An AI-based HRCT image analysis method and system for interstitial lung disease

Abstract

This invention relates to the field of medical informatics technology, providing an artificial intelligence-based method and system for HRCT image analysis of interstitial lung disease. The method includes: segmenting the lung parenchyma from input HRCT images to obtain lung parenchyma regions; segmenting the lung parenchyma regions into multiple lesion categories to obtain a segmentation mask; quantifying the lesion distribution characteristics based on the segmentation mask to obtain lesion distribution parameters; and classifying image patterns based on the segmentation mask and the lesion distribution parameters to obtain classification results. This invention improves the efficiency and accuracy of HRCT image analysis and expands its applicability.

Description

Technical Field

[0001] This invention belongs to the field of medical informatics technology, and in particular relates to an artificial intelligence-based HRCT image analysis method and system for interstitial lung disease. Background Technology

[0002] Interstitial lung disease (ILD) is a heterogeneous group of diseases characterized by chronic inflammation and fibrosis in the lung interstitium. Its diagnosis and assessment heavily rely on high-resolution CT (HRCT) imaging analysis of the chest. ILD chest HRCT imaging presents with complex features; common lesions include reticular formation, honeycomb formation, ground-glass opacities, and consolidation. Overall imaging patterns can be categorized into common patterns such as uncommon interstitial pneumonia (UIP), nonspecific interstitial pneumonia (NSIP), and organizing pneumonia (OP). Currently, traditional computer-aided lesion quantification tools are primarily based on quantitative computed tomography (QCT) technology, using image segmentation, texture analysis, and density measurement to identify and quantify lesions.

[0003] In recent years, artificial intelligence (AI) deep learning technology has begun to be applied to the analysis of ILD chest HRCT images, mainly including lesion identification, segmentation and quantification technology, using convolutional neural networks (CNN), U-Net and its improved versions to automatically identify and segment lesions and measure lesion volume from ILD chest HRCT images; and image pattern classification technology, using classification models such as support vector machines, random forests, deep learning models to classify the overall image pattern. Currently, there are studies that have applied deep learning models to classify UIP image patterns, with an accuracy similar to or slightly higher than that of radiologists.

[0004] However, the application of existing AI technologies in the field of ILD still has significant shortcomings: First, existing lesion segmentation models focus on single tasks, such as only being able to identify fibrotic lesions like honeycomb and reticular formations, lacking the ability to simultaneously detect other common ILD lesions such as ground-glass opacities and consolidation, and thus failing to provide comprehensive lesion information; second, existing models only perform simple detection and quantification of ILD lesions, failing to analyze the spatial distribution characteristics of lesions in the chest, such as upper and lower lung distribution, axial distribution, and subpleural distribution, which are crucial diagnostic information for ILD imaging diagnosis; finally, existing image pattern classification methods are mainly based on single deep learning or machine learning models, lacking rule engines that incorporate clinical diagnostic criteria, and are unable to handle rule conflicts, resulting in unstable classification results, insufficient clinical interpretability, and difficulty in being trusted and applied by clinicians. Summary of the Invention

[0005] In view of the shortcomings of the existing technology, the purpose of the invention is to provide an artificial intelligence-based HRCT image analysis method and system for interstitial lung disease.

[0006] This invention provides an artificial intelligence-based HRCT image analysis method for interstitial lung disease, comprising:

[0007] S1: Perform lung parenchyma segmentation on the input HRCT images to obtain the lung parenchyma region;

[0008] S2: Perform multi-category lesion segmentation on the lung parenchyma region to obtain a segmentation mask;

[0009] S3: Quantitatively calculate the lesion distribution characteristics based on the segmentation mask to obtain lesion distribution parameters;

[0010] S4: Perform image pattern classification based on the segmentation mask and the lesion distribution parameters to obtain the classification result.

[0011] In another optional embodiment, the AI-based HRCT image analysis method for interstitial lung disease of the present invention may further include:

[0012] S2': An encoder based on a pre-trained general model for medical image segmentation extracts three-dimensional features from the lung parenchyma region to obtain slice features;

[0013] S4': Based on the slice features, end-to-end image pattern classification of 3D HRCT is performed through an image pattern classification network to obtain patient-level image pattern classification results.

[0014] Step S2' includes: based on multiple HRCT slices of the patient's lung parenchyma region, inputting them into a pre-trained general model encoder for medical image segmentation to extract features, thereby obtaining slice features corresponding to multiple slices.

[0015] Step S4' includes:

[0016] S4'1: Perform patient-level feature aggregation on the slice features obtained in step S2' to obtain a patient-level representation vector;

[0017] S4'2: Input the patient-level representation vector into the image pattern classification network for classification processing to obtain the patient-level image pattern classification result.

[0018] According to the artificial intelligence-based HRCT image analysis method for interstitial lung disease provided by the present invention, step S1 further includes:

[0019] S11: The HRCT image is binarized using a preset density threshold to obtain an initial lung parenchyma mask;

[0020] S12: The initial lung parenchyma mask is smoothed and optimized by morphological opening and closing operations to obtain the lung parenchyma region.

[0021] According to the artificial intelligence-based HRCT image analysis method for interstitial lung disease provided by the present invention, step S2 further includes:

[0022] S21: Detect candidate regions of the lung parenchyma region using a prompt generator to obtain bounding box prompts;

[0023] S22: Based on the medical image segmentation model, the lung parenchyma region is segmented according to the bounding box prompts to obtain a segmentation mask; wherein, the medical image segmentation model is the MedSAM2 segmentation network.

[0024] According to the artificial intelligence-based HRCT image analysis method for interstitial lung disease provided by the present invention, the prompt generator in step S21 includes:

[0025] A first processing path and a second processing path are configured in parallel.

[0026] The first processing path automatically extracts bounding boxes using labeled masks to obtain precise bounding box hints; the second processing path performs coarse bounding box generation on unlabeled data using density threshold filtering and connected component analysis to obtain candidate bounding box hints.

[0027] According to the present invention, an artificial intelligence-based HRCT image analysis method for interstitial lung disease is provided, wherein the segmentation mask in step S22 includes: honeycomb shadow, mesh shadow, ground-glass shadow, and solid shadow.

[0028] According to the artificial intelligence-based HRCT image analysis method for interstitial lung disease provided by the present invention, the lesion distribution parameters in step S3 include: upper and lower lung distribution parameters, axial distribution parameters, and subpleural distribution parameters.

[0029] According to the artificial intelligence-based HRCT image analysis method for interstitial lung disease provided by the present invention, step S3 further includes:

[0030] The calculation method for the upper and lower lung distribution parameters includes:

[0031] S311: Using the cephalopod bisection method, the upper and lower lung regions are divided according to the segmentation mask to obtain the proportion of upper lung lesions and the proportion of lower lung lesions.

[0032] S312: The distribution parameters of the upper and lower lungs are calculated from the proportion of lesions in the upper lung and the proportion of lesions in the lower lung;

[0033] The calculation method for the axial distribution parameter includes:

[0034] S321: Using the Euclidean distance calculation method, the chest wall distance is measured based on the lesion area in the segmentation mask to obtain the shortest distance from the lesion to the chest wall;

[0035] S322: A first distance threshold is preset, and the axial distribution type is determined based on the shortest distance to obtain the axial distribution parameters;

[0036] The calculation method for the subpleural distribution parameters includes:

[0037] S331: Perform statistical analysis of the lung contour's adjacent region based on the segmentation mask to obtain the lesion percentage within the target region; the target region is the region whose distance from the lung contour is less than or equal to a second distance threshold;

[0038] S332: The subpleural distribution parameters are obtained by determining the proportion of the lesions.

[0039] According to the artificial intelligence-based HRCT image analysis method for interstitial lung disease provided by the present invention, step S4 further includes:

[0040] S41: Based on the segmentation mask and the lesion distribution parameters, perform preliminary classification of image patterns using a preset clinical imaging mode to obtain preliminary judgment results;

[0041] S42: When there is a rule conflict in the preliminary judgment result, the image pattern is optimized and classified by the Logistic regression model to obtain the optimized judgment result;

[0042] S43: Integrate the preliminary judgment result and the optimized judgment result, and output the classification result of the image modes including UIP mode, NSIP mode and OP mode.

[0043] According to the present invention, a method for analyzing interstitial lung disease using high-resolution computed tomography (HRCT) images based on artificial intelligence, step S41 specifically includes:

[0044] When reticular and honeycomb shadows are detected but no ground-glass opacities or consolidation are found, the UIP pattern is determined by combining peripheral distribution, lower lung dominance, and subpleural signs to obtain a preliminary determination result of the UIP pattern.

[0045] When a reticular pattern without honeycomb or consolidation is detected, the NSIP pattern is determined by combining the lower lung dominance and subpleural omission signs, and a preliminary determination result of the NSIP pattern is obtained.

[0046] When ground glass shadow or solid shadow is detected but there is no grid shadow or honeycomb shadow, OP mode determination is performed to obtain the preliminary determination result of OP mode.

[0047] This invention also provides an artificial intelligence-based HRCT image analysis system for interstitial lung disease, comprising:

[0048] The first segmentation module is used to segment the lung parenchyma of the input HRCT images to obtain the lung parenchyma region.

[0049] The second segmentation module is used to segment the lung parenchyma region into multiple categories of lesions to obtain a segmentation mask.

[0050] Calculation module: used to quantify the lesion distribution characteristics based on the segmentation mask to obtain lesion distribution parameters;

[0051] Classification module: used to classify image patterns based on the segmentation mask and the lesion distribution parameters, and obtain classification results.

[0052] This invention provides an AI-based HRCT image analysis method and system for interstitial lung disease (ILD). First, it segments the lung parenchyma using a density thresholding method combined with morphological manipulation, accurately removing interference from non-lung tissues such as the chest wall and mediastinum. This provides accurate regions of interest for subsequent lesion analysis, significantly reducing the false positive rate and greatly improving the accuracy of lesion identification. Second, this invention employs the MedSAM2 segmentation network for multi-category lesion segmentation, achieving simultaneous identification and accurate segmentation of four common ILD lesions: honeycomb, reticular, ground-glass opacities, and consolidation. This overcomes the limitations of existing technologies that can only identify single or partial lesion types, greatly expanding the scope of clinical application. It allows for comprehensive lesion information to be obtained in a single test, significantly improving diagnostic efficiency and reducing the risk of missed diagnoses. Furthermore, this invention... The distribution parameters of the upper and lower lungs, axial distribution parameters, and subpleural distribution parameters obtained through overquantitative calculations compensate for the significant deficiency of existing technologies in lacking spatial distribution analysis of lesions, providing clinicians with crucial localization information. These parameters are of great value in identifying ILD subtypes and can significantly improve the accuracy and reliability of diagnosis. Finally, this invention also uses a fusion classification method combining a medical rule engine and a probabilistic reasoning network to ensure that the classification results conform to clinical diagnostic standards and logic. Furthermore, probabilistic reasoning effectively handles rule conflicts, making the classification results more stable and reliable. At the same time, the generated interpretable diagnostic report contains detailed quantitative analysis evidence, greatly enhancing the transparency and clinical credibility of AI decision-making, effectively solving the black box problem of traditional AI, and promoting the widespread application of artificial intelligence technology in clinical practice and its acceptance by doctors. Attached Figure Description

[0053] The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Throughout the drawings, the same reference numerals denote the same parts. It is obvious that the drawings described below are merely some embodiments of the present invention, and those skilled in the art can obtain other drawings based on these drawings.

[0054] Figure 1 A schematic diagram of a method for analyzing interstitial lung disease using high-resolution computed tomography (HRCT) images based on artificial intelligence, provided in an embodiment of the present invention;

[0055] Figure 2 A schematic diagram of a method for analyzing interstitial lung disease using high-resolution computed tomography (HRCT) images based on artificial intelligence, provided as another embodiment of the present invention;

[0056] Figure 3 This is a schematic diagram of an AI-based HRCT image analysis system for interstitial lung disease, provided as an embodiment of the present invention.

[0057] Figure label:

[0058] 100. First segmentation module; 200. Second segmentation module; 300. Calculation module; 400. Classification module. Detailed Implementation

[0059] To enable those skilled in the art to better understand the technical solutions in the embodiments of the present invention, the technical solutions 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, not all embodiments. It should be understood that these descriptions are merely exemplary and are not intended to limit the scope of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0060] Furthermore, descriptions of well-known structures and techniques are omitted in the following description to avoid unnecessarily obscuring the concepts disclosed in this invention.

[0061] In the description of this invention, it should be noted that, unless otherwise explicitly specified and limited, the terms "center," "upper," "lower," "left," "right," "vertical," "horizontal," "inner," and "outer," etc., indicating orientation or positional relationships based on the orientation or positional relationships shown in the accompanying drawings, are only for the convenience of describing the invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of the invention. Furthermore, the terms "first," "second," and "third" are used for descriptive purposes only and should not be construed as indicating or implying relative importance. The terms "installed," "connected," and "linked" should be interpreted broadly; for example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal communication of two components. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0062] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numerals in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with the present invention. Rather, they are merely examples of methods and systems consistent with some aspects of the invention as detailed in the appended claims.

[0063] To better understand this invention, the terms used in this invention will be explained in detail below.

[0064] ILD (Interstitial Lung Disease): A heterogeneous group of diseases characterized by chronic inflammation and fibrosis in the interstitial lungs.

[0065] HRCT (High-Resolution Computed Tomography): A CT technique used for imaging the fine structures of the lungs.

[0066] UIP (Usual Interstitial Pneumonia): An imaging pattern of interstitial lung disease characterized by honeycomb or reticular shadows, often distributed subpleurally.

[0067] NSIP (Non-Specific Interstitial Pneumonia): Another imaging modality of interstitial lung disease, characterized by ground-glass opacities and reticular formations.

[0068] OP (Organizing Pneumonia): One of the imaging patterns of interstitial lung disease, characterized by consolidation and ground-glass opacities.

[0069] HP (Hypersensitivity Pneumonitis): an interstitial lung disease caused by inhalation of organic dust.

[0070] MedSAM2: A medical image segmentation model based on the SAM (Segment Anything Model) architecture.

[0071] GGO (Ground Glass Opacity): A lesion on CT images that shows a slight increase in density but still allows for the observation of vascular patterns.

[0072] RT (reticulation): The lung parenchyma shows crisscrossing linear high-density shadows, resembling a reticular pattern, suggesting thickening of the interlobular septa or interstitial fibrosis.

[0073] HC (honeycombing): Subpleural clusters of 3–10 mm thick-walled cysts stacked one on top of another, reflecting end-stage fibrosis (UIP characteristic).

[0074] CS (consolidation): The acini are filled with exudate or cells, resulting in a uniform increase in density, which obscures blood vessels and bronchi and is often accompanied by air bronchogram signs.

[0075] 3D HRCT: Three-dimensional high-resolution computed tomography scan, which can provide three-dimensional imaging information of the lungs.

[0076] ROI (Region of Interest): The region that needs to be analyzed in image processing.

[0077] The embodiments of the present invention are described below with reference to the figures.

[0078] like Figure 1 As shown, this invention provides an artificial intelligence-based HRCT image analysis method for interstitial lung disease, comprising:

[0079] S1: Perform lung parenchyma segmentation on the input HRCT image to obtain the lung parenchyma region.

[0080] Furthermore, HRCT (High-Resolution Computed Tomography) is a CT technology used for imaging the fine structure of the lungs, which can provide three-dimensional image information of the lungs. In step S1 of this invention, lung parenchyma segmentation is performed on the input three-dimensional HRCT image, with the goal of accurately extracting the lung parenchyma region from the original CT image containing various tissue structures such as the chest wall, mediastinum, and heart.

[0081] Step S1 further includes:

[0082] S11: The HRCT image is binarized and segmented using a preset density threshold to obtain the initial lung parenchyma mask.

[0083] In step S11, the input three-dimensional HRCT image data is first read, which is a volumetric image in DICOM format. Then, each voxel in the image is traversed to determine whether its density value falls within the range of [-1024, -200]. If the voxel density value is within this range, it is marked as 1 (foreground, representing lung tissue); if it is outside this range, it is marked as 0 (background, representing non-lung tissue), and finally the initial lung parenchyma mask is formed.

[0084] S12: The initial lung parenchyma mask is smoothed and optimized by morphological opening and closing operations to obtain the lung parenchyma region.

[0085] Morphological operations are image processing methods based on set theory, used to improve the shape features of binary images. Opening operation involves performing erosion followed by dilation, primarily used to remove small noise points and fill internal holes. Closing operation involves performing dilation followed by erosion, primarily used to connect neighboring regions and smooth boundaries.

[0086] In the processing, a cubic convolution kernel of the structuring element is first defined. An opening operation is performed on the initial lung parenchyma mask. For each voxel position in the mask, an erosion operation is performed within a 3×3×3 neighborhood centered on that position. Specifically, it checks if all voxel values ​​in that neighborhood are 1; if so, the current voxel value is kept at 1, otherwise it is set to 0. After erosion, a dilation operation is performed on the result. If at least one voxel with a value of 1 exists within the 3×3×3 neighborhood of each voxel, the central voxel is set to 1. Next, a closing operation is performed. The opening result is first dilated, then eroded, with the data processing method reversed compared to the opening operation. Finally, a smoothed and optimized lung parenchyma mask is obtained. The obtained mask removes small noise regions generated by density threshold segmentation, fills in small voids inside the lung parenchyma, and forms a continuous and complete lung parenchyma contour.

[0087] S2: Perform multi-category lesion segmentation on the lung parenchyma region to obtain a segmentation mask.

[0088] Furthermore, ILD (Interstitial Lung Disease) is a group of heterogeneous diseases characterized by chronic inflammation and fibrosis in the lung interstitium. In this step, we perform multi-category lesion identification and segmentation on the lung parenchyma region obtained in S1 and extract the spatial distribution information of different types of lesions.

[0089] Step S2 further includes:

[0090] S21: The candidate regions of the lung parenchyma are detected by the prompt generator to obtain bounding box prompts.

[0091] Furthermore, the cue generator is a preprocessing module that provides initial localization information for the segmentation network. It guides the subsequent accurate segmentation process by generating bounding box cue, and the obtained bounding box cue is the smallest rectangle containing the target region.

[0092] The prompt generator in step S21 includes:

[0093] A first processing path and a second processing path are configured in parallel. The first processing path automatically extracts bounding boxes using labeled masks to obtain precise bounding box hints. The second processing path performs coarse bounding box generation on unlabeled data using density threshold filtering and connected component analysis to obtain candidate bounding box hints.

[0094] For the first processing path, when the input data contains expert-annotated lesion masks, this path is used to generate precise bounding box cues. Specifically, the annotated lesion mask data is first read. Different lesion types in the mask are labeled with different values ​​(e.g., honeycomb shadows are labeled as 1, mesh shadows as 2, ground-glass shadows as 3, and solid shadows as 4). For each lesion type, the set of voxel coordinates of all values ​​equal to the specified values ​​in the mask is extracted. The minimum and maximum values ​​of the x-coordinate, y-coordinate, and z-coordinate in this set are calculated to generate a 3D bounding box as a precise bounding box cue for that lesion type.

[0095] For the second processing path, when the input data lacks expert annotation, this path automatically generates candidate bounding boxes based on image features. During density threshold filtering, threshold ranges are set according to the density characteristics of different lesion types: ground-glass opacity density range is [-700, -300] Ω, mesh and honeycomb opacity density range is [-200, 200] Ω, and solid opacity density range is [-100, 100] Ω. First, each voxel in the lung parenchyma region is traversed to determine whether its density value falls within the threshold range of a specific lesion type, generating a preliminary candidate region mask. Subsequently, connected component analysis is performed. During this process, the 26-adjacency connectivity rule is used to mark the candidate region mask as a connected component. Each foreground voxel in the mask is traversed, and a breadth-first search algorithm is used to find all voxels connected to it, forming a connected component. The volume (number of voxels contained) of each connected component is calculated, and a minimum volume threshold of 50 voxels is set. Connected components smaller than this threshold are filtered out to reduce noise. For each retained connected component, its minimum bounding cube is calculated as a candidate bounding box cue.

[0096] S22: Based on the MedSAM2 segmentation network, the lung parenchyma region is segmented according to the bounding box prompts to obtain a segmentation mask.

[0097] Furthermore, MedSAM2 is a medical image segmentation model based on the SAM (Segment Anything Model) architecture. It uses the Transformer neural network architecture for end-to-end image segmentation. The MedSAM2 model was fine-tuned and trained on more than 2,400 labeled images containing cellular shadows, mesh shadows, ground-glass shadows, and solid shadows.

[0098] In the segmentation, the lung parenchyma region image obtained in S1 and the bounding box cue generated in S21 are input into the network. First, the image is standardized by mapping the voxel density values ​​from the original range to the [0,1] interval. The 3D image is resampled to a uniform resolution. The input data with different layer thicknesses are processed using trilinear interpolation. For each bounding box cue, the corresponding region of interest is extracted from the image as the network input.

[0099] The subsequent forward propagation involves three main components of the MedSAM2 network: an image encoder, a cue encoder, and a mask decoder. The image encoder uses a Vision Transformer architecture, which divides the input 3D ROI into several patches, flattens each patch into a one-dimensional vector and adds position encoding, and inputs it into a multi-layer Transformer block for feature extraction.

[0100] During extraction, the cue encoder processes bounding box cue information, encoding the bounding box coordinates into a fixed-length feature vector. The mask decoder fuses image features and cue features, generating a segmentation prediction result with the same size as the input ROI through multiple convolutional and upsampling operations.

[0101] Finally, multi-category output is performed. The network output layer contains four independent segmentation heads, corresponding to four lesion types: cellular shadow, reticular shadow, ground-glass opacity, and consolidation. Each segmentation head outputs a probability map. This invention applies a threshold of 0.5 to binarize the probability maps to generate the final segmentation mask. After generation, the segmentation results of the four categories are merged to form a comprehensive segmentation mask.

[0102] The segmentation mask in step S22 includes: honeycomb shadow, mesh shadow, ground glass shadow, and solid transformation shadow.

[0103] Furthermore, the four lesion types included in the segmentation mask are defined as follows: honeycomb (HC) is a cluster of 3-10mm thick-walled cystic cavities superimposed under the pleura, reflecting end-stage fibrosis; reticular (RT) is a crisscrossing linear high-density shadow in the lung parenchyma, suggesting thickening of interlobular septa or interstitial fibrosis; ground-glass opacity (GGO) is a lesion with slightly increased density on CT images but still showing vascular texture; and consolidation (CS) is a uniform increase in density caused by acini filled with exudate or cells, obscuring blood vessels and bronchi, and often accompanied by air bronchograms.

[0104] S3: Quantitatively calculate the lesion distribution characteristics based on the segmentation mask to obtain lesion distribution parameters.

[0105] The lesion distribution parameters in step S3 include: upper and lower lung distribution parameters, axial distribution parameters, and subpleural distribution parameters.

[0106] Furthermore, lesion distribution characteristics are a key imaging indicator for the diagnosis of ILD. Different types of interstitial lung diseases exhibit specific distribution patterns in the lungs. This step uses the segmentation mask obtained by S2 to extract the spatial distribution pattern of lesions through mathematical calculation methods, generating quantitative lesion distribution parameters. The lesion distribution parameters include quantitative indicators in three dimensions: upper and lower lung distribution parameters, axial distribution parameters, and subpleural distribution parameters.

[0107] The calculation method for the upper and lower lung distribution parameters includes:

[0108] S311: Using the cephalopod bisection method, the upper and lower lung regions are divided according to the segmentation mask to obtain the proportion of upper lung lesions and the proportion of lower lung lesions.

[0109] Furthermore, the cephalopod direction is the anatomical direction of the human body from head to feet in medical imaging, corresponding to the Z-axis direction of CT images. The bisection method is a mathematical method that divides the lungs into two equal-volume regions along the cephalopod direction for analysis.

[0110] In the specific data processing, the lung parenchyma region mask obtained from S1 is read. This mask is a three-dimensional binary matrix. The boundary coordinates of the lung parenchyma region in the Z-axis direction are calculated. All voxels on each Z-level of the mask are traversed to find the minimum and maximum Z coordinates containing lung tissue. The overall height of the lung is calculated, and the midpoint position is determined.

[0111] Based on the midpoint location, the lung parenchyma region is divided into two parts: the upper lung region is defined as all lung tissue voxels within the Z-coordinate range, and the lower lung region is defined as all lung tissue voxels within the Z-coordinate range, generating upper and lower lung masks.

[0112] Read the comprehensive segmentation mask obtained from S2, calculate the lesion distribution in each region. For the upper lung lesion mask, extract all lesions in the upper lung region through voxel-by-voxel logical AND operation. For the lower lung lesion mask, first extract all lesions in the lower lung region, count the total number of voxels of upper lung lesions and lower lung lesions, and finally calculate the proportion of upper lung lesions and the proportion of lower lung lesions.

[0113] S312: The distribution parameters of the upper and lower lungs are calculated from the proportion of lesions in the upper lung and the proportion of lesions in the lower lung.

[0114] Based on the proportion of upper and lower lung lesions, quantitative upper and lower lung distribution parameters are obtained through mathematical calculations, and a judgment threshold of 0.6 is set as the critical value for the dominant distribution.

[0115] In the data processing logic, if the proportion of upper lung lesions is greater than the threshold, the upper and lower lung distribution parameter is set to "upper lung dominance" with a numerical code of 2. If the proportion of lower lung lesions is greater than the threshold, the upper and lower lung distribution parameter is set to "lower lung dominance" with a numerical code of 1. If 0.4 ≤ upper lung lesions ≤ 0.6 and 0.6 and 0.4 ≤ lower lung lesions ≤ 0.6, the upper and lower lung distribution parameter is set to "uniform distribution" with a numerical code of 0. At the same time, the upper and lower lung distribution ratio is calculated. This ratio reflects the degree of distribution preference of lesions in the vertical direction. When the ratio is > 1, it indicates that there are more upper lung lesions. When the ratio is < 1, it indicates that there are more lower lung lesions. When the ratio is = 1, it indicates that the upper and lower lung lesions are evenly distributed.

[0116] The calculation method for the axial distribution parameter includes:

[0117] S321: Using the Euclidean distance calculation method, the distance from the lesion to the chest wall is measured based on the lesion area in the segmentation mask to obtain the shortest distance from the lesion to the chest wall.

[0118] Furthermore, step S321, which aims to measure the distance to the chest wall, involves calculating the shortest spatial distance from each voxel within the lesion region to the lung contour boundary. Specifically, during data processing, the lung contour boundary is first extracted based on a lung parenchyma mask. Then, a set of boundary point coordinates is extracted, containing the three-dimensional coordinates of all lung contour boundary voxels. Subsequently, the above calculation is performed on all lesion voxels in the segmentation mask to generate a distance map, where each lesion voxel's position corresponds to its shortest distance to the chest wall. Finally, the minimum distance value is selected as the shortest distance from the lesion voxel to the chest wall.

[0119] S322: A first distance threshold is preset, and the axial distribution type is determined based on the shortest distance to obtain the axial distribution parameters.

[0120] A first distance threshold of 30mm is preset as the dividing standard between peripheral and central distributions, and the threshold is determined based on clinical imaging standards.

[0121] During data processing, the number of lesion voxels within different distance ranges in the distance map is counted. The number of peripheral lesion voxels and the number of central lesion voxels are calculated, and then the proportion of peripheral lesions and the proportion of central lesions are calculated.

[0122] The distribution type determination threshold is set to 0.7. In determining the axial distribution parameter, if the proportion of peripheral lesions is greater than the distribution type determination threshold, then the axial distribution parameter = "peripheral type", with a numerical code of 1; if the proportion of central lesions is greater than or equal to the distribution type determination threshold, then the axial distribution parameter = "central type", with a numerical code of 2; if the proportion of peripheral lesions is less than the distribution type determination threshold and the proportion of central lesions is less than the distribution type determination threshold, then the axial distribution parameter = "diffuse type", with a numerical code of 0. Simultaneously, the average chest wall distance is calculated.

[0123] The calculation method for the subpleural distribution parameters includes:

[0124] S331: Perform statistical analysis of the lung contour's adjacent region based on the segmentation mask to obtain the lesion percentage within the target region; the target region is the region whose distance from the lung contour is less than or equal to a second distance threshold.

[0125] Furthermore, subpleural distribution refers to the distribution pattern of lesions mainly located in areas close to the lung surface. It is a characteristic imaging manifestation of diseases such as UIP (ultra-interstitial pneumonia). The second distance threshold of 3 mm is preset as the definition standard for the subpleural region.

[0126] During data processing, the subpleural region mask is first defined based on the distance map calculated by S321. Then, the distribution of lesions in the subpleural region is statistically analyzed, the subpleural lesion mask is calculated, and the number of subpleural lesion voxels, the total number of lesion voxels, and the proportion of lesions in the target region are calculated. The proportion of lesions in the target region reflects the degree of concentration of lesions in the subpleural region.

[0127] S332: The subpleural distribution parameters are obtained by determining the proportion of the lesions.

[0128] Furthermore, in step S332, the present invention obtains the subpleural distribution parameters based on the proportion of lesions through a threshold determination method. The subpleural dominance determination threshold is set to 0.3, which is determined according to the statistical characteristics of subpleural distribution in clinical studies.

[0129] In the logic for determining the subpleural distribution parameter, if the proportion of lesions in the target area is greater than or equal to the subpleural dominance threshold, then the subpleural distribution parameter = "subpleural sign", with a numerical code of 1, indicating that the lesion has obvious subpleural distribution characteristics; if the proportion of lesions in the target area is less than the subpleural dominance threshold, then the subpleural distribution parameter = "subpleural omission", with a numerical code of 0, indicating that the lesion does not show a subpleural dominance distribution.

[0130] The subpleural distribution intensity index is then calculated, which numerically describes the degree of subpleural distribution. A subpleural distribution intensity index ≥ 1 indicates significant subpleural distribution, with higher values ​​indicating more concentrated distribution; a subpleural distribution intensity index < 1 indicates insignificant subpleural distribution. The final output is a vector of lesion distribution parameters, containing qualitative and quantitative descriptions of upper and lower lung distribution, axial distribution, and subpleural distribution.

[0131] S4: Perform image pattern classification based on the segmentation mask and the lesion distribution parameters to obtain the classification result.

[0132] Step S4 further includes:

[0133] S41: Based on the segmentation mask and the lesion distribution parameters, perform preliminary classification of image patterns using a preset clinical imaging mode to obtain preliminary judgment results.

[0134] Specifically, step S41 includes:

[0135] When reticular and honeycomb shadows are detected but no ground-glass opacities or consolidation are present, the UIP pattern is determined by combining peripheral distribution, lower lung dominance, and subpleural signs to obtain a preliminary UIP pattern determination result; when reticular shadows are detected but no honeycomb or consolidation are present, the NSIP pattern is determined by combining lower lung dominance and subpleural omission signs to obtain a preliminary NSIP pattern determination result; when ground-glass opacities or consolidation are detected but no reticular or honeycomb shadows are present, the OP pattern is determined to obtain a preliminary OP pattern determination result.

[0136] S42: When there is a rule conflict in the preliminary judgment result, the image pattern is optimized and classified by the Logistic regression model to obtain the optimized judgment result.

[0137] When multiple patterns in step S41 simultaneously meet the judgment criteria, the rule conflict handling mechanism is triggered. In the Logistic Regression model, feature standardization is first performed, and a standardized value is calculated for each feature. The standardized feature vector is then input into the Logistic Regression model, and the probability is calculated using the softmax function. The pattern with the highest probability is selected as the optimized judgment result. Finally, the confidence score is evaluated based on the entropy value of the probability distribution. When the difference between the highest probability and the second highest probability is greater than 0.3, the confidence score is adjusted by an additional weight of 0.1.

[0138] S43: Integrate the preliminary judgment result and the optimized judgment result, and output the classification result of the image modes including UIP mode, NSIP mode and OP mode.

[0139] Finally, the results are fused using a weighted voting mechanism to process the outputs of rule-based decision-making and logistic regression. The rule-based decision-making results are assigned a weight of 0.7, reflecting the importance of clinical expert knowledge, while the logistic regression results are assigned a weight of 0.3. When the two methods output the same classification label, that result is output directly. When the outputs are inconsistent, the confidence differences between the two methods are compared. If the confidence score is < 0.2, the result with the higher probability is output and its confidence score is used. If the confidence score is ≥ 0.2, the result of the high-confidence method is output and its confidence score is used. The final output is structured in JSON format, containing the predicted_class field storing the classification label ("UIP", "NSIP", "OP", "Other", or "Normal"), the confidence_score field storing the confidence score value (retaining three decimal places), the lesion_analysis field storing a nested dictionary containing the proportions of the four lesions, the distribution_features field storing the distribution feature parameters, and the decision_path field storing the text description of the decision path.

[0140] Steps S1-S4 of this invention rely on fine pixel-level lesion segmentation results. However, in actual clinical applications, the accuracy of lesion segmentation is highly limited by the quality and completeness of the labeled data, and the imaging patterns of some interstitial lung diseases are diffuse and have blurred boundaries. Relying solely on local lesion masks may not be able to fully capture the macroscopic pattern features of the entire lung. Therefore, this invention also provides an end-to-end image pattern classification pathway based on deep learning feature representation.

[0141] The following is combined Figure 2 This invention describes a parallel, end-to-end image pattern classification pathway based on deep learning feature representation. This pathway does not directly rely on pixel-level lesion segmentation results. Instead, it first extracts deep, three-dimensional slice-level features from the lung parenchyma region through the encoder of a pre-trained general medical image segmentation model. These features, after being abstracted by the encoder, can simultaneously encode the structural information of the lung parenchyma and the image pattern information related to potential lesions, providing richer and more advanced semantic representations.

[0142] S101: Perform lung parenchyma segmentation on the input HRCT image to obtain the lung parenchyma region.

[0143] S102: An encoder based on a pre-trained general model for medical image segmentation extracts three-dimensional features from the lung parenchyma region to obtain slice features.

[0144] Specifically, based on the HRCT slices of the patient's lung parenchyma region, each HRCT slice is input into the encoder of a pre-trained general medical image segmentation model for feature encoding and semantic representation extraction, generating slice-level feature representations that can characterize lung parenchyma structural information and lesion-related image pattern information, thereby obtaining slice features corresponding to each HRCT slice, providing basic input for subsequent patient-level feature aggregation and image pattern determination.

[0145] S103: Based on the slice features, perform end-to-end image pattern classification on the three-dimensional HRCT using an image pattern classification network to obtain patient-level image pattern classification results.

[0146] Step S103 includes:

[0147] S1031: Aggregate patient-level features on the slice features obtained in step S102 to obtain a patient-level representation vector; S1032: Input the patient-level representation vector into the image pattern classification network for classification processing to obtain the patient-level image pattern classification result.

[0148] Specifically, firstly, patient-level feature aggregation is performed on the slice features obtained in step S102. Based on the integrated slice features, a patient-level overall representation is constructed to obtain a patient-level representation vector that reflects the overall imaging features of the patient's lungs. Subsequently, the patient-level representation vector is input into an image pattern classification network for classification processing, and the image pattern classification result of the patient is output. The image pattern classification network can realize patient-level image pattern recognition based on discriminative learning of the patient-level representation vector, and can simultaneously output the confidence level of the image pattern determination to improve the stability and clinical applicability of the image pattern determination.

[0149] like Figure 3 As shown, the present invention also provides an artificial intelligence-based HRCT image analysis system for interstitial lung disease, comprising:

[0150] First segmentation module 100: used to perform lung parenchyma segmentation on the input HRCT image to obtain the lung parenchyma region;

[0151] Second segmentation module 200: used to segment the lung parenchyma region into multiple types of lesions to obtain a segmentation mask;

[0152] Calculation module 300: used to quantify the lesion distribution characteristics based on the segmentation mask to obtain lesion distribution parameters;

[0153] Classification module 400: used to classify image patterns according to the segmentation mask and the lesion distribution parameters, and obtain classification results.

[0154] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0155] This invention provides an AI-based HRCT image analysis method and system for interstitial lung disease, which improves model performance and accuracy, expands the applicability of the model, increases the dimensions of model analysis, and enhances the transparency and interpretability of model classification.

[0156] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the embodiments of the present invention, and are not intended to limit them. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the protection scope of the present invention.

Claims

1. An artificial intelligence-based interstitial lung disease HRCT image analysis method, characterized by, include: S1: Perform lung parenchyma segmentation on the input HRCT images to obtain the lung parenchyma region; S2: Perform multi-category lesion segmentation on the lung parenchyma region to obtain a segmentation mask; Step S2 further includes: S21: Detect candidate regions of the lung parenchyma region using a prompt generator to obtain bounding box prompts; The prompt generator in step S21 includes: a first processing path and a second processing path, wherein the first processing path and the second processing path are set in parallel; The first processing path automatically extracts bounding boxes using labeled masks to obtain precise bounding box hints; the second processing path performs coarse bounding box generation on unlabeled data through density threshold filtering and connected component analysis to obtain candidate bounding box hints. S22: Based on the medical image segmentation model, the lung parenchyma region is segmented according to the bounding box prompts to obtain a segmentation mask; S3: Quantitatively calculate the lesion distribution characteristics based on the segmentation mask to obtain lesion distribution parameters; the lesion distribution parameters in step S3 include: upper and lower lung distribution parameters, axial distribution parameters, and subpleural distribution parameters; S4: Perform image pattern classification based on the segmentation mask and the lesion distribution parameters to obtain the classification result; Step S4 further includes: S41: Based on the segmentation mask and the lesion distribution parameters, perform preliminary classification of image patterns using a preset clinical imaging mode to obtain preliminary judgment results; S42: When there is a rule conflict in the preliminary judgment result, the image pattern is optimized and classified by the Logistic regression model to obtain the optimized judgment result; S43: Integrate the preliminary judgment result and the optimized judgment result, and output the classification result of the image modes including UIP mode, NSIP mode and OP mode.

2. The method for analyzing interstitial lung disease using HRCT images based on artificial intelligence according to claim 1, characterized in that, Step S1 further includes: S11: Based on a preset density threshold, the HRCT image is binarized using a lung parenchyma segmentation model to obtain an initial lung parenchyma mask; S12: The initial lung parenchyma mask is smoothed and optimized by morphological opening and closing operations to obtain the lung parenchyma region.

3. The method for analyzing interstitial lung disease using HRCT images based on artificial intelligence according to claim 1, characterized in that, The segmentation mask in step S22 includes: honeycomb mask, mesh mask, ground glass mask, and solid transformation mask.

4. The method for analyzing interstitial lung disease using HRCT images based on artificial intelligence according to claim 1, characterized in that, Step S3 further includes: The calculation method for the upper and lower lung distribution parameters includes: S311: Using the cephalopod bisection method, the upper and lower lung regions are divided according to the segmentation mask to obtain the proportion of upper lung lesions and the proportion of lower lung lesions. S312: The distribution parameters of the upper and lower lungs are calculated from the proportion of lesions in the upper lung and the proportion of lesions in the lower lung; The calculation method for the axial distribution parameter includes: S321: Using the Euclidean distance calculation method, the chest wall distance is measured based on the lesion area in the segmentation mask to obtain the shortest distance from the lesion to the chest wall; S322: A first distance threshold is preset, and the axial distribution type is determined based on the shortest distance to obtain the axial distribution parameters; The calculation method for the subpleural distribution parameters includes: S331: Perform statistical analysis of the lung contour's adjacent region based on the segmentation mask to obtain the lesion percentage within the target region; the target region is the region whose distance from the lung contour is less than or equal to a second distance threshold; S332: The subpleural distribution parameters are obtained by determining the proportion of the lesions.

5. The method for analyzing interstitial lung disease using HRCT images based on artificial intelligence according to claim 1, characterized in that, Step S41 specifically includes: When reticular and honeycomb shadows are detected but no ground-glass opacities or consolidation are found, the UIP pattern is determined by combining peripheral distribution, lower lung dominance, and subpleural signs to obtain a preliminary determination result of the UIP pattern. When a reticular pattern without honeycomb or consolidation is detected, the NSIP pattern is determined by combining the lower lung dominance and subpleural omission signs, and a preliminary determination result of the NSIP pattern is obtained. When ground glass shadow or solid shadow is detected but there is no grid shadow or honeycomb shadow, OP mode determination is performed to obtain the preliminary determination result of OP mode.

6. An AI-based HRCT image analysis system for interstitial lung disease, characterized in that, include: The first segmentation module is used to segment the lung parenchyma of the input HRCT images to obtain the lung parenchyma region. The second segmentation module is used to segment the lung parenchyma region into multiple categories of lesions to obtain a segmentation mask. The second segmentation module is further configured to: detect candidate regions of the lung parenchyma region through a cue generator to obtain bounding box cue; Based on the medical image segmentation model, the lung parenchyma region is segmented according to the bounding box prompts to obtain a segmentation mask; The prompt generator includes: a first processing path and a second processing path, wherein the first processing path and the second processing path are set in parallel. The first processing path automatically extracts bounding boxes using an labeled mask to obtain precise bounding box prompts; the second processing path performs coarse bounding box generation on unlabeled data using density threshold filtering and connected component analysis to obtain candidate bounding box prompts; the calculation module is used to quantify the lesion distribution features based on the segmentation mask to obtain lesion distribution parameters; the lesion distribution parameters include: upper and lower lung distribution parameters, axial distribution parameters, and subpleural distribution parameters. Classification module: used to classify image patterns based on the segmentation mask and the lesion distribution parameters, and obtain classification results; The classification module is further used to: perform preliminary classification of image patterns according to the segmentation mask and the lesion distribution parameters through a preset clinical image pattern to obtain a preliminary judgment result; when there is a rule conflict in the preliminary judgment result, perform optimized classification of image patterns through a Logistic regression model to obtain an optimized judgment result; and fuse the preliminary judgment result and the optimized judgment result to output the classification result of image patterns including UIP mode, NSIP mode and OP mode.

Images