Chronic sinusitis CT image automatic typing and severity assessment method and system based on deep learning

Abstract

The invention discloses a chronic sinusitis CT (Computed Tomography) image automatic typing and severity assessment method and a chronic sinusitis CT image automatic typing and severity assessment system based on deep learning. The method comprises the following steps: acquiring paranasal sinus CT image data of a chronic sinusitis patient and an associated clinical tag; carrying out standardization processing on the CT image, wherein the standardization processing comprises gray level normalization, denoising and region-of-interest segmentation; a multi-task deep learning model is constructed, an improved 3D Re < s> Ne < t-50 > architecture is adopted by a typing sub-model to output an immune endotype classification result, and a severity evaluation sub-model segments a lesion area based on a U-Ne < t + + > structure and automatically calculates a severity level; integrating model output to generate individualized treatment suggestions; and superposing a labeling result through the thermodynamic diagram and embedding the labeling result into a hospital PACS system. Accurate typing and evaluation of chronic nasosinusitis are achieved, and diagnosis efficiency and clinical decision support capability are improved.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of medical artificial intelligence, and particularly relates to a deep learning-based chronic rhinosinusitis CT image automatic typing and severity evaluation method and system, which is suitable for clinical auxiliary decision-making in otolaryngology. BACKGROUND

[0002] Chronic rhinosinusitis (CRS) is a common chronic inflammatory disease of the upper respiratory tract, with clinical manifestations of nasal congestion, rhinorrhea, facial pain, and hyposmia, which seriously affects the quality of life of patients. At present, the diagnosis and typing of CRS mainly rely on nasal sinus CT images and clinical evaluation, such as Lund-Mackay score and SNOT-22 scale. However, the traditional method has the following problems:

[0003] Strong subjectivity: image evaluation depends on the experience of doctors, and there may be differences in interpretation between different observers, affecting the accuracy of typing and severity evaluation. Low efficiency: manual labeling of the nasal sinus lesion area is time-consuming and difficult to achieve accurate quantitative analysis. Limited typing: the immune type of CRS (such as Th2 type and non-Th2 type) is crucial for treatment strategy selection, but traditional imaging methods are difficult to directly correlate with immunohistochemical results, resulting in inaccurate typing. Lack of personalized treatment decision-making: the existing evaluation system is difficult to integrate multi-modal data (such as images, clinical scores, and immune typing) to provide individualized treatment recommendations.

[0004] In recent years, deep learning has shown great potential in medical image analysis, such as U-Net for lesion segmentation and ResNet for classification tasks. However, existing methods mainly focus on single tasks (such as classification or segmentation), lack of integration of CRS multi-dimensional evaluation (typing + severity), and insufficient model interpretability, which limits clinical applicability. Therefore, there is an urgent need for an automated, high-precision multi-task deep learning system to achieve accurate typing, severity evaluation, and personalized treatment recommendation for CRS. SUMMARY

[0005] The application proposes a deep learning-based chronic rhinosinusitis CT image automatic typing and severity evaluation method and system, which integrates an improved 3DResNet-50 typing sub-model and a U-Net++ evaluation sub-model through the construction of a multi-task deep learning model, innovatively introduces a space-channel attention mechanism and a dynamic threshold segmentation optimization module, and develops a clinical decision support system to realize full-process automation from image analysis to treatment recommendation, improve diagnostic accuracy, and improve analysis efficiency.

[0006] In a first aspect, the embodiments of the present application provide a deep learning-based chronic rhinosinusitis CT image automatic typing and severity evaluation method, which comprises:

[0007] S1, obtaining CT image data of a patient with chronic rhinosinusitis and associating clinical labels, the clinical labels including immune phenotype classification, Lund-Mackay score and SNOT-22 scale results, wherein the immune phenotype classification is verified by immunohistochemical staining results of postoperative tissue specimens;

[0008] S2, standardizing the CT image, including gray scale normalization, denoising, and region of interest segmentation, and extractable analysis, extracting the sinus and polyp regions;

[0009] S3, constructing a multi-task deep learning model, wherein the multi-task deep learning model includes a classification sub-model and a severity assessment sub-model, the classification sub-model uses an improved 3DResNet-50 architecture, inputs the pre-processed CT image, and outputs the immune phenotype classification result; the severity assessment sub-model segments the sinus lesion region based on a U-Net++ structure, and automatically calculates the severity grade in combination with the Lund-Mackay scoring rule;

[0010] S4, integrating the outputs of the classification sub-model and the severity assessment sub-model to generate individualized treatment recommendations;

[0011] S5, labeling the classification results, severity scores and treatment recommendations on the original CT image in the form of a heat map overlay, supporting DICOM format export, and embedding the hospital PACS system for interactive review.

[0012] Optionally, in an implementation form of the first aspect of the application, the standardization processing of the CT image includes gray scale normalization, denoising, and region of interest segmentation, and extractable analysis, including the following sub-steps:

[0013] Map the CT value to the standard Hounsfield unit and unify the resolution;

[0014] Adopt a non-local mean filtering algorithm to eliminate image noise;

[0015] Segment the sinus cavity and polyp region based on the U-Net network, and the segmentation result is reviewed and corrected by a radiologist;

[0016] Visualize the model attention area by gradient weighted class activation mapping (Grad-CAM), generate a heat map, and perform spatial overlap evaluation with the gold standard labeled area.

[0017] Optionally, in an implementation form of the first aspect of the application, the specific implementation of the classification sub-model using the improved 3DResNet-50 architecture includes the following steps:

[0018] Input processing module: Includes 3D convolutional layers, 3D batch normalization layers, and ReLU activation function, used to process 64×64×64 voxel CT image blocks;

[0019] Spatial-channel attention module: Embedded before each residual block, it includes: Spatial attention branch: Generates spatial weight map using 3D convolutional kernel; Channel attention branch: Generates channel weight vector through 3D global average pooling and fully connected layers.

[0020] Optimized 3D residual structure: includes: basic residual block: composed of 1×1×1, 3×3×3, and 1×1×1 3D convolutional layers; bottleneck residual block: introduces dilated convolution and adds Dropout layers;

[0021] The classification output module includes a global average pooling layer, a fully connected layer, a LayerNorm layer, a GeLU activation function, and a final classification layer.

[0022] Optionally, in one implementation of the first aspect of the present invention, the training method of the fractal sub-model includes:

[0023] The transfer learning technique is adopted, and the model parameters are initialized based on the pre-trained 3D ResNet-50 weights. The network input layer is adapted to the CT image data specifications by expanding the convolution kernel.

[0024] During the training process, an adversarial sample enhancement mechanism is introduced. Perturbation images are generated through a gradient attack algorithm, and the original images and adversarial samples are mixed and trained according to a preset ratio.

[0025] A composite loss function is adopted, including: label smoothing cross-entropy loss, which sets a smoothing factor to adjust the classification boundary; and a regularization term to control model complexity and prevent overfitting.

[0026] The AdamW optimizer is used for parameter updates, and a dynamic learning rate scheduling strategy is configured, where: the optimizer sets the initial learning rate, weight decay coefficient, and momentum parameter; gradient pruning technique is used to limit the gradient update magnitude; and a periodic learning rate adjustment strategy is implemented to make the learning rate change according to a preset pattern during training.

[0027] The model's decision-making basis is verified regularly through visualization interpretation technology. When the spatial overlap between the model's area of ​​interest and the standard labeled area is lower than a preset threshold, the model parameter adjustment mechanism is automatically triggered.

[0028] The loss function for the fractal sub-model is:

[0029]

[0030] Among them, L LSThe label smoothing cross-entropy loss is represented by ε, where ε is the smoothing factor, F and f represent the number of categories and their corresponding indices, respectively, and y f This represents the one-hot encoding of the actual label.

[0031] Optionally, in one implementation of the first aspect of the present invention, the severity assessment sub-model adopts an improved U-Net++ architecture, and the specific implementation process includes:

[0032] A multi-level encoder-decoder structure is set up. The encoder consists of multiple downsampling stages, each of which includes 3D convolution operations and residual connections. The decoder integrates multi-scale features through dense skip connections and uses transposed convolution to achieve upsampling. A deep supervision mechanism is introduced in each decoding stage to output intermediate segmentation results.

[0033] An attention mechanism is set at the connection between the encoder and decoder to dynamically adjust the feature weights by analyzing the spatial and channel correlation of the feature maps, thereby achieving adaptive feature enhancement.

[0034] Through a multi-task output layer, a lesion probability map of each sinus anatomical region is generated using three-dimensional convolution. Based on clinical scoring criteria, the probability map is converted into a lesion severity score for each region. The scores of each region are then combined to obtain the overall severity assessment result.

[0035] A composite loss function combining region overlap and classification error is used, along with an adaptive optimization algorithm and a dynamic learning rate adjustment strategy. Cross-validation is implemented to ensure model robustness.

[0036] The formula for the composite loss function of the severity assessment sub-model is as follows:

[0037]

[0038] Where K represents the number of sinus anatomical regions, k represents the region index, and i represents the position index in three-dimensional voxel space. This represents the predicted probability value of the k-th anatomical region at voxel i. The corresponding true label indicates whether the voxel belongs to the k-th partition, α represents the balancing weight, ε is the smoothing term, N represents the total number of samples, and S n This represents the predicted overall score for the nth sample. Let λ represent the actual score, λ represent the multi-task weights, and ∑ i This indicates that the values ​​at all voxel locations are summed to calculate the global similarity or error.

[0039] Optionally, in one implementation of the first aspect of the present invention, the severity assessment sub-model further includes a dynamic threshold segmentation optimization module, the implementation method of which includes:

[0040] A spatial location coding matrix is ​​established based on the three-dimensional anatomical atlas of the sinuses, and CT images are divided into multiple anatomical functional areas. Different initial segmentation thresholds are set for different functional areas.

[0041] Based on the local features of the image, the tissue density distribution is dynamically calculated, and combined with the spatial continuity constraints of adjacent slices, the segmentation threshold boundaries of each functional area are optimized in real time.

[0042] By integrating grayscale features, texture features, and morphological features, a multi-dimensional discrimination criterion for polyps and mucosal hyperplasia is established, and the final segmentation boundary is optimized through a cascaded classifier.

[0043] Three-dimensional connected component analysis is applied to eliminate isolated noise points, morphological closing operations are used to fill small voids, and abnormal segmentation regions are corrected based on anatomical constraints.

[0044] Optionally, in one implementation of the first aspect of the present invention, the specific implementation of generating individualized treatment recommendations in step S4 includes:

[0045] A multidimensional feature space mapping relationship between the typing results and the severity score was established. The immune endogenous type classification and Lund-Mackay score were associated and matched through a knowledge graph. A rule-based reasoning engine was used to generate preliminary treatment suggestions.

[0046] By integrating treatment guidelines from evidence-based medicine, a decision tree is constructed that includes drug therapy, surgical indications, and physical therapy, and the weights of treatment plans are dynamically adjusted according to patient characteristics.

[0047] Based on the patient's historical treatment records and complication history, the cost-effectiveness of each treatment option is evaluated, and a list of individualized treatment recommendations is generated in order of priority.

[0048] Generates explanations of recommendations with clinical evidence, visualizes key decision-making factors, and provides comparative analysis of alternative options;

[0049] We continuously learn from newly published treatment guidelines, optimize recommendation algorithms based on treatment feedback, and regularly update treatment plans in the knowledge base.

[0050] Optionally, in one implementation of the first aspect of the present invention, the visualization output system of S5 includes:

[0051] Intelligent annotation engine: It uses multi-layer fusion technology to overlay the classification results with the original CT images, uses a differentiated color coding scheme to distinguish different immune intratype regions, and uses transparency adjustment to achieve a contrast display between lesion areas and normal tissues;

[0052] Interactive visualization module: Supports multi-plane reconstruction MPR display according to DICOM standard, provides three-dimensional volume rendering function of lesion area, and realizes dynamic annotation and prompts of key scoring indicators;

[0053] PACS Integration Interface: A network communication protocol compliant with the DICOM standard, supporting bidirectional data interaction with the hospital's PACS system to achieve automatic matching and association of patient information;

[0054] Clinical report generator: Automatically generates structured assessment reports, including screenshots and annotations of key imaging features, and outputs PDF documents that conform to clinical standards;

[0055] Mobile adaptation components: Responsive design adapts to different terminal devices, supports touch gesture operation and zooming, and provides offline viewing and annotation functions.

[0056] Secondly, embodiments of this application provide a deep learning-based automatic classification and severity assessment system for CT images of chronic sinusitis, applied to the deep learning-based automatic classification and severity assessment method for CT images of chronic sinusitis as described in the first aspect. The system includes:

[0057] Data acquisition module: used to acquire sinus CT image data of patients with chronic sinusitis and associate them with clinical labels, which include immune endotype classification, Lund-Mackay score and SNOT-22 scale results, wherein the immune endotype classification is verified by immunohistochemical staining results of postoperative tissue specimens;

[0058] Preprocessing module: Standardizes CT images, including grayscale normalization, noise reduction and region of interest segmentation, interpretability analysis, and extraction of sinus and polyp regions;

[0059] Model building module: Constructs a multi-task deep learning model, which includes a sub-model for genotyping and a sub-model for severity assessment. The sub-model for genotyping adopts an improved 3DResNet-50 architecture, takes preprocessed CT images as input, and outputs immune intratype classification results. The sub-model for severity assessment is based on the U-Net++ structure to segment sinus lesion areas and automatically calculates the severity level by combining the Lund-Mackay scoring rule.

[0060] Fusion Decision Module: Integrates the outputs of the subtyping submodel and the severity assessment submodel to generate individualized treatment recommendations;

[0061] Visualization output module: Severity scores and treatment recommendations are annotated on the original CT images in the form of heatmap overlays, supporting DICOM format export and embedding into the hospital PACS system for interactive viewing.

[0062] Thirdly, embodiments of this application provide an electronic device, including:

[0063] processor;

[0064] Memory used to store processor-executable instructions;

[0065] The processor is configured to implement the deep learning-based automatic classification and severity assessment method for CT images of chronic sinusitis as described in the first aspect when executing the instructions.

[0066] Fourthly, embodiments of this application provide a computer-readable storage medium storing a program that instructs a device to execute the deep learning-based automatic classification and severity assessment method for CT images of chronic sinusitis as described in the first aspect.

[0067] This invention discloses a method and system for automatic classification and severity assessment of CT images of chronic sinusitis based on deep learning. The method includes: acquiring sinus CT image data of patients with chronic sinusitis and associated clinical labels (immune intratype classification, Lund-Mackay score, and SNOT-22 scale results); standardizing the CT images, including grayscale normalization, denoising, and region-of-interest segmentation; constructing a multi-task deep learning model, wherein the classification sub-model uses an improved 3DResNet-50 architecture to output immune intratype classification results, and the severity assessment sub-model segments the lesion region based on the U-Net++ structure and automatically calculates the severity level; integrating the model output to generate individualized treatment suggestions; and overlaying the results with heatmaps and embedding them into a hospital PACS system. The system includes modules for data acquisition, preprocessing, model construction, fusion decision-making, and visualization output. This invention achieves accurate classification and assessment of chronic sinusitis, improving diagnostic efficiency and clinical decision support capabilities.

[0068] Beneficial effects:

[0069] 1. Improve diagnostic accuracy and consistency. An improved 3DResNet-50 submodel and a U-Net++ severity assessment submodel are employed, combined with multi-task learning, to achieve precise immune intratype classification and lesion region segmentation for chronic sinusitis, reducing subjective differences in human interpretation and improving the objectivity and repeatability of diagnosis.

[0070] 2. Enhance model interpretability. By visualizing the model's regions of interest through Gradient Weighted Class Activation Mapping (Grad-CAM), a heatmap is generated and its spatial overlap with the gold standard labeled regions is evaluated. This allows clinicians to intuitively understand the model's decision-making basis and improves the credibility of the results.

[0071] 3. Automation and Efficiency. The automated processing workflow based on deep learning can quickly complete standardized preprocessing, lesion segmentation, and scoring of CT images, significantly shortening diagnostic time and improving clinical work efficiency.

[0072] 4. Personalized Treatment Recommendations. Integrating multi-dimensional data such as immune intratypic classification, Lund-Mackay score, and SNOT-22 scale, personalized treatment suggestions are generated through knowledge graphs and rule reasoning engines to optimize clinical decision-making and improve treatment outcomes.

[0073] 5. Highly clinically applicable. Supports DICOM format export and can be embedded into hospital PACS systems, achieving seamless integration with existing medical information systems. The interactive visualization module provides multiplanar reconstruction (MPR) and 3D rendering functions, facilitating doctors' observation of lesions from multiple angles.

[0074] 6. Model robustness and generalization ability. Transfer learning, adversarial example enhancement, and dynamic learning rate adjustment techniques are employed to optimize model training, enhancing its adaptability to CT data from different sources and ensuring stable performance in real-world clinical environments.

[0075] 7. Continuous optimization capability. The system supports continuous learning of new treatment guidelines and feedback on treatment effects, and can dynamically update the knowledge base to ensure that recommended treatments always conform to the latest medical advancements. Attached Figure Description

[0076] Figure 1 This is a schematic diagram of a deep learning-based method for automatic classification and severity assessment of CT images of chronic sinusitis, provided as an embodiment of this application.

[0077] Figure 2 This is a structural diagram of a sub-model provided in an embodiment of this application.

[0078] Figure 3 This is a diagram illustrating the architecture of a severity assessment sub-model provided in one embodiment of this application.

[0079] Figure 4 A system architecture diagram for automatic classification and severity assessment of CT images of chronic sinusitis based on deep learning, provided in one embodiment of this application.

[0080] Figure 5 A schematic diagram of an electronic device provided in an embodiment of this application. Detailed Implementation

[0081] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them.

[0082] It should be noted that, in the embodiments of this application, "at least one" refers to one or more, and "more than one" refers to two or more. Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used in the specification of this application is for the purpose of describing particular embodiments only and is not intended to be limiting of this application.

[0083] It should be noted that in the embodiments of this application, the terms "first," "second," etc., are used only for descriptive purposes and should not be construed as indicating or implying relative importance, nor as indicating or implying order. Features specified as "first" or "second" may explicitly or implicitly include one or more of the stated features. In the description of the embodiments of this application, words such as "exemplary" or "for example" are used to indicate examples, illustrations, or explanations. Any embodiment or design scheme described as "exemplary" or "for example" in the embodiments of this application should not be construed as being more preferred or advantageous than other embodiments or design schemes. Specifically, the use of words such as "exemplary" or "for example" is intended to present the relevant concepts in a concrete manner.

[0084] Based on the embodiments described in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0085] Example 1

[0086] Figure 1 This is a schematic diagram of a deep learning-based method for automatic classification and severity assessment of CT images for chronic sinusitis, provided as an embodiment of this application. Figure 1 As shown, a method for automatic classification and severity assessment of CT images of chronic sinusitis based on deep learning includes:

[0087] S1. Obtain sinus CT imaging data from patients with chronic sinusitis and associate them with clinical labels, including immune endotype classification, Lund-Mackay score, and SNOT-22 scale results. The immune endotype classification is validated by immunohistochemical staining results of postoperative tissue specimens. The Lund-Mackay score is an important method for assessing the severity of chronic sinusitis and is used in clinical and imaging evaluations. The SNOT-22 scale is used to assess patients' quality of life and is one of the important tools for assessing chronic sinusitis.

[0088] S2. Standardize the CT images, including grayscale normalization, noise reduction and region of interest segmentation, interpretability analysis, and extract sinus and polyp regions.

[0089] Specifically, in this embodiment, the standardization processing of CT images includes grayscale normalization, noise reduction and region of interest segmentation, and interpretability analysis, comprising the following sub-steps:

[0090] CT values ​​are mapped to standard Hounsfield units (HUs) and the resolution is standardized. Hounsfield units (HUs) are units that measure the absorption of X-rays by tissues and are widely used in medical image processing. Normalization (such as grayscale standardization) can reduce the impact of differences in scanning equipment or parameters.

[0091] Nonlocal mean filtering is employed to eliminate image noise. In existing technologies, Gaussian filtering and median filtering are commonly used denoising methods to reduce random noise in images, while nonlocal mean filtering is an advanced denoising method that can effectively preserve image details while removing noise.

[0092] The U-Net network was used to automatically segment the sinus cavity and polyp regions. The segmentation results were reviewed and corrected by a radiologist. The U-Net network is an architecture for medical image segmentation, particularly suitable for medical image segmentation tasks. Radiologists reviewed the segmentation results to ensure the accuracy of the model's output.

[0093] Gradient-weighted class activation mapping (Grad-CAM) is used to visualize the model's regions of interest, generating heatmaps and evaluating their spatial overlap with the gold-standard labeled regions. Grad-CAM helps interpret model decisions by visualizing these regions and aids in interpreting model decisions. Model performance is evaluated based on the spatial overlap between the heatmaps and the gold-standard labeled regions.

[0094] S3. Construct a multi-task deep learning model, wherein the multi-task deep learning model includes a sub-model for genotyping and a sub-model for severity assessment. The sub-model for genotyping adopts an improved 3DResNet-50 architecture, takes preprocessed CT images as input, and outputs immune intratype classification results. The sub-model for severity assessment is based on the U-Net++ structure to segment sinus lesion areas and automatically calculates the severity level by combining the Lund-Mackay scoring rule.

[0095] ResNet-50 is a deep convolutional neural network that addresses the vanishing and exploding gradient problems in deep network training through residual modules, making the network deeper and easier to train. ResNet-50 can be used as a pre-trained model and applied to other tasks, such as image classification and semantic segmentation, through transfer learning. U-Net is a fully convolutional network widely used in medical image segmentation, capable of handling pixel-level segmentation tasks. U-Net++ is an improved version of U-Net, used to improve segmentation accuracy or handle more complex structures. Multi-task models include two sub-tasks: classification and severity assessment. Multi-task learning in medical image analysis handles multiple tasks simultaneously, improving the model's generalization ability. For example, multi-task learning in medical image analysis can perform disease classification and severity assessment simultaneously.

[0096] This method combines deep learning models (such as ResNet-50 and U-Net++) with multi-task learning for classification and severity assessment in medical image analysis, demonstrating high feasibility and clinical value.

[0097] Figure 2 This is a structural diagram of a sub-model provided in one embodiment of this application. Specifically, as shown... Figure 2 As shown, in this embodiment, the specific implementation of the fractal sub-model using the improved 3DResNet-50 architecture includes the following steps:

[0098] The specific implementation of the sub-model using the improved 3DResNet-50 architecture includes the following steps:

[0099] Input Processing Module: Contains 3D convolutional layers, 3D batch normalization layers, and a ReLU activation function to process 64×64×64 voxel CT image blocks. It receives preprocessed 3D CT image blocks (64×64×64 voxels), first performing preliminary feature extraction through a 3D convolutional layer (kernel size = 7×7×7, stride = 2×2×2, padding = 3×3×3), followed by a 3D batch normalization layer (BatchNorm3d) and the ReLU activation function. Specifically, the 3D convolutional layer processes 3D data (such as CT images) and extracts spatial features. The 3D batch normalization layer accelerates training and improves model stability. The ReLU activation function introduces non-linearity, enhancing the model's expressive power.

[0100] The Spatial-Channel Attention Module (SCSE) is embedded before each residual block and includes: a Spatial Attention Branch, which generates a spatial weight map using 3D convolutional kernels; and a Channel Attention Branch, which generates channel weight vectors through 3D global average pooling and fully connected layers. Specifically, the Spatial Attention Branch generates a spatial weight map using 3D convolutional kernels, enhancing the model's attention to spatial features. The Channel Attention Branch generates channel weight vectors through 3D global average pooling and fully connected layers, enhancing the model's ability to perceive channel features. The SCSE module is embedded before each residual block in the ResNet-50. The Spatial Attention Branch uses 3D convolutional kernels (1×1×1) to compress the channel dimension and generates a spatial weight map using the Sigmoid function. The Channel Attention Branch obtains channel statistics through 3D global average pooling, generates channel weight vectors through two fully connected layers (with the intermediate dimension reduced to C / 16, where C is the number of input channels), and the Sigmoid function. Finally, the feature map is dynamically calibrated through element-wise multiplication of the spatial weight map and the channel weight vectors.

[0101] Optimized 3D residual structure: Includes: Basic residual block: composed of 1×1×1, 3×3×3, and 1×1×1 3D convolutional layers; Bottleneck residual block: introduces dilated convolutions and adds Dropout layers. The basic residual block, composed of 1×1×1, 3×3×3, and 1×1×1 3D convolutional layers, is used to extract multi-scale features. The bottleneck residual block introduces dilated convolutions and Dropout layers to enhance the model's generalization ability and reduce overfitting. Residual structures in deep learning, such as the ResNet series models. Residual structure optimization: The 2D convolutions of the original ResNet-50 are replaced with 3D convolutions, and a Dropout layer (p=0.3) is added after the third convolutional layer in each residual block. The basic residual block consists of three 3D convolutional layers (kernel sizes of 1×1×1, 3×3×3, and 1×1×1, respectively); dilated convolutions (dilation=2) are introduced in the bottleneck residual block to expand the receptive field.

[0102] The classification output module includes a global average pooling layer, a fully connected layer, a LayerNorm layer, a GeLU activation function, and a final classification layer. The global average pooling layer extracts global features. The fully connected layer is used for classification. The LayerNorm layer normalizes the feature representation. The GeLU activation function enhances the model's non-linear expressive power. The classification head design follows the global average pooling layer with a fully connected layer (512-dimensional → 256-dimensional), followed by a LayerNorm layer and GeLU activation; the final classification layer (256-dimensional → 3-dimensional) outputs the predicted probability distribution for Type-1 / 2 / 3.

[0103] Specifically, in this embodiment, the training method for the fractal sub-model includes:

[0104] This study employs transfer learning techniques, initializing model parameters based on pre-trained 3DResNet-50 weights, and adapting the network input layer to CT image data specifications through convolutional kernel expansion. Transfer learning in deep learning, especially in image classification tasks, can significantly improve model performance through pre-trained models (such as ResNet-50). ResNet-50 is a deep neural network architecture that is efficient and scalable in image classification tasks.

[0105] Adversarial example enhancement is introduced during training. Perturbation images are generated using a gradient attack algorithm, and the original images are mixed with adversarial examples at a preset ratio during training. Adversarial examples are a technique used to enhance model robustness; generating adversarial examples can improve the model's robustness to perturbations. Adversarial training is a common method to enhance model robustness; by introducing adversarial examples during training, the model's generalization ability can be improved.

[0106] A composite loss function is employed, including: label smoothing cross-entropy loss, which sets a smoothing factor to adjust the classification boundary; and a regularization term to control model complexity and prevent overfitting. Label smoothing can reduce the model's overfitting to the training data and improve its generalization ability. Regularization terms (such as L2 regularization) can control model complexity and prevent overfitting.

[0107] The AdamW optimizer is used for parameter updates, configuring a dynamic learning rate scheduling strategy. This strategy involves: setting the initial learning rate, weight decay coefficient, and momentum parameter; employing gradient pruning to limit the gradient update magnitude; and implementing a periodic learning rate adjustment strategy to ensure the learning rate changes according to a pre-defined pattern during training. AdamW is an optimizer that combines the Adam optimizer and weight decay for large-scale deep learning tasks. Dynamic learning rate scheduling (such as cosine annealing) can improve the model's convergence speed and performance.

[0108] Regularly validate the model's decision-making basis using visualization techniques. When the spatial overlap between the model's region of interest and the standard labeled region falls below a preset threshold, an automatic model parameter adjustment mechanism is triggered. Visualization (such as Grad-CAM) can explain the model's decision-making process, improving its interpretability. Model validation and adjustment are crucial steps in ensuring model performance, continuously optimizing the model through feedback mechanisms.

[0109] Loss function and optimizer configuration: The smoothing factor ε = 0.1 is set in the label smoothing cross-entropy loss function; the hyperparameters of the AdamW optimizer are set as follows: learning rate lr = 3e-4, weight decay weight_decay = 1e-2, betas = (0.9, 0.999); a cosine annealing learning rate scheduler is used, with an initial learning rate of 3e-4, a minimum learning rate of 1e-5, and a period T_max = 50.

[0110] For traditional cross-entropy loss, given: the model output logits vector z∈R 3 (3-class classification task), the one-hot encoding of the real label is y = [y1, y2, y3], where y f =1 (true class), the rest are 0. The predicted probability distribution p = softmax(z), that is: The standard cross-entropy loss is:

[0111] With a smoothing factor set to 0.1, the true label distribution is adjusted as follows:

[0112]

[0113] The specific calculation (assuming the true class is f=1) yields the smoothed loss function as follows:

[0114]

[0115] Expanded form (taking the real class f=1 as an example):

[0116] L=-(0.933logp1+0.033logp2+0.033p3),

[0117] Finally, the loss function for the fractal sub-model is:

[0118]

[0119] Among them, L LS The label smoothing cross-entropy loss is represented by ε, where ε is the smoothing factor, F and f represent the number of categories and their corresponding indices, respectively, and y f This represents the one-hot encoding of the actual label.

[0120] Traditional cross-entropy loss forces the model to predict a probability close to 1 for the correct label and close to 0 for other labels. This can lead to: overfitting: the model becomes overconfident in the training data, resulting in decreased generalization ability; and sensitivity to noisy labels: if the training data contains incorrect labels (noise), the model may struggle to correct them.

[0121] Label smoothing introduces a smoothing factor ε, which makes the target probability of the correct label 1-ε+ε / k (where k is the number of classes) and the target probability of the incorrect label ε / k (instead of 0). This prevents the model from over-relying on extreme probabilities and improves generalization ability.

[0122] In cosine annealing learning rate scheduling, the parameters are set as follows: Initial learning rate 3e-4: consistent with AdamW's initial learning rate. Minimum learning rate 1e-5: the lower limit for the learning rate decrease, avoiding training stagnation due to an excessively small learning rate. It is typically set to 1 / 10 to 1 / 100 of the initial learning rate. Period T_max = 50: represents the half-cycle length of cosine annealing (unit: epochs). Within 50 epochs, the learning rate smoothly decreases from 3e-4 to 1e-5, then re-increases (periodic restart). This is suitable for training of moderate lengths (e.g., 50–200 epochs).

[0123] Specifically, in this embodiment, label smoothing affects optimization in the following ways: Preventing overconfidence: The target probability of the true class decreases from 1 to 0.933, forcing the model to retain uncertainty. Regularization effect: The target probability of the incorrect class increases from 0 to 0.033, avoiding complete suppression of negative samples by the gradient.

[0124] Figure 3 This is a diagram illustrating the architecture of a severity assessment sub-model provided in one embodiment of this application. Figure 3 As shown, the severity assessment sub-model adopts an improved U-Net++ architecture, and the specific implementation process includes:

[0125] A multi-level encoder-decoder structure is adopted. The encoder consists of multiple downsampling stages, each containing 3D convolution operations and residual connections. The decoder integrates multi-scale features through dense skip connections and uses transposed convolutions for upsampling. Deep supervision mechanisms are introduced at each decoding stage to output intermediate segmentation results. The encoder consists of multiple downsampling stages, each containing 3D convolution operations and residual connections. The decoder integrates multi-scale features through dense skip connections and uses transposed convolutions for upsampling. The U-Net++ architecture is an improved version of U-Net, improving segmentation performance by introducing dense skip connections and deep supervision mechanisms. Dense skip connections enable the model to perform more thorough information fusion between feature maps at different levels, enhancing the expressive power of features. Deep supervision mechanisms, by outputting segmentation results at multiple stages of the decoder and supervising the training of these results, enable the model to better learn feature representations at different levels.

[0126] An attention mechanism is implemented at the encoder-decoder junction. By analyzing the spatial and channel correlations of the feature maps, feature weights are dynamically adjusted to achieve adaptive feature enhancement. This attention mechanism improves the model's focus on important regions. Specifically, the spatial attention module and the channel attention module enhance feature channels related to lesions, thereby improving segmentation accuracy.

[0127] Through a multi-task output layer, 3D convolution is used to generate lesion probability maps for each sinus anatomical region. Based on clinical scoring criteria, the probability maps are converted into lesion severity scores for each region, and the overall severity assessment result is obtained by summing the scores of each region. Multi-task learning can handle multiple tasks simultaneously, improving the model's generalization ability. The application of clinical scoring criteria can improve the clinical applicability of the model, making its output results more in line with clinical needs.

[0128] A composite loss function combining region overlap and classification error is employed, along with an adaptive optimization algorithm and a dynamic learning rate adjustment strategy, and cross-validation is implemented to ensure model robustness. The composite loss function balances the model's training objectives and improves its generalization ability. The adaptive optimization algorithm and dynamic learning rate adjustment strategy enhance the model's convergence speed and performance.

[0129] Specifically, in this embodiment, the model needs to process multiple task outputs simultaneously (regional lesion probability map + overall score), therefore a composite loss is used:

[0130] L total =L seg +λL score ,

[0131] Among them, the segmentation loss L seg Combining region overlap (Dice) and classification error (cross-entropy) makes it suitable for multi-anatomical partition segmentation:

[0132]

[0133] Where K represents the number of anatomical regions of the sinuses (e.g., K=4 corresponds to the maxillary sinus, ethmoid sinus, etc.), P k G represents the predicted probability map of the k-th partition. k α represents the corresponding real segmentation mask, and α represents the balancing weight.

[0134] Dice coefficient formula:

[0135]

[0136] Here, ε is a smoothing term to prevent division by zero.

[0137] Rating loss L score The formula is:

[0138]

[0139] Among them, S n This represents the predicted overall score for the nth sample. λ represents the actual score (such as the Lund-Mackay score annotated by the clinician), and λ represents the multi-task weight (suggested initial value of 0.1).

[0140] Summary of loss function formulas:

[0141]

[0142] Where K represents the number of sinus anatomical regions, k represents the region index, and i represents the position index in three-dimensional voxel space. This represents the predicted probability value of the k-th anatomical region at voxel i. The corresponding true label indicates whether the voxel belongs to the k-th partition, α represents the balancing weight, ε is the smoothing term, N represents the total number of samples, and S n This represents the predicted overall score for the nth sample. Let λ represent the actual score, λ represent the multi-task weights, and ∑ i This indicates that the values ​​at all voxel locations are summed to calculate the global similarity or error.

[0143] The severity assessment sub-model also includes a dynamic threshold segmentation optimization module, the implementation of which includes:

[0144] A spatial location coding matrix is ​​established based on the three-dimensional anatomical atlas of the sinuses, and CT images are divided into multiple anatomical functional areas. Different initial segmentation thresholds are set for different functional areas.

[0145] Based on the local features of the image, the tissue density distribution is dynamically calculated, and combined with the spatial continuity constraints of adjacent slices, the segmentation threshold boundaries of each functional area are optimized in real time.

[0146] By integrating grayscale features, texture features, and morphological features, a multi-dimensional discrimination criterion for polyps and mucosal hyperplasia is established, and the final segmentation boundary is optimized through a cascaded classifier.

[0147] Three-dimensional connected component analysis is applied to eliminate isolated noise points, morphological closing operations are used to fill small voids, and abnormal segmentation regions are corrected based on anatomical constraints.

[0148] S4. Integrate the outputs of the subtyping submodel and the severity assessment submodel to generate individualized treatment recommendations.

[0149] The specific implementation of generating personalized treatment recommendations in step S4 includes: establishing a multidimensional feature space mapping relationship between subtyping results and severity scores; associating and matching immune intratyping classifications with Lund-Mackay scores through a knowledge graph; and generating preliminary treatment recommendations using a rule-based inference engine. Multidimensional feature space mapping: By mapping subtyping results (e.g., immune intratyping) with severity scores (e.g., Lund-Mackay scores), a multidimensional representation of patient characteristics is constructed, which helps the model understand the complexity of the patient's condition. Knowledge graph: By constructing a knowledge graph, clinical data of different dimensions (e.g., immune intratyping, Lund-Mackay scores, SNOT-22 scores, etc.) are associated to form a structured knowledge network, supporting more accurate reasoning and decision-making. Rule-based inference engine: Preliminary treatment recommendations are generated through preset rules (e.g., "If the immune intratyping is type A and the Lund-Mackay score ≥ 3, then drug treatment is recommended"). Evidence-based medicine guidelines: Treatment guidelines from authoritative medical guidelines (e.g., NICE, guidelines, expert consensus) are integrated to ensure the scientific validity and authority of the recommendations.

[0150] This study integrates treatment guidelines from evidence-based medicine to construct a decision tree encompassing drug therapy, surgical indications, and physical therapy, dynamically adjusting the weights of treatment options based on patient characteristics. Dynamic adjustment: The weights of treatment options are dynamically adjusted based on individual patient characteristics (such as age, complications, and drug tolerance) to achieve personalized treatment. The cost-effectiveness ratio of each treatment option is evaluated by combining the patient's historical treatment records and complication history, outputting a priority-ranked list of individualized treatment recommendations. Cost-effectiveness analysis: The costs (such as medication costs, hospital stay, and complication risks) of treatment options are quantified and weighed against benefits (such as symptom relief and improved quality of life). Prioritization: A priority-ranked list of treatment recommendations is generated based on individual patient characteristics (such as disease severity, economic status, and treatment preferences).

[0151] Generates recommendations with clinical evidence, visualizes key decision-making factors, and provides comparative analysis of alternative treatments. Clinical Evidence: Provides clear clinical evidence for each treatment recommendation (e.g., "Drug therapy is recommended based on Lund-Mackay score and immune endotype classification"). Visualization: Displays key decision-making and influencing factors through charts, heatmaps, decision trees, and other visualization methods to enhance transparency and interpretability. Comparison of Alternative Treatments: Provides a comparison of the advantages and disadvantages of different treatment options to help patients and physicians make more comprehensive decisions.

[0152] We continuously learn from newly published treatment guidelines, optimize our recommendation algorithm based on treatment feedback, and regularly update treatment plans in our knowledge base. Our continuous learning mechanism involves constantly updating treatment recommendations through machine learning models (such as reinforcement learning and online learning) to adapt to evolving medical knowledge. Our knowledge base is regularly updated with treatment guidelines, clinical evidence, and patient data to ensure the timeliness and accuracy of our recommendations.

[0153] Specifically, the outputs of the two sub-models are integrated to generate a structured report, including: if the classification result is Type-2 and the Lund-Mackay score is ≥12, anti-IL-4Rα monoclonal antibody combined with endoscopic nasal surgery is recommended; if the classification result is Type-1 / 3 and the Lund-Mackay score is ≤8, nasal irrigation combined with short-term glucocorticoids is recommended.

[0154] S5. The classification results, severity scores, and treatment recommendations are annotated on the original CT images in the form of a heatmap overlay. It supports export in DICOM format and can be embedded into the hospital's PACS system for interactive viewing.

[0155] Specifically, in this embodiment, the visualization output system of S5 includes: an intelligent annotation engine: employing multi-layer fusion technology to overlay the classification results with the original CT images, using a differentiated color coding scheme to distinguish different immunophenotypic regions, and adjusting transparency to achieve a contrast display between lesion areas and normal tissue. Multi-layer fusion technology and color coding are used to enhance image readability and information delivery. Transparency adjustment helps doctors more clearly observe the contrast between lesion areas and normal tissue, improving diagnostic accuracy.

[0156] Interactive visualization module: Supports DICOM standard multiplanar reconstruction (MPR) display, provides 3D volumetric rendering of lesion areas, and enables dynamic annotation and prompts for key scoring indicators. DICOM is a standard protocol in medical image processing used for the storage and transmission of medical images. Multiplanar reconstruction (MPR) and 3D volumetric rendering provide more comprehensive lesion information. Dynamic annotation and prompts help doctors quickly identify and assess lesion areas.

[0157] PACS Integration Interface: A DICOM-compliant network communication protocol that supports bidirectional data exchange with the hospital's PACS system, enabling automatic matching and association of patient information. The DICOM standard is a core standard in medical image processing, ensuring compatibility and interoperability between different systems. Bidirectional data exchange can improve the efficiency of hospital information systems and reduce repetitive work.

[0158] Clinical Report Generator: Automatically generates structured assessment reports, including screenshots and annotations of key imaging features, and outputs PDF documents that conform to clinical standards. Structured assessment reports improve the standardization and consistency of clinical reports. PDF documents are a commonly used format for medical reports, facilitating storage and sharing.

[0159] Mobile adaptation components: Responsive design adapts to different terminal devices, supports touch gestures and zooming, and provides offline viewing and annotation functions. Responsive design ensures a smooth user experience across various devices. Touch gestures and offline viewing enhance the user experience.

[0160] Example 2

[0161] like Figure 4 As shown, this application provides an architecture diagram of a deep learning-based automatic classification and severity assessment system for CT images of chronic sinusitis, which is applied to the deep learning-based automatic classification and severity assessment system for CT images of chronic sinusitis as described in Embodiment 1. The system includes a data acquisition module 11, a preprocessing module 12, a model building module 13, a fusion decision module 14, and a visualization output module 15.

[0162] Data acquisition module 11: used to acquire sinus CT image data of patients with chronic sinusitis and associate them with clinical labels, including immune endotype classification, Lund-Mackay score and SNOT-22 scale results, wherein the immune endotype classification is verified by immunohistochemical staining results of postoperative tissue specimens.

[0163] Preprocessing module 12: Standardizes CT images, including grayscale normalization, denoising and region of interest segmentation, interpretability analysis, and extraction of sinus and polyp regions.

[0164] Model building module 13: Constructs a multi-task deep learning model, which includes a sub-model for genotyping and a sub-model for severity assessment. The sub-model for genotyping adopts an improved 3DResNet-50 architecture, takes preprocessed CT images as input, and outputs immune intratype classification results. The sub-model for severity assessment is based on the U-Net++ structure to segment sinus lesion areas and automatically calculates the severity level by combining the Lund-Mackay scoring rule.

[0165] Fusion Decision Module 14: Integrates the outputs of the subtyping submodel and the severity assessment submodel to generate individualized treatment recommendations.

[0166] Visualization output module 15: Severity scores and treatment recommendations are annotated on the original CT images in the form of heatmap overlays, supporting DICOM format export and embedding into the hospital PACS system for interactive viewing.

[0167] Figure 5 This is an electronic device provided in one embodiment of this application. For example... Figure 5 As shown, the electronic device includes at least the following components: processor 101 and memory 100, communication interface 103, and bus 102.

[0168] In this embodiment of the application, memory 100 is used to store executable instructions of processor 101, which, when configured to execute instructions, implements the method as described in the first aspect.

[0169] In embodiments of this application, a computer-readable storage medium includes instructions that instruct a device to perform the method as described in the first aspect. For example, the instructions instruct the device to perform... Figure 1 The method is shown in the process steps.

[0170] In one embodiment of this application, the program operating in the electronic device may be a program that controls a central processing unit (CPU) or similar device to achieve the functions of the above-described embodiments of the present invention (a program that enables the computer to function). Information processed by these devices is then temporarily stored in random access memory (RAM) during processing, and subsequently stored in various ROMs such as read-only memory (FlashROM) and hard disk drives (HDDs), and read, corrected, and written by the CPU as needed.

[0171] It should be noted that a portion of the electronic device described above can also be implemented using a computer. In this case, the program for implementing the control function can be recorded on a computer-readable recording medium, and the program recorded on the recording medium can be read into the computer and executed.

[0172] It should be noted that the term "computer" as used here refers to a computer built into an electronic device, employing hardware including an operating system and peripheral devices. Furthermore, "computer-readable recording media" refers to removable media such as floppy disks, magneto-optical disks, ROMs, and CD-ROMs, as well as storage devices such as hard drives built into a computer.

[0173] Furthermore, a "computer-readable recording medium" can include: a medium that dynamically stores a program for a short period of time, such as a communication line used when transmitting a program via a network such as the Internet or a communication line such as a telephone line; or a medium that stores a program for a fixed period of time, such as volatile memory inside a computer that serves as a server or client in this case. In addition, the aforementioned program can be a program used to implement the above-mentioned functions, or it can be a program that can implement the above-mentioned functions by combining with programs already recorded in the computer.

[0174] Furthermore, the electronic device in the above embodiments can also be implemented as an assembly (device group) composed of multiple devices. Each device constituting the device group can possess some or all of the functions or functional blocks of the electronic device in the above embodiments. As a device group, it is sufficient to have all the functions or functional blocks of the electronic device.

[0175] Those skilled in the art should recognize that the above embodiments are only used to illustrate this application and are not intended to limit this application. Any appropriate changes and variations made to the above embodiments within the essential spirit and scope of this application fall within the scope of protection claimed in this application.

Claims

1. A method for automatic classification and severity assessment of CT images of chronic sinusitis based on deep learning, characterized in that, The method includes: S1. Obtain sinus CT image data of patients with chronic sinusitis and associate them with clinical labels, including immune endotype classification, LundMackay score and SNOT-22 scale results, wherein the immune endotype classification is verified by immunohistochemical staining results of postoperative tissue specimens; S2. Standardize the CT images, including grayscale normalization, noise reduction and region of interest segmentation, interpretability analysis, and extract sinus and polyp regions. S3. Construct a multi-task deep learning model, wherein the multi-task deep learning model includes a sub-model for genotyping and a sub-model for severity assessment. The sub-model for genotyping adopts an improved 3DResNet-50 architecture, takes preprocessed CT images as input, and outputs immune intratyping classification results. The sub-model for severity assessment is based on the U-Net++ structure to segment the sinus lesion area and automatically calculates the severity level by combining the LundMacAay scoring rule. S 4. Integrate the outputs of the subtyping submodel and the severity assessment submodel to generate personalized treatment recommendations; S5. The classification results, severity scores, and treatment recommendations are annotated on the original CT images in the form of a heatmap overlay. It supports export in DIC OM format and can be embedded into the hospital PACS system for interactive viewing.

2. The method for automatic classification and severity assessment of chronic sinusitis CT images based on deep learning according to claim 1, characterized in that, The standardization processing of CT images, including grayscale normalization, noise reduction and region of interest segmentation, and interpretability analysis, includes the following sub-steps: Map CT values ​​to standard Hounsfield units and unify the resolution; A nonlocal mean filtering algorithm is used to eliminate image noise. The nasal sinus cavity and polyp area were automatically segmented based on the U-Net network, and the segmentation results were reviewed and corrected by radiologists. The model's region of interest is visualized using gradient-weighted class activation mapping (G ra d C AM), and a heatmap is generated. Spatial overlap with the gold standard labeled region is then evaluated.

3. The method for automatic classification and severity assessment of chronic sinusitis CT images based on deep learning according to claim 1, characterized in that, The specific implementation of the fractal sub-model using the improved 3DRe s Net-50 architecture includes the following steps: Input processing module: Includes 3D convolutional layers, 3D batch normalization layers, and ReLU activation function, used to process 64×64×64 voxel CT image blocks; Spatial-channel attention module: Embedded before each residual block, it includes: Spatial attention branch: Generates spatial weight map using 3D convolutional kernel; Channel attention branch: Generates channel weight vector through 3D global average pooling and fully connected layers. Optimized 3D residual structure: includes: basic residual block: composed of 1×1×1, 3×3×3, and 1×1×1 3D convolutional layers; bottleneck residual block: introduces dilated convolution and adds dropout layers; The classification output module includes a global average pooling layer, a fully connected layer, a Layer Norm layer, a GeLU activation function, and a final classification layer.

4. The method for automatic classification and severity assessment of chronic sinusitis CT images based on deep learning according to claim 3, characterized in that, The training method for the sub-model includes: The transfer learning technique was adopted, and the model parameters were initialized based on the pre-trained 3D ResNet-50 weights. The network input layer was adapted to the CT image data specifications by expanding the convolution kernel. During the training process, an adversarial sample enhancement mechanism is introduced. Perturbation images are generated through a gradient attack algorithm, and the original images and adversarial samples are mixed and trained according to a preset ratio. A composite loss function is adopted, including: label smoothing cross-entropy loss, which sets a smoothing factor to adjust the classification boundary; and a regularization term to control model complexity and prevent overfitting. The Adam W optimizer is used for parameter updates, and a dynamic learning rate scheduling strategy is configured, wherein: the optimizer sets the initial learning rate, weight decay coefficient, and momentum parameter; gradient pruning technique is used to limit the gradient update magnitude; and a periodic learning rate adjustment strategy is implemented to make the learning rate change according to a preset pattern during training. The model's decision-making basis is verified periodically using visualization techniques. When the spatial overlap between the model's region of interest and the standard labeled region is lower than a preset threshold, the model parameter adjustment mechanism is automatically triggered. The loss function for the fractal sub-model is: Among them, L LS The label smoothing cross-entropy loss is represented by ε, where ε is the smoothing factor, F and f represent the number of categories and their corresponding indices, respectively, and y f This represents the on eh ot encoding of the actual tag.

5. The method for automatic classification and severity assessment of chronic sinusitis CT images based on deep learning according to claim 1, characterized in that, The severity assessment sub-model adopts an improved U-Net++ architecture, and the specific implementation process includes: A multi-level encoder-decoder structure is set up. The encoder consists of multiple downsampling stages, each of which includes 3D convolution operations and residual connections. The decoder integrates multi-scale features through dense skip connections and uses transposed convolution to achieve upsampling. A deep supervision mechanism is introduced in each decoding stage to output intermediate segmentation results. An attention mechanism is set at the connection between the encoder and decoder to dynamically adjust the feature weights by analyzing the spatial and channel correlation of the feature maps, thereby achieving adaptive feature enhancement. Through a multi-task output layer, a lesion probability map of each sinus anatomical region is generated using three-dimensional convolution. Based on clinical scoring criteria, the probability map is converted into a lesion severity score for each region. The scores of each region are then combined to obtain the overall severity assessment result. A composite loss function combining region overlap and classification error is used, along with an adaptive optimization algorithm and a dynamic learning rate adjustment strategy. Cross-validation is implemented to ensure model robustness. The formula for the composite loss function of the severity assessment sub-model is as follows: Where K represents the number of sinus anatomical regions, k represents the region index, and i represents the position index in three-dimensional voxel space. This represents the predicted probability value of the k-th anatomical region at voxel i. The corresponding true label indicates whether the voxel belongs to the k-th partition, α represents the balancing weight, ε is the smoothing term, N represents the total number of samples, and S n This represents the predicted overall score for the nth sample. Let λ represent the actual score, λ represent the multi-task weights, and ∑ i This indicates that the values ​​at all voxel locations are summed to calculate the global similarity or error.

6. The method for automatic classification and severity assessment of chronic sinusitis CT images based on deep learning according to claim 5, characterized in that, The severity assessment sub-model also includes a dynamic threshold segmentation optimization module, the implementation of which includes: A spatial location coding matrix is ​​established based on the three-dimensional anatomical atlas of the sinuses, and CT images are divided into multiple anatomical functional areas. Different initial segmentation thresholds are set for different functional areas. Based on the local features of the image, the tissue density distribution is dynamically calculated, and combined with the spatial continuity constraints of adjacent slices, the segmentation threshold boundaries of each functional area are optimized in real time. By integrating grayscale features, texture features, and morphological features, a multi-dimensional discrimination criterion for polyps and mucosal hyperplasia is established, and the final segmentation boundary is optimized through a cascaded classifier. Three-dimensional connected component analysis is applied to eliminate isolated noise points, morphological closing operations are used to fill small voids, and abnormal segmentation regions are corrected based on anatomical constraints.

7. The method for automatic classification and severity assessment of chronic sinusitis CT images based on deep learning according to claim 1, characterized in that, The specific implementation of generating personalized treatment recommendations in step S4 includes: A multidimensional feature space mapping relationship between the typing results and the severity score was established. The immune endogenous type classification and the LundMacAay score were associated and matched through a knowledge graph. A rule-based reasoning engine was used to generate preliminary treatment suggestions. By integrating treatment guidelines from evidence-based medicine, a decision tree is constructed that includes drug therapy, surgical indications, and physical therapy, and the weights of treatment plans are dynamically adjusted according to patient characteristics. Based on the patient's historical treatment records and complication history, the cost-effectiveness of each treatment option is evaluated, and a list of individualized treatment recommendations is generated in order of priority. Generates explanations of recommendations with clinical evidence, visualizes key decision-making factors, and provides comparative analysis of alternative options; We continuously learn from newly published treatment guidelines, optimize recommendation algorithms based on treatment feedback, and regularly update treatment plans in the knowledge base.

8. The method for automatic classification and severity assessment of chronic sinusitis CT images based on deep learning according to claim 7, characterized in that, The visualization output system of S5 includes: Intelligent annotation engine: It uses multi-layer fusion technology to overlay the classification results with the original CT images, uses a differentiated color coding scheme to distinguish different immune intratype regions, and uses transparency adjustment to achieve a contrast display between lesion areas and normal tissues; Interactive visualization module: Supports multi-plane reconstruction MPR display according to DIC OM standard, provides three-dimensional volume rendering function of lesion area, and realizes dynamic annotation and prompts of key scoring indicators; PACS Integration Interface: A network communication protocol compliant with the DIC OM standard, supporting bidirectional data interaction with the hospital's PACS system to achieve automatic matching and association of patient information; Clinical report generator: Automatically generates structured assessment reports, including screenshots and annotations of key imaging features, and outputs PDF documents that conform to clinical standards; Mobile adaptation components: Responsive design adapts to different terminal devices, supports touch gesture operation and zooming, and provides offline viewing and annotation functions.

9. A deep learning-based automatic classification and severity assessment system for CT images of chronic sinusitis, applied to the deep learning-based automatic classification and severity assessment method for CT images of chronic sinusitis as described in any one of claims 1 to 8, characterized in that, The system includes: Data acquisition module: used to acquire sinus CT image data of patients with chronic sinusitis and associate them with clinical labels, including immune endotype classification, LundMackay score and SNOT-22 scale results, wherein the immune endotype classification is verified by immunohistochemical staining results of postoperative tissue specimens; Preprocessing module: Standardizes CT images, including grayscale normalization, noise reduction and region of interest segmentation, interpretability analysis, and extraction of sinus and polyp regions; Model building module: Constructs a multi-task deep learning model, which includes a sub-model for genotyping and a sub-model for severity assessment. The sub-model for genotyping adopts an improved 3DResNet-50 architecture, takes preprocessed CT images as input, and outputs immune intratyping classification results. The sub-model for severity assessment segments sinus lesion areas based on the U-Net++ structure and automatically calculates the severity level using the LundMacAay scoring rule. Fusion Decision Module: Integrates the outputs of the subtyping submodel and the severity assessment submodel to generate individualized treatment recommendations; Visualization output module: Severity scores and treatment recommendations are annotated on the original CT images in the form of heatmap overlays, supporting export in DIC OM format and embedding into the hospital PACS system for interactive viewing.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a program that instructs the device to perform the deep learning-based automatic classification and severity assessment method for CT images of chronic sinusitis as described in any one of claims 1 to 8.

Images