Intelligent asthma ct image typing system based on deep learning

Abstract

The present application relates to a deep learning-based asthma CT image intelligent classification system, which comprises an image data acquisition module, a data preprocessing module, a constraint extraction model construction module, a classification model construction module and an asthma image classification module. The present application obtains original data through data collection; adopts a data preprocessing method of HU normalization, lung parenchyma segmentation, dual-phase rigid registration, lung region cropping and data set division; adopts an improved residual network model as the constraint extraction model, combines deep feature learning with mechanical displacement field estimation, compresses the non-Euclidean airway topological information into compact constraint features through structured output, and improves the representation ability of features to individual differences; adopts an improved conditional diffusion model as the classification model, embeds the physical motion equation as a conditional constraint into the generation process, ensures that the generation and reasoning process conforms to the basic law of respiratory mechanics, and enhances the generalization ability of the model to new patients.

Description

Technical Field

[0001] This invention relates to the field of intelligent classification technology for asthma CT images, specifically to an intelligent classification system for asthma CT images based on deep learning. Background Technology

[0002] Asthma CT imaging classification is a computer-aided process that uses high-resolution CT images of patients to automatically classify and analyze different types of asthma. By combining image processing algorithms and machine learning techniques, it can extract features of lung structure and integrate them with clinical data to achieve accurate classification of asthma patients' conditions. This helps doctors develop personalized treatment plans and improve patients' treatment outcomes.

[0003] However, traditional asthma CT imaging classification systems rely on static, single-moment image features during the inspiratory or expiratory phases, neglecting pathological information about airway deformation during the respiratory cycle. This results in classification based solely on morphological surface findings and a lack of effective modeling of the complex topological structure of the airways. Consequently, they struggle to integrate discrete airway segment motion information into individualized features, leading to technical problems such as insufficient stability and repeatability. Furthermore, traditional asthma CT imaging classification systems lack physical consistency constraints, which may cause the model to deviate from the actual respiratory motion mechanism when learning image generation or feature extraction. They also ignore the differences in respiratory dynamic parameters between individuals, making it impossible to personalize parameter adaptations for new cases. This results in highly unstable performance of the model when facing patients with different disease severity or comorbidities. Summary of the Invention

[0004] To address the above issues and overcome the shortcomings of existing technologies, this invention provides a deep learning-based intelligent classification system for asthma CT images. Traditional asthma CT image classification systems rely on static, single-moment image features from the inspiratory or expiratory phases, neglecting pathological information about airway deformation during the respiratory cycle. This results in classification based solely on morphological surface-level features, a lack of effective modeling of complex airway topology, difficulty in integrating discrete airway segment motion information into individualized features, and insufficient stability and repeatability. This solution creatively employs an improved residual network model as a constraint extraction model, combining deep feature learning with mechanical displacement field estimation. While retaining high-level semantic extraction capabilities, it compresses non-Euclidean airway topological information into a compact structure through structured output. The improved conditional diffusion model is used as the classification model. It embeds the physical motion equation as a conditional constraint into the generation process. This allows the model to fit the statistical distribution and provide richer distribution information for classification, while ensuring that the generation and inference processes conform to the basic laws of respiratory mechanics. This enhances the model's ability to generalize to new patients.

[0005] The technical solution adopted by the present invention is as follows: The intelligent classification system for asthma CT images based on deep learning provided by the present invention includes an image data acquisition module, a data preprocessing module, a constraint extraction model construction module, a classification model construction module, and an asthma image classification module;

[0006] The image data acquisition module obtains the raw dataset of asthma subtypes through image acquisition and sends the raw dataset of asthma subtypes to the data preprocessing module.

[0007] The data preprocessing module uses HU normalization, lung parenchyma segmentation, biphasic rigid registration, lung region trimming, and dataset partitioning as data preprocessing methods to obtain a preliminary dataset to be classified, a preliminary training set, and a preliminary test set. The preliminary dataset to be classified, the preliminary training set, and the preliminary test set are then sent to the constraint extraction model construction module.

[0008] The constraint extraction model construction module obtains a constraint extraction model by constructing an improved residual network model. It takes the preliminary dataset to be classified, the preliminary training set, and the preliminary test set as input to obtain the dataset to be classified, the classification training set, and the classification test set. The constraint extraction model and the dataset to be classified are sent to the asthma image classification module, and the classification training set and the classification test set are sent to the classification model construction module.

[0009] The classification model construction module obtains a classification model by constructing an improved conditional diffusion model, and sends the classification model to the asthma imaging classification module;

[0010] The asthma image classification module performs real-time asthma CT image classification by combining the constraint extraction model and the classification model, and obtains the final asthma CT image classification result.

[0011] Furthermore, in the image data acquisition module, the original dataset for asthma subtyping specifically includes a historical original dataset for asthma subtyping and a current original dataset for asthma subtyping. Both the historical original dataset for asthma subtyping and the current original dataset for asthma subtyping contain biphasic CT image data. The historical original dataset for asthma subtyping also includes asthma subtyping label data. The biphasic CT image data specifically includes inspiratory CT image data and expiratory CT image data. The inspiratory CT image data specifically refers to high-resolution CT images of the lungs acquired at the end of inspiration for the patient, and the expiratory CT image data specifically refers to high-resolution CT images of the lungs acquired at the end of expiration for the patient. The inspiratory CT images and expiratory CT images acquired for the same patient constitute a data pair.

[0012] Furthermore, in the data preprocessing module, the HU normalization is used to eliminate differences in voxel spacing and grayscale range caused by different CT equipment and scanning protocols. Specifically, it involves resampling all CT image sequences to an isotropic 1mm. 3 Voxels were used, and the HU values ​​were truncated to the lung window range and then linearly mapped to the interval [0,1].

[0013] The lung parenchyma segmentation is used to remove non-lung-related structures and focus on the lung region. Specifically, it uses a thresholding method combined with connected component analysis and morphological closing operation to extract the lung region and obtain a binary lung mask.

[0014] The biphasic rigid registration is used to correct global misalignment caused by patient positional movement. Specifically, it uses the inspiratory phase as a fixed image and the expiratory phase as a floating image, and maximizes the normalized cross-correlation through a six-degree-of-freedom rigid body transformation to obtain a spatially aligned expiratory phase image.

[0015] The lung region cropping is used to increase the diversity of training samples. Specifically, it uses the minimum bounding box of the binary lung mask to crop the biphasic CT image, and then divides it into blocks of size 64×64×64 with a 50% overlapping sliding window to obtain local block pairs.

[0016] The dataset partitioning is used to obtain training data and test data, specifically by splitting the original historical asthma classification dataset.

[0017] The current asthma subtyping dataset is preprocessed using HU normalization, lung parenchyma segmentation, biphasic rigid registration, and lung region pruning to obtain a preliminary dataset to be subtyped. The historical asthma subtyping dataset is preprocessed using HU normalization, lung parenchyma segmentation, biphasic rigid registration, lung region pruning, and dataset partitioning to obtain a preliminary training set and a preliminary test set.

[0018] Furthermore, in the constraint extraction model construction module, a model is used to construct the individualized respiratory mechanics features of the patient from biphasic CT images. Specifically, an improved residual network model is constructed as the constraint extraction model. The improved residual network model extracts individual dynamic constraint features by predicting the three-dimensional displacement vector of each voxel from the inspiratory phase to the expiratory phase and constructing a graph structure.

[0019] The constraint extraction model construction module specifically includes continuous displacement field prediction, motion topology skeleton extraction, strain rate calculation, constraint feature extraction, and model construction and training.

[0020] The continuous displacement field prediction is used to predict the three-dimensional displacement vector of voxels from the preprocessed dual-phase CT block, and includes:

[0021] Biphasic feature extraction is used to extract high-level semantic features from biphasic CT blocks. Specifically, two weight-shared 3D ResNet-34s are used to extract features of inspiratory and expiratory local blocks, respectively. After global average pooling, inspiratory and expiratory feature maps are obtained. Then, coarse fused features are obtained by element-wise difference fusion and upsampling to restore resolution.

[0022] Continuous implicit decoding is used to convert the displacement field on discrete voxels into a function representation of continuous coordinates. Specifically, for any continuous coordinate within a block, local features are first obtained from coarse fusion feature bilinear interpolation, and then input into a multilayer perceptron along with the position encoding. The output is the three-dimensional displacement vector of that coordinate, thus obtaining the three-dimensional displacement field.

[0023] The motion topology skeleton extraction is used to automatically extract the graph structure of the airway centerline from the displacement field, and includes:

[0024] Displacement field divergence calculation is used to quantify the local volume change rate of each voxel. Specifically, it is obtained by calculating the sum of the partial derivatives of each component of the three-dimensional displacement field to obtain the divergence volume.

[0025] Zero-crossing surface processing is used to extract the airway wall from the divergence volume and simplify it into a centerline graph. Specifically, the traveling cube algorithm is used to extract isosurfaces with divergence of 0, and then the topology refinement algorithm is used to erode them to a single pixel width and remove short branches to obtain an undirected graph of the airway structure.

[0026] The airway attribute assignment is used to assign motion-related attributes to each airway segment. Specifically, five points are uniformly sampled along each edge of the undirected graph of the airway structure, and the displacement amplitude and divergence value of each point are read from the three-dimensional displacement field to obtain the airway enhancement undirected graph.

[0027] The strain rate calculation, used to quantify the degree of radial deformation of each airway segment during the respiratory cycle, includes:

[0028] Radial tube wall distance extraction is used to measure the airway inner and outer diameters at each sampling point. Specifically, eight rays are emitted from each sampling point along the radial direction perpendicular to each side of the undirected graph of the airway structure. These rays intersect with the zero divergence isosurface to obtain the nearest intersection distance in the inspiratory phase and the nearest intersection distance in the expiratory phase, which are used as the radial radius in the inspiratory phase and the radial radius in the expiratory phase, respectively.

[0029] Radial strain rate calculation is used to convert the change in tube diameter into a dimensionless strain rate index. Specifically, the radial radius of 5 sampling points on each side of the undirected diagram of the airway structure is averaged, and then the relative change rate of the exhalation phase relative to the inspiratory phase is calculated to obtain the radial strain rate value of each side of the undirected diagram of the airway structure.

[0030] The constraint feature extraction is used to compress the non-Euclidean airway structure undirected graph and radial strain rate sequence into a fixed-length vector, which includes:

[0031] Graph attention message passing is used to preserve the hierarchical structure information of the undirected graph of the airway structure in the graph neural network. Specifically, when passing messages in each layer of the graph neural network, the edge attributes and bifurcation differences of the undirected graph of the airway structure are incorporated as biases into the attention weight calculation to obtain the node latent vector after each layer is updated. The edge attributes of the undirected graph of the airway structure specifically include the edge length and the edge radial strain rate value.

[0032] Differentiable graph pooling is used to aggregate a variable-size set of nodes into a fixed-size graph-level vector and force the compression process to preserve the original topology. Specifically, differentiable graph pooling is used to aggregate the nodes of the undirected graph of the airway structure after graph attention message passing processing into 32 super nodes and sum them globally to obtain preliminary constraint features and pooled coarsened graph. The edit distance between the coarsened graph and the undirected graph of the airway structure is calculated as the topology regularization loss.

[0033] Constraint feature acquisition is used to reduce the dimensionality of high-dimensional graph-level vectors to obtain a compact representation. Specifically, the initial constraint features are further reduced in dimensionality through two fully connected layers to obtain individual dynamic constraint features.

[0034] The construction and training of the model specifically involves constructing an improved residual network model by integrating the continuous displacement field prediction, the motion topology skeleton extraction, the strain rate calculation, and the constraint feature extraction. The model is then trained and its performance is verified based on a preliminary training set and a preliminary test set to obtain the improved residual network model, which serves as the constraint extraction model.

[0035] The preliminary dataset to be classified, the preliminary training set, and the preliminary test set are used as inputs to the constraint extraction model to obtain the dataset to be classified, the classification training set, and the classification test set.

[0036] Furthermore, in the subtyping model construction module, the model required to construct the individual dynamic constraint features based on inspiratory phase CT images and upstream output asthma subtyping labels is specifically constructed by building an improved conditional diffusion model as the subtyping model.

[0037] The classification model construction module specifically includes conditional diffusion model construction, optimal physical parameter inversion, disease manifold space pre-calibration, final classification output, and model construction and training.

[0038] The conditional diffusion model is constructed to establish a generative model that can both generate realistic CT images and enforce physical consistency. The model includes:

[0039] The forward diffusion noise addition process is used to gradually destroy the coding results of the inspiratory phase CT images into standard Gaussian noise. Specifically, cosine noise scheduling is used to gradually add Gaussian noise with increasing variance to the coding results of the inspiratory phase CT images to obtain noisy coding features.

[0040] Physical consistency constraints are used to design a reaction-diffusion equation embedded in a forward diffusion noisy process. Specifically, an equation incorporating diffusion, reaction, and respiratory drivers is defined, and the respiratory driver term is modulated using individual dynamic constraint features. The predicted expiratory coding features are obtained by solving the equation and compared with expiratory phase features. Figure 1 Same as calculating physical consistency loss;

[0041] Conditional denoising network is used to train a denoising network with individual dynamic constraint features as conditions. Specifically, a three-dimensional U-Net is used as the denoising network, and FiLM modulation is used at each skip connection to inject individual dynamic constraint features as conditions to obtain the conditional denoising network.

[0042] The optimal physical parameter inversion is used to solve the reaction-diffusion equation parameters that best explain the CT images and respiratory movements of patients to be classified. The content includes:

[0043] Initialization is used to prepare the starting point for inversion. Specifically, the trained conditional diffusion model is fixed, and the inspiratory phase CT images of the patients to be classified are processed. The parameters of the reaction-diffusion equation are initialized to the mean of the training set.

[0044] Joint optimization with physical constraints is used to search for optimal reaction-diffusion equation parameters under a fixed model. Specifically, it involves alternating the denoising step of the conditional denoising network and the gradient descent of the reaction-diffusion equation parameters for patients to be classified, so as to obtain the reaction-diffusion equation parameters that make the generated image output by the conditional diffusion model best match the real image.

[0045] The disease manifold spatial pre-calibration is used to build a continuous spectral map of parameters on the training set, and includes:

[0046] The training set parameter set dimensionality reduction is used to reduce the reaction-diffusion equation parameters of the training set samples to a two-dimensional manifold. Specifically, it calculates the high-dimensional adjacency probability for the reaction-diffusion equation parameters of all training set samples, and obtains the two-dimensional manifold of each training set sample by minimizing the KL divergence with the low-dimensional adjacency probability.

[0047] Basin typology clustering is used to automatically divide the clustered regions of asthma types represented by asthma typology labels on a two-dimensional manifold. Specifically, it performs K-means clustering on the two-dimensional manifold of the training set samples using Wasserstein distance as a metric to obtain multiple attractor basins based on asthma typology labels.

[0048] The final classification output is used to project the patient to be classified onto a pre-calibrated manifold space and output the clinical classification. Specifically, it uses the same method as the dimensionality reduction of the training set parameter set to calculate the two-dimensional manifold of the response-diffusion equation parameters of the patient to be classified, and calculates the Euclidean distance to the center of all attractor basins. The distance is then converted into a normalized membership vector through the softmax function to obtain the continuous confidence score of the asthma classification label. The asthma classification label with the maximum confidence score greater than a set threshold is taken as the output.

[0049] The model construction and training specifically involves constructing an improved conditional diffusion model by integrating the conditional diffusion model construction, the optimal physical parameter inversion, the disease manifold space pre-calibration, and the final genotyping output. The model is then trained and its performance is verified based on the genotyping training set and the genotyping test set to obtain the improved conditional diffusion model, which serves as the genotyping model.

[0050] Furthermore, in the asthma image classification module, the dataset to be classified is used as the input of the classification model to perform asthma CT image classification, and the asthma classification label output by the classification model is used as the final asthma CT image classification result.

[0051] The beneficial effects achieved by the present invention using the above solution are as follows:

[0052] (1) In view of the technical problems of traditional asthma CT imaging classification system, which relies on static image features of a single moment of inspiratory or expiratory phase and ignores the pathological information of airway deformation during the respiratory cycle, the classification basis is only on the morphological surface and lacks effective modeling of complex airway topology. It is difficult to integrate discrete airway segment motion information into individualized features, and the stability and repeatability are insufficient. This solution creatively adopts an improved residual network model as a constraint extraction model, which combines deep feature learning with mechanical displacement field estimation. While retaining the ability to extract high-level semantics, it compresses non-Euclidean airway topology information into compact constraint features through structured output, thereby improving the ability of features to represent individual differences.

[0053] (2) In view of the technical problems of traditional asthma CT image classification system, which lacks physical consistency constraints, the model may deviate from the real respiratory motion mechanism when learning image generation or feature extraction, and ignores the differences in respiratory dynamic parameters between individuals, making it impossible to perform personalized parameter adaptation for new cases, resulting in the model performing extremely unstablely when facing patients with different disease severity or comorbidities, this solution creatively adopts an improved conditional diffusion model as the classification model, embedding the physical motion equation as a conditional constraint into the generation process, so that the model fits the statistical distribution to provide richer distribution information for classification, while ensuring that the generation and inference process conforms to the basic laws of respiratory mechanics, and enhancing the model's generalization ability to new patients. Attached Figure Description

[0054] Figure 1 A schematic diagram of the modules of the intelligent asthma CT image classification system based on deep learning provided by the present invention;

[0055] Figure 2 This is a flowchart illustrating the data preprocessing module.

[0056] Figure 3 A flowchart illustrating the process of building the constraint extraction model module;

[0057] Figure 4 A flowchart illustrating the process of building a fractal model module.

[0058] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the embodiments of the invention to explain the invention and do not constitute a limitation thereof. Detailed Implementation

[0059] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0060] In the description of this invention, it should be understood that the terms "upper", "lower", "front", "rear", "left", "right", "top", "bottom", "inner", "outer", etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.

[0061] Example 1, see Figure 1 The intelligent asthma CT image classification system based on deep learning provided by the present invention includes an image data acquisition module, a data preprocessing module, a constraint extraction model construction module, a classification model construction module, and an asthma image classification module.

[0062] The image data acquisition module obtains the raw dataset of asthma subtypes through image acquisition and sends the raw dataset of asthma subtypes to the data preprocessing module.

[0063] The data preprocessing module uses HU normalization, lung parenchyma segmentation, biphasic rigid registration, lung region trimming, and dataset partitioning as data preprocessing methods to obtain a preliminary dataset to be classified, a preliminary training set, and a preliminary test set. The preliminary dataset to be classified, the preliminary training set, and the preliminary test set are then sent to the constraint extraction model construction module.

[0064] The constraint extraction model construction module obtains a constraint extraction model by constructing an improved residual network model. It takes the preliminary dataset to be classified, the preliminary training set, and the preliminary test set as input to obtain the dataset to be classified, the classification training set, and the classification test set. The constraint extraction model and the dataset to be classified are sent to the asthma image classification module, and the classification training set and the classification test set are sent to the classification model construction module.

[0065] The classification model construction module obtains a classification model by constructing an improved conditional diffusion model, and sends the classification model to the asthma imaging classification module;

[0066] The asthma image classification module performs real-time asthma CT image classification by combining the constraint extraction model and the classification model, and obtains the final asthma CT image classification result.

[0067] Example 2, see Figure 1 This embodiment is based on the above embodiment. In the image data acquisition module, the original dataset for asthma subtyping specifically includes a historical original dataset for asthma subtyping and a current original dataset for asthma subtyping. Both the historical original dataset for asthma subtyping and the current original dataset for asthma subtyping contain biphasic CT image data. The historical original dataset for asthma subtyping also includes asthma subtyping label data. The biphasic CT image data specifically includes inspiratory CT image data and expiratory CT image data. The inspiratory CT image data specifically refers to high-resolution CT images of the lungs acquired at the end of inspiration for the patient. The expiratory CT image data specifically refers to high-resolution CT images of the lungs acquired at the end of expiration for the patient. The inspiratory CT images and expiratory CT images acquired for the same patient constitute a data pair.

[0068] Example 3, see Figure 1 and Figure 2 This embodiment is based on the above embodiment. In the data preprocessing module, the HU normalization is used to eliminate the differences in voxel spacing and grayscale range caused by different CT equipment and scanning protocols. Specifically, it involves resampling all CT image sequences to an isotropic 1mm. 3 Voxels were used, and the HU values ​​were truncated to the lung window range and then linearly mapped to the interval [0,1].

[0069] The lung parenchyma segmentation is used to remove non-lung-related structures and focus on the lung region. Specifically, it uses a thresholding method combined with connected component analysis and morphological closing operation to extract the lung region and obtain a binary lung mask.

[0070] The biphasic rigid registration is used to correct global misalignment caused by patient positional movement. Specifically, it uses the inspiratory phase as a fixed image and the expiratory phase as a floating image, and maximizes the normalized cross-correlation through a six-degree-of-freedom rigid body transformation to obtain a spatially aligned expiratory phase image.

[0071] The lung region cropping is used to increase the diversity of training samples. Specifically, it uses the minimum bounding box of the binary lung mask to crop the biphasic CT image, and then divides it into blocks of size 64×64×64 with a 50% overlapping sliding window to obtain local block pairs.

[0072] The dataset partitioning is used to obtain training data and test data, specifically by splitting the original historical asthma classification dataset.

[0073] The current asthma subtyping dataset is preprocessed using HU normalization, lung parenchyma segmentation, biphasic rigid registration, and lung region pruning to obtain a preliminary dataset to be subtyped. The historical asthma subtyping dataset is preprocessed using HU normalization, lung parenchyma segmentation, biphasic rigid registration, lung region pruning, and dataset partitioning to obtain a preliminary training set and a preliminary test set.

[0074] Example 4, see Figure 1 and Figure 3 This embodiment is based on the above embodiment. In the constraint extraction model construction module, a model is used to construct the model required to extract the patient's personalized respiratory mechanical characteristics from biphasic CT images. Specifically, an improved residual network model is constructed as the constraint extraction model. The improved residual network model obtains individual dynamic constraint characteristics by predicting the three-dimensional displacement vector of each voxel from the inspiratory phase to the expiratory phase and constructing a graph structure to extract features.

[0075] The constraint extraction model construction module specifically includes continuous displacement field prediction, motion topology skeleton extraction, strain rate calculation, constraint feature extraction, and model construction and training.

[0076] The continuous displacement field prediction is used to predict the three-dimensional displacement vector of voxels from the preprocessed dual-phase CT block, and includes:

[0077] Biphasic feature extraction is used to extract high-level semantic features from biphasic CT blocks. Specifically, two weighted 3D ResNet-34 arrays are used to extract features from inspiratory and expiratory local blocks, respectively. Global average pooling is then applied to obtain inspiratory and expiratory feature maps, which are then fused through element-wise difference. After upsampling to restore resolution, coarse fused features are obtained. The formula used is as follows:

[0078] ;

[0079] In the formula, This represents the characteristic diagram of the inhalation phase. This represents the expiratory phase characteristic diagram. This represents the global average pooling function. This represents a three-dimensional ResNet-34 function for processing the upgassing phase. This represents a three-dimensional ResNet-34 function for processing the expiratory phase. This represents the local block of the input intake phase. This represents a local block of the expiratory phase of the input. Indicates coarse fusion features, Represents the upsampling function. This represents a 3D convolution function of size 1×1. Indicates feature concatenation operation;

[0080] Continuous implicit decoding is used to convert the displacement field on discrete voxels into a functional representation of continuous coordinates. Specifically, for any continuous coordinate within a block, local features are first obtained from coarse fusion feature bilinear interpolation. Then, along with the position encoding, these features are input into a multilayer perceptron, which outputs the three-dimensional displacement vector of that coordinate, thus obtaining the three-dimensional displacement field. The formula used is as follows:

[0081] ;

[0082] In the formula, x represents the coordinate independent variable. Represents the three-dimensional displacement vector at coordinate x. This represents the multilayer perceptron function used to output three-dimensional displacement vectors. This represents the position encoding function. Represents the local features of coordinate x;

[0083] The motion topology skeleton extraction is used to automatically extract the graph structure of the airway centerline from the displacement field, and includes:

[0084] Displacement field divergence calculation is used to quantify the local volume change rate of each voxel. Specifically, it is obtained by calculating the sum of the partial derivatives of each component of the three-dimensional displacement field to obtain the divergence volume.

[0085] Zero-crossing surface processing is used to extract the airway wall from the divergence volume and simplify it into a centerline graph. Specifically, the traveling cube algorithm is used to extract isosurfaces with divergence of 0, and then the topology refinement algorithm is used to erode them to a single pixel width and remove short branches to obtain an undirected graph of the airway structure.

[0086] The airway attribute assignment is used to assign motion-related attributes to each airway segment. Specifically, five points are uniformly sampled along each edge of the undirected graph of the airway structure, and the displacement amplitude and divergence value of each point are read from the three-dimensional displacement field to obtain the airway enhancement undirected graph.

[0087] The strain rate calculation, used to quantify the degree of radial deformation of each airway segment during the respiratory cycle, includes:

[0088] Radial tube wall distance extraction is used to measure the airway inner and outer diameters at each sampling point. Specifically, eight rays are emitted from each sampling point along the radial direction perpendicular to each side of the undirected graph of the airway structure. These rays intersect with the zero divergence isosurface to obtain the nearest intersection distance in the inspiratory phase and the nearest intersection distance in the expiratory phase, which are used as the radial radius in the inspiratory phase and the radial radius in the expiratory phase, respectively.

[0089] Radial strain rate calculation is used to convert the change in tube diameter into a dimensionless strain rate index. Specifically, it involves averaging the radial radii of five sampling points on each side of the undirected diagram of the airway structure, and then calculating the relative rate of change of the expiratory phase relative to the inspiratory phase to obtain the radial strain rate value of each side of the undirected diagram of the airway structure. The formula used is as follows:

[0090] ;

[0091] In the formula, This represents the radial strain rate value of the a-th edge of the undirected graph of the airway structure. This represents the average expiratory radial radius of the a-th edge in the undirected graph of the airway structure. This represents the average inspiratory phase radial radius of the a-th edge in the undirected graph of the airway structure;

[0092] The constraint feature extraction is used to compress the non-Euclidean airway structure undirected graph and radial strain rate sequence into a fixed-length vector, which includes:

[0093] Graph attention message passing is used to preserve the hierarchical structure information of the undirected graph of the airway structure in the graph neural network. Specifically, when passing messages in each layer of the graph neural network, the edge attributes and bifurcation differences of the undirected graph of the airway structure are incorporated as biases into the attention weight calculation to obtain the node latent vector after each layer is updated. The edge attributes of the undirected graph of the airway structure specifically include the edge length and the edge radial strain rate value.

[0094] Differentiable graph pooling is used to aggregate variable-sized node sets into fixed-size graph-level vectors, and forces the compression process to preserve the original topology. Specifically, differentiable graph pooling is used to cluster the nodes of the undirected graph of the airway structure after graph attention message passing processing into 32 super nodes and globally sum them to obtain preliminary constraint features and a pooled coarsened graph. The edit distance between the coarsened graph and the undirected graph of the airway structure is calculated as the topology regularization loss. The formula used is as follows:

[0095] ;

[0096] In the formula, This represents the topological regularization loss value. This represents the graph editing distance calculation function, where AG represents the undirected graph of airway structure. Represents the coarsening plot. Indicates preliminary constraint features, This represents the node aggregation features after graph attention message passing processing;

[0097] Constraint feature acquisition is used to reduce the dimensionality of high-dimensional graph-level vectors to obtain a compact representation. Specifically, the initial constraint features are further reduced in dimensionality through two fully connected layers to obtain individual dynamic constraint features.

[0098] The construction and training of the model specifically involves constructing an improved residual network model by integrating the continuous displacement field prediction, the motion topology skeleton extraction, the strain rate calculation, and the constraint feature extraction. The model is then trained and its performance is verified based on a preliminary training set and a preliminary test set to obtain the improved residual network model, which serves as the constraint extraction model.

[0099] The preliminary dataset to be classified, the preliminary training set, and the preliminary test set are used as inputs to the constraint extraction model to obtain the dataset to be classified, the classification training set, and the classification test set.

[0100] By performing the above operations, this solution addresses the technical problems of traditional asthma CT imaging classification systems, which rely on static image features from a single moment of inspiratory or expiratory phases, neglecting pathological information about airway deformation during the respiratory cycle. This results in classification based solely on morphological surface-level features, a lack of effective modeling of complex airway topology, difficulty in integrating discrete airway segment motion information into individualized features, and insufficient stability and repeatability. This solution creatively adopts an improved residual network model as a constraint extraction model, combining deep feature learning with mechanical displacement field estimation. While retaining high-level semantic extraction capabilities, it compresses non-Euclidean airway topological information into compact constraint features through structured output, thereby improving the feature's ability to represent individual differences.

[0101] Example 5, see Figure 1 and Figure 4 This embodiment is based on the above embodiment. In the subtyping model construction module, a model is used to construct the asthma subtyping label based on the individual dynamic constraint features of inspiratory phase CT images and upstream output. Specifically, an improved conditional diffusion model is constructed as the subtyping model.

[0102] The classification model construction module specifically includes conditional diffusion model construction, optimal physical parameter inversion, disease manifold space pre-calibration, final classification output, and model construction and training.

[0103] The conditional diffusion model is constructed to establish a generative model that can both generate realistic CT images and enforce physical consistency. The model includes:

[0104] The forward diffusion noise addition process is used to progressively destroy the coding results of the inspiratory phase CT images into standard Gaussian noise. Specifically, cosine noise scheduling is used to progressively add Gaussian noise with increasing variance to the coding results of the inspiratory phase CT images, resulting in noisy coding features. The formula used is as follows:

[0105] ;

[0106] In the formula, This indicates the coding features of inspiratory phase CT images. This represents the pre-trained convolutional neural network function used to acquire the encoded results of inspiration phase CT images. This represents the noisy coding features at step t of the forward process. This represents the noisy coded features at step t-1 of the forward process. Indicates that in a given Under the conditions, The distribution that it follows Let I represent the cosine noise intensity at step t of the forward process, and let I denote the identity matrix. Indicates the forward process of the first... The cosine noise of the step is adjusted to control the noise intensity. Indicates standard noise;

[0107] Physical consistency constraints are used to design a reaction-diffusion equation embedded in a forward diffusion noisy process. Specifically, an equation incorporating diffusion, reaction, and respiratory drivers is defined, and the respiratory driver term is modulated using individual dynamic constraint features. The predicted expiratory coding features are obtained by solving the equation and compared with expiratory phase features. Figure 1 The formula used to calculate the physical consistency loss is as follows:

[0108] ;

[0109] In the formula, t represents the index of the forward noise addition process, and Dc represents the diffusion coefficient. Represents the reaction rate coefficient. Indicates the reaction strength coefficient. This represents a pre-trained multilayer perceptron function that predicts breathing drive terms based on individual dynamic constraint features. Represents the dynamic constraint characteristics of an individual. This represents the physical consistency loss value. This represents the weights of the linear mapping of expiratory coding features. This represents the predicted expiratory coding features;

[0110] Conditional denoising network is used to train a denoising network with individual dynamic constraint features as conditions. Specifically, a three-dimensional U-Net is used as the denoising network, and FiLM modulation is used at each skip connection to inject individual dynamic constraint features as conditions to obtain the conditional denoising network.

[0111] The optimal physical parameter inversion is used to solve the reaction-diffusion equation parameters that best explain the CT images and respiratory movements of patients to be classified. The content includes:

[0112] Initialization is used to prepare the starting point for inversion. Specifically, it involves fixing the trained conditional diffusion model and processing the inspiratory phase CT images of the patients to be subtyped. The reaction-diffusion equation parameters are initialized to the training set mean. These reaction-diffusion equation parameters specifically include Dc, , and respiratory drive characteristics;

[0113] Joint optimization with physical constraints is used to search for optimal reaction-diffusion equation parameters under a fixed model. Specifically, it involves alternating the denoising step of the conditional denoising network and the gradient descent of the reaction-diffusion equation parameters for patients to be classified, so as to obtain the reaction-diffusion equation parameters that make the generated image output by the conditional diffusion model best match the real image.

[0114] The disease manifold spatial pre-calibration is used to build a continuous spectral map of parameters on the training set, and includes:

[0115] The training set parameter set dimensionality reduction is used to reduce the reaction-diffusion equation parameters of the training set samples to a two-dimensional manifold. Specifically, it calculates the high-dimensional adjacency probability for the reaction-diffusion equation parameters of all training set samples, and obtains the two-dimensional manifold of each training set sample by minimizing the KL divergence with the low-dimensional adjacency probability.

[0116] Basin typology clustering is used to automatically divide the clustered regions of asthma types represented by asthma typology labels on a two-dimensional manifold. Specifically, it performs K-means clustering on the two-dimensional manifold of the training set samples using Wasserstein distance as a metric to obtain multiple attractor basins based on asthma typology labels.

[0117] The final classification output is used to project the patient to be classified onto a pre-calibrated manifold space and output the clinical classification. Specifically, it uses the same method as the dimensionality reduction of the training set parameter set to calculate the two-dimensional manifold of the response-diffusion equation parameters of the patient to be classified, and calculates the Euclidean distance to the center of all attractor basins. The distance is then converted into a normalized membership vector through the softmax function to obtain the continuous confidence score of the asthma classification label. The asthma classification label with the maximum confidence score greater than a set threshold is taken as the output.

[0118] The model construction and training specifically involves constructing an improved conditional diffusion model by integrating the conditional diffusion model construction, the optimal physical parameter inversion, the disease manifold space pre-calibration, and the final genotyping output. The model is then trained and its performance is verified based on the genotyping training set and the genotyping test set to obtain the improved conditional diffusion model, which serves as the genotyping model.

[0119] By performing the above operations, this solution addresses the technical problems of traditional asthma CT image classification systems, which lack physical consistency constraints, causing the model to deviate from the actual respiratory motion mechanism when learning image generation or feature extraction, and ignore the differences in respiratory dynamic parameters between individuals, making it impossible to personalize parameter adaptation for new cases and resulting in highly unstable model performance when facing patients with different disease severity or comorbidities. This solution creatively adopts an improved conditional diffusion model as the classification model, embedding physical motion equations as conditional constraints into the generation process. This allows the model to fit statistical distributions to provide richer distribution information for classification while ensuring that the generation and inference processes conform to the basic laws of respiratory mechanics, thus enhancing the model's generalization ability to new patients.

[0120] Example 6, see Figure 1 This embodiment is based on the above embodiment. In the asthma image classification module, the dataset to be classified is used as the input of the classification model to perform asthma CT image classification, and the asthma classification label output by the classification model is used as the final asthma CT image classification result.

[0121] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0122] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention.

[0123] The present invention and its embodiments have been described above. This description is not restrictive, and the accompanying drawings are only one embodiment of the present invention; the actual structure is not limited thereto. In conclusion, if those skilled in the art are inspired by this description and design similar structures and embodiments without departing from the spirit of the invention, such designs should fall within the protection scope of the present invention.

Claims

1. A deep learning-based intelligent classification system for asthma CT images, characterized in that: The system includes an image data acquisition module, a data preprocessing module, a constraint extraction model construction module, a classification model construction module, and an asthma image classification module; The image data acquisition module obtains the original dataset of asthma classification through data acquisition. The original dataset of asthma classification specifically includes the historical original dataset of asthma classification and the current original dataset of asthma classification. The data preprocessing module employs HU normalization, lung parenchyma segmentation, biphasic rigid registration, lung region trimming, and dataset partitioning to obtain a preliminary dataset to be classified, a preliminary training set, and a preliminary test set. The constraint extraction model construction module is used to construct the model required to extract the patient's personalized respiratory mechanics features from dual-phase CT images. Specifically, it constructs an improved residual network model as the constraint extraction model and processes it with the preliminary dataset to be classified, the preliminary training set, and the preliminary test set as inputs to obtain the dataset to be classified, the classification training set, and the classification test set. The improved residual network model extracts individual dynamic constraint features by predicting the three-dimensional displacement vector of each voxel from the inspiratory phase to the expiratory phase and constructing a graph structure. The classification model construction module is used to construct the model required for outputting asthma classification labels based on inspiratory phase CT images and individual dynamic constraint features output from upstream. Specifically, it constructs an improved conditional diffusion model as the classification model. The asthma image classification module specifically uses the dataset to be classified as the input of the classification model to perform asthma CT image classification, and the asthma classification label output by the classification model is used as the final asthma CT image classification result.

2. The intelligent asthma CT image classification system based on deep learning according to claim 1, characterized in that: The constraint extraction model construction module specifically includes continuous displacement field prediction, motion topology skeleton extraction, strain rate calculation, constraint feature extraction, and model construction and training.

3. The intelligent asthma CT image classification system based on deep learning according to claim 2, characterized in that: The continuous displacement field prediction is used to predict the three-dimensional displacement vector of voxels from the preprocessed dual-phase CT block, and includes: Biphasic feature extraction is used to extract high-level semantic features from biphasic CT blocks. Specifically, two weight-shared 3D ResNet-34s are used to extract features of inspiratory and expiratory local blocks, respectively. After global average pooling, inspiratory and expiratory feature maps are obtained. Then, coarse fused features are obtained by element-wise difference fusion and upsampling to restore resolution. Continuous implicit decoding is used to convert the displacement field on discrete voxels into a function representation of continuous coordinates. Specifically, for any continuous coordinate within a block, local features are first obtained from coarse fusion feature bilinear interpolation, and then input into a multilayer perceptron along with the position encoding. The output is the three-dimensional displacement vector of that coordinate, thus obtaining the three-dimensional displacement field. The motion topology skeleton extraction is used to automatically extract the graph structure of the airway centerline from the displacement field, and includes: Displacement field divergence calculation is used to quantify the local volume change rate of each voxel. Specifically, it is obtained by calculating the sum of the partial derivatives of each component of the three-dimensional displacement field to obtain the divergence volume. Zero-crossing surface processing is used to extract the airway wall from the divergence volume and simplify it into a centerline graph. Specifically, the traveling cube algorithm is used to extract isosurfaces with divergence of 0, and then the topology refinement algorithm is used to erode them to a single pixel width and remove short branches to obtain an undirected graph of the airway structure. The airway attribute assignment is used to assign motion-related attributes to each airway segment. Specifically, five points are uniformly sampled along each edge of the undirected graph of the airway structure, and the displacement amplitude and divergence value of each point are read from the three-dimensional displacement field to obtain the airway enhancement undirected graph. The strain rate calculation, used to quantify the degree of radial deformation of each airway segment during the respiratory cycle, includes: Radial tube wall distance extraction is used to measure the airway inner and outer diameters at each sampling point. Specifically, eight rays are emitted from each sampling point along the radial direction perpendicular to each side of the undirected graph of the airway structure. These rays intersect with the zero divergence isosurface to obtain the nearest intersection distance in the inspiratory phase and the nearest intersection distance in the expiratory phase, which are used as the radial radius in the inspiratory phase and the radial radius in the expiratory phase, respectively. Radial strain rate calculation is used to convert the change in tube diameter into a dimensionless strain rate index. Specifically, the radial radius of 5 sampling points on each side of the undirected diagram of the airway structure is averaged, and then the relative change rate of the exhalation phase relative to the inspiratory phase is calculated to obtain the radial strain rate value of each side of the undirected diagram of the airway structure. The constraint feature extraction is used to compress the non-Euclidean airway structure undirected graph and radial strain rate sequence into a fixed-length vector, which includes: Graph attention message passing is used to preserve the hierarchical structure information of the undirected graph of the airway structure in the graph neural network. Specifically, when passing messages in each layer of the graph neural network, the edge attributes and bifurcation differences of the undirected graph of the airway structure are incorporated as biases into the attention weight calculation to obtain the node latent vector after each layer is updated. The edge attributes of the undirected graph of the airway structure specifically include the edge length and the edge radial strain rate value. Differentiable graph pooling is used to aggregate a variable-size set of nodes into a fixed-size graph-level vector and force the compression process to preserve the original topology. Specifically, differentiable graph pooling is used to aggregate the nodes of the undirected graph of the airway structure after graph attention message passing processing into 32 super nodes and sum them globally to obtain preliminary constraint features and pooled coarsened graph. The edit distance between the coarsened graph and the undirected graph of the airway structure is calculated as the topology regularization loss. Constraint feature acquisition is used to reduce the dimensionality of high-dimensional graph-level vectors to obtain a compact representation. Specifically, the initial constraint features are further reduced in dimensionality through two fully connected layers to obtain individual dynamic constraint features. The construction and training of the model specifically involves constructing an improved residual network model by integrating the continuous displacement field prediction, the motion topology skeleton extraction, the strain rate calculation, and the constraint feature extraction. The model is then trained and its performance is verified based on a preliminary training set and a preliminary test set to obtain the improved residual network model, which serves as the constraint extraction model. The preliminary dataset to be classified, the preliminary training set, and the preliminary test set are used as inputs to the constraint extraction model to obtain the dataset to be classified, the classification training set, and the classification test set.

4. The intelligent asthma CT image classification system based on deep learning according to claim 1, characterized in that: The classification model construction module specifically includes conditional diffusion model construction, optimal physical parameter inversion, disease manifold space pre-calibration, final classification output, and model construction and training.

5. The intelligent asthma CT image classification system based on deep learning according to claim 4, characterized in that: The conditional diffusion model is constructed to establish a generative model that can both generate realistic CT images and enforce physical consistency. The model includes: The forward diffusion noise addition process is used to gradually destroy the coding results of the inspiratory phase CT images into standard Gaussian noise. Specifically, cosine noise scheduling is used to gradually add Gaussian noise with increasing variance to the coding results of the inspiratory phase CT images to obtain noisy coding features. Physical consistency constraints are used to design a reaction-diffusion equation embedded in the forward diffusion noise process. Specifically, an equation is defined that includes diffusion, reaction, and respiratory drive, and the respiratory drive term is modulated using individual dynamic constraint features. The predicted expiratory coding features are obtained by solving the equation, and the physical consistency loss is calculated together with the expiratory phase feature map. Conditional denoising network is used to train a denoising network with individual dynamic constraint features as conditions. Specifically, a three-dimensional U-Net is used as the denoising network, and FiLM modulation is used at each skip connection to inject individual dynamic constraint features as conditions to obtain the conditional denoising network. The optimal physical parameter inversion is used to solve the reaction-diffusion equation parameters that best explain the CT images and respiratory movements of patients to be classified. The content includes: Initialization is used to prepare the starting point for inversion. Specifically, the trained conditional diffusion model is fixed, and the inspiratory phase CT images of the patients to be classified are processed. The parameters of the reaction-diffusion equation are initialized to the mean of the training set. Joint optimization with physical constraints is used to search for optimal reaction-diffusion equation parameters under a fixed model. Specifically, it involves alternating the denoising step of the conditional denoising network and the gradient descent of the reaction-diffusion equation parameters for patients to be classified, so as to obtain the reaction-diffusion equation parameters that make the generated image output by the conditional diffusion model best match the real image. The disease manifold spatial pre-calibration is used to build a continuous spectral map of parameters on the training set, and includes: The training set parameter set dimensionality reduction is used to reduce the reaction-diffusion equation parameters of the training set samples to a two-dimensional manifold. Specifically, it calculates the high-dimensional adjacency probability for the reaction-diffusion equation parameters of all training set samples, and obtains the two-dimensional manifold of each training set sample by minimizing the KL divergence with the low-dimensional adjacency probability. Basin typology clustering is used to automatically divide the clustered regions of asthma types represented by asthma typology labels on a two-dimensional manifold. Specifically, it performs K-means clustering on the two-dimensional manifold of the training set samples using Wasserstein distance as a metric to obtain multiple attractor basins based on asthma typology labels. The final classification output is used to project the patient to be classified onto a pre-calibrated manifold space and output the clinical classification. Specifically, it uses the same method as the dimensionality reduction of the training set parameter set to calculate the two-dimensional manifold of the response-diffusion equation parameters of the patient to be classified, and calculates the Euclidean distance to the center of all attractor basins. The distance is then converted into a normalized membership vector through the softmax function to obtain the continuous confidence score of the asthma classification label. The asthma classification label with the maximum confidence score greater than a set threshold is taken as the output. The model construction and training specifically involves constructing an improved conditional diffusion model by integrating the conditional diffusion model construction, the optimal physical parameter inversion, the disease manifold space pre-calibration, and the final genotyping output. The model is then trained and its performance is verified based on the genotyping training set and the genotyping test set to obtain the improved conditional diffusion model, which serves as the genotyping model.

6. The intelligent asthma CT image classification system based on deep learning according to claim 1, characterized in that: Both the historical asthma classification dataset and the current asthma classification dataset contain biphasic CT image data. The historical asthma classification dataset also includes asthma classification label data. The biphasic CT image data specifically includes inspiratory CT image data and expiratory CT image data. The inspiratory CT image data specifically refers to high-resolution CT images of the lungs acquired at the end of inspiration, and the expiratory CT image data specifically refers to high-resolution CT images of the lungs acquired at the end of expiration. Inspiratory CT images and expiratory CT images acquired from the same patient constitute a data pair.

7. The intelligent asthma CT image classification system based on deep learning according to claim 1, characterized in that: In the data preprocessing module, HU normalization is used to eliminate differences in voxel spacing and grayscale range caused by different CT equipment and scanning protocols. Specifically, it involves resampling all CT image sequences to an isotropic 1mm. 3 Voxels were used, and the HU values ​​were truncated to the lung window range and then linearly mapped to the interval [0,1]. The lung parenchyma segmentation is used to remove non-lung-related structures and focus on the lung region. Specifically, it uses a thresholding method combined with connected component analysis and morphological closing operation to extract the lung region and obtain a binary lung mask. The biphasic rigid registration is used to correct global misalignment caused by patient positional movement. Specifically, it uses the inspiratory phase as a fixed image and the expiratory phase as a floating image, and maximizes the normalized cross-correlation through a six-degree-of-freedom rigid body transformation to obtain a spatially aligned expiratory phase image. The lung region cropping is used to increase the diversity of training samples. Specifically, it uses the minimum bounding box of the binary lung mask to crop the biphasic CT image, and then divides it into blocks of size 64×64×64 with a 50% overlapping sliding window to obtain local block pairs. The dataset partitioning is used to obtain training data and test data, specifically by splitting the original historical asthma classification dataset. The current asthma subtyping dataset is preprocessed using HU normalization, lung parenchyma segmentation, biphasic rigid registration, and lung region pruning to obtain a preliminary dataset to be subtyped. The historical asthma subtyping dataset is preprocessed using HU normalization, lung parenchyma segmentation, biphasic rigid registration, lung region pruning, and dataset partitioning to obtain a preliminary training set and a preliminary test set.

Images