A Method and System for Interstitial Lung Imaging Analysis Based on Clinical Prior Guidance Feature Fusion

Abstract

This invention discloses a method and system for lung interstitial imaging analysis based on clinically prior-guided feature fusion. The method includes: acquiring chest CT images and corresponding clinical data at multiple time points from the user, performing data preprocessing and region segmentation; extracting features from the automatically segmented chest CT images and corresponding clinical data using specific indicators; encoding specific time points into temporal embedding vectors, projecting the total image feature vectors through a linear layer to the same dimension as the temporal embedding vectors to generate temporal encoding fusion; actively "querying" and "weighting" the most relevant CT image follow-up time points using clinical risk factors, performing temporal image feature fusion, and outputting the weighted fused features; calculating the progression probability using the fused features, and predicting the risk of progression in the next year based on the calculation results. This invention achieves dynamic temporal feature selection driven by clinical prior knowledge.

Description

Technical Field

[0001] This invention relates to the fields of medical artificial intelligence and radiomics technology, specifically to a method and system for lung interstitial imaging analysis based on clinically prior-guided feature fusion. Background Technology

[0002] Interstitial lung abnormalities (ILA) are lung parenchymal abnormalities detected on high-resolution CT scans that may represent early-stage interstitial lung disease, particularly common in the elderly and smokers. While some ILA segments remain stable, others gradually progress to fibrotic ILA or even idiopathic pulmonary fibrosis (IPF), leading to irreversible lung function decline and a poor prognosis. Current technologies lack objective and quantifiable tools for predicting ILA progression. Radiologists' subjective assessments are inconsistent and struggle to capture subtle quantitative changes. Although AI research has attempted end-to-end prediction using deep learning, it generally suffers from problems such as being a "black box," poor interpretability, and reliance on large-scale labeled data. Summary of the Invention

[0003] The purpose of this invention is to provide a method and system for lung interstitial imaging analysis based on clinically prior-guided feature fusion, so as to solve the problems mentioned in the background art.

[0004] The specific technical solution provided by this invention is as follows: A lung interstitial imaging analysis method based on clinically prior-guided feature fusion, comprising the following operational steps:

[0005] Step S1: Acquire chest CT images and their corresponding clinical data from multiple time points of the user, and perform preprocessing, including region segmentation of the acquired chest CT images.

[0006] Preferably, preprocessing of the acquired chest CT images, including region segmentation, includes:

[0007] Step S11: Obtain the user's chest CT images at multiple time points and corresponding clinical data including age, gender, smoking history, lung function, and symptoms;

[0008] Step S12: Perform standardization processing on the chest CT image sequence, including grayscale normalization and resampling to a uniform voxel spacing;

[0009] Step S13: Use a pre-trained 3D U-Net model to automatically segment a pair of chest CT images and generate a lung lobe mask including background, left upper lobe, left lower lobe, right upper lobe, right middle lobe, and right lower lobe.

[0010] Step S14: Simultaneously use another pre-trained 3D U-Net model II, take the CT image within the generated lung lobe partition mask as input, and output the binary mask of the ILA lesion area including the background and the lesion.

[0011] Step S15: Using the lung lobe partition mask, create a left lung mask and a right lung mask in two directions, left and right respectively. Perform morphological erosion on the left lung mask and the right lung mask respectively, using a 10mm diameter spherical structure for erosion. Subtract the left lung mask and the right lung mask after one morphological erosion from the original left lung mask and the right lung mask without morphological erosion to obtain a 10mm subpleural mask for the left lung and a 10mm subpleural mask for the right lung.

[0012] Step S2: Use specific indicators to extract features from the automatically segmented chest CT images and their corresponding clinical data.

[0013] Preferably, feature extraction includes: classifying chest CT images into four feature categories: lesion burden features, spatial distribution features, density heterogeneity features, and temporal variation features; and using different indicators to calculate and extract quantitative features of CT images for each feature category to obtain the total image feature vector.

[0014] Preferably, the corresponding clinical data is divided into five feature categories, including age, gender, smoking index, lung function indicators, and symptom presentation. Different indicators are used to calculate and extract clinical quantitative features for each feature category, transforming the corresponding clinical data into clinical quantitative features. The clinical feature vectors, including age, gender, smoking index, lung function indicators, and symptom presentation, are calculated and extracted using specific indicators to obtain the total clinical feature vector.

[0015] Step S3: Use sine-cosine position coding to encode specific time points into temporal embedding vectors, and project the total image feature vector through a linear layer to the same dimension as the temporal embedding vector to generate temporal coding fusion.

[0016] Preferably, encoding specific time points into time embedding vectors includes: defining time points as the number of days since the baseline scan. Encoded as a temporal embedding vector , indicating the first Time-encoded vectors at each time point , Indicates the number of dimensions of the time-encoded embedding. The 3D real space is represented as We employ sine-cosine positional encoding for time-point encoding. Through this time-point encoding, we obtain a unique and smoothly changing vector representation for different time points, which is used for the model to learn time-dependent patterns.

[0017] Preferably, the total image feature vector is represented as The temporal embedding vector is represented as They are represented as follows:

[0018]

[0019]

[0020] By introducing a linear projection layer Its parameters are trainable matrices With bias vector The total image feature vector Mapped to In 3D space, the total image feature vector after projection alignment is obtained. The resulting total image feature vector With temporal embedding vector Element-wise addition generates time-coded fusion features Based on the generated time-coded fusion features Obtain the time-coded feature sequence , , T Indicates total T Each time point This indicates the number of dimensions of the time-encoded features.

[0021] Step S4: Use the extracted clinical features as global context information to perform weighted aggregation of time-series image features, and output the weighted aggregation time-series image features as fusion features.

[0022] Preferably, the temporal image feature fusion includes: using the total clinical feature vector as global context information for weighted aggregation of temporal image features; using clinical features as the query vector and time-coded feature sequences as Key / Value pairs for clinical feature-guided temporal feature selection; generating the query vector by projecting the total clinical feature vector onto a trainable matrix; generating the Key matrix and Value matrix by projecting the time-coded feature sequences onto the Key projection matrix and Value projection matrix, respectively; calculating the attention weights using the scaling dot product softmax function; and finally outputting the fused features representing the aggregated temporal image features under clinical weighting.

[0023] Step S5: Calculate the probability of progress using fusion features, and predict the risk of progress in the next year based on the calculation results. The low risk / high risk output of the prediction is represented by a probability of 0 to 1.

[0024] Preferably, the progress probability calculation includes: inputting the output fused features into a fully connected neural network, and calculating the output progress probability after sigmoid activation. Progress probability The closer the value is to 1, the higher the risk of progress; the closer it is to 0, the lower the risk.

[0025] On the other hand, the lung interstitial imaging analysis system based on clinical prior-guided feature fusion includes a three-layer architecture: a basic data layer, a business logic layer, and a data management and interaction layer, to execute the aforementioned lung interstitial imaging analysis method based on clinical prior-guided feature fusion.

[0026] Compared with existing technologies, the beneficial effects achieved by this invention are as follows: This invention uses highly interpretable quantitative radiomics features (volume, density, regional proportion, etc.) as core inputs, combined with a longitudinal temporal coding mechanism to capture lesion evolution patterns; it introduces multimodal interactive modeling to integrate clinical risk factors, allowing clinical risk factors (such as age, smoking index, FVC index, clinical symptoms) to actively "query" and "weight" the most relevant image follow-up time points, using image temporal features as Key / Value (content library), which are quantitative change features at all time points, extracted according to clinical needs; for the first time in ILA progression prediction, clinical risk factors are used as attention queries, rather than passive splicing or post-fusion, to achieve dynamic temporal feature selection driven by clinical prior knowledge, which is more in line with clinical decision-making logic and has better interpretability in clinical practice. This invention employs unidirectional, asymmetric Cross-Attention, with clinical information as the Query and imaging as the Key / Value, ensuring a clear information flow: the clinic "asks," and the imaging "answers." The output is a single vector (1×d) that can be directly used for prediction, resulting in a concise and efficient structure. Furthermore, the constructed lightweight, deployable, and verifiable prediction model is suitable for real-world clinical follow-up scenarios. Attached Figure Description

[0027] Figure 1 This is a flowchart of the operational steps of the lung interstitial imaging analysis method based on clinical prior-guided feature fusion provided in this embodiment of the invention;

[0028] Figure 2 This is a schematic diagram of the original chest CT image provided in an embodiment of the present invention;

[0029] Figure 3 This is a schematic diagram of the 10mm subpleural region provided in an embodiment of the present invention;

[0030] Figure 4 This is a schematic diagram of ILA lesion segmentation provided in an embodiment of the present invention;

[0031] Figure 5 This is a schematic diagram of lung lobe segmentation provided in an embodiment of the present invention;

[0032] Figure 6 This is an architecture diagram of the lung interstitial imaging analysis system based on clinical prior-guided feature fusion provided in an embodiment of the present invention. Detailed Implementation

[0033] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention are within the scope of protection of the present invention.

[0034] Example 1:

[0035] Combination Figures 1-5 As shown in the figure, the lung interstitial imaging analysis method based on clinically prior-guided feature fusion described in this embodiment includes the following steps:

[0036] Step S1: Acquire chest CT images and their corresponding clinical data from multiple time points of the user, and perform preprocessing, including region segmentation of the acquired chest CT images.

[0037] In this embodiment, after acquiring the user's chest CT images and corresponding clinical data, the data preprocessing and region segmentation specifically include the following steps:

[0038] Step S11: Obtain user data at multiple time points Chest CT images and corresponding clinical data including age, gender, smoking history, lung function, symptoms, etc.;

[0039] Step S12: Standardize the acquired chest CT image sequence, including grayscale normalization (window width: 1500HU, window level: -600HU) and resampling to a uniform voxel spacing (1mm×1mm×1mm).

[0040] Step S13: Use a pre-trained 3D U-Net model to automatically segment chest CT images, generating a lung lobe partition mask including background, left upper lobe, left lower lobe, right upper lobe, right middle lobe, and right lower lobe; the 3D U-Net model adopts an encoder-decoder structure, the encoder part uses 3D convolution and max pooling for downsampling, the decoder part uses transposed convolution for upsampling, skip connections are used to preserve detail information, softmax activation is used, and the loss function is DiceLoss.

[0041] Step S14: Simultaneously use another pre-trained 3D U-Net model II, take the CT image within the generated lung lobe partition mask as input, and output the binary mask of the ILA lesion region including the background and the lesion; the 3D U-Net model II is trained on the CT image dataset labeled with ILA regions, and the loss function is DiceLoss+FocalLoss.

[0042] Step S15: Using the lung lobe partition mask, create a left lung mask and a right lung mask in two directions, left and right respectively. Perform morphological erosion on the left lung mask and the right lung mask respectively, using a 10mm diameter spherical structure for erosion. Subtract the left lung mask and the right lung mask after one morphological erosion from the original left lung mask and the right lung mask without morphological erosion to obtain a 10mm subpleural mask for the left lung and a 10mm subpleural mask for the right lung.

[0043] For example, obtaining the user at a specific time point Thin-slice (less than 2mm slice thickness) chest CT image sequences and corresponding clinical data including age, gender, smoking history, lung function, symptoms, etc. were collected. The above data were standardized and preprocessed as follows: the CT image format was defined as DICOM format, slice thickness ≤2mm, and simultaneous normalization and resampling were performed; the acquisition time of clinical data was synchronized with the scanning time of chest CT images, and the error was set not to exceed ±7 days. Then, a three-dimensional U-Net network was used to automatically segment the CT images at each time point to obtain the lobar region mask and the ILA lesion region mask.

[0044] Step S2: Use specific indicators to extract features from the automatically segmented CT images and their corresponding clinical data.

[0045] In this embodiment, chest CT images are categorized into four feature categories: lesion burden features, spatial distribution features, density heterogeneity features, and temporal variation features. Quantitative features of the CT images are calculated and extracted for each feature category using different indicators. The specific indicators for each category are shown below:

[0046] The specific indicators corresponding to the disease burden characteristics include:

[0047] (1) The number of voxels in the lesion mask The lesion volume (mL) in each lung lobe, left and right lungs, and whole lung region calculated by single voxel volume.

[0048] (2) The mean density of the lesion (HU) is expressed by the mean of the CT values ​​of all voxels in the lesion area.

[0049] (3) The proportion of lesions obtained by dividing the lesion volume by the lung volume of the affected area.

[0050] The specific indicators corresponding to the spatial distribution characteristics include: lesion volume (mL) and average density (HU) within 10 mm below the pleura.

[0051] The specific indicators corresponding to density heterogeneity characteristics include: standard deviation, skewness, kurtosis, and entropy within the lesion regions of each lung lobe.

[0052] The specific indicators corresponding to the time-series change characteristics include: the rate of volume change and the amount of density change compared to the previous time period.

[0053] For example, based on each time point chest CT The segmentation results of automatic image segmentation, calculated and extracted using specific indicators, include: lesion burden characteristics. Spatial distribution characteristics Density heterogeneity characteristics 、 Time-series variation characteristics Interpretable quantitative features of CT images , This constitutes the total image feature vector. , ,in, They represent the first Extracted from each time point The quantitative CT image feature vector is represented by... This indicates the number of feature dimensions (number of features) in a chest CT image. The 3D real space is represented as .

[0054] In this embodiment, the corresponding clinical data are simultaneously categorized into five feature categories: age, gender, smoking index, lung function indicators, and symptom presentation. Different indicators are then used to calculate and extract clinical quantitative features for each feature category. The specific indicators for each category are shown below:

[0055] Specific indicators corresponding to age include: using the current age in months / (100) 12);

[0056] The specific indicators corresponding to gender include: male and female;

[0057] The specific indicators corresponding to the smoking index include: number of packs smoked per day. Years of smoking;

[0058] The specific indicators corresponding to lung function indicators include: lung function predicted values ​​calculated based on age, height, and gender using the GLI-2012 Southeast Asia equation, and the calculation of (user's measured lung function value / predicted value). 100%;

[0059] The specific indicators corresponding to the symptoms include: no symptoms, no change in symptoms, worsening, and relief.

[0060] For example, the corresponding clinical data is transformed into clinical quantitative features, and specific indicators are used to calculate and extract features including: age. ,gender Smoking Index Lung function indicators Symptoms Clinical feature vector , This constitutes the overall clinical feature vector. , , ,in, This indicates that it was extracted from the current clinical data. The clinical feature vector is represented. This indicates the number of clinical feature dimensions (number of features). The 3D real space is represented as .

[0061] Step S3: Use sine-cosine positional encoding to encode specific time points into temporal embedding vectors, and project the total image feature vector through a linear layer onto the vector. The same dimension generates time-encoded fusion.

[0062] In this embodiment, to capture the impact of follow-up time intervals on progression, time points are... (Defined as the number of days since the baseline scan) is encoded as a time embedding vector That is, it is represented as the first Time-encoded vectors at each time point , Indicates the number of dimensions of the time-encoded embedding. The 3D real space is represented as The formula for calculating time encoding using sine-cosine position is as follows:

[0063]

[0064]

[0065] in, Indicates the first position in the sequence The position, the th position of its encoded vector Values ​​for an even number of dimensions. Indicates the first position in the sequence The position, the th position of its encoded vector Values ​​for an odd number of dimensions, dimension index Used to generate paired sine and cosine components. , For the embedded dimension.

[0066] For example, when Represented as the first The number of days between the first CT scan and the baseline scan (in days). If the first scan is at time zero, then... =0, the second scan is 1 year later, then =365, The preset time embedding dimension (e.g., 64) is used, and k is an integer index starting from 0, used to traverse half of the embedding dimension. Through this encoding, a unique and smoothly changing vector representation is obtained at different time points, which facilitates the model to learn time dependency patterns.

[0067] In this embodiment, the total image feature vector is... Projected onto the temporal embedding vector through a linear layer Same dimensions:

[0068]

[0069]

[0070] in, Represents the total image feature vector after projection alignment, obtained by transforming the original... 3D features linearly mapped to Dimension, so as to be added with the temporal embedding, due to the original quantitative feature dimension. With time embedding dimension There may be inconsistencies; this invention introduces a linear projection layer (MLP fully connected layer). Its parameters are trainable matrices With bias vector ,Will Mapped to 3D space, to obtain ; will be obtained With time embedding Element-by-element addition generates fusion features Based on fusion features Obtain the time-coded feature sequence , T represents a total of T time points. This represents the number of dimensions of the time-encoded features, and its real space is represented as .

[0071] Step S4: Use the extracted clinical features as global context information to perform weighted aggregation of time-series image features, and output the weighted aggregation time-series image features as fusion features.

[0072] In this embodiment, the total clinical feature vector is... As global contextual information, it is used for weighted aggregation of time-series image features. Clinical features are used as the query vector, and time-encoded feature sequences are used as the key / value pair to achieve clinical feature-guided time-series feature selection.

[0073]

[0074]

[0075]

[0076]

[0077]

[0078] in, Represents the query vector Query. Represents the Query projection matrix. The Key matrix is ​​the real-space representation of the Query projection matrix. Index keys representing image features at each time point. Represents a time-coded feature sequence. Represents the Key projection matrix. The Key projection matrix is ​​represented in real space, and the Value matrix is... This represents the content values ​​to be weighted and aggregated at each point in time. Represents the Value projection matrix. Let the dimension of the projection matrix of Value be the real space representation. Attention key / query dimension to control the computational complexity of attention; This is the attention weight vector, representing the importance weight of each time point in the clinical context. , The transpose of the key matrix is ​​the query-key dot product matrix (attention score). This indicates the degree of match between the clinical query and the key at each time point. The scaling factor used in the Softmax function is to prevent the softmax gradient from vanishing due to excessively large dot product values. The attention scores are normalized into a weight distribution. This is a multimodal fusion feature vector, which serves as the final input feature after weighted aggregation for the final prediction.

[0079] For example, the total clinical feature vector Trainable matrix Projection to generate query vector Time-coded feature sequences Passing through the matrix respectively , Projection to generate the key matrix With Value Matrix Attention weights The fusion features, calculated using the scaled dot product softmax function, are ultimately obtained as a clinically weighted temporal image feature aggregation representation. .

[0080] In this embodiment, the present invention does not allow the model to blindly learn which time points are important. Instead, it allows clinical risk factors (such as age, smoking index, FVC index, and clinical symptoms) to actively "query" and "weight" the most relevant image follow-up time points. The image temporal features serve as the Key / Value (content library), which are quantitative change features at all time points. They are extracted and weighted according to clinical needs. For the first time in ILA progression prediction, clinical risk factors are used as attention queries, rather than passively splicing or later fusion. This achieves dynamic temporal feature selection driven by clinical prior knowledge, which is more in line with clinical decision-making logic and has better interpretability in clinical practice. Meanwhile, because clinical features are static, low-dimensional, global, and highly semantic, while image features are dynamic, high-dimensional (temporal), local, and quantitative, their modal heterogeneity, information granularity, and temporal attributes make them unsuitable for conventional symmetric Cross-Attention structures. This invention employs a unidirectional, asymmetric Cross-Attention mechanism, with clinical features as the Query and image features as the Key / Value, ensuring a clear information flow: clinical "questions," and image "answers." The output is a single vector (1×d) directly used for prediction, resulting in a concise and efficient structure. The weighted fusion of global feature vectors reflects the most important temporal image feature evolution patterns in the context of clinical risk.

[0081] Step S5: Calculate the probability of progress using fusion features, and predict the risk of progress in the next year based on the calculation results. Low risk / high risk is represented by a probability of 0 to 1.

[0082] In this embodiment, the output fusion features The input to a fully connected neural network (MLP) is processed by sigmoid activation to calculate the output progression probability. :

[0083]

[0084]

[0085] in, It is the Sigmoid activation function. This is the specific definition of the Sigmoid activation function. This is the weight matrix of a fully connected neural network (MLP). The weight matrix is ​​represented in real space by its dimension. It is the bias of the MLP network and the probability of progress. The closer the value is to 1, the higher the risk of progress; the closer it is to 0, the lower the risk.

[0086] Example 2:

[0087] Combination Figure 6 As shown, the lung interstitial imaging analysis system based on clinical prior-guided feature fusion described in this embodiment includes a three-layer architecture: a basic data layer, a business logic layer, and a data management and interaction layer, to execute the lung interstitial imaging analysis method based on clinical prior-guided feature fusion of Embodiment 1.

[0088] In this embodiment, the basic data layer stores CT images and clinical data from multiple time points obtained from the data source, as well as the feature categories corresponding to the CT images and clinical data.

[0089] In the business logic layer: a 3D U-Net network for automatic segmentation of each CT scan is integrated, and highly interpretable quantitative radiomics features (volume, density, regional proportion, etc.) are used as core inputs. The longitudinal temporal coding mechanism is combined to capture the lesion evolution pattern. Multimodal interactive modeling is introduced to integrate clinical risk factors. At the same time, a lightweight, deployable, and verifiable prediction model is constructed for real clinical follow-up scenarios.

[0090] In the data management and interaction layer: a real-time database for storing real-time CT images and a relational database for retrieving corresponding clinical data from CT images were constructed, and a feedback mechanism for real-time interaction with the user was established.

[0091] It should be noted that, in this invention, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus.

[0092] Finally, it should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A lung interstitial imaging analysis method based on clinically prior-guided feature fusion, characterized in that: The following steps are included: Step S1: Acquire chest CT images and their corresponding clinical data at multiple time points from the user, perform preprocessing, and segment the acquired chest CT images into regions. Step S2: Use specific indicators to extract features from the automatically segmented chest CT images and their corresponding clinical data; Feature extraction includes: classifying chest CT images into four feature categories: lesion burden, spatial distribution, density heterogeneity, and temporal variation; calculating and extracting quantitative features of CT images for each feature category using different indicators to obtain a total image feature vector; classifying corresponding clinical data into five feature categories: age, gender, smoking index, lung function indicators, and symptom presentation; calculating and extracting quantitative clinical features for each feature category using different indicators to obtain a total clinical feature vector. Specific indicators corresponding to lesion burden characteristics include: lesion volume in lobes, left and right lobes, and the entire lung region; average lesion density; and lesion percentage. Specific indicators corresponding to spatial distribution characteristics include: lesion volume and average density within 10 mm of the subpleural region; Specific indicators corresponding to density heterogeneity characteristics include: standard deviation, skewness, kurtosis, and entropy within the lesion regions of each lung lobe; The specific indicators corresponding to the time-series change characteristics include: the rate of volume change and the amount of density change compared to the previous time period; Step S3: Use sine-cosine position coding to encode specific time points into temporal embedding vectors, and project the total image feature vector through a linear layer to the same dimension as the temporal embedding vector to generate temporal coding fusion; Encoding a specific time point into a time embedding vector using sine-cosine position coding includes: using the number of days from the baseline scan as the time point for encoding, and encoding the time point using sine-cosine position coding to obtain a time embedding vector; Step S4: Use the extracted clinical features as global context information to perform weighted aggregation on the temporal image features, and output the weighted aggregation temporal image features as fusion features; Using clinical features as query vectors Query The time-encoded feature sequence is used as the key / value pair for clinical feature-guided temporal feature selection, and is extracted with weights according to clinical needs. The total clinical feature vector is projected onto a trainable matrix to generate a query vector. The time-encoded feature sequence is projected onto the key projection matrix and the value projection matrix respectively to generate the key matrix and the value matrix. The attention weight is calculated by the scaling dot product softmax function, and the final output is the fusion feature of the time-series image feature aggregation representation under clinical weighting. Step S5: Calculate the probability of progress using the fusion features of the output, and predict the risk of progress in the next year based on the calculation results. The low risk / high risk of the predicted output is represented by a probability of 0 to 1.

2. The lung interstitial imaging analysis method based on clinically prior-guided feature fusion according to claim 1, characterized in that: In step S1, preprocessing is performed on the acquired chest CT images, including region segmentation: Step S11: Obtain the user's chest CT images at multiple time points and corresponding clinical data including age, gender, smoking history, lung function, and symptoms; Step S12: Perform standardization processing on the chest CT image sequence, including grayscale normalization and resampling to a uniform voxel spacing; Step S13: Use a pre-trained 3D U-Net model to automatically segment a pair of chest CT images and generate a lung lobe mask including background, left upper lobe, left lower lobe, right upper lobe, right middle lobe, and right lower lobe. Step S14: Simultaneously use another pre-trained 3D U-Net model II, take the CT image within the generated lung lobe partition mask as input, and output the binary mask of the ILA lesion area including the background and the lesion. Step S15: Using the lung lobe partition mask, create a left lung mask and a right lung mask in two directions, left and right respectively. Perform morphological erosion on the left lung mask and the right lung mask respectively, using a 10mm diameter spherical structure for erosion. Subtract the left lung mask and the right lung mask after one morphological erosion from the original left lung mask and the right lung mask without morphological erosion to obtain a 10mm subpleural mask for the left lung and a 10mm subpleural mask for the right lung.

3. The lung interstitial imaging analysis method based on clinically prior-guided feature fusion according to claim 2, characterized in that: Specific indicators corresponding to age include: using the current age in months / (100) 12); The specific indicators corresponding to gender include: male and female; The specific indicators corresponding to the smoking index include: number of packs smoked per day. Years of smoking; The specific indicators corresponding to lung function indicators include: based on age, height, and gender, the predicted lung function value is calculated using the GLI-2012 Southeast Asia equation, calculated as: (user's measured lung function value / predicted value). 100%; The specific indicators corresponding to the symptoms include: no symptoms, no change in symptoms, worsening, and relief.

4. The lung interstitial imaging analysis method based on clinically prior-guided feature fusion according to claim 3, characterized in that: The total image feature vector is represented as The temporal embedding vector is represented as They are represented as follows: By introducing a linear projection layer Its parameters are trainable matrices With bias vector The total image feature vector Mapped to In 3D space, the total image feature vector after projection alignment is obtained. The resulting total image feature vector With temporal embedding vector Element-wise addition generates time-coded fusion features Based on the generated time-coded fusion features Obtain the time-coded feature sequence , .

5. The lung interstitial imaging analysis method based on clinically prior-guided feature fusion according to claim 4, characterized in that: Step S5, the progress probability calculation, includes: inputting the output fused features into a fully connected neural network, and calculating the output progress probability after sigmoid activation. Progress probability The closer the value is to 1, the higher the risk of progress; the closer it is to 0, the lower the risk.

6. A lung interstitial imaging analysis system based on clinically prior-guided feature fusion, characterized in that: The system comprises a three-layer architecture consisting of a basic data layer, a business logic layer, and a data management and interaction layer, to execute the lung interstitial imaging analysis method based on clinical prior-guided feature fusion as described in any one of claims 1-5.

Images