radiomics feature set to remove features with insufficient stability and redundant correlations, thereby obtaining a candidate radiomics feature set. Based on the feature importance evaluation method, the discriminative ability of the candidate radiomics feature set is evaluated, and the preset number of features with the highest discriminative ability are selected as the final radiomics features used for pneumoconiosis staging identification.

5. The method for staging pneumoconiosis based on multimodal artificial imaging feature fusion according to claim 1, characterized in that, The numerical characteristics of the microfoci, the first-order entropy characteristics, and the radiomics characteristics are standardized and fused in a unified feature space to construct a multimodal artificial imaging feature vector, specifically including: The numerical characteristics of the microfoci, the first-order entropy characteristics, and the radiomics characteristics are standardized. An early fusion strategy was adopted to construct a unified high-dimensional feature vector by concatenating three types of standardized features in columns, thus obtaining a multimodal artificial imaging feature vector.

6. A staging system for pneumoconiosis based on multimodal artificial imaging feature fusion, characterized in that, include: The preprocessing module is used to acquire chest X-ray images, perform grayscale conversion, pixel normalization and contrast enhancement on the chest X-ray images, and segment the lung field regions to obtain standardized lung field images. The microfocal feature extraction module is used to detect and count microfocals in the standardized lung field image based on prior knowledge of clinical imaging, and extract the microfocal number feature reflecting the number and spatial distribution characteristics of microfocals. The entropy feature extraction module is used to perform first-order entropy feature calculation on the standardized lung field image and extract first-order entropy features to characterize the complexity and irregularity of lung parenchyma texture. The radiomics feature extraction module is used to extract radiomics features that characterize gray-scale distribution, texture structure, and spatial heterogeneity from the standardized lung field image based on radiomics analysis methods. The feature fusion module is used to standardize the feature of the number of microfoci, the first-order entropy feature and the radiomics feature, and fuse them in a unified feature space to construct a multimodal artificial imaging feature vector. The staging diagnosis module is used to input the multimodal artificial imaging feature vectors into the classification model for feature training and output the preset pneumoconiosis classification results; the pneumoconiosis classification results include no pneumoconiosis, stage I pneumoconiosis, stage II pneumoconiosis, and stage III pneumoconiosis; the classification model adopts a classification model based on gradient boosting decision tree.

7. The pneumoconiosis staging identification system based on multimodal artificial imaging feature fusion according to claim 6, characterized in that, The microfocal feature extraction module specifically includes: The microfocal enhancement unit is used to highlight the microfocal signal in the standardized lung field image by employing contrast-limited adaptive histogram equalization and the Laplacian operator to obtain a microfocal signal enhancement image. The region segmentation unit is used to segment the region based on the enhanced image of the micro foci signal using a superpixel segmentation algorithm, introduce morphological opening operation to remove rib structures, and perform binarization processing in the segmented candidate regions using a multi-threshold strategy to generate preliminary candidate micro foci regions. The region traversal unit is used to apply morphological opening operation again to remove isolated noise points for the preliminary candidate microfocal regions to obtain the final microfocal detection region. An adaptive traversal mechanism is used to combine the contrast enhancement coefficient, binarization threshold and morphological kernel size to search and obtain the number and corresponding location of microfocals. The feature extraction unit is used to extract features based on the number and corresponding location of the microfocals to obtain microfocal quantity features that reflect the number and spatial distribution characteristics of the microfocals.

8. The pneumoconiosis staging identification system based on multimodal artificial imaging feature fusion according to claim 6, characterized in that, The entropy feature extraction module specifically includes: The image dimensionality reduction unit is used to perform spatial scan mapping on the standardized lung field image and to convert the two-dimensional image into a one-dimensional signal sequence using Hilbert cross scan transformation; A symbolization processing unit is used to perform distribution mapping and symbolization processing on the one-dimensional signal sequence to generate a symbol sequence for entropy calculation; The entropy calculation unit is used to calculate Shannon entropy, sample entropy and multi-scale entropy based on the symbol sequence, respectively, to obtain first-order entropy features for characterizing the complexity and irregularity of lung parenchyma texture.

9. The pneumoconiosis staging identification system based on multimodal artificial imaging feature fusion according to claim 6, characterized in that, The radiomics feature extraction module specifically includes: The initial radiomics feature set construction unit is used to extract high-throughput artificial imaging features from the standardized lung field images based on radiomics analysis methods to construct an initial radiomics feature set; wherein, the radiomics features include, but are not limited to, morphological features, first-order gray-level statistical features, gray-level co-occurrence matrix features, gray-level run matrix features, gray-level region size matrix features, gray-level dependency matrix features, and neighborhood gray-level tone difference matrix features. The candidate radiomics feature set construction unit is used to perform robust screening and correlation analysis on the initial radiomics feature set, remove features with insufficient stability and redundant correlation, and obtain the candidate radiomics feature set. The feature evaluation unit is used to evaluate the discriminative power of the candidate radiomics feature set based on the feature importance evaluation method, and select the preset number of features with the highest discriminative power as the final radiomics features used for pneumoconiosis staging identification.

10. The pneumoconiosis staging identification system based on multimodal artificial imaging feature fusion according to claim 6, characterized in that, The feature fusion module specifically includes: A standardization processing unit is used to standardize the numerical characteristics of the microfoci, the first-order entropy characteristics, and the radiomics characteristics. The multi-feature stitching unit is used to construct a unified high-dimensional feature vector by stitching together three types of standardized features through column fusion using an early fusion strategy, thereby obtaining a multimodal artificial imaging feature vector.

Images