CRSwNP paranasal sinus automatic segmentation and intrinsic type prediction method and system based on CT image

Abstract

The invention discloses a CRSwNP paranasal sinus automatic segmentation and intrinsic type prediction method and system based on a CT image, and the method is characterized in that the method comprises the steps: collecting a paranasal sinus CT image comprising an axial position layer, a coronal position layer and a sagittal position layer, marking a paranasal sinus region, marking an intrinsic type label, generating and converting the paranasal sinus region into a paranasal sinus 3D image, and forming a CRSwNP image database; the intrinsic type labels comprise eosinophilic type labels and non-eosinophilic type labels; designing a 3D multi-task classification segmentation network model for receiving a 3D paranasal sinus CT image and predicting an intrinsic type category and a paranasal sinus region; training the 3D multi-task classification segmentation network model by using a CRSwNP image database to obtain a prediction model; and predicting the paranasal sinus area and the intrinsic type of nasosinusitis with nasal polyp in the 3D paranasal sinus CT image by using the prediction model. According to the method, automatic segmentation of the paranasal sinus 3D model based on the CT image is realized, and accurate prediction of the CRSwNP intrinsic type is carried out on the basis of the paranasal sinus 3D model, so that a clinician can be guided to carry out individualized accurate treatment.

Description

Technical Field

[0001] The present invention relates to the field of medical and engineering intersection, and particularly to a method and system for automatic segmentation and endotype prediction of CRSwNP sinuses based on CT images. Background Art

[0002] Chronic rhinosinusitis (CRS) is a common heterogeneous inflammatory disease of the upper respiratory tract, which seriously affects the quality of life of patients due to its high incidence (about 10%), high recurrence rate (about 50%) and high medical cost (about $33 billion / year). Clinically, in order to implement individualized management for CRS patients, they are divided into two phenotypic categories: chronic rhinosinusitis with nasal polyps (CRSwNP) and chronic rhinosinusitis without nasal polyps (CRSsNP). Research shows that 33.7% of CRSwNP patients and 20.97% of CRSsNP patients still have refractory / recurrent clinical outcomes after 1 year of standardized treatment, and the recurrence rate of CRSwNP patients can be as high as 55.3% after 2 years of surgery. Thus, there is a high degree of heterogeneity in the treatment response of patients. Obviously, this traditional phenotypic classification method cannot meet the original intention of precision treatment, and in-depth exploration of the underlying immune-inflammatory patterns and molecular mechanisms of CRS has become the breakthrough point for achieving precise treatment, and thus the concept of "endotype" has been proposed. Currently, domestic and foreign clinical guidelines classify CRS patients into eosinophilic chronic rhinosinusitis (eCRS) and non-eosinophilic CRS (non-eCRS) according to their pathophysiological mechanisms. Therefore, it is very crucial to clarify the endotype of patients before treatment.

[0003] Although pathological biopsy is the "gold standard" for diagnosing the intrinsic type of CRS, it still has the following deficiencies: preoperative biopsy can only obtain the inflammatory response of local tissues and cannot represent the overall inflammatory state; during intraoperative biopsy, the cell tissues of the biopsy section are swollen and prone to ice crystals, seriously interfering with the classification of inflammatory cells; although the pathological intrinsic type can be accurately obtained after surgery, the precise treatment opportunities before and during surgery have been missed. Sinus CT, as an important objective examination tool for preoperative assessment of the severity of CRS patients, evaluation of treatment effects, and formulation of individualized precise surgical strategies, has drawn a precise "digital blueprint" for otolaryngologists to successfully diagnose and treat CRS patients. There have been relevant research reports on the intrinsic type prediction model based on sinus CT, but these studies all require a large amount of manual data statistics. The data capture only stays at the shallow features of the images and has not been verified by external data, making it difficult to obtain a reliable precise prediction model. Therefore, how to construct a higher-order prediction model has become a key scientific problem that urgently needs to be solved.

[0004] In recent years, artificial intelligence led by deep learning, as a cutting-edge technology, has made great progress. Among them, deep learning technology has the technical characteristics of powerful automatic image segmentation and automatic feature extraction, especially good at capturing microscopic image features in complex images. At present, the combination of deep learning technology and CT images has made remarkable research progress in the medical field, especially in the clinical transformation application of lung nodule screening using chest CT. However, in the field of otolaryngology, the research on using deep learning technology to realize the intrinsic type prediction model based on sinus CT is still scarce. Recently, Hua et al. constructed an intrinsic prediction model of CRS based on sinus CT using the U-net neural network, and the prediction accuracies of image intrinsic type annotation and patient intrinsic type annotation were 76.2% and 89.3% respectively. However, this study lacks the management of a high-quality sinus CT image database. The constructed model cannot distinguish whether the patient has nasal polyps. At the same time, limited by the width of the single-channel data of the U-net neural network, it is impossible to obtain high-precision prediction results, and there is no further verification by external data, and the prediction reliability is still highly controversial. Therefore, the development of a 3D multi-task classification and segmentation network will provide a key technical tool for accurately predicting the intrinsic type diagnosis of CRSwNP in sinus CT images. Summary of the Invention

[0005] The purpose of the present invention is to solve the problems in the prior art.

[0006] The technical solution adopted by the present invention to solve its technical problems is: to provide a method for automatic segmentation and intrinsic type prediction of CRSwNP sinuses based on CT images, including the following steps:

[0007] Collect sinus CT images including axial, coronal, and sagittal planes, label the sinus region and assign intrinsic type labels, generate and transform them into 3D sinus images to form a CRSwNP image database; the intrinsic type labels are divided into eosinophilic and non-eosinophilic;

[0008] Design a 3D multi-task classification and segmentation network model for receiving 3D sinus CT images, learning the microscopic deep features of the images, and thus predicting the intrinsic type category and the sinus region;

[0009] Use the CRSwNP image database to train the 3D multi-task classification and segmentation network model to obtain a prediction model;

[0010] Use the prediction model to predict the sinus region and the intrinsic type of sinusitis with nasal polyps in 3D sinus CT images.

[0011] Preferably, the 3D multi-task classification and segmentation network model includes:

[0012] A feature extraction module for receiving 3D sinus CT images, extracting and outputting image features;

[0013] A feature reconstruction module that receives the image features and reconstructs a high-resolution spatial feature map based on the image features;

[0014] A feature classification module that receives the image features and determines the intrinsic type category according to the image features;

[0015] A feature segmentation module that receives the high-resolution spatial feature map and segments the sinus region by pixel-by-pixel classification.

[0016] Preferably, the structure of the feature extraction module is a residual network module, a channel attention module, and four repeated blocks connected in sequence, where the structure of each repeated block is a max pooling layer and a residual network module connected in sequence.

[0017] Preferably, the channel attention module generates dynamic weights through a channel attention sub-module and calibrates the features, specifically including the following steps:

[0018] The channel attention sub-module receives the input feature map After that, it uses the global average pooling layer to compress along the spatial dimensions D×H×W to generate a channel description vector Use the first fully connected layer to reduce the dimension and perform non-linear mapping on Z, activate it through ReLU; use the second fully connected layer to restore the channel dimension of Z, and then generate a 1D channel attention map through Sigmoid activation

[0019] Multiply the input feature map by the channel attention map M c(F) is multiplied channel - by - channel with the input feature map F to obtain the feature map F ′ as the output; the whole process is expressed as:

[0020]

[0021] wherein, represents element - wise multiplication.

[0022] Preferably, the structure of the residual network module is a first convolutional layer, a Relu activation function, a second convolutional layer, an addition, and a Relu activation function connected in sequence, wherein the output end of the addition is also connected to the first convolutional layer through a 1×1×1 convolution; the convolutional kernels of both convolutional layers are 3×3×3.

[0023] Preferably, the feature reconstruction module is composed of four identical reconstruction sub - modules connected in series in sequence. Each reconstruction sub - module includes:

[0024] A transposed convolutional layer, using a 2×2×2 convolutional kernel and a stride of 2, to perform 2 - fold upsampling on the input feature map;

[0025] Skip - connection fusion, splicing the feature map of the corresponding level of the encoder and the upsampling result channel - by - channel;

[0026] A depth - wise separable convolution module for feature refinement.

[0027] Preferably, the structure of the depth - wise separable convolution module is a first depth - wise separable convolutional layer, a batch normalization layer, a Relu activation function layer, a second depth - wise separable convolutional layer, a batch normalization layer, and a Relu activation function connected in sequence;

[0028] Both the first depth - wise separable convolutional layer and the second depth - wise separable convolutional layer are composed of a depth convolution and a point - wise convolution; wherein, the convolutional kernels of the depth convolution are all 3×3×3, and the convolutional kernels of the point - wise convolution are all 1×1×1.

[0029] Preferably, the feature classification module includes:

[0030] A dual - branch pooling unit, receiving the input feature performing max - pooling and average - pooling respectively to obtain the feature vectors and splicing F max and F avg along the channel dimension to generate a fused feature

[0031] A feature compression unit, using the third fully - connected layer to map F cat to a low - dimensional space and output Among them, r is the compression ratio, and it is further activated by Relu; the fourth fully connected layer is used to restore H to the target category dimension and output. Among them, K is the number of intrinsic categories.

[0032] The probability prediction unit uses Softmax to normalize S and generate the probability distribution of each category.

[0033] Preferably, the feature segmentation module is composed of a convolution, and finally the Softmax layer outputs the category probability distribution of each pixel.

[0034] The present invention also provides a system for automatic segmentation and intrinsic prediction of CRSwNP sinuses based on CT images, including:

[0035] The database construction module collects sinus CT images including axial, coronal, and sagittal planes, annotates the sinus region and assigns intrinsic labels, generates and converts them into 3D sinus images to form a CRSwNP image database; the intrinsic labels are divided into eosinophilic and non-eosinophilic.

[0036] The model construction module designs a 3D multi-task classification and segmentation network model for receiving 3D sinus CT images and learning the microscopic deep features of the images to predict the intrinsic category and sinus region.

[0037] The model training module uses the CRSwNP image database to train the 3D multi-task classification and segmentation network model to obtain a prediction model.

[0038] The prediction module uses the prediction model to predict the sinus region and the intrinsic type of sinusitis with nasal polyps in 3D sinus CT images.

[0039] The present invention has the following beneficial effects:

[0040] (1) The present invention adopts a 3D multi-task classification and segmentation model, fully utilizes the three-dimensional spatial information of CT images, and simultaneously performs segmentation and classification tasks, making it show better performance in sinus CT image analysis, realizing accurate segmentation of the sinus region and effective prediction of the intrinsic category.

[0041] (2) The present invention adopts a multi-task learning mechanism of feature sharing, and co-optimizes the hard parameters of the shared encoder through classification and segmentation tasks, realizing mutual promotion between the two, and improving the model's learning ability for sinus region and category features; at the same time, the network structure is optimized, reducing computational redundancy and significantly reducing the computational cost of training and inference.

[0042] (3) The present invention adopts a residual structure and depthwise separable convolution, avoiding the problem of gradient disappearance in network training, enabling the model training to require less computational effort, the bandwidth required for the model to be downloaded from the cloud to the application side to be smaller, and it can be more flexibly deployed on hardware with limited memory.

[0043] The present invention will be further described in detail below in conjunction with the accompanying drawings and embodiments, but the present invention is not limited to the embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] Figure 1 is a flowchart of the method steps of an embodiment of the present invention;

[0045] Figure 2 is a schematic flowchart of an embodiment of the present invention;

[0046] Figure 3 is a schematic structural diagram of a channel attention module (SE Block), a residual network module (Residual Block), and a depthwise separable convolution module (DS Block) of an embodiment of the present invention;

[0047] Figure 4 is a structural diagram of a 3D multi-task classification and segmentation network model of an embodiment of the present invention;

[0048] Figure 5 is the classification prediction result of the 3D multi-task classification and segmentation network model of an embodiment of the present invention for the CRSwNP standardized professional database, where, Figure 5 (a) is the prediction result of the training group, Figure 5 (b) is the prediction result of the internal testing group, Figure 5 (c) is the prediction result of the external testing group;

[0049] Figure 6 is the segmentation prediction result of the 3D multi-task classification and segmentation network model of an embodiment of the present invention for the CRSwNP standardized professional database;

[0050] Figure 7 is a system structure diagram of an embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0051] Referring to Figure 1 and Figure 2 shown, which are a flowchart of the method steps and a schematic flowchart of an embodiment of the present invention, including:

[0052] S101. Collect sinus CT images including axial, coronal, and sagittal planes, label the sinus region and assign an endotype label, generate a converted 3D sinus image, and form a CRSwNP image database. The endotype labels are divided into eosinophilic and non-eosinophilic.

[0053] S102. Design a 3D multi-task classification and segmentation network model to receive 3D sinus CT images, learn the microscopic deep features of the images, and thus predict the endotype category and sinus region.

[0054] S103. Use the CRSwNP image database to train the 3D multi-task classification and segmentation network model to obtain a prediction model.

[0055] S104. Use the prediction model to predict the sinus region and the endotype of chronic rhinosinusitis with nasal polyps in 3D sinus CT images.

[0056] Specifically, in S101, collect sinus CT images including axial, coronal, and sagittal planes, label the sinus region in the CT images, generate a converted 3D sinus image and perform data normalization on the 3D sinus image, and finally assign an endotype label to form a CRSwNP image database. The endotype labels are divided into eosinophilic and non-eosinophilic.

[0057] The database used in the embodiments of the present invention included 227 patients with chronic rhinosinusitis with nasal polyps (CRSwNP) from two different years. The inclusion criteria were: diagnosed according to the International Chronic Rhinosinusitis Diagnosis and Treatment Guidelines, no previous nasal surgery, normal liver and kidney functions, and no use of oral corticosteroids or macrolide antibiotics within 1 month before treatment. The exclusion criteria included: patients under 18 years old, fungal rhinosinusitis, posterior nasal polyps, nasal or sinus tumors, primary ciliary dyskinesia, cystic fibrosis, patients without sinus CT examinations or without surgical treatment for rhinosinusitis.

[0058] After performing spiral CT scans to obtain sinus images, the data of each patient was converted into volume data using the PACS (picture archiving and communication systems) platform and output in DICOM (Digital Imaging and Communications in Medicine) format, which can be viewed in sagittal, coronal, and axial slices. Subsequently, a dedicated code developed by the CranioCatch software team in the Python programming language, along with the Pydicom, Nibabel, and Opencv libraries, was used to convert the DICOM files into NIfTI (Neuroimaging Informatics Technology Initiative) format. The database information is shown in Table-1. To establish a 3D multi-task classification and segmentation model, 227 CRSwNP patients were classified, including eosinophilic (eosinophilic chronic rhinosinusitis, eCRS) and non-eosinophilic CRS (non-eosinophilic chronic rhinosinusitis, non-eCRS); among them, the training group consisted of 185 patients, and the external test group consisted of 42 patients. Five-fold cross-validation was performed on the training group, in which each fold had the same proportion of positive sample recurrence to further process the data and establish the model.

[0059] Table-1: Information of the CRSwNP standardized professional database

[0060]

[0061] Specifically, the construction process of the 3D multi-task classification and segmentation network model is as follows:

[0062] (1) The channel attention module is a lightweight attention mechanism module, whose motivation is to enhance the network's representational ability through the attention mechanism: dynamically assign weights to the features of different channels, thereby focusing on important features and suppressing irrelevant features. Since the convolution operation fuses cross-channel information, and each channel may contain different feature importances, this module emphasizes the features that are meaningful in the channel dimension. The structure of the channel attention module is as Figure 3 (a) shown. Dynamic weights are generated and features are calibrated through the channel attention sub-module, specifically including:

[0063] (a) Channel attention sub-module: After receiving the input feature map , the following operations are performed in sequence:

[0064] Global Average Pooling layer (GAP), compresses along the spatial dimensions D×H×W to generate a channel description vector

[0065] The first fully connected layer (FC1) performs dimensionality reduction and non-linear mapping on Z, activated by ReLU;

[0066] The second fully connected layer (FC2) restores the channel dimension, generates a 1D channel attention map through Sigmoid activation

[0067]

[0068] (b) Feature calibration: Multiply the channel attention map M c (F) with the input feature map F channel by channel to obtain the feature map F ′ as the output; The whole process is expressed as:

[0069]

[0070] where represents element-wise multiplication.

[0071] (2) The residual network module aims to solve the problem of gradient vanishing in the process of network training of deep neural networks. When there is only one channel in the neural network for backpropagation, as the number of network layers increases, the continuous product will cause the gradient to vanish. Therefore, if there are two channels, it can be changed into a summation form, avoiding gradient vanishing, and then a deeper network architecture can be built to improve the model learning performance. Its structure is as Figure 3 (b) shown, specifically including a first convolutional layer, a Relu activation function, a second convolutional layer, an addition, and a Relu activation function connected in sequence; among them, the output end of the addition is also connected to the first convolutional layer through a 1×1×1 convolution; the convolutional kernels of the two convolutional layers are both 3×3×3; specifically, the outputs of the first convolutional layer and the second convolutional layer also go through batch processing before output.

[0072] (3) The depthwise separable convolution module includes a first depthwise separable convolutional layer, a batch normalization layer (BN), a Relu activation function layer, a second depthwise separable convolutional layer, a batch normalization layer (BN), and a Relu activation function connected in sequence. Depthwise separable convolution is composed of depth convolution and pointwise convolution, separating feature extraction and channel fusion, thereby effectively reducing the computational complexity and the number of parameters. Specifically, depth convolution separately performs spatial feature extraction on each channel, while pointwise convolution is responsible for feature fusion between channels. This separated operation form reduces computational redundancy, enables lightweight design, and maintains high model performance under limited computing resources. The specific structure is as Figure 3As shown in (c), the convolutional kernels of the depthwise convolutions are all 3×3×3, and the convolutional kernels of the pointwise convolutions are all 1×1×1.

[0073] (4) Based on the above structure, a 3D multi-task classification and segmentation network architecture is constructed. The input of the model is a 3D sinus CT image. After passing through the feature extraction module, the corresponding feature map is output. The output feature maps are respectively sent into the feature reconstruction module, the classification network, and the feature classification module. After passing through the feature reconstruction module, the extracted features are reconstructed into a high-resolution spatial feature map. After passing through the feature classification module, the intrinsic type category is judged according to the extracted features. Finally, the high-resolution spatial feature map is input into the feature segmentation module, and the sinus region is segmented through per-pixel classification. The network structure of the 3D multi-task classification and segmentation network architecture is as Figure 4 shown, including:

[0074] A feature extraction module (encoder), which consists of a residual network module, a channel attention module, four max pooling layers, and four residual network modules;

[0075] A feature reconstruction module (decoder), which consists of a transposed convolution layer, skip connections, and four depthwise separable convolution modules;

[0076] A feature classification module (classification head), which consists of a max pooling layer, an average pooling layer, and a fully connected layer, and finally outputs the intrinsic type category of CRS patients through a Softmax layer;

[0077] A feature segmentation module (segmentation head), which consists of a convolution, and finally outputs the class probability distribution of each pixel through a Softmax layer.

[0078] Relevant experiments are conducted on the embodiments of the present invention to explore its detection effect.

[0079] (1) Evaluation of the classification prediction performance of the 3D multi-task classification and segmentation model

[0080] Based on the constructed sinus CT dataset, the 3D multi-task classification and segmentation model is trained and tested. The AUC curve and confusion matrix during the model training process are as Figure 5 shown in (a) of, and the AUC curve and confusion matrix in the internal test are as Figure 5 shown in (b) of, and the AUC curve and confusion matrix in the external test set are as Figure 5 shown in (c) of. The area under the curve (AUC), accuracy, sensitivity, specificity, false positive rate (FPR), and false negative rate (FNR) indicators are used to evaluate the classification prediction performance, and the results are shown in Table-2. The results show that the AUC of the 3D multi-task classification and segmentation network model of the embodiments of the present invention reaches 0.805 in the test set.

[0081] Table - 2: Classification Prediction Performance of 3D Multi - task Classification and Segmentation Model

[0082]

[0083] (2) Evaluation of the Segmentation Prediction Performance of the 3D Multi - task Classification and Segmentation Model

[0084] Based on the constructed sinus CT dataset, the 3D multi - task classification and segmentation model is trained and tested. The visual segmentation results of the model in the external test set are as Figure 6 shown. The Dice similarity coefficient (DSC), Jaccard similarity coefficient (JSC), precision (PRE), and recall (REC) metrics are used to evaluate the segmentation prediction performance, and the results are shown in Table - 3. The results show that the DSC of the 3D multi - task classification and segmentation network model of the embodiment of the present invention reaches 0.8937 in the test set.

[0085] Table - 3: Segmentation Prediction Performance of 3D Multi - task Classification and Segmentation Model

[0086]

[0087] Specifically, referring to Figure 7 shown, which is the system structure diagram of the embodiment of the present invention, including:

[0088] Database construction module 701, which collects sinus CT images including axial, coronal, and sagittal planes, annotates the sinus region and assigns intrinsic type labels, generates and converts them into sinus 3D images, and forms a CRSwNP image database; the intrinsic type labels are divided into eosinophilic and non - eosinophilic;

[0089] Model construction module 702, which designs a 3D multi - task classification and segmentation network model for receiving 3D sinus CT images, learning the microscopic deep - layer features of the images, and thus predicting the intrinsic type and sinus region;

[0090] Model training module 703, which uses the CRSwNP image database to train the 3D multi - task classification and segmentation network model to obtain a prediction model;

[0091] Prediction module 704, which uses the prediction model to predict the sinus region and the intrinsic type of chronic rhinosinusitis with nasal polyps in 3D sinus CT images.

[0092] It can be seen that the present invention not only realizes the automatic segmentation of the sinus 3D model based on CT images, but also accurately predicts the intrinsic type of CRSwNP on the basis of this sinus 3D model, thus helping to guide clinicians to perform individualized and precise treatment.

[0093] The above are only the preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of the present invention shall be included within the protection scope of the present invention.

Claims

1. A method for automatic segmentation and endotype prediction of CRSwNP sinuses based on CT images, characterized in that It includes the following steps: Collect sinus CT images including axial, coronal, and sagittal planes, annotate the sinus region and assign intrinsic type labels, generate and convert them into 3D sinus images to form a CRSwNP image database; the intrinsic type labels are divided into eosinophilic and non-eosinophilic; Design a 3D multi-task classification and segmentation network model for receiving 3D sinus CT images, learning the microscopic deep features of the images, and thus predicting the intrinsic type category and the sinus region; Use the CRSwNP image database to train the 3D multi-task classification and segmentation network model to obtain a prediction model; Use the prediction model to predict the sinus region and the intrinsic type of nasal polyps with sinusitis in 3D sinus CT images.

2. The method for automatic segmentation and endotype prediction of CRSwNP sinuses based on CT images according to claim 1, characterized in that, The 3D multi-task classification and segmentation network model includes: A feature extraction module for receiving 3D sinus CT images, extracting and outputting image features; A feature reconstruction module that receives the image features and reconstructs a high-resolution spatial feature map based on the image features; A feature classification module that receives the image features and determines the intrinsic type category according to the image features; A feature segmentation module that receives the high-resolution spatial feature map and segments the sinus region by pixel-by-pixel classification.

3. The method for automatic segmentation and endotype prediction of CRSwNP sinuses based on CT images according to claim 2, wherein The structure of the feature extraction module is a residual network module, a channel attention module, and four repeated blocks connected in sequence. Among them, the structure of each repeated block is a max pooling layer and a residual network module connected in sequence.

4. The method for automatic segmentation and endotype prediction of CRSwNP sinuses based on CT images according to claim 3, characterized in that, The channel attention module generates dynamic weights through a channel attention sub-module and calibrates the features, specifically including the following steps: The channel attention sub-module receives the input feature map After that, a global average pooling layer is used to compress along the spatial dimensions D×H×W to generate a channel description vector A first fully connected layer is used to reduce the dimension and perform non-linear mapping on Z, activated by ReLU; a second fully connected layer is used to restore the channel dimension of Z, and then a 1D channel attention map is generated through Sigmoid activation Multiply the channel attention map M c (F) with the input feature map F channel by channel to obtain the feature map F ′ as the output; the whole process is expressed as: Among them, represents element multiplication.

5. The method for automatic segmentation and endotype prediction of CRSwNP sinuses based on CT images according to claim 3, wherein The structure of the residual network module is a first convolutional layer, a Relu activation function, a second convolutional layer, an addition, and a Relu activation function connected in sequence. The output end of the addition is also connected to the first convolutional layer through a 1×1×1 convolution; the convolutional kernels of both convolutional layers are 3×3×3.

6. The method for automatic segmentation and endotype prediction of CRSwNP sinuses based on CT images according to claim 2, characterized in that, The feature reconstruction module is composed of four identical reconstruction sub-modules connected in series in sequence. Each reconstruction sub-module includes: A deconvolution layer that uses a 2×2×2 convolutional kernel and a stride of 2 to perform 2-fold upsampling on the input feature map; Skip connection fusion that performs channel splicing on the feature map of the corresponding level of the encoder and the upsampling result; A depthwise separable convolution module for feature refinement.

7. The method for automatic segmentation and endotype prediction of CRSwNP sinuses based on CT images according to claim 6, wherein The structure of the depthwise separable convolution module is a first depthwise separable convolutional layer, a batch normalization layer, a Relu activation function layer, a second depthwise separable convolutional layer, a batch normalization layer, and a Relu activation function connected in sequence; Both the first depthwise separable convolutional layer and the second depthwise separable convolutional layer are composed of a depthwise convolution and a pointwise convolution; among them, the convolutional kernels of the depthwise convolution are all 3×3×3, and the convolutional kernels of the pointwise convolution are all 1×1×1.

8. The method for automatic segmentation and endotype prediction of CRSwNP sinuses based on CT images according to claim 2, characterized in that, The feature classification module includes: Dual-branch pooling unit, receiving input features Perform max pooling and average pooling respectively to obtain feature vectors and Concatenate F max and F avg along the channel dimension to generate fused features Feature compression unit, using the third fully connected layer to map F cat to a low-dimensional space and output where r is the compression ratio, and then through Relu activation; using the fourth fully connected layer to restore H to the target class dimension and output where K is the number of intrinsic classes; A probability prediction unit that normalizes S using Softmax to generate the probability distribution of each category.

9. The method for automatic segmentation and endotype prediction of CRSwNP sinuses based on CT images according to claim 2, wherein, The feature segmentation module is composed of a convolution and finally outputs the category probability distribution of each pixel through a Soft max layer.

10. A system for automatic segmentation and endotype prediction of CRSwNP sinuses based on CT images, characterized in that, It includes: Database construction module, which collects sinus CT images including axial, coronal, and sagittal planes, annotates the sinus region and assigns an endotype label, generates and converts them into 3D sinus images, and forms a CRSwNP image database; the endotype labels are divided into eosinophilic and non-eosinophilic; Model construction module, which designs a 3D multi-task classification and segmentation network model for receiving 3D sinus CT images, learning the microscopic deep features of the images, and thus predicting the endotype category and sinus region; Model training module, which uses the CRSwNP image database to train the 3D multi-task classification and segmentation network model to obtain a prediction model; Prediction module, which uses the prediction model to predict the sinus region and the endotype of nasal polyps with sinusitis in 3D sinus CT images.

Images