Training method of lung classification model, lung image classification method and device thereof

Abstract

The present disclosure relates to a lung classification model training method, a lung image classification method and a device thereof. The training method comprises: performing regional feature extraction on each first lung image to obtain a first multi-region feature set, and performing regional dimension feature correlation analysis on the first multi-region feature set to obtain a first fusion feature set containing a non-redundant feature set and a correlation feature set; and performing iterative training of a lung classification model according to training label data and a plurality of first fusion feature sets, wherein the non-redundant feature set represents independent feature information with independent distinguishability of a plurality of lung regions, the correlation feature set represents complementary feature information between the plurality of lung regions, and the fusion feature set has precise representation capability of independent features of a single lung region and complementary features across regions, thereby breaking through the adaptation limitation of a fixed application scenario and improving the accuracy and robustness of the lung classification model.

Description

Technical Field

[0001] This disclosure relates to the field of medical image processing technology, and in particular to a training method for a lung classification model, a classification method for lung images, and an apparatus thereof. Background Technology

[0002] With the rapid development of medical imaging technology, various imaging devices can accurately capture the fine structural information of lung tissue. By extracting features and performing quantitative analysis on lung images, a technical foundation is provided for the accurate quantitative identification of lung tissue.

[0003] However, most mainstream lung image classification methods are limited to single-dimensional feature input, which not only makes it difficult to accurately reflect the complex characteristics of lung pathological changes and comprehensively depict the overall state of lung tissue, but also makes them susceptible to interference from external factors such as differences in equipment hardware and individual physiological differences, resulting in deviations in feature distribution and ultimately causing serious inconsistencies and instability in classification results.

[0004] Therefore, traditional lung image classification methods cannot meet the needs of practical applications in terms of accuracy and robustness. Summary of the Invention

[0005] This disclosure provides a training method for a lung classification model, a classification method for lung images, and an apparatus thereof, to address the problem of incomplete information representation of single-dimensional features, and to improve the robustness of lung image classification while ensuring the accuracy of lung image classification.

[0006] This disclosure provides a method for training a lung classification model, the method comprising: Obtain a first lung image set and corresponding training label data for the first lung image set, wherein the first lung image set contains multiple first lung images; The first lung image is divided into regions for feature extraction to obtain a first multi-region feature set; Perform feature correlation analysis on the first multi-region feature set at the region dimension to obtain the first fusion feature set corresponding to the first lung image; Each of the first fused feature sets is used as input data for the untrained lung classification model, and the model is iteratively trained based on the training label data to obtain the trained lung classification model. The first fusion feature set includes a non-redundant feature set and a correlated feature set. The non-redundant feature set represents independent feature information of multiple lung regions, and the correlated feature set represents complementary feature information between multiple lung regions.

[0007] One aspect of this disclosure provides a method for classifying lung images, the method comprising: Obtain the second lung image and the feature processing parameters corresponding to the trained lung classification model; Based on the feature processing parameters, feature processing is performed on the second lung image to obtain a second fusion feature set; The second fused feature set is input into the lung classification model to obtain the image classification label corresponding to the output second lung image; The lung classification model is obtained by training the lung classification model according to any embodiment of the present disclosure. The second fusion feature set includes a non-redundant feature set and a related feature set. The non-redundant feature set represents independent feature information of multiple lung regions, and the related feature set represents complementary feature information between multiple lung regions.

[0008] Another aspect of this disclosure provides a training apparatus for a lung classification model, the apparatus comprising: The first lung image set acquisition module is used to acquire a first lung image set and training label data corresponding to the first lung image set. The first lung image set contains multiple first lung images. The first multi-region feature set determination module is used to extract features from the first lung image by region to obtain the first multi-region feature set; The first fusion feature set determination module is used to perform feature correlation analysis on the first multi-region feature set in the region dimension to obtain the first fusion feature set corresponding to the first lung image. The lung classification model training module is used to take each of the first fused feature sets as input data for the untrained lung classification model, and perform iterative training of the model based on the training label data to obtain the trained lung classification model. The first fusion feature set includes a non-redundant feature set and a correlated feature set. The non-redundant feature set represents independent feature information of multiple lung regions, and the correlated feature set represents complementary feature information between multiple lung regions.

[0009] Another aspect of this disclosure provides a classification apparatus for lung images, the apparatus comprising: The second lung image acquisition module is used to acquire the second lung image and the feature processing parameters corresponding to the trained lung classification model. The second fusion feature set determination module is used to perform feature processing on the second lung image according to the feature processing parameters to obtain a second fusion feature set; The image classification label determination module is used to input the second fused feature set into the lung classification model to obtain the image classification label corresponding to the output second lung image; The lung classification model is obtained by training the lung classification model according to any embodiment of the present disclosure. The second fusion feature set includes a non-redundant feature set and a related feature set. The non-redundant feature set represents independent feature information of multiple lung regions, and the related feature set represents complementary feature information between multiple lung regions.

[0010] Another aspect of this disclosure provides an electronic device comprising: At least one processor; and A memory communicatively connected to the at least one processor; wherein, The memory stores a computer program that can be executed by the at least one processor, the computer program being executed by the at least one processor to enable the at least one processor to perform the training method of the lung classification model described in any embodiment of the present disclosure and / or the classification method of the lung image described in any embodiment of the present disclosure.

[0011] Another aspect of this disclosure provides a computer-readable storage medium storing computer instructions that, when executed by a processor, implement the training method for a lung classification model according to any embodiment of this disclosure and / or the classification method for lung images according to any embodiment of this disclosure.

[0012] Another aspect of this disclosure provides a computer program product, including a computer program that, when executed by a processor, implements the training method for the lung classification model described in any embodiment of this disclosure and / or the classification method for lung images described in any embodiment of this disclosure.

[0013] The technical solution of this disclosure involves extracting features from lung images at the region dimension to obtain a multi-region feature set, and then performing feature correlation analysis on the multi-region feature set at the region dimension to obtain a fused feature set containing a non-redundant feature set and a correlated feature set. The non-redundant feature set represents independent feature information of multiple lung regions with independent discriminative power, while the correlated feature set represents complementary feature information between multiple lung regions. The full feature correlation space enables the fused feature set to have both specific and complementary feature representation capabilities. At the same time, the decoupled training paradigm of feature extraction and model training not only reduces the computational complexity of iterative training but also avoids the technical defects of coupled training paradigms, such as parameter coordination bias and slow convergence. The lung classification model can learn a more discriminative decision boundary, thereby reducing the sensitivity of the lung classification model to differences in different application scenarios and improving the accuracy and robustness of the lung classification model.

[0014] It should be understood that the description in this section is not intended to identify key or essential features of the embodiments of this disclosure, nor is it intended to limit the scope of this disclosure. Other features of this disclosure will become readily apparent from the following description. Attached Figure Description

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

[0016] Figure 1 A flowchart illustrating a training method for a lung classification model provided in one embodiment of this disclosure; Figure 2 A flowchart illustrating another method for training a lung classification model provided in one embodiment of this disclosure; Figure 3 A flowchart illustrating another method for training a lung classification model provided in one embodiment of this disclosure; Figure 4 A flowchart illustrating a method for classifying lung images according to an embodiment of this disclosure; Figure 5 A flowchart illustrating a specific example of a lung image classification method provided in one embodiment of this disclosure; Figure 6 This is a schematic diagram of the structure of a training device for a lung classification model provided in one embodiment of the present disclosure; Figure 7 This is a schematic diagram of the structure of a lung image classification device provided in one embodiment of the present disclosure; Figure 8 This is a schematic diagram of the structure of an electronic device provided in one embodiment of the present disclosure. Detailed Implementation

[0017] To enable those skilled in the art to better understand the present disclosure, the technical solutions of the present disclosure will be clearly and completely described below with reference to the accompanying drawings of the embodiments. Obviously, the described embodiments are only some embodiments of the present disclosure, and not all embodiments. Based on the embodiments of the present disclosure, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present disclosure.

[0018] It should be noted that the terms "first," "second," "third," "fourth," etc., used in the specification, claims, and accompanying drawings of this disclosure are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this disclosure described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0019] Figure 1 This is a flowchart illustrating a training method for a lung classification model according to an embodiment of this disclosure. This embodiment is applicable to training a lung classification model for classifying lung images. The method can be executed by a training device for the lung classification model, which can be implemented in hardware and / or software and can be configured in a terminal device. Figure 1 As shown, the method includes: S110. Obtain the first lung image set and the training label data corresponding to the first lung image set.

[0020] In this embodiment, the first lung image set includes multiple first lung images. Each first lung image contains complete image information of the entire lung tissue of the training samples. For example, the data modality of the first lung images includes, but is not limited to, computed tomography (CT) images, magnetic resonance imaging (MRI) images, X-ray images, and other medical images that can reflect the structure of lung tissue. In an optional embodiment, the data modality of the first lung images is a CT image. CT images have high spatial and density resolution, clearly presenting the anatomical structure, subtle textures, and density differences of lung tissue, providing an accurate and reliable image basis for subsequent feature extraction, structural analysis, and processing of the lung region.

[0021] In one optional embodiment, obtaining a first lung image set includes: obtaining a raw lung image set acquired by a medical imaging device, and determining a first lung image set based on image quality conditions and the raw lung image set.

[0022] In one specific embodiment, determining a first lung image set based on image quality conditions and the original lung image set includes: filtering the original lung image set to obtain the first lung image set based on image quality conditions.

[0023] For example, image quality conditions may include image artifact parameters meeting artifact quality ranges and image coverage parameters meeting coverage quality ranges. The image artifact parameters can be signal-to-noise ratio, contrast-to-noise ratio, or mean square error, used to determine whether there are obvious artifacts in the original lung image. These artifacts include motion artifacts, metal artifacts, and breathing artifacts, etc. The presence of artifacts can interfere with the normal display of lung tissue, leading to deviations in subsequent feature extraction. The image coverage parameters can be lung anatomical coverage, scan slice thickness, or effective Z-axis scan length, used to determine whether the original lung image covers the complete lung tissue, ensuring the integrity of subsequent feature extraction.

[0024] In another specific embodiment, determining the first lung image set based on image quality conditions and the original lung image set includes: filtering the original lung image set based on image quality conditions, and preprocessing the filtered original lung image set to obtain the first lung image set.

[0025] For example, preprocessing includes, but is not limited to, grayscale normalization, threshold filtering, and Gaussian filtering. Grayscale normalization is used to unify the grayscale distribution differences between different imaging devices, scanning parameters, and training samples. For instance, it maps the original grayscale values ​​of the original lung image to a normalized range of [-1024HU, 400HU], where HU is a Henlein unit. Threshold filtering is used to remove interference information from non-lung tissues in the original lung image. Gaussian filtering is used to reduce the interference of random noise in the original lung image, which can blur the depiction of the lung's fine structures. For example, a Gaussian convolution kernel can be used for smoothing filtering. Combining this with the texture characteristics of the original lung image, the kernel size can be set to 3×3, and the standard deviation to 0.8, filtering noise while preserving the edge and detail features of the lung tissue to the greatest extent possible.

[0026] Specifically, the preset classification labels in the training label data correspond one-to-one with the first lung images in the first lung image set. The preset classification labels are the quantitative grading labels of the first lung images, representing the quantitative level of the lung tissue corresponding to the first lung image in the pathological dimension. For example, when the preset classification labels are used to quantify density distribution, the classification label system can represent the emphysema level of the lung tissue; when the preset classification labels are used to quantify texture regularity, the classification label system can represent the fibrosis level of the lung tissue.

[0027] S120. Perform feature extraction on the first lung image by region to obtain a first multi-region feature set.

[0028] Specifically, the first multi-region feature set includes lung feature data corresponding to at least two lung regions. In an optional embodiment, the lung region combination includes a whole lung region, a lung parenchyma region, and a target region. The target region represents a region of interest in the lung tissue that is prone to pathological abnormalities. For example, when a preset classification label is used to quantify density distribution, the target region can be a low-density region; when a preset classification label is used to quantify ventilation uniformity, the target region can be an airway structure region; and when a preset classification label is used to quantify texture regularity, the target region can be a lung interstitial region.

[0029] In this embodiment, the step of performing regional feature extraction on the first lung image to obtain a first multi-region feature set includes: performing regional segmentation on the first lung image to obtain at least two lung region images from a whole lung region image, a lung parenchyma region image, and a target region image; and performing feature extraction on each lung region image to obtain the first multi-region feature set.

[0030] The whole-lung region image encompasses complete lung contour information, reflecting the overall morphology, size, and spatial distribution of lung tissue, providing global feature support for model input. The lung parenchyma region image covers the core area of ​​the lung parenchyma, providing local feature support related to the lung parenchyma for model input. When the target region image is a low-density region image, by setting a density threshold range for normal lung tissue, low-density regions below the density threshold range can be separated from the lung parenchyma region image. When the target region image is an airway region image, the airway region image can be segmented from the whole-lung region image based on the grayscale difference and morphological features of the airway wall and lumen. When the target region image is a lung interstitial region image, the lung interstitial region image can be extracted from the surrounding area image of the lung parenchyma region through texture features and spatial distribution features.

[0031] For example, the images of each lung region are uniformly stored as a standardized pixel matrix, and the resolution can be set to 512×512 pixels. This resolution can ensure the clarity of the lung region images, preserve the fine structural features of the lung tissue, and control the amount of image data, avoiding the increase in computational complexity due to excessive resolution, and ensuring the efficiency and accuracy of subsequent feature extraction and model iteration training.

[0032] S130. Perform feature correlation analysis on the first multi-region feature set at the region dimension to obtain the first fusion feature set corresponding to the first lung image.

[0033] Specifically, before executing S130, outlier cleaning and normalization are performed on each of the first multi-region feature sets. Outlier cleaning employs methods such as statistical data analysis, business scenario parameter constraints, or anomaly detection models to remove discrete noise values ​​from the first multi-region feature sets and replace them with standardized feature values. For example, outlier cleaning can use the 3-standard-deviation principle, which identifies outliers in the feature set that deviate from the mean by more than 3 standard deviations and replaces the identified outliers with the median. Normalization is used to uniformly map the range of lung features to the [0,1] interval to eliminate the influence of differences in feature dimensions on correlation. For example, normalization includes, but is not limited to, min-max normalization, Z-score standardization, mean normalization, L2 norm normalization, nonlinear normalization, or quantile normalization. For instance, min-max normalization satisfies the following formula: in, This represents the lung characteristics after normalization. The minimum value representing lung characteristics. The maximum value representing lung characteristics. This represents the lung characteristics before normalization.

[0034] In this embodiment, the first fused feature set includes a non-redundant feature set and a correlated feature set. The non-redundant feature set represents independent feature information of multiple lung regions, and the correlated feature set represents complementary feature information between multiple lung regions.

[0035] Specifically, the non-redundant feature set contains at least one non-redundant lung feature corresponding to each lung region, and the associated feature set contains associated features between the lung features corresponding to the two lung regions respectively. The associated features represent the complementary information between the two lung features.

[0036] In an optional embodiment, performing feature correlation analysis at the region dimension on the first multi-region feature set to obtain a first fused feature set corresponding to the first lung image includes: performing feature filtering on the first multi-region feature set to obtain a non-redundant feature set; performing feature pairing at the region dimension on the non-redundant feature set to obtain an initial paired feature set; determining the correlation coefficient corresponding to each lung feature pair in the initial paired feature set; filtering the initial paired feature set according to each correlation coefficient to obtain a target paired feature set; and performing association processing on each lung feature pair in the target paired feature set to obtain an associated feature set.

[0037] Specifically, feature filtering is used to remove redundant lung features from the first multi-region feature set. Feature filtering methods can include clustering, feature thresholding, feature importance quantification algorithms, or feature redundancy quantification algorithms, but are not limited to the given examples. For instance, when the feature filtering method is clustering, the first multi-region feature set is clustered to obtain at least two clusters. Feature filtering is then performed on each cluster to obtain a non-redundant feature set.

[0038] Specifically, the initial paired feature set contains multiple lung feature pairs, where the two lung features in each pair are derived from two lung region images.

[0039] For example, the feature selection parameters for clusters can be feature importance, which can be determined based on parameters such as feature variance, feature importance score, and physiological contribution value; the correlation coefficient can be Pearson correlation coefficient, Spearman correlation coefficient, or mutual information; the association processing can be feature concatenation, feature product, feature ratio, feature difference, weighted summation, or attention-weighted fusion, etc., but is not limited to the examples given above.

[0040] S140. Use each of the first fusion feature sets as input data for the untrained lung classification model, and perform iterative training of the model based on the training label data to obtain the trained lung classification model.

[0041] Specifically, before iterative training of the model, outlier cleaning and normalization are performed on each first fused feature set. For example, normalization includes, but is not limited to, min-max normalization, Z-score standardization, mean normalization, L2 norm normalization, nonlinear normalization, or quantile normalization.

[0042] The lung classification model is used to predict the category of lung images. For example, the lung classification model can adopt deep learning network architectures such as convolutional neural networks, Transformer networks, residual networks (ResNet) or combinations thereof, but is not limited to the given example.

[0043] In one optional embodiment, the lung classification model consists of a convolutional neural network and a fully connected layer. The convolutional neural network is used to extract abstract vector information of non-redundant or related features from the first fusion feature set, and the fully connected layer is used to perform weighted superposition calculation on the extracted feature vectors and output the image classification result that matches the classification label system.

[0044] For example, the convolutional neural network includes three convolutional layers and two pooling layers. The convolutional kernel size is set to 3×3, the stride is set to 1, and the padding method uses the same mode to ensure that the size of the feature map output by the convolutional layer is consistent with that of the input. The pooling layers can be max pooling layers or average pooling layers to reduce the dimensionality of the feature maps output by the convolutional layers. The fully connected layer can contain an input layer, a hidden layer, and an output layer. The number of neurons in the input layer and the hidden layer are set to 128 and 64, respectively. The number of neurons in the output layer is the same as the number of labels in the labeling system. For example, the training labels are level 0, level 1, level 2, level 3, or level 4, which represent the quantitative levels of lung tissue with no abnormalities, mild, moderate, severe, and extremely severe lung diseases, respectively.

[0045] In an optional embodiment, the step of iteratively training the model based on the training label data to obtain a trained lung classification model includes: dividing the first fused feature set and the training label data into a training set, a validation set, and a test set; iteratively training the untrained lung classification model based on the training set to obtain a lung classification model at the end of the current training round; evaluating the training effect of the lung classification model at the end of the current training round based on the validation set; if the training evaluation result does not meet the training termination condition, adjusting the hyperparameters of the lung classification model to obtain a new untrained lung classification model; repeating the step of iteratively training the untrained lung classification model based on the training set to obtain a lung classification model at the end of the current training round; until the training evaluation result meets the training termination condition, using the lung classification model at the end of the current training round as the initially trained lung classification model; and performing performance testing on the initially trained lung classification model based on the test set; if the performance test parameters meet the performance test conditions, using the initially trained lung classification model as the final trained lung classification model.

[0046] For example, the ratio of training set, validation set and test set is 7:2:1, 8:1:1 or 6:2:2. The loss function in the iterative training process can be the cross-entropy loss function or the mean squared error loss function. The optimizer can be the Adam optimizer, SGD optimizer, RMSprop optimizer or Adagrad optimizer. The learning rate of the preset optimizer can be 0.001 and the decay rate can be 0.9, but it is not limited to the example given above.

[0047] For example, the training termination conditions include, but are not limited to, at least one of the following: the loss value of the lung classification model on the validation set is lower than a preset loss threshold, the validation accuracy is higher than a preset accuracy threshold, the number of iterations reaches a preset number, or an early stop strategy is triggered. For example, if the training evaluation results of the lung classification model for 5 consecutive rounds do not meet the training termination conditions, the model iteration training process is stopped to prevent overfitting of the lung classification model. Performance testing conditions include, but are not limited to, accuracy ≥ 92% and Kappa coefficient ≥ 0.85.

[0048] The technical solution of this embodiment extracts multi-region feature sets from lung images by performing region-dimensional feature extraction. Then, it performs region-dimensional feature correlation analysis on these multi-region feature sets to obtain a fused feature set containing non-redundant and associated feature sets. The non-redundant feature set represents independent feature information with independent discriminative power for multiple lung regions, while the associated feature set represents complementary feature information between multiple lung regions. The full feature correlation space enables the fused feature set to possess both specific and complementary feature representation capabilities. Furthermore, the decoupled training paradigm of feature extraction and model training not only reduces the computational complexity of iterative training but also avoids the technical defects of coupled training paradigms, such as parameter coordination bias and slow convergence. The lung classification model can learn more discriminative decision boundaries, thereby reducing the sensitivity of the lung classification model to differences in different application scenarios and improving the accuracy and robustness of the lung classification model.

[0049] Figure 2 This is a flowchart of another training method for a lung classification model provided in one embodiment of this disclosure. This embodiment further defines the "regional feature extraction of the first lung image to obtain a first multi-region feature set" in the above embodiment. In this embodiment, the regional feature extraction of the first lung image to obtain a first multi-region feature set includes: performing region segmentation on the first lung image to obtain a whole-lung region image, a lung parenchyma region image, and a low-density region image, and performing feature extraction on each lung region image to obtain the first multi-region feature set. For example... Figure 2 As shown, the method includes: S210. Obtain the first lung image set and the training label data corresponding to the first lung image set.

[0050] S210 in this embodiment is the same as or similar to that in the above embodiment, and will not be described again here.

[0051] S220. Perform region segmentation on the first lung image to obtain a whole lung region image, a lung parenchyma region image, and a low-density region image.

[0052] For example, based on a first grayscale threshold, the pixels in the first lung image are divided into lung region points and non-lung region points to obtain a lung mask image. Connected component analysis is performed on the lung mask image, and the two connected components with the largest connected area are retained to obtain the whole lung region image. The two connected components correspond to the left lung connected component and the right lung connected component, respectively.

[0053] Based on the grayscale distribution range, a lung parenchyma mask image is obtained by distinguishing lung parenchyma from non-parenchymal structures such as trachea and blood vessels from the whole lung region image. Morphological closing operations are performed on the lung parenchyma mask image to fill in tiny cavities within the lung parenchyma, resulting in a lung parenchyma region image. A low-density mask image is obtained by pixel classification of the lung parenchyma region image based on a second grayscale threshold. Connected component analysis is then performed on the low-density mask image to remove noise points with connected areas smaller than a preset threshold, resulting in a low-density region image. The first grayscale threshold is greater than the second grayscale threshold.

[0054] This section only provides illustrative examples of segmentation methods for whole lung region images, lung parenchyma region images, and low-density region images, and does not limit them. In some specific embodiments, deep learning models such as the U-NET model, region growing methods, edge detection methods, or atlas matching segmentation methods can also be used to achieve region segmentation of the first lung image.

[0055] S230. Extract features from each lung region image to obtain the first multi-region feature set.

[0056] Specifically, the first multi-region feature set includes whole-lung feature data corresponding to the whole-lung region, lung parenchyma feature data corresponding to the lung parenchyma region, and low-density feature data corresponding to the low-density region.

[0057] Among them, the whole lung feature data represents global quantitative information on the morphological structure and spatial distribution of lung tissue. In an optional embodiment, the whole lung feature data includes at least one of the following: whole lung morphological features, lung density distribution features, lung anatomical zonation features, whole lung texture features, airway association features, lung morphological irregularity features, hilar morphological features, lung parenchyma thickness distribution features, lobar morphological features, and whole lung symmetry features.

[0058] Among them, whole-lung morphological features represent the quantitative information of the overall morphology of lung tissue in three-dimensional space, calculated based on the physical dimensions of voxels calibrated by medical imaging equipment. For example, whole-lung morphological features include, but are not limited to, whole-lung volume, whole-lung long diameter / wide diameter, lobe volume ratio, left-right lung volume ratio, and lung parenchyma ratio. Among them, the whole-lung long diameter is the three-dimensional straight-line distance from the apex to the base of the lung tissue, and the whole-lung wide diameter is the three-dimensional straight-line distance at the widest point of the left and right lungs; the lobe volume ratio is the ratio of the volume of a lobe to the total volume of the whole lung, where the lobe volume is the volume of the left upper lobe, left lower lobe, right upper lobe, right middle lobe, or right lower lobe; the left-right lung volume ratio is the ratio of the total volume of the left lung to the total volume of the right lung; and the lung parenchyma ratio is the ratio of the lung parenchyma volume to the total lung bounding box volume.

[0059] Lung density distribution features characterize the overall distribution of CT values ​​in the lung parenchyma within the lung tissue. This can be obtained by statistically analyzing a distribution histogram constructed from the CT values ​​of all pixels in a whole-lung region image. For example, lung density distribution features include, but are not limited to, the mean, median, standard deviation, skewness, kurtosis, volume percentage below -950 HU, volume percentage between -910 and -950 HU, volume percentage between -800 and -910 HU, and the whole-lung coefficient of variation, all statistically derived from the distribution histogram. Here, -910 to -950 HU represents a typical low-density range, -800 to -910 HU represents a slightly abnormal low-density range, and the whole-lung coefficient of variation, the ratio of the standard deviation to the mean, reflects the density uniformity of the entire lung region.

[0060] The lung anatomical partition features characterize the spatial distribution of 18 standard anatomical partitions in the lung tissue and the quantitative information on the differences between these partitions. These 18 standard anatomical partitions include 8 lung segments in the left lung region and 10 lung segments in the right lung region. For example, the lung anatomical partition features include, but are not limited to, the density difference coefficient between lung segments, the volume percentage of each lung segment, the average CT value, and the distribution percentage of low-density areas. The density difference coefficient represents the ratio of the standard deviation of the average CT value of all lung segments to the average CT value of the entire lung region. The distribution percentage of low-density areas is the ratio of the volume of the low-density area within a lung segment to the volume of the lung segment it contains.

[0061] The whole-lung texture features characterize the overall distribution of grayscale texture in lung tissue and can be extracted using the gray-level co-occurrence matrix and gray-level run-length matrix corresponding to the whole-lung region image. For example, the whole-lung texture features include, but are not limited to, energy, entropy, contrast, correlation, and homogeneity extracted based on the gray-level co-occurrence matrix, and long-run emphasis, short-run emphasis, gray-level non-uniformity, and run-length non-uniformity extracted based on the gray-level run-length matrix. The window size of the gray-level co-occurrence matrix can be 20×20 with a step size of 1. For example, gray-level co-occurrence matrices are constructed in four directions: 0°, 45°, 90°, and 135°. The energy, entropy, contrast, correlation, and homogeneity in the whole-lung texture features are the average values ​​of the texture features extracted from the four gray-level co-occurrence matrices.

[0062] Among them, airway association features characterize the morphological distribution of airway structures in lung tissue. For example, airway association features include, but are not limited to, total airway volume / total lung volume, main bronchus diameter / corresponding lung lobe transverse diameter, average distance from the segmental bronchus opening to the lung segment edge, and airway branch density, where airway branch density represents the number of airway branch points per unit volume.

[0063] Among them, the irregular features of lung morphology characterize the differences between the three-dimensional contour of the whole lung and the standard lung contour, as well as the variations in its own shape. For example, the irregular features of lung morphology include, but are not limited to, the fitting deviation of the whole lung contour, the tilt angle of the lung apex-base contour, and the roughness of the lung edge. The fitting deviation of the whole lung contour is the average distance between each edge point on the three-dimensional contour of the whole lung and the fitted ellipse. The tilt angle of the lung apex-base contour is the angle between the line connecting the lung apex and the lung base and the sagittal plane of the human body. The roughness of the lung edge is the standard deviation of the gray-level gradient of the three-dimensional contour of the whole lung, reflecting the degree of irregularity of the edge of the three-dimensional contour of the whole lung.

[0064] Among them, the hilar morphological features characterize the overall distribution of the hilar region. For example, the hilar morphological features include, but are not limited to, the ratio of hilar volume to total lung volume, the distance from the hilar center to the lung apex / the distance from the hilar center to the lung base, the hilar transverse diameter / total lung transverse diameter, and the boundary ambiguity between the hilar region and the surrounding lung parenchyma, where the boundary ambiguity represents the mean gray-level gradient of each edge point of the hilum.

[0065] Among them, the lung parenchyma thickness distribution characteristics characterize the quantitative information of lung parenchyma thickness in lung tissue. For example, lung parenchyma thickness distribution characteristics include, but are not limited to, the average thickness of the entire lung parenchyma, the standard deviation of thickness, the ratio of parenchyma thickness between lung lobes, and the volume ratio of thin lung parenchyma, where thin lung parenchyma refers to lung parenchyma with a thickness of <2 mm.

[0066] Among them, the morphological features of the lung lobes represent the quantitative information of the morphology of the lung lobes in the lung tissue. For example, the morphological features of the lung lobes include, but are not limited to, the long diameter / transverse diameter of each lung lobe, the roundness of the lung lobe shape, the concavity and convexity of the lung lobe contour, and the average width of the interlobar space, etc. Among them, the roundness of the lung lobe shape is determined based on the lung lobe area and the lung lobe perimeter, and the concavity and convexity of the lung lobe contour is the ratio of the convex bulge area of ​​the lung lobe contour to the actual area.

[0067] The whole-lung symmetry feature characterizes the degree of symmetry or offset between the left and right lungs, defined by the midsagittal plane of the human body as the axis of symmetry. For example, the whole-lung symmetry feature includes, but is not limited to, the morphological symmetry coefficient, the coefficient of variation of the volume difference between the left and right lung segments, the difference in lung apex height, and the center offset distance. The morphological symmetry coefficient is the ratio of the overlapping area of ​​the left and right lung contours to the total area of ​​the whole lung, and the center offset distance is the difference in distance between the geometric centers of the left and right lungs and the axis of symmetry.

[0068] Specifically, lung parenchymal feature data represents quantitative information on the morphological structure and functional attributes of the lung parenchyma. In one optional embodiment, lung parenchymal feature data includes lung parenchymal structural features and / or lung parenchymal functional features. Lung parenchymal structural features represent quantitative information on the morphological structure of the lung parenchyma. For example, lung parenchymal structural features include, but are not limited to, surface area / volume, mean alveolar wall thickness, lung parenchymal texture features, lobular structural integrity, and interstitial texture roughness. The window size of the gray-level co-occurrence matrix corresponding to the lung parenchymal texture features can be 16×16. Lobular structural integrity is the ratio of the number of lobules in the lung parenchyma region to the total number of lobules in the entire lung region. Interstitial texture roughness is the high-frequency texture component of the interstitial region extracted using frequency domain transformation. Lung parenchymal functional features represent quantitative information on the functional attributes of the lung parenchyma. Lung parenchymal functional features include, but are not limited to, perfusion uniformity, ventilatory reserve characteristics, elasticity coefficient, and ventilation efficiency of each lung segment. Among them, perfusion uniformity is the variance of CT values ​​in the lung parenchyma region, ventilation reserve characteristics are the rate of change of parenchymal volume in the inspiratory phase and the parenchymal volume in the expiratory phase, elasticity coefficient is the rate of change of parenchymal density in the inspiratory phase and the parenchymal density in the expiratory phase, and ventilation efficiency is the coefficient of variation of the rate of change of lung segment volume in the inspiratory phase and the lung segment volume in the expiratory phase.

[0069] Specifically, low-density feature data represents quantitative information about the morphological structure and functional attributes of low-density regions. In one optional embodiment, the lung feature data corresponding to the low-density region includes at least one of target area morphological features, density distribution features, and physiological correlation features. For example, target area morphological features include, but are not limited to, the number of low-density regions, maximum region volume, region distribution density, morphological roundness, three-dimensional sphericity, fusion degree, vertical distribution skewness, and horizontal distribution skewness. Among them, the regional distribution density is the number of low-density regions distributed per unit volume; the three-dimensional sphericity is determined based on the volume and surface area of ​​the low-density regions; the fusion degree represents the degree of aggregation of low-density regions, with low-density regions whose number of pixels reaches the threshold being regarded as fused regions, and low-density regions whose number of pixels does not reach the threshold being regarded as independent regions; the fusion degree is the ratio of the number of fused regions to the total number of all low-density regions; the vertical distribution skewness is the difference between the volume proportion of low-density regions in the upper lung region and its volume proportion in the lower lung region; the horizontal distribution skewness is the difference between the volume proportion of low-density regions in the left lung region and its volume proportion in the right lung region; density distribution features include, but are not limited to, average CT value, minimum CT value, CT value distribution range, CT difference between low-density regions and adjacent normal density regions, CT standard deviation, and edge CT gradient, etc., where the edge CT gradient is the average value of the CT difference between edge pixels and internal pixels of low-density regions; physiological correlation features include, but are not limited to, pixel grayscale gradient, alveolar cavity expansion coefficient, and lung parenchyma destruction index, etc. The pixel grayscale gradient reflects the severity of alveolar wall damage. The alveolar cavity expansion coefficient is the ratio of the average voxel spacing of the low-density region to the average voxel spacing of the normal lung parenchyma. The lung parenchyma damage index is the ratio of the volume of the low-density region to the volume of the lung segment in which it is located.

[0070] It should be noted that the " / " in the lung features given in the above embodiments represents the ratio of the two parameters.

[0071] This embodiment achieves comprehensive and refined feature extraction of the entire lung region, lung parenchyma region, and low-density region, realizing full-dimensional feature coverage of macroscopic anatomy, mesoscopic structure, and microscopic pathology. Furthermore, the refined definition of each lung feature improves the objectivity and repeatability of lung features.

[0072] S240. Perform feature correlation analysis on the first multi-region feature set at the region dimension to obtain the first fusion feature set corresponding to the first lung image.

[0073] In this embodiment, the feature correlation analysis at the regional dimension includes three correlation logics: correlation analysis between whole lung feature data and lung parenchymal feature data, correlation analysis between whole lung feature data and low-density feature data, and correlation analysis between lung parenchymal feature data and low-density feature data.

[0074] Specifically, the first matching feature set represents the feature pairing information of the first multi-region feature set in the region dimension. The first matching feature set may contain all the lung feature pairs of the first multi-region feature set, or it may contain some of the lung feature pairs.

[0075] For example, the full set of lung feature pairs can be obtained by arranging and combining the first matching feature set in terms of region dimension, while some lung feature pairs can be customized according to the correlation between the two lung features in terms of physiological meaning, physical meaning and spatial distribution.

[0076] Taking the correlation between the features of the whole lung region and the lung parenchyma region as an example, there is a direct negative correlation between the whole lung volume in the whole lung feature data and the surface area / volume in the lung parenchyma feature data in a physical sense; that is, the more obvious the expansion of the lung parenchyma, the larger the whole lung volume and the lower the surface area / volume. There is a direct positive correlation between the mean whole lung CT value in the whole lung feature data and the perfusion uniformity in the lung parenchyma feature data in spatial distribution; that is, the higher the mean whole lung CT value, the more uniform the density distribution of the whole lung, and correspondingly, the more uniform the perfusion distribution of the lung parenchyma. There is a direct positive correlation between the airway branch density in the whole lung feature data and the ventilation efficiency in the lung parenchyma feature data in a physiological sense. The denser the airway branches in the whole lung region, the higher the ventilation efficiency of the lung parenchyma. The proportion of abnormal HU range in the whole lung feature data is directly negatively correlated with the integrity of lobular structure in the lung parenchyma feature data in a physiological sense. That is, the more pixels in the whole lung region that are in the abnormal HU range, the more severe the damage to the lobules and the worse the integrity of the lobular structure. The entropy value in the whole lung texture feature data is directly positively correlated with the texture roughness of the lung interstitium in the lung parenchyma feature data in a physiological sense. That is, the more chaotic the whole lung texture, the more severe the fibrosis of the lung interstitium and the higher its texture roughness.

[0077] Taking the feature correlation between the whole lung region and low-density regions as an example, the proportion of low-density region volume in the whole lung feature data is directly positively correlated with the number of regions and the degree of fusion in the low-density feature data in terms of physical or physiological significance; the left-right lung volume ratio in the whole lung feature data is directly correlated with the left-right distribution skewness in the low-density feature data in terms of physical significance, that is, the greater the left-right distribution skewness, the more obvious the expansion of the volume of one lung; the total airway volume / total lung volume in the whole lung feature data is directly negatively correlated with the regional distribution density in the low-density feature data in terms of physical significance, that is, the denser the distribution of low-density regions, the more severe the airway damage and the smaller the total airway volume; the density difference coefficient between lung segments in the whole lung feature data is directly positively correlated with the lung parenchyma damage index in the low-density feature data in terms of spatial distribution. The degree of feature correlation between the whole lung feature data and the low-density feature data also positively affects the quantitative grading of the first lung image.

[0078] Taking the correlation between lung parenchyma and low-density regions as an example, the mean alveolar wall thickness in lung parenchyma feature data is directly negatively correlated with alveolar cavity enlargement coefficient and lung parenchyma destruction index in low-density feature data in a physiological sense. That is, the thinner the alveolar wall, the more severe the alveolar cavity destruction. The ventilation reserve feature in lung parenchyma feature data is directly negatively correlated with the degree of fusion in low-density feature data in a physiological sense. That is, the more severe the fusion of low-density regions, the more obvious the lung function impairment and the weaker the ventilation reserve capacity of lung parenchyma. The integrity of lobular structure in lung parenchyma feature data is directly negatively correlated with the three-dimensional sphericity in low-density feature data in a physiological sense. That is, the closer the three-dimensional sphericity is to 1, the higher the difficulty of fusion of low-density regions and the more intact the lobular structure. The perfusion homogeneity in lung parenchyma feature data is directly negatively correlated with the CT standard deviation in low-density feature data in a physical sense. The elasticity coefficient in lung parenchyma feature data is directly negatively correlated with the alveolar cavity enlargement coefficient in low-density feature data in a physiological sense. That is, the enlargement of alveolar cavities leads to a decrease in the elasticity of lung parenchyma. The correlation between low-density feature data and lung parenchymal feature data is mainly physiological. The appearance of low-density regions can damage the alveolar walls of the lung parenchyma and cause alveolar cavity enlargement, leading to physiological changes in key anatomical structures. The feature correlation between the two is the core data support for the quantitative grading of the first lung image.

[0079] In one optional embodiment, the number of feature pairs corresponding to the subset of matching features in the first matching feature set is determined according to the feature association weights of the corresponding association logic. The feature association weights for whole lung region associated with lung parenchyma region, whole lung region associated with low-density region, and lung parenchyma region associated with low-density region increase sequentially. The subset of matching features contains lung feature pairs under the association logic, and the feature association weights represent the importance of the regional feature associations.

[0080] The advantage of this setup is that there is a strong physiological correlation between low-density regions and lung parenchyma regions. The feature correlation between the two can accurately capture the physiological impact of the formation of low-density regions on lung parenchyma. The feature correlation between the whole lung region and low-density regions reflects the macroscopic localization correlation of the target area in lung tissue, which can positively affect the quantitative grading of the first lung image. Furthermore, there is a large amount of inherent overlap in the imaging features of the whole lung region and lung parenchyma regions. The feature correlation between the two carries the anatomical background information of lung tissue. By configuring the gradient number, the effective feature correlation mining between lung regions can be strengthened, while the interference of redundant feature correlations can be weakened, thereby improving the accuracy and effectiveness of feature correlation.

[0081] S250. Use each of the first fused feature sets as input data for the untrained lung classification model, and perform iterative training of the model based on the training label data to obtain the trained lung classification model.

[0082] S250 in this embodiment is the same as or similar to that in the above embodiment, and will not be described again here.

[0083] The technical solution of this embodiment improves the targeting and effectiveness of feature extraction by explicitly setting the target region image as a low-density region image, focusing on the pathological change areas in the lung image. This provides a purer and more discriminative feature foundation for the lung classification model, thereby ensuring the image classification ability of the lung classification model. Through feature correlation analysis guided by low-density regions, the removal effect of redundant lung features is improved, and more systematic and complementary feature support is provided for the lung classification model. This enables the lung classification model to capture the core feature patterns related to the image classification label more quickly and accurately during the training process, thereby improving the stability, discriminativeness and interpretability of the lung classification model.

[0084] Figure 3 This is a flowchart of another training method for a lung classification model provided in one embodiment of this disclosure. This embodiment further refines the "performing region-dimensional feature correlation analysis on the first multi-region feature set to obtain a first fusion feature set corresponding to the first lung image" in the above embodiment. In this embodiment, the step of performing region-dimensional feature correlation analysis on the first multi-region feature set to obtain a first fusion feature set corresponding to the first lung image includes: performing region-dimensional feature pairing on the first multi-region feature set to obtain a first matching feature set, and determining the correlation coefficient matrix corresponding to the first matching feature set; performing redundant feature filtering based on the first matching feature set and the correlation coefficient matrix to obtain a non-redundant feature set and a second matching feature set, wherein the second matching feature set contains at least one lung feature pair that was not filtered in the first matching feature set; and performing association processing on each lung feature pair in the second matching feature set to obtain an associated feature set. Figure 3 As shown, the method includes: S310. Obtain the first lung image set and the training label data corresponding to the first lung image set.

[0085] S320. Perform feature extraction on the first lung image by region to obtain a first multi-region feature set.

[0086] In this embodiment, S310-S320 are the same as or similar to any of the above embodiments, and will not be described again here.

[0087] S330. Perform feature pairing at the region dimension on the first multi-region feature set to obtain a first matching feature set, and determine the correlation coefficient matrix corresponding to the first matching feature set.

[0088] In this embodiment, the method for determining the first matching feature set is the same as described above. Figure 2 The methods for determining the first matching feature set given in the illustrated embodiments are the same or similar, and will not be repeated here.

[0089] Specifically, the matrix elements in the correlation coefficient matrix are the feature correlation coefficients of lung feature pairs in the first matching feature set, and correspond one-to-one with the lung feature pairs in the first matching feature set.

[0090] In one optional embodiment, the matrix elements in the correlation coefficient matrix correspond to the same type of correlation coefficient. For example, the correlation coefficient type can be Pearson correlation coefficient, Spearman correlation coefficient, or mutual information, but is not limited to the given examples.

[0091] In another alternative embodiment, the correlation coefficient matrix is ​​determined based on the association type to which the lung feature pairs in the first matching feature set belong, wherein the association type characterizes the nature and form of the interaction between the two lung features in the lung feature pair.

[0092] In this embodiment, the association type is a linear association conforming to a normal distribution, a linear association not conforming to a normal distribution, or a non-linear association. A linear association conforming to a normal distribution indicates that both lung features are continuous features that approximately follow a normal distribution and are linearly correlated, such as the mean value of whole-lung CT scans and the mean value of alveolar wall thickness. A linear association not conforming to a normal distribution indicates that both lung features are continuous features that do not follow a normal distribution, or that both lung features are ordered features and are linearly correlated, such as lesion fusion degree and lobular structural integrity. A non-linear association indicates that two lung features are non-linearly correlated, such as a curvilinear relationship or other complex relationships, such as whole-lung texture features and regional density features of low-density areas.

[0093] For example, the correlation coefficient types corresponding to linear associations conforming to a normal distribution include, but are not limited to, Pearson correlation coefficients or linear regression coefficients. Taking the Pearson correlation coefficient as an example, the Pearson correlation coefficient ranges from [-1, 1]. The larger the absolute value of the Pearson correlation coefficient, the stronger the linear association. For a pair of lung features, it corresponds to the feature sequences of the two lung features. and characteristic sequences , respectively represented as and , This represents the number of training samples, and the Pearson correlation coefficient. Satisfy the following formula: in, Represents the feature sequence The Middle The feature values ​​corresponding to each training sample Represents the feature sequence The Middle The feature values ​​corresponding to each training sample Represents the feature sequence The mean, Represents the feature sequence The mean, and the numerator represent the characteristic sequence. and characteristic sequences The corresponding covariance, where the denominator represents the characteristic sequence. and characteristic sequences The product of the corresponding standard deviations.

[0094] For example, correlation coefficient types corresponding to linear associations that do not conform to a normal distribution include, but are not limited to, Spearman correlation coefficient, quantile regression coefficient, or robust regression coefficient. Taking the Spearman correlation coefficient as an example, its principle is to determine the degree of association between two lung feature pairs by using the feature ranks corresponding to the two lung features. This can avoid the influence of feature distribution on feature correlation, and its value range is also [-1, 1]. The feature sequence... and characteristic sequences Sort them in ascending order and assign them feature ranks respectively, denoted as follows: and Spearman correlation coefficient Satisfy the following formula: in, Indicates the rank of the feature The Middle The rank corresponding to each training sample Indicates the rank of the feature The Middle The rank of each training sample is represented by the numerator, which is the sum of squared differences in the ranks of all training samples, and the denominator is the standardization coefficient of the ranks of the two features.

[0095] For example, the correlation coefficient types corresponding to nonlinear associations include, but are not limited to, multinomial regression coefficients, mutual information, or smooth curve fitting coefficients. Taking mutual information as an example, its principle is to reflect the degree of association between two lung feature pairs by the amount of information shared between them, accurately measuring the association relationship between lung features of any data type. Statistical feature sequences Marginal probability distribution Feature sequences Marginal probability distribution and the joint probability distribution corresponding to the two feature sequences mutual information Satisfy the following formula: In one specific embodiment, the mutual information of lung feature pairs is subjected to maximum normalization processing, and the mutual information of lung feature pairs is mapped to [0,1]. The larger the value of the normalized mutual information, the stronger the nonlinear correlation between the two lung features.

[0096] For example, the maximum normalization process satisfies the following formula: ; ; ; in, Represents the normalized mutual information value. Representing the characteristic sequence Information entropy Representing the characteristic sequence Information entropy.

[0097] The advantage of this setup is that it enables the correlation coefficient matrix to more accurately and comprehensively characterize the degree of correlation between lung features under different correlation types, avoids the loss of feature correlation information caused by uniform correlation coefficients, effectively improves the pertinence and accuracy of feature correlation measurement, and thus improves the effectiveness and reliability of feature correlation analysis.

[0098] S340. Redundant feature filtering is performed based on the first matching feature set and the correlation coefficient matrix to obtain a non-redundant feature set and a second matching feature set.

[0099] In an optional embodiment, the step of filtering redundant features based on the first matching feature set and the correlation coefficient matrix to obtain a non-redundant feature set and a second matching feature set includes: filtering the first matching feature set according to feature association conditions and the correlation coefficient matrix to obtain a third matching feature set; filtering the third matching feature set according to feature redundancy conditions and the correlation coefficient matrix to obtain a second matching feature set; and removing redundancy from each of the filtered lung feature pairs to obtain a non-redundant feature set.

[0100] The third matching feature set includes lung feature pairs that satisfy the feature association conditions in the first matching feature set. In embodiments with different correlation coefficient types, the feature association conditions include at least two first coefficient thresholds corresponding to different association types. The feature correlation coefficient or its absolute value corresponding to the lung feature pair in the third matching feature set is greater than or equal to the first coefficient threshold of its corresponding association type. For example, the first coefficient thresholds corresponding to the Pearson correlation coefficient, Spearman correlation coefficient, and mutual information are 0.65, 0.6, or 0.6, respectively, but are not limited to the given examples.

[0101] The second matching feature set includes lung feature pairs that do not meet the feature redundancy condition. In embodiments with different correlation coefficient types, the feature redundancy condition includes at least two second coefficient thresholds corresponding to different association types. The feature correlation coefficient or its absolute value corresponding to the lung feature pair in the second matching feature set is less than the second coefficient threshold of its corresponding association type, and the first coefficient threshold is less than the second coefficient threshold. For example, the first coefficient thresholds corresponding to the Pearson correlation coefficient, Spearman correlation coefficient, and mutual information are 0.8, 0.75, or 0.75, respectively, but are not limited to the given examples.

[0102] For example, one lung feature can be randomly removed from the filtered lung feature pair, and the remaining lung features can be used as non-redundant features. Alternatively, the feature importance of each lung feature in the filtered lung feature pair can be determined, and the lung feature with higher feature importance can be used as a non-redundant feature. Feature importance can be determined through feature variance, feature importance score, and physiological importance. For instance, when the lung feature pair consists of "region volume percentage below -950 HU" and "number of pixels in the region below -950 HU", the "region volume percentage below -950 HU" is used as a non-redundant feature because it more intuitively and stably reflects the relative physiological structure and tissue distribution characteristics of the corresponding lung region and has a clearer physiological representation meaning.

[0103] In another optional embodiment, the step of performing redundant feature filtering based on the first matching feature set and the correlation coefficient matrix to obtain a non-redundant feature set and a second matching feature set includes: filtering the first matching feature set according to feature association conditions and the correlation coefficient matrix to obtain a third matching feature set; performing physiological verification filtering on the third matching feature set to obtain a fourth matching feature set, and obtaining the correlation weight matrix of the physiological dimension corresponding to the fourth matching feature set; filtering the correlation coefficient matrix according to the fourth matching feature set to obtain a correlation coefficient sub-matrix, and performing redundant feature filtering on the fourth matching feature set according to the correlation coefficient sub-matrix and the correlation weight matrix to obtain a non-redundant feature set and a second matching feature set.

[0104] Specifically, physiological verification screening refers to selecting lung feature pairs with physiological correlation from the third matching feature set and filtering out lung feature pairs without physiological correlation. Physiological correlation means that there is a clear physiological mechanism relationship between the two lung features, that they can jointly characterize a certain physiological function of the lungs, and that the correlation between the features conforms to the normal physiological structure and function of the lungs. For example, lung feature pairs that meet the physiological correlation conditions can be the proportion of the area below -950 HU and the degree of alveolar destruction, the degree of fusion and ventilation efficiency, and the density of airway branches and the distribution density of low-density areas. Among them, the higher the proportion of the area below -950 HU, the more severe the alveolar destruction and the worse the integrity of the normal alveolar structure, and the physiological correlation between the two is positive. The higher the degree of fusion, the more obstructed the airway and alveolar ventilation channels and the lower the ventilation efficiency, and the physiological correlation between the two is negative. The lower the density of airway branches and the higher the distribution density of low-density areas, the weaker the lung gas exchange function, and the physiological correlation between the two is negative.

[0105] Specifically, the matrix elements in the association weight matrix represent the physiological weight coefficients of the lung feature pairs in the fourth matching feature set, and correspond one-to-one with the lung feature pairs in the fourth matching feature set. The physiological weight coefficients represent the degree of association between the lung feature pairs in the physiological dimension. For example, the range of the physiological weight coefficients can be [0.7, 1].

[0106] Among them, the matrix elements in the correlation coefficient submatrix are the feature correlation coefficients of the lung feature pairs in the fourth matching feature set, and correspond one-to-one with the lung feature pairs in the fourth matching feature set.

[0107] In an optional embodiment, redundant feature filtering is performed on the fourth matching feature set to obtain a non-redundant feature set and a second matching feature set based on the correlation coefficient submatrix and the association weight matrix. This includes: for lung feature pairs in the fourth matching feature set, obtaining the feature correlation coefficient and physiological association weight corresponding to the lung feature pair from the correlation coefficient submatrix and the association weight matrix, respectively; if the feature correlation coefficient and physiological association weight satisfy the feature redundancy condition, the lung feature pair is treated as a redundant feature pair, and one of the lung features in the redundant feature pair is retained as a non-redundant feature and added to the non-redundant feature set; if the feature correlation coefficient and physiological association weight do not satisfy the feature redundancy condition, the lung feature pair is added to the second matching feature set.

[0108] In embodiments with different correlation coefficient types, the feature redundancy condition includes an association weight threshold and a second coefficient threshold corresponding to at least two association types. The physiological association weight of the lung feature pair in the second matching feature set is less than the association weight threshold, and the feature correlation coefficient or its absolute value is less than the second coefficient threshold of its corresponding association type. For example, the association weight threshold can be 0.8.

[0109] In another optional embodiment, redundant feature filtering is performed on the fourth matching feature set to obtain a non-redundant feature set and a second matching feature set based on the correlation coefficient submatrix and the association weight matrix. This includes: for lung feature pairs in the fourth matching feature set, obtaining the feature correlation coefficient and physiological association weight corresponding to the lung feature pair from the correlation coefficient submatrix and the association weight matrix, respectively; determining the feature association score based on the feature correlation coefficient and physiological association weight; if the feature association score is greater than a score threshold, treating the lung feature pair as a redundant feature pair and retaining one of the lung features in the redundant feature pair as a non-redundant feature added to the non-redundant feature set; if the feature association score is less than or equal to the score threshold, adding the lung feature pair to the second matching feature set. Here, the feature association score represents the weighted sum of the feature correlation coefficient and the physiological association weight.

[0110] The advantage of setting up physiological verification and screening, as well as the association weight matrix, is that it solves the problem that the feature correlation coefficient at the mathematical and statistical level cannot reflect the physiological rationality and actual representational value of feature association, and is prone to false screening that is statistically correlated but physiologically meaningless. It can filter out feature pairs that have no physiological correlation significance, retain lung feature pairs that can truly reflect the physiological structure, function or physiological state of the lungs, and ensure the physiological effectiveness of the fused feature set. It can also assign differentiated weights to different lung feature pairs through physiological correlation coefficients, so that the redundant filtering process is both in line with mathematical and statistical laws and fits the physiological mechanism of the lungs. The final fused feature set is more targeted and practical.

[0111] In another optional embodiment, redundant feature filtering is performed on the fourth matching feature set to obtain a non-redundant feature set and a second matching feature set based on the correlation coefficient submatrix and the association weight matrix, including: redundant feature filtering is performed on the fourth matching feature set to obtain a non-redundant feature set and a fifth matching feature set based on the correlation coefficient submatrix and the association weight matrix; an association type set corresponding to the fifth matching feature set is obtained, the association type set containing at least two association types; for each lung feature pair in the fifth matching feature set, a correlation coefficient sequence corresponding to the lung feature pair is determined based on at least one association type in the association type set other than the association type to which the lung feature pair belongs; the feature correlation coefficient and physiological association weight corresponding to the lung feature pair are obtained from the correlation coefficient submatrix and the association weight matrix, respectively; and a feature association score of the lung feature pair is determined based on the feature correlation coefficient, physiological association weight, and correlation coefficient sequence; and the fifth matching feature set is filtered based on each of the feature association scores to obtain the second matching feature set.

[0112] The correlation coefficient sequence includes correlation coefficients corresponding to at least one association type other than the association type to which the lung feature belongs.

[0113] In an optional embodiment, determining the feature association score of the lung feature pair based on the feature correlation coefficient, physiological association weight, and correlation coefficient sequence includes: multiplying the feature correlation coefficient, physiological association weight, correlation coefficient sequence, and their respective weights, and summing all the product results to obtain the feature association score of the lung feature pair.

[0114] In another optional embodiment, determining the feature association score of the lung feature pair based on the feature correlation coefficient, physiological association weight, and correlation coefficient sequence includes: determining the feature association strength of the lung feature pair based on the feature correlation coefficient and correlation coefficient sequence; and determining the feature association score of the lung feature pair based on the feature association strength and the physiological association weight.

[0115] The feature association strength is used to characterize the degree of association of the lung feature with itself and the degree of synergistic association with other associated features.

[0116] For example, feature association score Satisfy the following formula: in, , and Indicates the coefficient weight. This represents the Pearson correlation coefficient. Represents the Spearman correlation coefficient. Represents the normalized mutual information value. This represents the weight of physiological associations. For example, , and The values ​​can be 0.4, 0.3, and 0.3 respectively.

[0117] Based on the above embodiments, optionally, the fifth matching feature set is filtered to obtain the second matching feature set according to the feature association scores of each of the above-described features, including: filtering the fifth matching feature set to obtain the second matching feature set according to the feature filtering conditions and the feature association scores of each of the above-described features.

[0118] The feature selection criteria can be feature selection score, feature selection ratio, or feature selection quantity, etc. For example, lung feature pairs with feature association scores greater than or equal to the feature selection score in the fifth matching feature set are added to the second matching feature set. For example, the feature selection score can be 0.6.

[0119] A single correlation type cannot comprehensively evaluate the correlation value of lung features within the entire feature system, and is prone to selection biases where a single feature has a high correlation coefficient but poor synergy with other feature types. This embodiment achieves a multi-dimensional and comprehensive evaluation of the correlation value of lung features by setting feature correlation coefficients, breaking the limitations of relying solely on a single correlation coefficient, improving the effectiveness and relevance of the fused feature set, providing higher-quality feature input for subsequent model iteration training, and further improving the classification performance and robustness of the lung classification model.

[0120] S350. Perform association processing on each lung feature pair in the second matching feature set to obtain an associated feature set.

[0121] S360. Use each of the first fusion feature sets as input data for the untrained lung classification model, and perform iterative training of the model based on the training label data to obtain the trained lung classification model.

[0122] In this embodiment, S350-S360 are the same as or similar to those in the above embodiment, and will not be described again here.

[0123] The technical solution of this embodiment solves the problem of lost feature associations between regions caused by redundant filtering followed by feature pairing by defining a second matching feature set containing at least one lung feature pair that was not filtered in the first matching feature set. It fully preserves the regional association rules of lung feature pairs in the feature correlation analysis process, accurately removes redundant features and retains complementary features based on the regional association rules, improves the relevance and representation ability of the associated feature set, and optimizes the training basis of the lung classification model from the root, thereby reducing the training cost and generalization ability of the lung classification model, making the trained lung classification model more suitable for complex lung image classification scenarios.

[0124] Figure 4 This is a flowchart illustrating a lung image classification method according to an embodiment of this disclosure. This embodiment is applicable to the classification of lung images. The method can be executed by a lung image classification device, which can be implemented in hardware and / or software and can be configured in a terminal device. Figure 4 As shown, the method includes: S410. Obtain the second lung image and the feature processing parameters corresponding to the trained lung classification model.

[0125] Specifically, the second lung image is the lung image that needs to be classified and identified. Its data modality is consistent with the data modality of the first lung image set used in the lung classification model. For example, it is a computed tomography image, magnetic resonance imaging image or X-ray image, to ensure the consistency and comparability of input features.

[0126] For example, the second lung image is obtained by taking the original lung image from medical imaging equipment and then performing the same preprocessing operation as the model iterative training process. The consistency of the preprocessing method is the key to avoid introducing additional feature bias and ensuring the accuracy of subsequent classification. Such preprocessing includes, but is not limited to, grayscale standardization, threshold filtering and Gaussian filtering.

[0127] In this embodiment, the lung classification model is obtained using the training method of the lung classification model described in any embodiment of this disclosure. The feature processing parameters can be directly read from the parameter file of the trained lung classification model. These parameters define the rules followed by the model in processing the input image during the inference phase, including image segmentation parameters, feature extraction parameters, and feature correlation parameters. The image segmentation parameters define multiple lung regions corresponding to the lung image, and segmentation parameters corresponding to each lung region, such as segmentation grayscale threshold, contour constraints, and connectivity constraints. The feature extraction parameters include extraction parameters for each lung feature corresponding to each lung region image, such as convolution kernel size, scale recovery parameters, and histogram construction parameters. The feature correlation parameters are used to define the relationships between lung features and the redundancy determination criteria, such as feature association conditions and feature redundancy conditions.

[0128] S420. Based on the feature processing parameters, perform feature processing on the second lung image to obtain a second fusion feature set.

[0129] Specifically, the second fusion feature set maintains consistency with the first fusion feature set generated during the model iterative training phase in terms of feature composition and correlation, thereby ensuring that the fusion feature set can be correctly decoded and classified by the lung classification model. In this embodiment, the second fusion feature set includes a non-redundant feature set and a correlated feature set. The non-redundant feature set represents independent feature information of multiple lung regions, and the correlated feature set represents complementary feature information between multiple lung regions.

[0130] Specifically, based on the feature processing parameters, feature processing is performed on the second lung image to obtain a second fusion feature set, including: segmenting the second lung image into regions according to region division parameters to obtain a whole-lung region image, a lung parenchyma region image, and a target region image; extracting features from each lung region image according to feature extraction parameters to obtain a second multi-region feature set; and determining the second fusion feature set based on the feature correlation parameters and the second multi-region feature set. The feature correlation parameters include feature pairing parameters, feature redundancy conditions, feature association conditions, and feature association parameters. The feature pairing parameters instruct the second multi-region feature set to perform feature pairing; the feature redundancy conditions instruct the filtering out of redundant lung features in the second multi-region feature set; the feature association conditions instruct the retention of non-redundant lung feature pairs; and the feature association parameters instruct the specific association method between two lung features.

[0131] S430. Input the second fused feature set into the lung classification model to obtain the image classification label corresponding to the output second lung image.

[0132] Specifically, the lung classification model outputs a predicted probability value for each preset classification label. The predicted probability value represents the confidence level of determining that the second lung image belongs to the preset classification label. The preset classification label with the highest predicted probability value is taken as the image classification label corresponding to the second lung image.

[0133] In one specific embodiment, the method further includes: generating an image analysis report based on the image classification label, the predicted probability value corresponding to each classification label, and the contribution of key features. The contribution of key features includes one or more lung features that have a significant impact on the classification decision. For example, the influence weight of each lung feature is determined using interpretability algorithms, feature importance ranking algorithms, and gradient boosting decision tree algorithms.

[0134] Optionally, based on the above embodiments, the method further includes: when the predicted probability value corresponding to the image classification label is less than a preset probability threshold, obtaining a training sample medical record dataset corresponding to the first lung image set; determining a third multi-region feature set corresponding to the first lung image, and determining a third fusion feature set based on the third multi-region feature set, and adding the training sample medical record data corresponding to the first lung image in the training sample medical record dataset to the third fusion feature set to obtain a fourth fusion feature set; and re-iteratio training the untrained lung classification model based on the training label data and each of the fourth fusion feature sets.

[0135] For example, the preset probability threshold can be 0.75. If the predicted probability value is less than the preset probability threshold, it means that the image classification label has low credibility. The lung classification model may have problems such as insufficient feature representation and insufficient model generalization ability, and the iterative training process needs to be restarted. If the predicted probability value is greater than or equal to the preset probability threshold, it means that the image classification label has high credibility. There is no need to restart the training process, and the image classification label can be output directly.

[0136] Specifically, the training sample medical record dataset contains sample medical record data that corresponds one-to-one with the training samples in the first lung image set. The sample medical record data contains a large amount of textual and numerical information related to lung physiology and pathology. For example, the sample medical record data includes, but is not limited to, basic sample information, vital sign parameters, pathology reports, years of smoking, lung function indicators, inflammatory indicators, and other information.

[0137] In an optional embodiment, the feature extraction parameters corresponding to the third multi-region feature set are the same as those corresponding to the first multi-region feature set, and the feature correlation parameters of the feature correlation analysis corresponding to the third fusion feature set are the same as those of the feature correlation analysis corresponding to the first fusion feature set. This embodiment optimizes the lung classification model's ability to classify similar feature inputs by adding medical record information to the training samples to supplement the feature dimensions without changing the feature extraction and correlation analysis rules.

[0138] In another optional embodiment, the feature extraction parameters corresponding to the third multi-region feature set are different from the feature extraction parameters corresponding to the first multi-region feature set, and / or the feature correlation parameters of the feature correlation analysis corresponding to the third fused feature set are different from the feature correlation parameters of the feature correlation analysis corresponding to the first fused feature set.

[0139] This embodiment improves feature extraction quality and feature selection accuracy by adjusting feature processing parameters to uncover more deep features in lung images.

[0140] Specifically, different feature extraction parameters represent differences in feature extraction methods, extraction dimensions, and extraction accuracy. For example, feature extraction parameters include, but are not limited to, matrix parameters, shape fitting parameters, and statistical parameters of the gray-level co-occurrence matrix; feature pairing parameters include, but are not limited to, physiological association libraries and pairing driving parameters; feature redundancy conditions include association weight thresholds and second coefficient thresholds for each association type; feature association conditions include first coefficient thresholds for each association type; and different feature association parameters represent different specific association methods.

[0141] Specifically, the fourth fusion feature set includes a non-redundant feature set, a related feature set, and training sample medical record data, providing richer feature inputs for iterative model training.

[0142] Specifically, the original lung classification model is replaced by the re-trained lung classification model for subsequent lung image classification inference, ensuring the accuracy and reliability of subsequent image classification. Simultaneously, the feature processing parameters generated during model iteration training are stored in a model parameter file to provide a reference for subsequent model optimization.

[0143] The benefit of setting up a retraining mechanism is that it improves the generalization ability and classification accuracy of the lung classification model, enabling it to adapt to more complex lung image scenarios.

[0144] Based on the above embodiments, optionally, the method further includes: determining a label classification bias based on the image classification label and the true classification label of the second lung image; adding the second lung image to an incremental lung image set when the label classification bias meets the classification bias condition; in response to the number of images corresponding to the incremental lung image set reaching a preset number threshold, obtaining a second fusion feature set and a true classification label corresponding to each second lung image in the incremental lung image set; and incrementally training the lung classification model based on each of the second fusion feature sets and each of the true classification labels to obtain an optimized lung classification model.

[0145] Specifically, the true classification label is the actual classification result of the second lung image obtained through clinical diagnosis, pathological examination, etc., which has extremely high accuracy and is the core basis for judging whether the model classification result is accurate. The label classification deviation indicates the degree of difference between the image classification label output by the model and the true classification label. The larger the deviation, the further the model's classification result deviates from the true result, and the weaker the model's ability to classify this type of lung image. For example, the label classification deviation includes, but is not limited to, the difference in rank or the difference in probability.

[0146] Specifically, the classification bias condition indicates that the image classification label deviates significantly from the true classification label. For example, the label classification bias can be a difference in level greater than 1 level or a difference in probability greater than 10 points, but it is not limited to the example given above.

[0147] Specifically, the preset number threshold can be set according to the actual application scenario. For example, the preset number threshold can be 50 cases. If the number of images in the incremental lung image set reaches this threshold, it means that there are enough samples for incremental training, which can effectively improve the model performance. If the preset number threshold is not reached, the second lung images with large label classification deviations will continue to be collected until the preset number threshold is reached.

[0148] For example, incremental training can be carried out by fine-tuning, using the original parameters of the lung classification model as initial parameters, freezing the bottom feature extraction layer of the lung classification model, and updating the parameters only for the top classification layer and some feature extraction layers. During incremental training, the stochastic gradient descent optimization algorithm can be used to adjust the model parameters of the lung classification model. The learning rate can be set to 1 / 10 of the original training learning rate, and the number of iterations can be set to 20-50 rounds to ensure the training efficiency of the lung classification model.

[0149] The benefits of setting up an incremental training mechanism are that it can improve the lung classification model's ability to classify lung image sets, specifically addressing the problem of large model classification bias, while avoiding the degradation of the original model's performance. It eliminates the need to retrain the entire model, significantly reducing training costs and shortening the training cycle. At the same time, it allows the model to continuously adapt to new types of lung images, improving the model's generalization ability and clinical adaptability, ensuring that the model can provide reliable support for clinical diagnosis in the long term.

[0150] Figure 5 This flowchart illustrates a specific example of a lung image classification method provided in one embodiment of this disclosure. Specifically, it involves acquiring lung CT images obtained from medical imaging equipment, preprocessing and segmenting the lung CT images to obtain whole-lung region images, lung parenchyma region images, and low-density region images. Multidimensional feature extraction is performed on each lung region image to obtain whole-lung feature data, lung parenchyma feature data, and low-density feature data. Feature correlation analysis is performed on each lung feature data to construct a fused feature set. A trained lung classification model is loaded, and the fused feature set is input into the lung classification model to obtain predictions corresponding to each preset classification level. The probability value is used to determine whether the maximum predicted probability value is greater than or equal to 0.75. If not, the lung classification model is retrained iteratively by supplementing medical record features in the fusion feature set and / or constructing a new fusion feature set. If so, the preset classification level corresponding to the predicted probability value is used as the image classification level. The level classification deviation between the image classification level and the true classification level corresponding to the lung CT image is obtained. The level classification deviation is determined to be greater than 1 level. If not, the lung image classification process ends. If so, the preprocessed lung CT image is added to the incremental lung image set, and the lung classification model is continuously incrementally trained periodically or quantitatively.

[0151] The technical solution of this embodiment, by calling the solidified feature processing parameters of the trained lung classification model, ensures that the second fusion feature set is highly consistent with the first fusion feature set generated during the model iteration training stage in terms of feature dimension, structure, and intrinsic correlation. This ensures the consistency of the lung classification model throughout its entire lifecycle from training to application, eliminates feature distribution shifts caused by differences in feature processing procedures, and improves the classification accuracy of lung images. The non-redundant feature set and the associated feature set provide the lung classification model with a multi-level and complementary feature perspective, enabling it to capture the fine structural information and cross-regional contextual relationships of lung images more comprehensively and accurately, thereby improving the robustness of lung image classification.

[0152] It should be noted that the collection, use, storage, sharing and transfer of user personal information involved in the technical solution disclosed herein all comply with the provisions of relevant laws and regulations, and require notification to users and obtaining their consent or authorization. Where applicable, user personal information has been subjected to de-identification and / or anonymization and / or encryption technical processing.

[0153] The following are embodiments of the lung image classification device provided in this disclosure. This device and the lung image classification method described above belong to the same disclosed concept. For details not described in detail in the embodiments of the lung image classification device, please refer to the content of the lung image classification method in the above embodiments.

[0154] Figure 6 This is a schematic diagram of the structure of a training device for a lung classification model provided in one embodiment of this disclosure. Figure 6 As shown, the device includes: a first lung image set acquisition module 510, a first multi-region feature set determination module 520, a first fusion feature set determination module 530, and a lung classification model training module 540.

[0155] The first lung image set acquisition module 510 is used to acquire a first lung image set and training label data corresponding to the first lung image set. The first lung image set contains multiple first lung images. The first multi-region feature set determination module 520 is used to extract features from the first lung image by region to obtain the first multi-region feature set; The first fusion feature set determination module 530 is used to perform feature correlation analysis on the first multi-region feature set in the region dimension to obtain the first fusion feature set corresponding to the first lung image. The lung classification model training module 540 is used to take each of the first fused feature sets as input data for the untrained lung classification model, and perform iterative training of the model based on the training label data to obtain the trained lung classification model. The first fusion feature set includes a non-redundant feature set and a correlated feature set. The non-redundant feature set represents independent feature information of multiple lung regions, and the correlated feature set represents complementary feature information between multiple lung regions.

[0156] The technical solution of this embodiment extracts multi-region feature sets from lung images by performing region-dimensional feature extraction. Then, it performs region-dimensional feature correlation analysis on these multi-region feature sets to obtain a fused feature set containing non-redundant and associated feature sets. The non-redundant feature set represents independent feature information with independent discriminative power for multiple lung regions, while the associated feature set represents complementary feature information between multiple lung regions. The full feature correlation space enables the fused feature set to possess both specific and complementary feature representation capabilities. Furthermore, the decoupled training paradigm of feature extraction and model training not only reduces the computational complexity of iterative training but also avoids the technical defects of coupled training paradigms, such as parameter coordination bias and slow convergence. The lung classification model can learn more discriminative decision boundaries, thereby reducing the sensitivity of the lung classification model to differences in different application scenarios and improving the accuracy and robustness of the lung classification model.

[0157] In an optional embodiment, the first fusion feature set determination module 530 includes: The correlation coefficient matrix determination unit is used to perform feature pairing at the region dimension on the first multi-region feature set to obtain a first matching feature set, and to determine the correlation coefficient matrix corresponding to the first matching feature set. The second matching feature set determination unit is used to perform redundant feature filtering based on the first matching feature set and the correlation coefficient matrix to obtain a non-redundant feature set and a second matching feature set. The second matching feature set contains at least one lung feature pair that was not filtered in the first matching feature set. The associated feature set determination unit is used to perform association processing on each lung feature pair in the second matching feature set to obtain an associated feature set.

[0158] In an optional embodiment, the second matching feature set determination unit includes: The third matching feature set determination subunit is used to filter the first matching feature set to obtain the third matching feature set based on the feature association conditions and the correlation coefficient matrix. The association weight matrix acquisition sub-unit is used to perform physiological verification and screening on the third matching feature set to obtain the fourth matching feature set, and to obtain the association weight matrix of the physiological dimension corresponding to the fourth matching feature set. The second matching feature set determination subunit is used to filter the correlation coefficient matrix according to the fourth matching feature set to obtain a correlation coefficient submatrix, and to perform redundant feature filtering on the fourth matching feature set according to the correlation coefficient submatrix and the association weight matrix to obtain a non-redundant feature set and a second matching feature set.

[0159] In an optional embodiment, the correlation coefficient matrix is ​​determined based on the association type to which the lung feature pairs in the first matching feature set belong; Accordingly, the second matching feature set determines the sub-unit, specifically used for: Based on the correlation coefficient submatrix and the association weight matrix, redundant feature filtering is performed on the fourth matching feature set to obtain a non-redundant feature set and a fifth matching feature set. Obtain the association type set corresponding to the fifth matching feature set, wherein the association type set contains at least two association types; For each lung feature pair in the fifth matching feature set, based on at least one association type in the association type set other than the association type to which the lung feature pair belongs, the correlation coefficient sequence corresponding to the lung feature pair is determined. The feature correlation coefficient and physiological association weight corresponding to the lung feature pair are obtained from the correlation coefficient submatrix and the association weight matrix, respectively. Based on the feature correlation coefficient, physiological association weight and correlation coefficient sequence, the feature association score of the lung feature pair is determined. Based on the correlation scores of each feature, the fifth matching feature set is filtered to obtain the second matching feature set.

[0160] In an optional embodiment, the second matching feature set determines the subunit, specifically for: The feature association strength of the lung feature pairs is determined based on the feature correlation coefficients and the correlation coefficient sequence. The feature association score of the lung feature pair is determined based on the feature association strength and the physiological association weight.

[0161] The training device for the lung classification model provided in this disclosure can execute the training method for the lung classification model provided in any embodiment of this disclosure, and has the corresponding functional modules and beneficial effects of the execution method.

[0162] Figure 7 This is a schematic diagram of the structure of a lung image classification device provided in one embodiment of this disclosure. Figure 7 As shown, the device includes: a second lung image acquisition module 610, a second fusion feature set determination module 620, and an image classification label determination module 630.

[0163] The second lung image acquisition module 610 is used to acquire the second lung image and the feature processing parameters corresponding to the trained lung classification model. The second fusion feature set determination module 620 is used to perform feature processing on the second lung image according to the feature processing parameters to obtain a second fusion feature set; The image classification label determination module 630 is used to input the second fused feature set into the lung classification model to obtain the image classification label corresponding to the output second lung image; The lung classification model is obtained by training the lung classification model according to any embodiment of the present disclosure. The second fusion feature set includes a non-redundant feature set and a related feature set. The non-redundant feature set represents independent feature information of multiple lung regions, and the related feature set represents complementary feature information between multiple lung regions.

[0164] The technical solution of this embodiment, by calling the solidified feature processing parameters of the trained lung classification model, ensures that the second fusion feature set is highly consistent with the first fusion feature set generated during the model iteration training stage in terms of feature dimension, structure, and intrinsic correlation. This ensures the consistency of the lung classification model throughout its entire lifecycle from training to application, eliminates feature distribution shifts caused by differences in feature processing procedures, and improves the classification accuracy of lung images. The non-redundant feature set and the associated feature set provide the lung classification model with a multi-level and complementary feature perspective, enabling it to capture the fine structural information and cross-regional contextual relationships of lung images more comprehensively and accurately, thereby improving the robustness of lung image classification.

[0165] In an optional embodiment, the device further includes: The model retraining module is used to obtain the training sample medical record dataset corresponding to the first lung image set when the predicted probability value corresponding to the image classification label is less than a preset probability threshold. A third multi-region feature set corresponding to the first lung image is determined, and a third fusion feature set is determined based on the third multi-region feature set. The training sample medical record data corresponding to the first lung image in the training sample medical record dataset is added to the third fusion feature set to obtain a fourth fusion feature set. Based on the training label data and each of the fourth fusion feature sets, the untrained lung classification model is retrained iteratively.

[0166] In an optional embodiment, the feature extraction parameters corresponding to the third multi-region feature set are different from the feature extraction parameters corresponding to the first multi-region feature set, and / or the correlation parameters of the feature correlation analysis corresponding to the third fused feature set are different from the correlation parameters of the feature correlation analysis corresponding to the first fused feature set.

[0167] In an optional embodiment, the device further includes: The model incremental training module is used to determine the label classification bias based on the image classification label and the true classification label of the second lung image; If the label classification deviation meets the classification deviation condition, the second lung image is added to the incremental lung image set; In response to the number of images corresponding to the incremental lung image set reaching a preset threshold, the second fusion feature set and the true classification label corresponding to each second lung image in the incremental lung image set are obtained; Based on each of the second fusion feature sets and each of the true classification labels, the lung classification model is incrementally trained to obtain an optimized lung classification model.

[0168] The lung image classification device provided in this disclosure can execute the lung image classification method provided in any embodiment of this disclosure, and has the corresponding functional modules and beneficial effects for executing the method.

[0169] Figure 8 This is a schematic diagram of an electronic device provided according to one embodiment of the present disclosure. The electronic device 10 is intended to represent various forms of digital computers, such as laptop computers, desktop computers, workbenches, servers, blade servers, mainframe computers, and other suitable computers. The electronic device may also represent various forms of mobile devices, such as personal digital assistants, cellular phones, smartphones, wearable devices (such as helmets, glasses, watches, etc.), and other similar computing devices. The components shown herein, their connections and relationships, and their functions are merely illustrative and are not intended to limit the implementation of the present disclosure described and / or claimed herein.

[0170] like Figure 8 As shown, the electronic device 10 includes at least one processor 11 and a memory, such as a read-only memory (ROM) 12 or a random access memory (RAM) 13, communicatively connected to the at least one processor 11. The memory stores computer programs executable by the at least one processor 11. The processor 11 can perform various appropriate actions and processes based on the computer program stored in the ROM 12 or loaded from storage unit 18 into the RAM 13. The RAM 13 can also store various programs and data required for the operation of the electronic device 10. The processor 11, ROM 12, and RAM 13 are interconnected via a bus 14. An input / output (I / O) interface 15 is also connected to the bus 14.

[0171] Multiple components in electronic device 10 are connected to I / O interface 15, including: input unit 16, such as keyboard, mouse, etc.; output unit 17, such as various types of displays, speakers, etc.; storage unit 18, such as disk, optical disk, etc.; and communication unit 19, such as network card, modem, wireless transceiver, etc. Communication unit 19 allows electronic device 10 to exchange information or data with other devices through computer networks such as the Internet and / or various telecommunications networks.

[0172] Processor 11 can be various general-purpose and / or special-purpose processing components with processing and computing capabilities. Some examples of processor 11 include, but are not limited to, a central processing unit (CPU), a graphics processing unit (GPU), various special-purpose artificial intelligence (AI) computing chips, various processors running machine learning model algorithms, digital signal processors (DSPs), and any suitable processor, controller, microcontroller, etc. Processor 11 performs the various methods and processes described above, such as the training method for the lung classification model and / or the classification method for lung images provided in the above embodiments.

[0173] In some embodiments, the training method for the lung classification model and / or the classification method for lung images provided in the above embodiments can be implemented as a computer program tangibly contained in a computer-readable storage medium, such as storage unit 18. In some embodiments, part or all of the computer program can be loaded and / or installed on the electronic device 10 via ROM 12 and / or communication unit 19. When the computer program is loaded into RAM 13 and executed by processor 11, one or more steps of the training method for the lung classification model and / or the classification method for lung images described above can be performed. Alternatively, in other embodiments, processor 11 can be configured to perform the training method for the lung classification model and / or the classification method for lung images by any other suitable means (e.g., by means of firmware).

[0174] In particular, according to embodiments of this disclosure, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments of this disclosure include a computer program product comprising a computer program carried on a non-transitory computer-readable medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via communication unit 19, or installed from storage unit 18, or installed from ROM 12. When the computer program is executed by processor 11, it performs the functions defined in the methods of embodiments of this disclosure.

[0175] Various embodiments of the systems and techniques described above herein can be implemented in the following systems or combinations thereof: digital electronic circuit systems, integrated circuit systems, field-programmable gate arrays (FPGAs), application-specific integrated circuits (ASICs), application-specific standard parts (ASSPs), system-on-chips (SOCs), complex programmable logic devices (CPLDs), computer hardware, firmware, software, and / or combinations thereof. These various embodiments may include implementations in one or more computer programs that can be executed and / or interpreted on a programmable system including at least one programmable processor, which may be a dedicated or general-purpose programmable processor, capable of receiving data and instructions from a storage system, at least one input device, and at least one output device, and transmitting data and instructions to the storage system, the at least one input device, and the at least one output device.

[0176] Computer programs used for implementing the training methods of the lung classification model and / or the classification methods of lung images of this disclosure can be written in any combination of one or more programming languages. These computer programs can be provided to a processor of a general-purpose computer, a special-purpose computer, or other programmable data processing device, such that when executed by the processor, the computer programs cause the functions / operations specified in the flowcharts and / or block diagrams to be implemented. The computer programs can be executed entirely on a machine, partially on a machine, as a standalone software package partially on a machine and partially on a remote machine, or entirely on a remote machine or server.

[0177] In the context of this application, a computer-readable storage medium can be a tangible medium that may contain or store a computer program for use by or in conjunction with an instruction execution system, apparatus, or device. A computer-readable storage medium can be, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatus, or devices, or any suitable combination of the foregoing. Alternatively, a computer-readable storage medium can be a machine-readable storage medium. Examples of machine-readable storage media include, based on an electrical connection of at least one wire, a portable computer disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM), flash memory, optical fiber, compact disc read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination of the foregoing.

[0178] To provide interaction with a user, the systems and techniques described herein can be implemented on a terminal device having: a display device for displaying information to the user (e.g., a cathode-ray tube (CRT) or liquid crystal display (LCD) monitor); and a keyboard and pointing device (e.g., a mouse or trackball) through which the user provides input to the terminal device. Other types of devices can also provide interaction with the user; for example, feedback provided to the user can be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and input from the user can be received in any form (including sound input, voice input, or tactile input).

[0179] The systems and technologies described herein can be implemented in computing systems that include backend components (e.g., as data servers), or middleware components (e.g., application servers), or frontend components (e.g., user computers with graphical user interfaces or web browsers through which users can interact with implementations of the systems and technologies described herein), or any combination of such backend, middleware, or frontend components. The components of the system can be interconnected via digital data communication of any form or medium (e.g., communication networks). Examples of communication networks include local area networks (LANs), wide area networks (WANs), blockchain networks, and the Internet.

[0180] A computing system can include clients and servers. Clients and servers are generally located far apart and typically interact through communication networks. The client-server relationship is created by computer programs running on the respective computers and having a client-server relationship with each other. The server can be a cloud server, also known as a cloud computing server or cloud host, which is a hosting product within the cloud computing service system. It addresses the shortcomings of traditional physical hosts and Virtual Private Server (VPS) services, such as high management difficulty and weak business scalability.

[0181] It should be understood that the various forms of processes shown above can be used to rearrange, add, or delete steps. For example, the steps described in this disclosure can be executed in parallel, sequentially, or in different orders, as long as the desired result of the technical solution of this disclosure can be achieved, and this is not limited herein.

[0182] The specific embodiments described above do not constitute a limitation on the scope of protection of this disclosure. Those skilled in the art should understand that various modifications, combinations, sub-combinations, and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this disclosure should be included within the scope of protection of this disclosure.

Claims

1. A training method for a lung classification model, characterized in that, include: Obtain a first lung image set and corresponding training label data for the first lung image set, wherein the first lung image set contains multiple first lung images; The first lung image is divided into regions for feature extraction to obtain a first multi-region feature set; Perform feature correlation analysis on the first multi-region feature set at the region dimension to obtain the first fusion feature set corresponding to the first lung image; Each of the first fused feature sets is used as input data for the untrained lung classification model, and the model is iteratively trained based on the training label data to obtain the trained lung classification model. The first fusion feature set includes a non-redundant feature set and a correlated feature set. The non-redundant feature set represents independent feature information of multiple lung regions, and the correlated feature set represents complementary feature information between multiple lung regions.

2. The training method according to claim 1, characterized in that, The step of performing feature correlation analysis on the first multi-region feature set at the region dimension to obtain the first fused feature set corresponding to the first lung image includes: Perform feature pairing at the region dimension on the first multi-region feature set to obtain a first matching feature set, and determine the correlation coefficient matrix corresponding to the first matching feature set; Redundant feature filtering is performed based on the first matching feature set and the correlation coefficient matrix to obtain a non-redundant feature set and a second matching feature set. The second matching feature set contains at least one lung feature pair that was not filtered in the first matching feature set. Each lung feature pair in the second matching feature set is associated with another feature to obtain an associated feature set.

3. The training method according to claim 2, characterized in that, The step of filtering redundant features based on the first matching feature set and the correlation coefficient matrix to obtain a non-redundant feature set and a second matching feature set includes: Based on the feature association conditions and the correlation coefficient matrix, the first matching feature set is filtered to obtain the third matching feature set; The third matching feature set is subjected to physiological verification and screening to obtain the fourth matching feature set, and the correlation weight matrix of the physiological dimension corresponding to the fourth matching feature set is obtained. The correlation coefficient matrix is ​​filtered according to the fourth matching feature set to obtain a correlation coefficient sub-matrix. Then, based on the correlation coefficient sub-matrix and the association weight matrix, redundant features are filtered from the fourth matching feature set to obtain a non-redundant feature set and a second matching feature set.

4. The training method according to claim 3, characterized in that, The correlation coefficient matrix is ​​determined based on the association type of the lung features in the first matching feature set; Accordingly, the step of performing redundant feature filtering on the fourth matching feature set to obtain a non-redundant feature set and a second matching feature set based on the correlation coefficient sub-matrix and the association weight matrix includes: Based on the correlation coefficient submatrix and the association weight matrix, redundant feature filtering is performed on the fourth matching feature set to obtain a non-redundant feature set and a fifth matching feature set. Obtain the association type set corresponding to the fifth matching feature set, wherein the association type set contains at least two association types; For each lung feature pair in the fifth matching feature set, based on at least one association type in the association type set other than the association type to which the lung feature pair belongs, the correlation coefficient sequence corresponding to the lung feature pair is determined. The feature correlation coefficient and physiological association weight corresponding to the lung feature pair are obtained from the correlation coefficient submatrix and the association weight matrix, respectively. Based on the feature correlation coefficient, physiological association weight and correlation coefficient sequence, the feature association score of the lung feature pair is determined. Based on the correlation scores of each feature, the fifth matching feature set is filtered to obtain the second matching feature set.

5. The training method according to claim 4, characterized in that, The step of determining the feature association score of the lung feature pair based on the feature correlation coefficient, physiological association weight, and correlation coefficient sequence includes: The feature association strength of the lung feature pairs is determined based on the feature correlation coefficients and the correlation coefficient sequence. The feature association score of the lung feature pair is determined based on the feature association strength and the physiological association weight.

6. A method for classifying lung images, characterized in that, include: Obtain the second lung image and the feature processing parameters corresponding to the trained lung classification model; Based on the feature processing parameters, feature processing is performed on the second lung image to obtain a second fusion feature set; The second fused feature set is input into the lung classification model to obtain the image classification label corresponding to the output second lung image; The lung classification model is obtained by training the lung classification model as described in any one of claims 1-5. The second fusion feature set includes a non-redundant feature set and a correlated feature set. The non-redundant feature set represents independent feature information of multiple lung regions, and the correlated feature set represents complementary feature information between multiple lung regions.

7. The classification method according to claim 6, characterized in that, The method further includes: If the predicted probability value corresponding to the image classification label is less than a preset probability threshold, obtain the training sample medical record dataset corresponding to the first lung image set; A third multi-region feature set corresponding to the first lung image is determined, and a third fusion feature set is determined based on the third multi-region feature set. The training sample medical record data corresponding to the first lung image in the training sample medical record dataset is added to the third fusion feature set to obtain a fourth fusion feature set. Based on the training label data and each of the fourth fusion feature sets, the untrained lung classification model is retrained iteratively.

8. The classification method according to claim 7, characterized in that, The feature extraction parameters corresponding to the third multi-region feature set are different from those corresponding to the first multi-region feature set, and / or the correlation parameters of the feature correlation analysis corresponding to the third fusion feature set are different from those of the feature correlation analysis corresponding to the first fusion feature set.

9. The classification method according to claim 6, characterized in that, The method further includes: Based on the image classification labels and the true classification labels of the second lung image, determine the label classification bias; If the label classification deviation meets the classification deviation condition, the second lung image is added to the incremental lung image set; In response to the number of images corresponding to the incremental lung image set reaching a preset threshold, the second fusion feature set and the true classification label corresponding to each second lung image in the incremental lung image set are obtained; Based on each of the second fusion feature sets and each of the true classification labels, the lung classification model is incrementally trained to obtain an optimized lung classification model.

10. A training device for a lung classification model, characterized in that, include: The first lung image set acquisition module is used to acquire a first lung image set and training label data corresponding to the first lung image set. The first lung image set contains multiple first lung images. The first multi-region feature set determination module is used to extract features from the first lung image by region to obtain the first multi-region feature set; The first fusion feature set determination module is used to perform feature correlation analysis on the first multi-region feature set in the region dimension to obtain the first fusion feature set corresponding to the first lung image. The lung classification model training module is used to take each of the first fused feature sets as input data for the untrained lung classification model, and perform iterative training of the model based on the training label data to obtain the trained lung classification model. The first fusion feature set includes a non-redundant feature set and a correlated feature set. The non-redundant feature set represents independent feature information of multiple lung regions, and the correlated feature set represents complementary feature information between multiple lung regions.

11. A lung image classification device, characterized in that, include: The second lung image acquisition module is used to acquire the second lung image and the feature processing parameters corresponding to the trained lung classification model. The second fusion feature set determination module is used to perform feature processing on the second lung image according to the feature processing parameters to obtain a second fusion feature set; The image classification label determination module is used to input the second fused feature set into the lung classification model to obtain the image classification label corresponding to the output second lung image; The lung classification model is obtained by training the lung classification model as described in any one of claims 1-5. The second fusion feature set includes a non-redundant feature set and a correlated feature set. The non-redundant feature set represents independent feature information of multiple lung regions, and the correlated feature set represents complementary feature information between multiple lung regions.

Images