A multi-modal intelligent identification system for child plastic bronchitis, which is fused with a three-dimensional airway deformation-scale synergistic enhancement module

Abstract

This invention relates to the field of intelligent medical imaging diagnosis technology, specifically a multimodal intelligent identification system for pediatric plastic bronchitis integrating a three-dimensional airway deformation-scale co-enhancement module. The system includes a data acquisition and preprocessing module, a multimodal feature encoding module, an early feature fusion module, a cross-modal attention module, a three-dimensional airway deformation-scale co-enhancement module, and a classification and decision module. Through multimodal feature encoding, spatial structural features of CT images are extracted and high-dimensional embedded representations of clinical data are obtained. After early fusion and cross-modal attention weighting adjustment, the three-dimensional airway deformation-scale co-enhancement module captures the three-dimensional irregular deformation and cross-scale diffusion features of lesions through dual-branch collaborative modeling, ultimately outputting a three-classification result and treatment suggestions. This invention achieves end-to-end intelligent identification, significantly improving diagnostic accuracy and consistency, reducing reliance on physician experience, adapting to the structural characteristics of children's airways, and providing reliable support for precise clinical diagnosis.

Description

Technical Field

[0001] This invention relates to the field of intelligent medical imaging diagnostic technology, specifically to a multimodal intelligent recognition system for pediatric plastic bronchitis that integrates a three-dimensional airway deformation-scale co-enhancement module. Background Technology

[0002] Plastic bronchitis is a serious respiratory disease in children, characterized by the formation of tree-like plastic structures within the bronchi. These structures obstruct the airways, causing symptoms such as difficulty breathing, wheezing, and cyanosis. In severe cases, it can lead to respiratory failure and even death. Children's lungs have delicate airway structures that are not fully developed, resulting in complex and variable lesions in plastic bronchitis. Early symptoms are often confused with common pneumonia, asthma, and other diseases, making clinical diagnosis challenging. Timely and accurate diagnosis is a prerequisite for developing an effective treatment plan, especially for children requiring urgent bronchoscopy for thrombectomy. Delayed diagnosis can significantly increase treatment risks and the probability of poor prognosis. Therefore, there is an urgent clinical need for efficient and precise identification techniques for plastic bronchitis in children.

[0003] Currently, the clinical diagnosis of plastic bronchitis mainly relies on lung CT imaging combined with the physician's clinical experience. Physicians observe changes in airway morphology, the location and density of lesions in lung CT images, and combine this with clinical information such as the child's age, symptoms, and signs to make a comprehensive judgment. However, this traditional diagnostic method has many limitations: firstly, the airway structure in children's lung CT images is delicate, and lesions often exhibit irregular three-dimensional deformation and cross-scale diffusion characteristics, making it difficult to accurately capture these subtle features through manual observation, easily leading to missed diagnoses and misdiagnoses; secondly, the diagnostic results are highly dependent on the physician's experience level, and different physicians have different standards for judging imaging features, resulting in poor diagnostic consistency, especially in primary healthcare institutions or among inexperienced physicians.

[0004] With the development of intelligent medical image processing technology, deep learning-based disease identification methods are gradually being applied to the field of lung disease diagnosis, providing new ideas for the auxiliary diagnosis of plastic bronchitis. Existing technologies include some lung disease identification models based on two-dimensional convolutional neural networks, which classify diseases by extracting features from two-dimensional CT images of the lungs. However, these models can only process two-dimensional image information and cannot effectively capture the three-dimensional spatial structural features of lesions. Since lesions in pediatric plastic bronchitis exhibit significant three-dimensional deformation characteristics, two-dimensional models struggle to accurately represent these features, thus limiting the accuracy of identification.

[0005] Furthermore, existing intelligent identification technologies largely rely solely on single CT image data, failing to fully integrate the diagnostic value of clinical data. Clinical data, including information such as the child's age, symptoms, signs, and laboratory test results, is highly complementary to CT image features, providing crucial clinical guidance for lesion identification. Simultaneously, current technologies lack feature enhancement designs tailored to the unique airway structure of children, making it difficult to adapt to the specific characteristics of pediatric lung imaging, further hindering performance improvement. Therefore, developing an intelligent identification system capable of integrating multimodal data and accurately capturing the three-dimensional lesion features of pediatric plastic bronchitis has become an urgent need in the development of current clinical diagnostic technologies. Summary of the Invention

[0006] The purpose of this invention is to provide a multimodal intelligent identification system for childhood plastic bronchitis that integrates a three-dimensional airway deformation-scale co-enhancement module, in order to solve the problems mentioned in the background art, such as the reliance on human experience in the diagnosis of childhood plastic bronchitis leading to missed diagnoses and misdiagnoses, the difficulty of existing intelligent models in accurately capturing the three-dimensional features of lesions, and the lack of sufficient integration of multimodal data.

[0007] To achieve the above objectives, the present invention provides the following technical solution:

[0008] A multimodal intelligent identification system for pediatric plastic bronchitis integrating a three-dimensional airway deformation-scale co-enhancement module includes the following modules:

[0009] The data acquisition and preprocessing module is used to receive and standardize lung 3D CT image data and structured clinical data;

[0010] The multimodal feature encoding module includes a 3D convolutional network branch and a multilayer perceptron branch, which are used to extract the spatial structural features of CT images and the high-dimensional embedding representation of clinical data, respectively.

[0011] The early feature fusion module maps the high-dimensional embedded representation of clinical data to a dimension consistent with the feature channels of CT images, and achieves feature-level fusion by element-wise summation.

[0012] The cross-modal attention module generates channel-level attention weights based on fused features and raw clinical data, and performs clinically guided adaptive weighting of CT image features.

[0013] The 3D-ASC (3D Airway Deformation-Scale Co-enhancement Module) collaboratively models the 3D irregular deformation and cross-scale diffusion characteristics of PB lesions through deformable convolution branches and multi-scale sensing branches, and integrates spatial attention and channel attention mechanisms. The calculation of the deformable convolution branch in the 3D-ASC module satisfies the following formula:

[0014] ;

[0015] In the formula, This is the current center position of the convolution kernel. The offset of the convolution kernel sampling points relative to the center. A learnable dynamic offset. The weights corresponding to the sampling points. The input feature map contains the pixel values ​​at the corresponding locations. This is the convolution output value at the current position;

[0016] The classification and decision module is used to output the three-category classification results of plastic bronchitis lesion types and bronchoscopy examination decision suggestions.

[0017] Preferably, the standardized processing flow of the data acquisition and preprocessing module is as follows: CT images undergo sequence sorting, 3D reconstruction, resampling, size standardization, and intensity normalization; clinical data undergoes encoding, missing value imputation, and standardization to finally obtain standardized 3D image tensors and clinical feature vectors; wherein the intensity normalization of CT images uses the following formula: In the formula, These are the pixel values ​​of the original CT image. The average value of the pixels. The standard deviation of pixel values. These are the normalized pixel values; the Z-score standardization for clinical data uses the following formula: In the formula, These are the original clinical characteristic values. The mean of this feature. The standard deviation of this feature. These are the standardized clinical characteristic values.

[0018] Preferably, the specific operation flow of the multimodal feature encoding module is as follows: using the 3DDenseNet backbone network as a branch of the three-dimensional convolutional network, the hierarchical spatial structure features of CT images are extracted; using the multilayer perceptron as a branch of the multilayer perceptron, the high-dimensional embedding representation of clinical data is extracted; wherein the 3DDenseNet backbone network is composed of multiple dense blocks and transition layers alternately, and the fusion and reuse of multi-scale features are achieved through feature concatenation within each dense block, and the transition layer uses 1×1×1 convolution to perform channel dimensionality reduction and feature smoothing.

[0019] Preferably, the specific operation procedure of the cross-modal attention module is as follows:

[0020] S1: Global average pooling is performed on the early-fused image features to obtain the global image vector:

[0021] ;

[0022] S2: Combine the global image vector with the original clinical feature vector By splicing together, a cross-modal joint quantity is formed:

[0023] ;

[0024] S3: Input the cross-modal joint vector into the multilayer perceptron to generate channel-level attention weights. This weight is used to adaptively weight the fused features, thereby adjusting the response intensity of each channel of the image features:

[0025] ;

[0026] S4: Apply these weights to each channel of the fused feature map to achieve adaptive feature enhancement based on clinical context:

[0027] ;

[0028] In the formula, For global image vectors, This is a global average pooling operation. The number of feature channels, This represents a vector in a C-dimensional real space. This is the original clinical feature vector. For cross-modal joint vectors, and For the learnable weights of the multilayer perceptron, For activation function, It is the Sigmoid activation function. For the generated channel-level attention weight vector, This represents the feature tensor of the early-stage fused 3D image. This is the attention-weighted enhanced feature tensor.

[0029] Preferably, the specific operation process of the 3D-ASC airway deformation-scale co-enhancement module is as follows: first, the 3D deformation features and cross-scale diffusion features of the lesion are extracted by deformable convolution branches and multi-scale perception branches respectively; then, the outputs of the two branches are weighted and fused; finally, the fused features are enhanced by spatial attention and channel attention mechanisms to obtain a highly recognizable lesion feature representation; the 3D-ASC airway deformation-scale co-enhancement module is inserted layer by layer into the middle and high-level dense blocks of the 3DDenseNet backbone network, specifically at DenseBlock2, DenseBlock3 and DenseBlock4. The insertion method is to not change the topology of the backbone network and directly use the module output as the output feature of the corresponding dense block to achieve airway morphology enhancement from shallow to deep.

[0030] Preferably, the feature fusion process of the multi-scale perception branch in the 3D-ASC three-dimensional airway deformation-scale co-enhancement module satisfies the following formula:

[0031] ;

[0032] In the formula, , , The results are shown for three-dimensional average pooling operations with step sizes of 1, 2, and 4, respectively. This represents a 1×1×1 convolution operation. This indicates a concatenation operation along the channel dimension. This is the output feature tensor of the multi-scale sensing branch.

[0033] Preferably, the weighted fusion of the dual-branch outputs in the 3D-ASC three-dimensional airway deformation-scale co-enhancement module satisfies the following formula:

[0034] ;

[0035] In the formula, The output feature tensor of the deformable convolution branch, The output feature tensor of the multi-scale sensing branch and These are learnable weighting coefficients. This is the feature tensor after the two branches are merged.

[0036] Preferably, the spatial attention weight in the 3D-ASC three-dimensional airway deformation-scale co-enhancement module satisfies the following formula:

[0037] ;

[0038] In the formula, The feature tensor after the two branches are fused. and These are respectively three-dimensional max pooling and average pooling operations. This represents a 7×7×7 convolution operation. This indicates a concatenation operation along the channel dimension. It is the Sigmoid activation function. This is the generated spatial attention weight map.

[0039] Preferably, the channel attention weights in the 3D-ASC three-dimensional airway deformation-scale co-enhancement module satisfy the following formula:

[0040] ;

[0041] In the formula, The feature tensor after the two branches are fused. This indicates a global average pooling operation. and For learnable weights, For activation function, It is the Sigmoid activation function. This is the generated channel attention weight vector.

[0042] Preferably, the specific operation process of the classification and decision module is as follows: global pooling and fully connected operations are performed on the enhanced features output by the 3D-ASC module for three-dimensional airway deformation-scale co-enhancement to output a three-class probability distribution of plastic bronchitis lesion types; bronchoscopy decision suggestions are generated according to preset rules, wherein the preset rules are: when the predicted lesion type is mucus plug or plastic, bronchoscopy is recommended; when the predicted lesion type is solid or no lesion, bronchoscopy is not recommended. The classification probability threshold is set to 0.5. When the highest category probability is lower than 0.5, a prompt message indicating that further clinical evaluation is required is output.

[0043] Compared with the prior art, the beneficial effects of the present invention are:

[0044] (1) This invention integrates three-dimensional CT images and structured clinical data to automatically classify three types of PB-related lesions (solidification, mucus plug, and plasticity), and outputs a binary decision result on whether bronchoscopy is required. This provides clinicians with an objective, standardized, and non-invasive auxiliary judgment basis, which helps to reduce unnecessary bronchoscopy operations and reduce the risk of trauma to children.

[0045] (2) Significantly improves the accuracy and consistency of PB-related lesion identification. The multimodal deep learning model of this invention has significantly better classification accuracy, AUC and F1 score on independent test sets than the single-modal CT model and the average recognition level of doctors. It can provide stable and reliable lesion classification results even when PB phenotypes are highly heterogeneous and imaging features are highly overlapping, thereby improving the diagnostic consistency among different doctors and institutions.

[0046] (3) The 3D-ASC module significantly enhances the modeling ability of PB lesions in three dimensions. By introducing deformable convolution, multi-scale pooling, and a spatial-channel attention combination structure, the 3D-ASC module enables the network to adaptively focus on typical dendritic tubular emboli and related airway regions in PB in three-dimensional space, and capture the cross-scale lesion distribution from large airways to terminal small airways. Experimental ablation results show that, under the same dataset and backbone network conditions, the model integrating the 3D-ASC module outperforms the baseline model without the module in terms of classification accuracy and AUC, proving that the present invention has achieved substantial technical effects in spatial feature enhancement.

[0047] (4) Multimodal fusion and cross-modal attention mechanisms improve the model’s adaptability to individualized pathological states. This invention enables the model to dynamically adjust the degree of attention to different imaging feature channels according to the specific clinical indicators of the child, through early feature-level fusion and clinically guided cross-modal channel attention, so as to better reflect the individual differences of PB in different underlying diseases and different inflammatory states, thereby improving the model’s generalization ability and clinical applicability.

[0048] In summary, the technical solution of this invention has significant improvements over the prior art in terms of structural design, performance, and clinical application value, and is suitable for widespread application in the diagnosis and treatment of pediatric respiratory diseases. Attached Figure Description

[0049] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are explained in detail together with the embodiments of the invention, but do not constitute a limitation thereof.

[0050] Figure 1 This is a flowchart of the core architecture of the multimodal pediatric plastic bronchitis intelligent identification system of the present invention;

[0051] Figure 2 This is a flowchart of the data preprocessing process of the present invention;

[0052] Figure 3 This is a ROC curve comparing the performance of the models in this invention.

[0053] Figure 4This is a schematic diagram of the internal structure of the 3D-ASC module of the present invention;

[0054] Figure 5 Bar charts and line graphs showing the performance comparison of different model architecture combinations of this invention;

[0055] Figure 6 This is a visual schematic diagram of the system diagnostic process of the present invention;

[0056] Figure 7 This is a comparison of one-to-many ROC curves for three types of lesions in different backbone networks according to the present invention; Detailed Implementation

[0057] The technical solutions of the present invention will be clearly and completely described below with reference to the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0058] like Figures 1-6 As shown, the multimodal intelligent identification system for childhood plastic bronchitis of this invention, which integrates a three-dimensional airway deformation-scale collaborative enhancement module, achieves end-to-end intelligent identification through the collaborative work of multiple modules. It can accurately capture the three-dimensional irregular deformation and cross-scale diffusion characteristics of childhood plastic bronchitis lesions, improving the accuracy and consistency of identification and providing reliable support for clinical diagnosis. The specific implementation of each module and the overall system workflow are described in detail below:

[0059] The core function of the data acquisition and preprocessing module is to receive and standardize 3D CT images of the lungs and structured clinical data, providing high-quality data input for subsequent feature extraction and recognition tasks.

[0060] The standardized processing of 3D CT images of the lungs involves the following steps: First, sequence sorting is performed, organizing the CT tomographic images according to the scan time order to ensure the continuity of the image sequence; next, 3D reconstruction is performed, using a volume rendering algorithm to reconstruct the 2D tomographic images into 3D image data, with the reconstruction resolution set to 1×1×1 mm; then, resampling is performed, resampling the 3D image data to a fixed size of 128×128×128, using linear interpolation to ensure the integrity of the details in the resampled image. The comparison of image sizes before and after resampling is shown in the table below:

[0061] Next, size standardization is performed by cropping or padding the resampled image data to a uniform size to avoid affecting model training and recognition performance due to differences in the original image size; finally, intensity normalization is performed using the following formula:

[0062]

[0063] In the formula, I represents the pixel value of the original CT image, μ represents the mean pixel value, σ represents the standard deviation of the pixel value, and Inorm represents the normalized pixel value. Intensity normalization maps the pixel values ​​of the CT image to the [-1,1] interval, reducing intensity differences caused by different scanning equipment and scanning parameters.

[0064] The specific steps for standardizing structured clinical data are as follows: First, encoding is performed. Categorical variables such as gender, age, and symptoms are converted to numerical data using one-hot encoding, while continuous variables such as body temperature and white blood cell count are directly retained as their original values. Next, missing values ​​are imputed. Missing continuous variables are imputed using the mean imputed method, and missing categorical variables are imputed using the mode imputed method. Finally, standardization is performed using the Z-score standardization formula.

[0065]

[0066] In the formula, f represents the original clinical characteristic value. The mean of this feature. The standard deviation of this feature. These are the standardized clinical feature values. Standardization ensures that all features in the clinical data are of the same magnitude, avoiding interference from differences in feature magnitudes during model training.

[0067] After the above standardization process, we obtain standardized 3D image tensors and clinical feature vectors that can be directly input into subsequent modules.

[0068] The multimodal feature encoding module includes a 3D convolutional network branch and a multilayer perceptron branch, which are used to extract the spatial structural features of CT images and the high-dimensional embedding representation of clinical data, respectively.

[0069] The 3D convolutional network branch is implemented using the 3DDenseNet backbone, which consists of four dense blocks and three transition layers alternatingly. Each dense block contains multiple 3D convolutional layers with a kernel size of 3×3×3, a stride of 1, and same padding to ensure that the feature map size remains unchanged after convolution. Within each dense block, multi-scale feature fusion and reuse are achieved through feature cascading; that is, the input to each convolutional layer is a cascade of the output feature maps of all preceding convolutional layers. The transition layers use 1×1×1 convolutions for channel dimensionality reduction and feature smoothing, with the number of 1×1×1 convolutional kernels set to 0.5 times the number of output channels of the previous dense block, reducing computational cost while preserving key features. The 3DDenseNet backbone performs layer-by-layer convolution operations on the 3D image tensor, ultimately outputting a hierarchical spatial structure feature with a dimension of 256×16×16×16.

[0070] The multilayer perceptron branch consists of three fully connected layers. The first fully connected layer takes the clinical feature vector as its input dimension and outputs a dimension of 128. The second fully connected layer takes the same input dimension as the clinical feature vector and outputs a dimension of 256. The third fully connected layer takes the same input dimension as the clinical feature vector and outputs a dimension of 256. Each fully connected layer is followed by a ReLU activation function and a Dropout layer with a Dropout probability of 0.5 to prevent overfitting. The multilayer perceptron maps the clinical feature vector layer by layer, ultimately outputting a high-dimensional embedding representation with a dimension of 256.

[0071] The core function of the early feature fusion module is to perform feature-level fusion of the high-dimensional embedded representation of clinical data with the spatial structural features of CT images. The specific implementation process is as follows: First, the high-dimensional embedded representation of clinical data is mapped to a dimension of 256, consistent with the feature channels of the CT images, using a 1×1×1 convolution. Then, the mapped clinical features are shape-expanded from a one-dimensional vector to a 256×16×16×16 tensor, with the expansion method being repeated filling along the spatial dimension. Finally, the expanded clinical features and the spatial structural features of the CT images are summed element-wise along the channel dimension to complete the early feature fusion, outputting the fused feature tensor. Through early feature fusion, clinical information can guide the extraction of image features at an early stage, improving the discriminative ability of the features.

[0072] The cross-modal attention module is used to generate channel-level attention weights based on the fused features and the original clinical data, and to perform clinically guided adaptive weighting of CT image features. The specific implementation process is as follows: before inputting the fused feature tensor into the deep network, it further enters the cross-modal attention mechanism module to dynamically adjust the importance distribution of the feature channels. This module consists of three parts: (1) global average pooling to extract image feature vectors; (2) concatenating them with clinical features to form a joint representation vector; (3) inputting it to a multilayer perceptron and outputting a channel-level attention weight vector W∈ .

[0073] Specifically, this module generates weight coefficients for each channel by calculating the dynamic correlation between global image features and clinical features, which are then used to adjust the response intensity of the image feature map. The specific implementation involves first processing the fused image features... Perform global average pooling to obtain the global image vector:

[0074]

[0075] Subsequently compared with the original clinical feature vector Concatenate to form a cross-modal joint vector:

[0076]

[0077] It is then fed into a multilayer perceptron (MLP) to generate channel-level attention weights. Used to adjust the response intensity of each channel of the image features:

[0078]

[0079] Finally, these weights are applied to each channel of the fused feature map to achieve adaptive feature enhancement based on clinical context:

[0080]

[0081] The main motivation for introducing cross-modal attention is that the imaging manifestations of plastic bronchitis exhibit high heterogeneity, and simple modal superposition is insufficient to adequately model inter-individual differences. Through a clinically guided channel weighting mechanism, the network can be encouraged to focus on important lesion-related regions based on the patient's specific physiological state, improving the consistency and discriminative power of feature representations.

[0082] In the formula, For global image vectors, This is a global average pooling operation. The number of feature channels, This represents a vector in a C-dimensional real space. This is the original clinical feature vector. For cross-modal joint vectors, and For the learnable weights of the multilayer perceptron, For activation function, It is the Sigmoid activation function. For the generated channel-level attention weight vector, This represents the feature tensor of the early-stage fused 3D image. This is the attention-weighted enhanced feature tensor.

[0083] The 3D-ASC module, also known as the three-dimensional airway deformation-scale co-enhancement module, is used to specifically enhance the network's ability to model irregular deformation, local structural mutations, and cross-scale diffusion characteristics of PB lesions in children in three-dimensional space.

[0084] The 3D-ASC module adopts a dual-branch structure, consisting of a deformable convolutional branch and a multi-scale perceptual branch. Spatial attention and channel attention are then concatenated after branch fusion to achieve fine-grained feature enhancement. This module contains two parallel sub-branches:

[0085] Branch 1: Deformable Convolution Branch. The deformable convolution branch introduces a learnable offset δ based on standard 3D convolution, dynamically adjusting the position of each sampling point. This allows the convolution kernel to adaptively fit the dendritic tubular emboli and irregular obstructive regions distributed along the bronchial tree. The computational form of this branch can be expressed as:

[0086] );

[0087] Branch Two: Multi-Scale Perception Branch. The multi-scale perception branch first performs channel dimensionality reduction on the input feature map using 1×1×1 convolutions to reduce computational overhead and compress redundant information. Then, it applies multiple sets of three-dimensional average pooling operations with different lengths (e.g., 1, 2, 4) to the dimensionality-reduced features to obtain feature representations at different spatial scales. Next, the features at each scale are upsampled to a uniform spatial size and concatenated along the channel dimension. Finally, 1×1×1 convolutions are used for channel fusion and dimensionality upscaling, forming a multi-scale feature tensor containing local details and global semantics. This process can be formally represented as:

[0088]

[0089] The outputs of the two sub-branches are weighted and fused (learning parameters α and β) to synthesize a fused feature tensor, which is then fed into a spatial attention module (7×7×7 convolution to generate a spatial weight map) and a channel attention module (SE structure) for dual feature enhancement. The fused output is formally represented as follows:

[0090]

[0091] The spatial attention weights are calculated as follows and used to generate a three-dimensional spatial attention map:

[0092]

[0093] The channel attention calculation (SE mechanism) is as follows, where Indicates global average pooling. , For learning weights:

[0094]

[0095] In the overall network structure, the 3D-ASC module is inserted hierarchically into the mid-to-high-level feature blocks (such as DenseBlock2, DenseBlock3, and DenseBlock4) of the 3D convolutional backbone network without altering the backbone topology. This enhances the feature representation capabilities at different depths and receptive fields. Through this hierarchical insertion approach, the network can collaboratively model the airway distribution patterns and 3D structural changes of PB lesions at multiple scales and levels. Experimental results show that, compared to the baseline model without the 3D-ASC module, integrating this module significantly improves classification accuracy, AUC, and F1 scores in both the PB three-class classification task and the bronchoscopy decision task.

[0096] In the formula, F is the feature tensor after dual-branch fusion, GAP represents the global average pooling operation, W1 and W2 are learnable weights, ReLU is the activation function, σ is the sigmoid activation function, and Mc is the generated channel attention weight vector. The channel attention weight vector is multiplied element-wise with the spatially augmented feature tensor along the channel dimension to obtain the final high-recognition lesion feature representation.

[0097] The three-dimensional airway deformation-scale co-enhancement module is layered and inserted into the middle and high-level dense blocks of the 3DDenseNet backbone network. Specifically, it is inserted at DenseBlock2, DenseBlock3, and DenseBlock4. The insertion method does not change the topology of the backbone network. The module output is directly used as the output feature of the corresponding dense block to achieve airway morphology enhancement from shallow to deep.

[0098] The classification and decision module outputs a three-dimensional classification result for plastic bronchitis lesion types and bronchoscopy decision recommendations. The specific implementation process is as follows: First, the enhanced features output by the 3D airway deformation-scale co-enhancement module are globally pooled, converting the 256×16×16×16 enhanced feature tensor into a 256-dimensional feature vector. Then, the feature vector is input into a fully connected layer (input dimension 256, output dimension 3), followed by a Softmax activation function to output a three-dimensional probability distribution for plastic bronchitis lesion types: consolidation, mucus plug, and plastic lesions. Finally, bronchoscopy decision recommendations are generated according to preset rules. The preset rules recommend bronchoscopy when the predicted lesion type is mucus plug or plastic lesion, and discourage bronchoscopy when the predicted lesion type is consolidation or no lesion. The classification probability threshold is set to 0.5; when the highest category probability is below 0.5, a suggestion for further clinical evaluation is output.

[0099] To facilitate understanding of the working mechanism of the system of the present invention, the above modules are now described in series from the perspective of the diagnostic process. The typical workflow of the multimodal pediatric PB intelligent recognition system includes the following steps:

[0100] (1) Step 1: Data Acquisition and Preprocessing: The system automatically acquires the CTDICOM images of the child's chest from the hospital's PACS system and simultaneously retrieves the corresponding structured clinical data from the HIS / LIS system. The data acquisition and preprocessing module performs sequence sorting, three-dimensional reconstruction, resampling, size standardization and intensity normalization on the CT images, and performs encoding, missing value imputation and standardization on the clinical data to form standardized three-dimensional image tensors and clinical feature vectors.

[0101] (2) Step 2: Modal Independent Feature Encoding: The preprocessed CT tensor is input into the three-dimensional convolution branch of the multimodal feature encoding module to extract hierarchical spatial structure features; the clinical feature vector is input into the multilayer perceptron branch to obtain a high-dimensional clinical embedding representation.

[0102] (3) Step 3: Early feature fusion: The early feature fusion module maps the clinical embedding to the same dimension as the image channel, and fuses it with the shallow image features by expansion and element-wise summation to obtain a preliminary fused three-dimensional feature map.

[0103] (4) Step 4: Cross-modal attention regulation: The fused feature map is globally averaged and pooled by the cross-modal attention module, and then concatenated with the original clinical feature vector and input into the multilayer perceptron to generate channel-level attention weights; the system uses these weights to adjust the features of each channel to form an enhanced feature map guided by clinical information.

[0104] (5) Step 5: 3D Deformation-Scale Co-enhancement: The enhanced feature map input is the 3D airway deformation-scale co-enhancement module (3D-ASC). In the deformable convolution branch, the 3D irregular deformation of the airway and embolus structure is modeled. In the multi-scale perception branch, the cross-scale lesion distribution from the large airway to the terminal small airway is captured. The response of key regions related to PB is further amplified through spatial and channel attention, and a highly recognizable 3D feature representation is output.

[0105] (6) Step Six: Classification Output and Bronchoscopy Decision: The classification and decision module performs global pooling and full connection operations on the above features to obtain the probability distribution of three lesion types: solidification, mucus plug and plasticity. Based on the preset rules, mucus plug and plasticity are classified as recommended bronchoscopy categories, and non-obstructive / solidification are classified as not recommended bronchoscopy categories. Finally, PB phenotype classification results and bronchoscopy examination decision suggestions are generated.

[0106] Through the above steps, this invention organically combines multi-source data acquisition, intelligent identification, and clinical decision-making to form a complete and standardized intelligent diagnostic process for pediatric PB.

[0107] Example

[0108] This embodiment presents a user interface for a pediatric lung CT-based intelligent plastic bronchitis auxiliary analysis system for use by doctors. It primarily facilitates the input, analysis, and prediction of patient image and laboratory data. Designed for clinical decision support scenarios, the interface employs a simple and intuitive web-based backend layout, allowing for visual demonstration of system functions in patent application materials and providing a front-end interaction framework reference for subsequent actual system development.

[0109] 1. Patient information entry function

[0110] The system interface includes a patient information area for entering or displaying basic patient information, including patient name, patient number, gender, age, and examination date. This function establishes a basic information profile for the subject to be analyzed, allowing physicians to simultaneously view patient identity and examination information before performing image analysis, thereby enhancing the standardization of the system interface and its clinical usability. This module also provides foundational data support for subsequent generation of analysis records, output of diagnostic reports, and case management.

[0111] 2. CT image data upload function

[0112] This system features a dedicated CT data upload area for receiving chest CT images of patients to be analyzed. Upload options include selecting single or multiple DICOM files, as well as directly selecting a folder containing multiple DICOM slices, accommodating the common practice of storing clinical imaging data in sequential folders. After the user completes the upload, the interface displays the names of the currently selected files or folders within the upload area, and the system generates the CT image input data to be analyzed. This function demonstrates the system's ability to access raw medical imaging data and is one of the core input components of the entire intelligent prediction process.

[0113] 3. Laboratory indicator data upload function

[0114] In addition to CT image data, the system interface also includes a laboratory indicator data upload function for importing patient-related tabular clinical indicator data, such as complete blood count, biochemical indicators, or inflammation-related test results. This module supports uploading Excel or CSV files, and the name of the currently selected table file can be displayed on the interface after uploading. Through this function, the system can demonstrate the joint analysis approach of image information and clinical indicator information at the interface level, providing an input entry point for subsequent multimodal model integration and enhancing the system's integrity and scalability.

[0115] 4. Slice quantity display function

[0116] After the system completes the upload of CT image data, it can display the number of slices corresponding to the current case on the interface. This function is used to intuitively reflect the scale and completeness of the imported image data, making it easy for doctors to quickly confirm whether the data has been successfully loaded. This display item plays an important interactive prompt role in the interface design, enhancing the data visualization of the system and providing a display location for automatically reading DICOM sequence information and counting the number of valid slices in the subsequent real system.

[0117] 5. CT image preview function

[0118] This embodiment of the system features a CT image preview area in the center, used to display a schematic cross-sectional view of the image to be analyzed and providing a candidate slice switching function. Users can switch the current preview level by clicking different candidate slice buttons, thus intuitively viewing the image level of interest to the system. Although the current patent demonstration version primarily uses illustrative images, this module has reserved an entry point for displaying actual images in the interface logic, which can be further expanded to include functions such as actual CT slice browsing, window width and level adjustment, and lesion visualization highlighting. This function demonstrates the system's support capability for image visualization interaction.

[0119] 6. Intelligent analysis startup function

[0120] This embodiment of the system features a "Start Intelligent Analysis" button below the CT image preview area. This button initiates the prediction process after the CT image data and laboratory indicator data have been uploaded. The button's activation logic is linked to the data upload status, allowing clicking only when the necessary input data meets the requirements, thus ensuring the integrity and rationality of the analysis process. After the user clicks the button, the system enters the analysis state and displays the processing progress on the interface. This function embodies the system's task triggering mechanism and serves as a crucial interactive node connecting data input and model output.

[0121] 7. Analysis progress display function

[0122] After the intelligent analysis is initiated, the system interface displays the analysis progress in real time through a progress bar and percentage values. This design provides clear feedback to the user on the current task execution status, avoiding any impact on the user experience due to the analysis process being invisible. In the patent disclosure materials, this function can be used to illustrate that the system not only has data input and result output capabilities but also provides status feedback and process visualization capabilities for the analysis process, demonstrating the completeness of the system's human-computer interaction design.

[0123] 8. Classification prediction results display function

[0124] The system features a prediction results area on the right, which outputs the model's classification predictions for the current patient's chest CT data. In the current design, the system identifies and distinguishes between three categories: lumen patency, mucus plug, and plasticity. After analysis, the interface highlights the main predicted category and simultaneously displays the corresponding probability values ​​or confidence bar charts for each category. This module is the core output area of ​​the entire system, used to intuitively present the AI ​​model's judgment results on the case images to doctors, thus providing auxiliary reference for clinical judgment.

[0125] 9. Auxiliary conclusion output function

[0126] In addition to the basic classification results, this embodiment also includes an auxiliary conclusion module in the system interface. This module generates textual prompts corresponding to the analysis results, including lesion indications and AI suggestions. This module summarizes the system's judgment of abnormal areas and provides suggestive prompts for the current case, expanding the result display from a single classification output to a combination of classification results and explanatory auxiliary conclusions, thus improving the clinical readability of the interface and the system's completeness. If a real-world model is subsequently integrated, it can be further expanded to include functions such as image region localization, analysis of abnormal laboratory indicators, and automatic report generation.

[0127] 10. Report export function

[0128] The system interface includes a button for exporting a demo report, indicating a future expansion of the system's ability to generate and export analysis results reports. This module provides an interface for future applications such as archiving case analysis results, printing, adding supplementary medical record notes, and integration with hospital systems. Although it primarily serves as a functional entry point in the current patent presentation, its placement demonstrates the system's capability for result recording and output expansion.

[0129] Key module effectiveness and ablation experiment

[0130] To further verify the effectiveness of the key structural design in the proposed multimodal model and clarify the specific contribution of each module to performance improvement, this patent study conducted module ablation experiments under a unified experimental setting. The focus was on evaluating the impact of the 3D hierarchical multiscale attention module (3D-HMA), cross-modal attention mechanism, and early fusion strategy on the performance of the three-class classification task. All ablation experiments in this section used the multimodal DenseNet-169 as the backbone network. While maintaining consistency in training data partitioning, optimizer and learning rate configuration, number of training epochs, evaluation metrics, and the five-fold cross-validation strategy, only the model structural components were adjusted using controlled variables to ensure that performance differences between different model versions could be primarily attributed to the presence or absence of corresponding modules.

[0131] As shown in the table below, the complete model (D169+HMA+Attn+EF) achieved the best performance in all metrics, including ACC, AUC, Precision, Recall, and F1-score. Specifically, AUC reached 0.76, ACC was 0.60, and F1-score was 0.60, indicating that introducing 3D-HMA, cross-modal attention, and early fusion on this backbone network can create a synergistic overall gain. Further comparison of different ablation versions reveals differences in the performance contribution of each module, exhibiting a relatively clear functional division.

[0132]

[0133] When the 3D-HMA module was removed, while cross-modal attention and early fusion were retained (Model B), the model's AUC decreased from 0.76 to 0.70, ACC from 0.60 to 0.55, and F1-score from 0.60 to 0.53, showing the most significant performance decline. This result indicates that 3D-HMA plays a crucial role in enhancing image feature representation in this task. Its local deformation modeling, multi-scale contextual convergence, and fine-grained attention control mechanisms for space and channels effectively improve the model's ability to perceive complex 3D lesion patterns, thus significantly improving overall discrimination ability and inter-class balance. This phenomenon is consistent with the task characteristics of the aforementioned difficulty in class discrimination: when images are highly similar and differences are mainly reflected in subtle spatial structures and distribution patterns, simply relying on a conventional 3D convolutional backbone is often insufficient to stably capture key differences, while enhancement modules with stronger structural adaptation capabilities can exert more significant benefits.

[0134] When the cross-modal attention mechanism was removed, while 3D-HMA and early fusion (Model C) were retained, the AUC decreased from 0.76 to 0.73, the ACC from 0.60 to 0.58, and the F1-score from 0.60 to 0.55. These results suggest that cross-modal attention makes a stable contribution to performance improvement. Its main role is to dynamically recalibrate the image channel response using clinical information, enabling the model to focus more on discriminative channels related to the patient's physiological state when facing individual differences and image nonspecificity, thereby improving the consistency and separability of the joint representation. Although the gain it brings is lower than that of 3D-HMA, the simultaneous decrease in comprehensive indicators such as Precision, Recall, and F1-score indicates that this module not only affects the overall discriminative ability but also has a positive effect on reducing inter-class misclassification and improving predictive balance.

[0135] When the early fusion strategy was removed, while 3D-HMA and cross-modal attention (model D) were retained, the AUC decreased from 0.76 to 0.74, the ACC decreased from 0.60 to 0.58, and the F1-score decreased from 0.60 to 0.57. This result indicates that the timing of fusion has a significant impact on the final performance. Early fusion allows clinical context to participate in image representation construction in the early stages of feature learning, thus providing continuous guidance for subsequent levels of semantic abstraction. When fusion is delayed or weakened, even with the retention of attention and enhancement modules, the model's efficiency in utilizing cross-modal complementary information still decreases, ultimately resulting in a decline in overall performance metrics. Compared to cross-modal attention, the decrease in AUC from removing early fusion was slightly smaller, but the F1-score still showed a significant decline, suggesting that the timing of fusion is more likely to affect the overall balance between categories, rather than just the discriminative ability at a single threshold.

[0136] Furthermore, as a simplified control, when 3D-HMA, cross-modal attention, and early fusion were completely removed, and only the DenseNet-169 backbone was retained and directly stitched together or equivalently simplified (Model E), the model performance was the lowest across all metrics, with AUC of 0.70, ACC of 0.53, and F1-score of 0.50. This result further illustrates that relying solely on the backbone network or simple modal overlay is insufficient to fully exploit the complementary relationship between image and clinical data, and also makes it difficult to stably form discriminative joint feature representations in the fine-grained three-class classification scenario of this task. In summary, the ablation experiments structurally validated the effectiveness of the multimodal framework of this invention. 3D-HMA is a key component in improving image representation capabilities and overall discriminative performance, while cross-modal attention and early fusion provide continuous gains in enhancing modal interaction efficiency and improving inter-class comprehensive balance. Together, these three elements constitute an effective structural combination for recognizing complex pediatric PB phenotypes.

[0137] Additional diagrams demonstrating other effects.

[0138] To further evaluate the model's ability to distinguish different bronchoscopic outcomes at the categorical level, this invention employs a one-vs-rest strategy to plot ROC curves for the three types of lesions and calculate the AUC value for each category. This analysis avoids masking the differences between easily identifiable and poorly identifiable categories by simply summarizing performance using overall indicators, thus more intuitively revealing the main sources of error and their potential causes in the three-class classification task. The one-vs-rest ROC curve results for each backbone network are shown below. Figure 7 As shown.

[0139] from Figure 7It can be observed that the model's discriminative ability varies significantly across different categories. For the category of luminal patency, the AUCs of each network are generally between 0.72 and 0.79, showing relatively stable performance. This suggests that the model can consistently capture the common features of this category in terms of imaging and clinical information, and can differentiate it from the other two categories of airway obstruction-related lesions to some extent. Among them, ResNet-18 and DenseNet-169 have AUCs of 0.79 and 0.77 for this category, respectively, while DenseNet-121 and DenseNet-201 have AUCs of 0.74 and 0.72, respectively. The overall differences are not significant, indicating that the model has a relatively reliable discriminative ability at the coarse-grained phenotypic level of whether there are significant airway obstruction-related changes.

[0140] For the shaping category, all networks achieved high AUCs and outperformed the other two categories overall. Figure 7 The results showed that ResNet-18 and DenseNet-121 both achieved a shaping AUC of 0.89, while DenseNet-169 and DenseNet-201 achieved 0.84 and 0.83, respectively, suggesting that the models were relatively more effective in identifying plasticity lesions. This result is also consistent with clinical experience, namely that plasticity lesions typically correspond to more pronounced and widespread airway obstruction and secondary pulmonary consolidation, atelectasis, or inflammatory exudation, thus exhibiting stronger separability on imaging and clinical indicators. After learning such significant combined phenotypes, the model can form a more stable discrimination boundary.

[0141] In contrast, the AUC of the mucus plug category is generally low across all networks and fluctuates significantly, making it the most challenging category in the three-class classification task of this invention. Figure 7 The results show that ResNet-18's AUC for mucus plugs is only 0.45, DenseNet-201's is 0.46, while DenseNet-121 and DenseNet-169, although relatively higher, only reach 0.60 and 0.61 respectively. The ROC curves for the mucus plug category are generally closer to the diagonal, suggesting that the model has high uncertainty in distinguishing this category and is easily confused with luminal patency or vascularization. This phenomenon indicates that, under the existing data and task settings, the identifiable features of mucus plugs at the imaging level are relatively subtle, and their phenotype may exhibit significant internal heterogeneity, making it difficult for the model to learn a stable and consistent discrimination pattern across different cases.

[0142] The combined results of the three categories show that the model of this invention is better at identifying outcomes with relatively more consistent or significant phenotypic features, namely, patency and plasticity. Mucus plugs, however, are the key bottleneck category limiting further improvement in overall multi-class classification performance. Nevertheless, differences in class discrimination can still be observed among different backbone networks from the overall ROC morphology and corresponding AUC levels. DenseNet-121 and DenseNet-169 achieved relatively higher AUCs in the mucus plug category, suggesting their superior ability to capture fine-grained differential features. This also corroborates the superior performance of the DenseNet series as the backbone in the aforementioned overall performance comparison.

[0143] It should be noted that although the multimodal model performs well in terms of overall AUC and accuracy in the three-class classification task, analysis at the category level reveals that this task still presents significant challenges in clinical applications. The three types of lesions exhibit some overlap in their CT imaging manifestations, especially in the early stages of the disease or when the obstruction is mild. Mucus plugs may not yet have caused typical complete luminal obstruction, and their imaging appearance may resemble patency or focal inflammatory changes, thus increasing the difficulty of differentiation. Furthermore, both mucus plugs and lesions can present as airway obstruction-related changes, and they often show similar indirect signs on imaging. Clinicians, lacking bronchoscopic evidence, find it difficult to accurately differentiate them based solely on visual interpretation. Therefore, the model needs to rely on high-dimensional image spatial details and contextual differences provided by clinical indicators to complete the discrimination, which places higher demands on data scale, feature representation capabilities, and model structure design. The ROC results at the category level mentioned above provide important evidence for subsequent human-machine comparative analysis and binary classification assessment for clinical decision-making. They also suggest that there is still room for improvement in data amplification, fine annotation of mucus plug-related phenotypes, or the introduction of more targeted representation learning strategies.

[0144] The table below details the core metrics such as ACC, AUC, and F1 for U-Net, ResNet series, and DenseNet series models in both unimodal and multimodal modes:

[0145]

[0146] The multimodal intelligent identification system for childhood plastic bronchitis integrating a three-dimensional airway deformation-scale collaborative enhancement module has the following advantages:

[0147] (1) This invention integrates three-dimensional CT images and structured clinical data to automatically classify three types of PB-related lesions (solidification, mucus plug, and plasticity), and outputs a binary decision result on whether bronchoscopy is required. This provides clinicians with an objective, standardized, and non-invasive auxiliary judgment basis, which helps to reduce unnecessary bronchoscopy operations and reduce the risk of trauma to children.

[0148] (2) Significantly improves the accuracy and consistency of PB-related lesion identification. The multimodal deep learning model of this invention has significantly better classification accuracy, AUC and F1 score on independent test sets than the single-modal CT model and the average recognition level of doctors. It can provide stable and reliable lesion classification results even when PB phenotypes are highly heterogeneous and imaging features are highly overlapping, thereby improving the diagnostic consistency among different doctors and institutions.

[0149] (3) The 3D-ASC module significantly enhances the modeling ability of PB lesion 3D structure. The performance differences of different architecture combinations (with and without 3D-ASC, Attn, EF) were compared through ablation experiments. By introducing deformable convolution, multi-scale pooling and spatial-channel attention combination structure, the 3D-ASC module enables the network to adaptively focus on typical dendritic tubular emboli and related airway regions of PB in 3D space, and capture the cross-scale lesion distribution from large airways to terminal small airways. The experimental ablation results show that, under the same dataset and backbone network conditions, the model integrating the 3D-ASC module is better than the baseline model without the module in terms of classification accuracy and AUC, proving that the present invention has achieved substantial technical effect in spatial feature enhancement.

[0150] (4) Multimodal fusion and cross-modal attention mechanisms improve the model’s adaptability to individualized pathological states. This invention enables the model to dynamically adjust the degree of attention to different imaging feature channels according to the specific clinical indicators of the child, through early feature-level fusion and clinically guided cross-modal channel attention, so as to better reflect the individual differences of PB in different underlying diseases and different inflammatory states, thereby improving the model’s generalization ability and clinical applicability.

[0151] In summary, the technical solution of this invention has significant improvements over the prior art in terms of structural design, performance, and clinical application value, and is suitable for widespread application in the diagnosis and treatment of pediatric respiratory diseases.

[0152] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely preferred examples and are not intended to limit the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A multimodal intelligent identification system for childhood plastic bronchitis integrating a three-dimensional airway deformation-scale co-enhancement module, characterized in that, Includes the following modules: The data acquisition and preprocessing module is used to receive and standardize lung 3D CT image data and structured clinical data; The multimodal feature encoding module includes a 3D convolutional network branch and a multilayer perceptron branch, which are used to extract the spatial structural features of CT images and the high-dimensional embedding representation of clinical data, respectively. The early feature fusion module maps the high-dimensional embedded representation of clinical data to a dimension consistent with the feature channels of CT images, and achieves feature-level fusion by element-wise summation. The cross-modal attention module generates channel-level attention weights based on fused features and raw clinical data, and performs clinically guided adaptive weighting of CT image features. The 3D-ASC (3D Airway Deformation-Scale Co-enhancement Module) collaboratively models the 3D irregular deformation and cross-scale diffusion characteristics of PB lesions through deformable convolution branches and multi-scale sensing branches, and integrates spatial attention and channel attention mechanisms. The calculation of the deformable convolution branch in the 3D-ASC module satisfies the following formula: ； In the formula, This is the current center position of the convolution kernel. The offset of the convolution kernel sampling points relative to the center. A learnable dynamic offset. The weights corresponding to the sampling points. The input feature map contains the pixel values ​​at the corresponding locations. This is the convolution output value at the current position; The classification and decision module is used to output the three-category classification results of plastic bronchitis lesion types and bronchoscopy examination decision suggestions.

2. The multimodal intelligent identification system for childhood plastic bronchitis integrating a three-dimensional airway deformation-scale co-enhancement module as described in claim 1, characterized in that, The standardized processing flow of the data acquisition and preprocessing module is as follows: CT images undergo sequence sorting, 3D reconstruction, resampling, size standardization, and intensity normalization; clinical data undergoes encoding, missing value imputation, and standardization to finally obtain standardized 3D image tensors and clinical feature vectors; the intensity normalization of CT images uses the following formula: In the formula, These are the pixel values ​​of the original CT image. The average value of the pixels. The standard deviation of pixel values. These are the normalized pixel values; the Z-score standardization for clinical data uses the following formula: In the formula, These are the original clinical characteristic values. The mean of this feature. The standard deviation of this feature. These are the standardized clinical characteristic values.

3. The multimodal intelligent identification system for childhood plastic bronchitis integrating a three-dimensional airway deformation-scale co-enhancement module as described in claim 1, characterized in that, The specific operation flow of the multimodal feature encoding module is as follows: the hierarchical spatial structure features of CT images are extracted by using the 3DDenseNet backbone network as a branch of the three-dimensional convolutional network; the high-dimensional embedding representation of clinical data is extracted by using the multilayer perceptron as a branch of the multilayer perceptron; the 3DDenseNet backbone network is composed of multiple dense blocks and transition layers alternately. Each dense block achieves the fusion and reuse of multi-scale features through feature concatenation. The transition layer uses 1×1×1 convolution to perform channel dimensionality reduction and feature smoothing.

4. The multimodal intelligent identification system for childhood plastic bronchitis integrating a three-dimensional airway deformation-scale co-enhancement module as described in claim 1, characterized in that, The specific operation procedure of the cross-modal attention module is as follows: S1: Global average pooling is performed on the early-fused image features to obtain the global image vector: ； S2: Combine the global image vector with the original clinical feature vector By splicing together, a cross-modal joint quantity is formed: ； S3: Input the cross-modal joint vector into the multilayer perceptron to generate channel-level attention weights. This weight is used to adaptively weight the fused features, thereby adjusting the response intensity of each channel of the image features: ； S4: Apply these weights to each channel of the fused feature map to achieve adaptive feature enhancement based on clinical context: ； In the formula, For global image vectors, This is a global average pooling operation. The number of feature channels, This represents a vector in a C-dimensional real space. This is the original clinical feature vector. For cross-modal joint vectors, and For the learnable weights of the multilayer perceptron, For activation function, It is the Sigmoid activation function. The generated channel-level attention weight vector, This represents the feature tensor of the early-stage fused 3D image. This is the attention-weighted enhanced feature tensor.

5. The multimodal intelligent identification system for childhood plastic bronchitis integrating a three-dimensional airway deformation-scale co-enhancement module according to claim 1, characterized in that, The specific operation process of the three-dimensional airway deformation-scale co-enhancement module 3D-ASC is as follows: First, the three-dimensional deformation features and cross-scale diffusion features of the lesion are extracted by deformable convolution branches and multi-scale perception branches respectively. Then, the outputs of the two branches are weighted and fused. Finally, the fused features are enhanced by spatial attention and channel attention mechanisms to obtain a highly recognizable lesion feature representation. The three-dimensional airway deformation-scale co-enhancement module 3D-ASC is inserted layer by layer into the middle and high-level dense blocks of the 3DDenseNet backbone network. Specifically, it is inserted at DenseBlock2, DenseBlock3, and DenseBlock4. The insertion method is to not change the topology of the backbone network and directly use the module output as the output feature of the corresponding dense block to achieve airway morphology enhancement from shallow to deep.

6. The multimodal intelligent identification system for childhood plastic bronchitis integrating a three-dimensional airway deformation-scale co-enhancement module as described in claim 5, is characterized in that, The feature fusion process of the multi-scale perception branch in the 3D-ASC three-dimensional airway deformation-scale co-enhancement module satisfies the following formula: ； In the formula, , , The results are shown for three-dimensional average pooling operations with step sizes of 1, 2, and 4, respectively. This represents a 1×1×1 convolution operation. This indicates a concatenation operation along the channel dimension. This is the output feature tensor of the multi-scale sensing branch.

7. The multimodal intelligent identification system for childhood plastic bronchitis integrating a three-dimensional airway deformation-scale co-enhancement module as described in claim 5, characterized in that, The weighted fusion of the dual-branch outputs in the 3D-ASC three-dimensional airway deformation-scale co-enhancement module satisfies the following formula: ； In the formula, The output feature tensor of the deformable convolution branch, The output feature tensor of the multi-scale sensing branch and These are learnable weighting coefficients. This is the feature tensor after the two branches are merged.

8. The multimodal intelligent identification system for childhood plastic bronchitis integrating a three-dimensional airway deformation-scale co-enhancement module as described in claim 5, characterized in that, The spatial attention weight in the 3D-ASC three-dimensional airway deformation-scale co-enhancement module is calculated according to the following formula: ； In the formula, The feature tensor after the two branches are fused. and These are respectively three-dimensional max pooling and average pooling operations. This represents a 7×7×7 convolution operation. This indicates a concatenation operation along the channel dimension. It is the Sigmoid activation function. This is the generated spatial attention weight map.

9. The multimodal intelligent identification system for childhood plastic bronchitis integrating a three-dimensional airway deformation-scale co-enhancement module as described in claim 5, characterized in that, The calculation of channel attention weights in the 3D-ASC three-dimensional airway deformation-scale co-enhancement module satisfies the following formula: ； In the formula, The feature tensor after the two branches are fused. This indicates a global average pooling operation. and For learnable weights, For activation function, It is the Sigmoid activation function. This is the generated channel attention weight vector.

10. The multimodal intelligent identification system for childhood plastic bronchitis integrating a three-dimensional airway deformation-scale co-enhancement module according to claim 1, characterized in that, The specific operation flow of the classification and decision module is as follows: global pooling and fully connected operations are performed on the enhanced features output by the 3D-ASC module for three-dimensional airway deformation-scale co-enhancement to output a three-class probability distribution of plastic bronchitis lesion types; bronchoscopy decision suggestions are generated according to preset rules, which are: when the predicted lesion type is mucus plug or plastic, bronchoscopy is recommended; when the predicted lesion type is solid or no lesion, bronchoscopy is not recommended. The classification probability threshold is set to 0.

5. When the highest category probability is lower than 0.5, a prompt message indicating that further clinical evaluation is required is output.

Images