Diffuse interstitial lung disease texture recognition model and implementation method thereof

Abstract

The invention discloses a diffuse interstitial lung disease texture recognition model and an implementation method thereof, and relates to the technical field of lung disease texture recognition, the diffuse interstitial lung disease texture recognition model comprises a PACS database, a training server, an application server, a plurality of data processing units and a user side; the training server is connected with the PACS database through the Ethernet, and meanwhile, the training server interacts with the user side through the Ethernet to sketch a volume of interest (VOI) for model training; the training server and the application server are connected with the plurality of data processing units through the Ethernet, and the data processing units are connected with the image reconstruction work stations one by one through the Ethernet; according to the invention, on the basis of a fully-supervised prior lung texture segmentation training set, classification fineness which cannot be reached by an unsupervised clustering scheme is realized; compared with an existing full supervision scheme, the method achieves the improvement of classification accuracy and segmentation precision through a deep learning network instead of a lung density quantitative segmentation scheme.

Description

Technical Field

[0001] The present invention relates to the technical field of lung disease texture recognition, and in particular to a diffuse interstitial lung disease texture recognition model and an implementation method thereof. Background Art

[0002] Interstitial lung disease is a general term for a class of highly heterogeneous diseases with complex causes, pathological changes, and clinical manifestations, which seriously hinders the diagnosis and evaluation of the disease. Chest CT imaging is an important method for diagnosing interstitial lung disease, but the current mainstream method is still subjective qualitative assessment. CT lung texture analysis uses local or global texture characteristics of the lung parenchyma to classify and segment different lung regions for quantitative analysis, making a huge contribution to the diagnosis and treatment of interstitial lung disease.

[0003] Existing texture analysis schemes mainly include unsupervised and fully supervised. The former is an unsupervised data-driven tissue texture analysis. By inputting unlabeled CT raw data, it clusters similar lung texture areas, thereby separating fibrous areas from normal tissue, and achieving quantitative measurement of pulmonary fibrosis. It can save manual labeling costs and is suitable for scenarios where data labels are scarce. It can identify hidden categories or abnormal structures in the data and is suitable for exploratory analysis of unknown problems. The latter relies on labeled training data and establishes a mapping relationship between features and labels through a supervised learning model. It divides the lungs into 15x15x15 voxel blocks, calculates their density histogram features, and compares them with the previously labeled training set to achieve voxel classification. It can achieve segmentation into normal, ground-glass opacity, grid shadow, honeycombing, and mild / moderate / severe low-density areas. It uses label information to learn complex texture patterns, and the classification accuracy is generally better than unsupervised methods. At the same time, its results directly correspond to predefined categories. Regularization, data augmentation and other techniques can reduce the impact of noise, and it is highly robust.

[0004] Although the above-mentioned existing methods have achieved the purpose of quantitative analysis of lung texture, they still have drawbacks; the unsupervised method currently only realizes the identification of fibrosis areas, and lacks prior knowledge guidance, making it difficult to perform more detailed lung texture classification and segmentation. Its clustering quality depends on feature selection and algorithm parameters, and the results are uncertain. At the same time, the clustering results may not have a clear correspondence with the actual category, and are easily sensitive to noise. Outliers or noisy data can easily lead to incorrect grouping; full supervision labeling is costly, and the labeling quality directly affects the model performance. It may fail when the data distribution changes (such as images collected by different devices), and the generalization ability is limited. In addition, although full supervision realizes the classification of different textures, there will still be misjudgments in the specific sign segmentation. For example, cystic bronchial dilatation and honeycomb sign often show similar density characteristics, and its segmentation resolution is limited by the minimum voxel block, making it difficult to obtain pixel-level segmentation results.

[0005] The Chinese patent publication number CN108460774A provides a method for automatic segmentation of lung medical images, but its high computational complexity and the combination of multi-step iteration and hyperspectral color visualization result in poor real-time performance of the algorithm, making it difficult to meet immediate clinical needs, low efficiency, and insufficient lesion segmentation accuracy, which cannot meet usage requirements.

[0006] To address the challenges faced by the existing technology, the present invention proposes a method and device for CT lung image texture analysis based on a 3D input convolutional neural network. The method realizes automatic segmentation of different lung texture areas through semantic segmentation, with the advantages of high recognition accuracy, fine segmentation, near real-time, and low computational burden. Summary of the Invention

[0007] To solve the above technical problems, the present invention provides a technical solution: a diffuse interstitial lung disease texture recognition model based on CT images and deep learning, including a PACS database, a training server, an application server, multiple data processing units and a user terminal; the training server is connected to the PACS database via Ethernet, and at the same time, the training server interacts with the user terminal via Ethernet to outline the volume of interest (VOI) for model training; the training server and the application server are connected to the multiple data processing units via Ethernet, and the data processing units are connected one by one to the image reconstruction workstation via Ethernet.

[0008] Preferably, a texture recognition model for diffuse interstitial lung disease based on CT images and deep learning, and a model training method thereof, comprise the following steps:

[0009] Step S1, obtaining initial training samples: Based on the existing interstitial lung disease data CT library, the user manually outlines the volume of interest (VOI) with different textures layer by layer to obtain initial training samples; the volume of interest (VOI) with different textures includes normal areas, ground-glass opacity areas, grid shadow areas, consolidation areas, honeycombing areas, and emphysema areas.

[0010] Step S2, model construction: In the Linux system, the Yolov11-seg model was built and trained using the PyTorch deep learning framework to perform semantic segmentation of different lung texture areas;

[0011] Step S3, thick-slice image reconstruction: In each training round, the initial training samples are simulated into CT images of different slice thicknesses according to the set probability, and the CT value arrays of the CT images of different slice thicknesses are reconstructed. The CT value arrays of multiple cross-sectional layers of adjacent slice thicknesses are overlapped to obtain a multi-channel 3D input array. The images are reconstructed using the lung window to obtain the current training set, completing one training round.

[0012] Step S4, model quantization: The obtained model is quantized and pruned through the TensorRT inference engine, and distributed and deployed in various data processing units.

[0013] In some embodiments, in step S1, the layer-by-layer delineation method includes the following steps:

[0014] S1.1. Use 3DSlicer on the user-side medical imaging workstation to load the CT sequences in the interstitial lung disease CT database.

[0015] S1.2. Use multi-layer orthogonal views to simultaneously display the transverse, coronal, and sagittal planes of the lung lobes;

[0016] S1.3. Use a semi-automatic segmentation tool to assist in outlining; the semi-automatic segmentation tool is: ActiveContour algorithm;

[0017] S1.4. On each cross-section, the ROI was initially delineated using the dynamic region growing algorithm, and the boundaries were manually adjusted to match the anatomical structure. Specifically, the fine brush tool was used for fine correction.

[0018] S1.5, maintain three-dimensional continuity by propagating functions across adjacent layers;

[0019] S1.6. Generate a three-dimensional VOI mask and perform topology verification.

[0020] In some embodiments, in step S2, the Yolov11-seg model specifically includes the following contents:

[0021] Set up a spatial input module containing 3D spatial information, using the CT value of the current cross-section as input channel 1, and the deformation fields based on the current layer and the upper and lower adjacent layers as the remaining two input channels;

[0022] The deformation field is used as a vector field to represent the displacement relationship between the pixels in the current layer and the previous and next layers:

[0023] φ:R 2 →R 2

[0024] For pixel coordinate x, this gives the displacement vector:

[0025] φ(x)=x+u(x)

[0026] The u(x) displacement field is the core of the deformation field;

[0027] Apply a pre-trained U-Net network f θ Make predictions about the deformation field:

[0028] u=f θ (Ifixed ,I moving )

[0029] Its loss function is defined as:

[0030]

[0031] In some embodiments, in step S3, the initial training samples are all thin-slice CT reconstruction sequences with a slice thickness of less than 2 mm. The method of simulating thick slices from thin slices for the initial training samples includes the following:

[0032] Set the original layer thickness to t and the target layer thickness to T. When T is less than or equal to the original layer thickness t, no data enhancement is necessary. When T is greater than t:

[0033] (1) Calculate the number of fusion layers n = ceil(T / t)

[0034] (2) Linear interpolation weighted fusion along the Z axis

[0035] (3) Apply the noise injection model: Where σ0 is the noise standard deviation of the original image.

[0036] In some embodiments, in step S3, the lung window image reconstruction specifically includes the following:

[0037] (1) Lung window used for reconstruction: window width W1500HU, window position L-400HU;

[0038] (2) Apply the following formula to convert the input CT value tensor S h Convert to lung window pixel Y:

[0039]

[0040] (3) where clip(x,a,b) is defined as a piecewise function.

[0041] In some embodiments, in step S4, the quantization and pruning operations include the following:

[0042] (1) Post-training quantization (PTQ): After the calibration set statistical parameters are calculated, its data format is converted from FP32 to INT8;

[0043] (2) Structured pruning: Calculate channel importance scores by output channel, remove those with scores < 0.1, and fine-tune the pruned model using the original dataset;

[0044] (3) Layer fusion optimization: According to the operator support of the target hardware device, the Conv+BN layer is fused and the matrix multiplication and addition operations are merged.

[0045] Preferably, a texture recognition model for diffuse interstitial lung disease based on CT images and deep learning, wherein the recognition process comprises the following steps:

[0046] Step (1), data preprocessing: using each data processing unit as an edge computing node, obtaining the reconstructed lung sequence from the CT post-processing workstation, performing pulmonary fissure integrity analysis, and storing the pulmonary fissure integrity analysis results in the application server for data aggregation;

[0047] Step (2), lung lobe integrity judgment: Based on the lung lobe segmentation results of step (1) and the lung fissure integrity analysis results, confirm whether the lung lobe segmentation is complete and valid; if the lung lobe segmentation is complete and valid, enter the texture recognition operation; if the lung lobe segmentation is incomplete, end the current texture recognition process;

[0048] Step (3), texture recognition: performing thick-slice image reconstruction on the lung lobe segmentation result of step (1) to obtain a reconstructed CT value array, overlapping the CT value arrays of multiple cross-sectional layers of adjacent slice thicknesses to obtain a multi-channel 3D input array, and then inputting the multi-channel 3D input array layer by layer into the recognition model to obtain lung texture inference results, and storing the lung texture inference results in the application server for data aggregation;

[0049] Step (4), volume of interest post-processing: obtain the volume of interest (VOI) of normal area, ground-glass shadow area, grid shadow area, consolidation shadow area, honeycomb sign area, and emphysema area through layer-by-layer fusion, discrete point suppression, and connected cavity analysis, and store the above volume of interest VOI in the application server for data aggregation;

[0050] Step (5), generating a structured report: aggregating and storing the data stored in the application server in steps (1)-(4), and running the recognition model to generate a structured report on interstitial lung disease.

[0051] In some embodiments, in step (1), the pulmonary fissure integrity analysis comprises the following steps:

[0052] (1.1) Surface mesh generation

[0053] (1.2) Calculate the Euler characteristic: χ = V - E + F;

[0054] (1.3) Fissure continuity detection: geodetic distance analysis is performed;

[0055] (1.4) Verification of connectivity of the hilar region;

[0056] (1.5) Volume rationality test: Single lobe volume > 5% of total lung volume.

[0057] In some embodiments, in step (2), the complete and effective segmentation of the lung lobes is specifically as follows:

[0058] (2.1) Pulmonary fissure detection completeness score > 0.85;

[0059] (2.2) The left and right lungs are completely separated without adhesion;

[0060] (2.3) Main bronchial connectivity is normal;

[0061] (2.4) The anatomical position has been verified and the positions of each lobe are correct.

[0062] In some embodiments, in step (4), layer-by-layer fusion is performed: the inference results of each layer are stacked into a VOI array of the same size as the original CT; discrete point suppression is performed: through 6-domain 3D connected domain analysis, the volume smaller than 4mm is removed. 3 isolated lesions.

[0063] The advantages of the present invention compared with the prior art are:

[0064] Based on a fully supervised prior lung texture segmentation training set, the present invention achieves a classification level that unsupervised clustering schemes cannot reach. Compared with existing fully supervised schemes, the present invention achieves improved classification accuracy and segmentation precision through a deep learning network rather than a lung density quantitative segmentation scheme. Through a distributed architecture, multi-scale data enhancement, 3D deformation field modeling, intelligent quality control and clinically guided post-processing, the present invention constructs an efficient, robust and easily scalable lung texture recognition system, providing a powerful tool for the accurate diagnosis of diffuse interstitial lung disease.

[0065] (1) The distributed data processing unit of the present invention is connected to the central server via Ethernet, realizing edge computing and cloud collaboration, effectively sharing the pressure of the central server, and supporting high-concurrency real-time reasoning; the data processing unit of the edge node is directly connected to the CT post-processing workstation, reducing data transmission delay and improving clinical application efficiency.

[0066] (2) The present invention uses TensorRT for quantization, pruning, and layer fusion, compressing the FP32 model to INT8, significantly reducing computing resource requirements while maintaining high precision, suitable for embedded device deployment; structured pruning and layer fusion optimization further improve inference speed.

[0067] (3) The present invention simulates common clinical scanning protocols through probabilistic layer thickness reconstruction technology, enhancing the robustness of the model to different layer thicknesses, noise levels and resolutions; noise injection and linear interpolation fusion technology simulate real clinical scenarios, avoid overfitting, and improve the generalization ability of the model.

[0068] (4) The present invention uses 3DSlicer software in combination with the ActiveContour algorithm, dynamic region growing, and fine brush tools to achieve efficient and accurate layer-by-layer VOI delineation; the adjacent layer propagation function and three-dimensional mask verification ensure the spatial continuity of the annotation and improve the quality of the training data.

[0069] (5) The present invention adopts the Yolov11-seg model to introduce the deformation field of the current layer and the adjacent layer as the input channel to capture the three-dimensional spatial structure changes and improve the accuracy of lung texture segmentation; pre-training U-Net predicts the deformation field to enhance the model's ability to understand the geometric deformation of the image.

[0070] (6) The present invention ensures the integrity and anatomical rationality of lung lobe segmentation through geodetic distance analysis, Euler characteristic number calculation and connectivity verification, thereby reducing subsequent analysis errors; multi-dimensional verification indicators (such as volume rationality and hilar connectivity) enhance the clinical credibility of the segmentation results.

[0071] (7) Multi-channel 3D input array combined with deep learning model to achieve semantic segmentation of multiple textures (ground glass shadow, grid shadow, etc.), supporting real-time analysis; layer-by-layer fusion and discrete point suppression technology (removing <4mm 3 Lesions) optimize segmentation results and reduce artifact interference.

[0072] (8) Generate structured reports including 3D texture display, bull's eye diagram, and lung lobe quantitative table, providing intuitive and comprehensive data support for clinical decision-making. BRIEF DESCRIPTION OF THE DRAWINGS

[0073] Figure 1 This is a schematic diagram of the device deployment of a diffuse interstitial lung disease texture recognition model and its implementation method of the present invention.

[0074] Figure 2 The present invention is a flow chart of a model training method for a diffuse interstitial lung disease texture recognition model and its implementation method.

[0075] Figure 3 This is a flow chart of the recognition process of a diffuse interstitial lung disease texture recognition model and its implementation method of the present invention. Figure 4 This is a schematic diagram of a structured report generated by a diffuse interstitial lung disease texture recognition model and its implementation method of the present invention. DETAILED DESCRIPTION

[0076] To make the objectives, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions of the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Generally, the components of the embodiments of the present invention described and shown in the drawings herein can be arranged and designed in various different configurations.

[0077] Example:

[0078] Combined with attachment Figure 1-3 A texture recognition model for diffuse interstitial lung disease based on CT images and deep learning includes a PACS database, a training server, an application server, multiple data processing units, and a client. The training server is connected to the PACS database via Ethernet. Simultaneously, the training server interacts with the client via Ethernet to outline volumes of interest (VOIs) for model training. The data processing unit provides real-time lung texture segmentation inference services for reconstructed lung CT sequences. The main hardware component is the AGX embedded platform launched by NVIDIA or other x86 platforms equipped with GeForce 10 Series or later graphics cards / computing cards. Distributed deployment shares the computing pressure of the central application server.

[0079] In addition, the training server and application server are connected to multiple data processing units via Ethernet, providing a unified user interaction interface and lung texture quantitative report generation. The data processing units are connected one by one to the image reconstruction workstations via Ethernet. The model of this embodiment realizes lung texture segmentation through deep learning, and it is easy to improve the generalization and robustness of the model through incremental learning as the data set increases.

[0080] In this embodiment, the model training method comprises the following steps: first, based on an existing interstitial lung disease data CT library, manually outlining volumes of interest with different textures layer by layer, specifically including normal areas, ground-glass opacity areas, grid shadow areas, consolidation areas, honeycombing areas, and emphysema areas, thereby obtaining initial training samples; then, through probabilistic layer thickness reconstruction, in each training round, the training samples are simulated with CT images of different layer thicknesses of 0.6 / 1.2 / 2.0 / 5.0 / 10.0 mm according to a certain probability, and a CT value array is reconstructed; then, adjacent multi-layer cross-sectional layers are overlapped to obtain a multi-channel 3D input array, and finally, an image is reconstructed using a lung window (window level -400HU, window width 1500HU) to obtain the current training set; a Yolov11-seg model is built and trained using the PyTorch deep learning framework on a Linux system to implement the semantic segmentation task of different lung texture areas; then, the obtained model is quantized, pruned, and other operations are performed through the TensorRT inference engine and deployed in the data processing unit.

[0081] The specific steps include:

[0082] Step S1, obtaining initial training samples: Based on an existing interstitial lung disease data CT library, the user manually outlines volumes of interest (VOIs) with different textures layer by layer to obtain initial training samples; the volumes of interest (VOIs) with different textures include normal areas, ground-glass opacity areas, grid-like shadow areas, consolidation areas, honeycombing areas, and emphysema areas;

[0083] The layer-by-layer delineation method includes the following steps:

[0084] S1.1. Use 3DSlicer software on the user-side medical imaging workstation to load the CT sequences in the interstitial lung disease CT database.

[0085] S1.2. Use multi-layer orthogonal views to simultaneously display the transverse, coronal, and sagittal planes of the lung lobes;

[0086] S1.3. Use a semi-automatic segmentation tool to assist in outlining; the semi-automatic segmentation tool is the ActiveContour algorithm;

[0087] S1.4. On each cross-section, the ROI was initially delineated using the dynamic region growing algorithm, and the boundaries were manually adjusted to match the anatomical structure. Specifically, the fine brush tool was used for fine correction.

[0088] S1.5, maintain three-dimensional continuity by propagating functions across adjacent layers;

[0089] S1.6. Generate a three-dimensional VOI mask and perform topology verification.

[0090] Step S2, model construction: In the Linux system, the Yolov11-seg model was built and trained using the PyTorch deep learning framework to perform semantic segmentation of different lung texture areas;

[0091] Among them, the construction of the Yolov11-seg model specifically includes the following contents:

[0092] S2.1 sets up a spatial input module containing 3D spatial information, using the CT value of the current cross-section as input channel 1, and the deformation fields based on the current layer and the upper and lower adjacent layers as the remaining two input channels;

[0093] S2.2. The deformation field is used as a vector field to represent the displacement relationship between the pixel in the current layer and the previous and next layers:

[0094] φ:R 2 →R 2

[0095] S2.3. For pixel coordinate x, this gives the displacement vector:

[0096] φ(x)=x+u(x)

[0097] The u(x) displacement field is the core of the deformation field;

[0098] S2.4. Apply a pre-trained U-Net network f θ Make predictions about the deformation field:

[0099] u=f θ (I fixed ,I moving )

[0100] S2.5. Define the loss function as:

[0101]

[0102] It should be noted that Yolov11-seg is only an implementation structure of semantic segmentation, and other semantic segmentation models similar to U-Net can also achieve the same purpose.

[0103] Step S3, thick-slice image reconstruction: In each training round, the initial training samples are simulated into CT images of different slice thicknesses according to the set probability, and the CT value arrays of the CT images of different slice thicknesses are reconstructed. The CT value arrays of multiple cross-sectional layers of adjacent slice thicknesses are overlapped to obtain a multi-channel 3D input array. The images are reconstructed using the lung window to obtain the current training set, completing one training round.

[0104] The samples included in this example are all thin-slice CT reconstruction sequences with a slice thickness of <2 mm. However, in actual scanning, some devices are still unable to provide thin-slice reconstruction, so thick-slice CT image simulations are added. The initial training samples are based on the set probability, which is based on the frequency statistics of clinical scan reconstruction protocols in our hospital database. A hierarchical weighted probability distribution is used, for example: 0.6 mm 10%, 1.0 mm 10%, 1.2 mm 25%, 2.0 mm 20%, 5.00 mm 25%, 10.0 mm 10%. The above probabilities are not true distributions and are only examples, because the actual probability distribution depends on the data distribution in the PACS at the time of training. The clinical significance of reconstruction with different slice thicknesses is: preserving the original high-resolution features, simulating conventional diagnostic slice thickness, balancing signal-to-noise ratio and resolution, simulating low-dose scans, and testing robustness in extreme cases.

[0105] In this embodiment, the method of simulating thick layers from thin layers in the initial training samples includes the following:

[0106] Set the original layer thickness to t and the target layer thickness to T. When T is less than or equal to the original layer thickness t, no data enhancement is necessary. When T is greater than t:

[0107] S3.1. Calculate the number of fusion layers n = ceil(T / t);

[0108] S3.2, linear interpolation weighted fusion along the Z axis;

[0109] S3.3. Application of noise injection model: Where σ0 is the noise standard deviation of the original image;

[0110] For lung window reconstruction images, the following are specifically included:

[0111] The lung window used for reconstruction is: window width W1500HU, window level L-400HU; the input CT value tensor S is converted to h Convert to lung window pixel Y:

[0112]

[0113] where clip(x,a,b) is defined as a piecewise function:

[0114] Step S4: Model quantization: The obtained model is quantized and processed by the TnoRt engine, and then distributed and deployed in each data processing unit;

[0115] Among them, quantization and pruning operations include the following:

[0116] S4.1. Post-training quantization (PTQ): After the calibration set statistical parameters are calculated, its data format is converted from FP32 to INT8;

[0117] S4.2, Structured Pruning: Calculate channel importance scores for each output channel, remove those with scores < 0.1, and fine-tune the pruned model using the original dataset;

[0118] S4.3. Layer fusion optimization: Based on the operator support of the target hardware device, the Conv+BN layer is fused and the matrix multiplication and addition operations are merged;

[0119] The recognition model of this embodiment includes the following steps:

[0120] Step (1) Data preprocessing: Each data processing unit is used as an edge computing node to obtain the reconstructed lung sequence from the CT post-processing workstation, perform lung fissure integrity analysis, and store the results of the lung fissure integrity analysis in the application server for data aggregation;

[0121] The CT post-processing workstation is a dedicated image processing system (such as GEAW or Siemens Syngo) configured in the medical imaging department. It provides 3D reconstruction functions such as MPR and MIP. In this system, it serves as a DICOM data source to output the sequences required for lung segmentation. The lung sequence is a DICOM image sequence reconstructed after a thin-slice (≤2 mm) chest CT scan, that is, the lung sequence automatically reconstructed by the CT machine when it is sent to the post-processing workstation.

[0122] The integrity analysis of the lung fissures includes the following steps:

[0123] (1.1) Surface mesh generation (MarchingCubes);

[0124] (1.2) Calculate the Euler characteristic: χ = V - E + F;

[0125] (1.3) Fissure continuity detection (geodesic distance analysis);

[0126] (1.4) Verification of connectivity of the hilar region;

[0127] (1.5) Volume rationality test (single lobe volume > 5% of total lung volume);

[0128] Geodesic distance analysis for fissure continuity detection is a method based on geometry and image processing, primarily used to assess the integrity and continuity of fissures (the separation structures between lung lobes) in 3D medical images (such as CT). Its specific content includes the following aspects:

[0129] ① Data preprocessing and lung fissure segmentation

[0130] Image enhancement: Reduce noise through filtering (such as Gaussian filtering), enhance the contrast of lung fissure areas, and highlight linear or curved structures.

[0131] Lung fissure segmentation: Use threshold segmentation, morphological operations, or deep learning models (such as U-Net) to extract the lung fissure region and generate a binary or probability map.

[0132] Surface reconstruction: Convert the segmentation results into a 3D surface or voxel model for subsequent geometric analysis.

[0133] ②Definition and calculation of geodesic distance

[0134] Geodesic distance principle: On a three-dimensional surface, the geodesic distance is the length of the shortest path along the surface between two points (as opposed to the straight-line distance of Euclidean distance).

[0135] ③Algorithm implementation:

[0136] Fast Marching Method: used to calculate the geodesic distance from the seed point to all other points.

[0137] Graph theory shortest path algorithm: voxels or grid vertices are used as nodes, edge weights are based on image gradients or local curvature, and the shortest path is solved using the Dijkstra algorithm.

[0138] ④ Constraints: The path must propagate along the surface of the lung fissure and avoid crossing the fracture area.

[0139] ⑤Continuity evaluation indicators:

[0140] Ratio of geodesic distance to Euclidean distance: If the geodesic distance of a path segment is significantly greater than the Euclidean distance, it may indicate a break or discontinuity.

[0141] Local continuity analysis: Geodesic distance sampling is performed on key points or areas on the surface of the lung fissure to detect abnormal jump points.

[0142] ⑥Threshold determination: Set a threshold for geodetic distance change, and mark areas exceeding the threshold as potential fractures.

[0143] Step (2) lung lobe integrity judgment: based on the lung lobe segmentation results of step (1) and the lung fissure integrity analysis results, confirm whether the lung lobe segmentation is complete and valid; if the lung lobe segmentation is complete and valid, enter the texture recognition operation; if the lung lobe segmentation is incomplete, end the current texture recognition process;

[0144] Among them, the lung lobe segmentation is complete and valid, which is confirmed based on the following analysis:

[0145] (2.1) Pulmonary fissure detection completeness score > 0.85;

[0146] (2.2) Complete separation of the left and right lungs (no adhesion);

[0147] (2.3) Main bronchial connectivity is normal;

[0148] (2.4) Anatomical position verification passed (the position of each lobe is correct);

[0149] Step (3), texture recognition: performing thick-slice image reconstruction on the lung lobe segmentation result of step (1) to obtain a reconstructed CT value array, overlapping the CT value arrays of multiple cross-sectional layers of adjacent slice thicknesses to obtain a multi-channel 3D input array, and then inputting the multi-channel 3D input array layer by layer into the recognition model to obtain lung texture inference results, and storing the lung texture inference results in the application server for data aggregation;

[0150] Step (4), volume of interest post-processing: obtain the volume of interest (VOI) of the normal area, ground-glass shadow area, grid shadow area, consolidation shadow area, honeycomb sign area, and emphysema area through layer-by-layer fusion and discrete point suppression, and store the above volume of interest VOI in the application server for data aggregation;

[0151] Among them, layer-by-layer fusion is to stack the inference results of each layer into a VOI array of the same size as the original CT; discrete point suppression is to remove the volume smaller than 4mm through 6-domain 3D connected domain analysis. 3 Solitary lesions;

[0152] Step (5), generating a structured report: the data stored in the application server in steps (1)-(4) are aggregated and stored, and the recognition model is run to generate a structured report of interstitial lung disease; the report content of the structured report of interstitial lung disease includes: basic information of the patient, scanning and reconstruction protocol information, 3D display of lung texture results, bull's eye diagram of texture longitudinal distribution, quantitative result table of texture analysis by lung lobe, screenshot of characteristic lesion level, and intelligent imaging classification diagnosis result.

[0153] The present invention and its embodiments are described above. This description is not restrictive. The drawings show only one embodiment of the present invention, and the actual structure is not limited thereto. In short, if a person skilled in the art is inspired by this and, without departing from the purpose of the present invention, designs structures and embodiments similar to this technical solution without inventiveness, they shall fall within the scope of protection of the present invention.

Claims

1. A texture recognition model for diffuse interstitial lung disease, characterized by: The system includes a PACS database, a training server, an application server, multiple data processing units, and a user terminal. The training server is connected to the PACS database via Ethernet. At the same time, the training server interacts with the user terminal via Ethernet to outline the volume of interest (VOI) for model training. The training server and the application server are connected to multiple data processing units via Ethernet. The data processing units are connected one by one to the image reconstruction workstation via Ethernet.

2. A texture recognition model for diffuse interstitial lung disease according to claim 1, wherein the model training method comprises the following steps: Step S1, obtaining initial training samples: Based on the existing interstitial lung disease data CT library, the user manually outlines the volume of interest (VOI) with different textures layer by layer to obtain initial training samples; the volume of interest (VOI) with different textures includes normal areas, ground-glass opacity areas, grid shadow areas, consolidation areas, honeycombing areas, and emphysema areas. Step S2, model construction: In the Linux system, the Yolov11-seg model was built and trained using the PyTorch deep learning framework to perform semantic segmentation of different lung texture areas; Step S3, thick-slice image reconstruction: In each training round, the initial training samples are simulated into CT images of different slice thicknesses according to the set probability, and the CT value arrays of the CT images of different slice thicknesses are reconstructed. The CT value arrays of multiple cross-sectional layers of adjacent slice thicknesses are overlapped to obtain a multi-channel 3D input array. The images are reconstructed using the lung window to obtain the current training set, completing one training round. Step S4, model quantization: The obtained model is quantized and pruned through the TensorRT inference engine, and distributed and deployed in various data processing units.

3. A diffuse interstitial lung disease texture recognition model according to claim 2, characterized in that: In step S1, the layer-by-layer delineation method includes the following steps: S1.

1. Use 3DSlicer on the user-side medical imaging workstation to load the CT sequences in the interstitial lung disease CT database. S1.

2. Use multi-layer orthogonal views to simultaneously display the transverse, coronal, and sagittal planes of the lung lobes; S1.

3. Use a semi-automatic segmentation tool to assist in outlining; the semi-automatic segmentation tool is: ActiveContour algorithm; S1.

4. On each cross-section, the ROI was initially delineated using the dynamic region growing algorithm, and the boundaries were manually adjusted to match the anatomical structure. Specifically, the fine brush tool was used for fine correction. S1.5, maintain three-dimensional continuity by propagating functions across adjacent layers; S1.

6. Generate a three-dimensional VOI mask and perform topology verification.

4. A diffuse interstitial lung disease texture recognition model according to claim 2, characterized in that: In step S2, the Yolov11-seg model specifically includes the following contents: Set up a spatial input module containing 3D spatial information, using the CT value of the current cross-section as input channel 1, and the deformation field based on the current layer and the upper and lower adjacent layers as the remaining two input channels; The deformation field is used as a vector field to represent the displacement relationship between the pixels in the current layer and the previous and next layers: f:R 2 →R 2 For pixel coordinate x, this gives the displacement vector: φ(x)=x+u(x) The u(x0 displacement field is the core of the deformation field; Apply a pre-trained U-Net network f θ Make predictions about the deformation field: u=f θ (I fixed ,I moving ) Its loss function is defined as:

5. The texture recognition model for diffuse interstitial lung disease according to claim 2, characterized in that: In step S3, the initial training samples are all thin-slice CT reconstruction sequences with a slice thickness of less than 2 mm. The method of simulating thick slices from thin slices for the initial training samples includes the following: Set the original layer thickness to t and the target layer thickness to T. When T is less than or equal to the original layer thickness t, no data enhancement is necessary. When T is greater than t: (1) Calculate the number of fusion layers n = ceil(T / t) (2) Linear interpolation weighted fusion along the Z axis (3) Apply the noise injection model: Where σ0 is the noise standard deviation of the original image. In step S3, the lung window image reconstruction specifically includes the following contents: (1) Lung window used for reconstruction: window width W1500HU, window position L-400HU; (2) Apply the following formula to convert the input CT value tensor S h Convert to lung window pixel Y: (3) where clip(x,a,b) is defined as a piecewise function.

6. A texture recognition model for diffuse interstitial lung disease according to claim 2, characterized in that: In step S4, the quantization and pruning operations include the following: (1) Post-training quantization (PTQ): After the calibration set statistical parameters are calculated, its data format is converted from FP32 to INT8; (2) Structured pruning: Calculate channel importance scores by output channel, remove those with scores < 0.1, and fine-tune the pruned model using the original dataset; (3) Layer fusion optimization: According to the operator support of the target hardware device, the Conv+BN layer is fused and the matrix multiplication and addition operations are merged.

7. The texture recognition model for diffuse interstitial lung disease according to claim 1, wherein the recognition process comprises the following steps: Step (1), data preprocessing: using each data processing unit as an edge computing node, obtaining the reconstructed lung sequence from the CT post-processing workstation, performing pulmonary fissure integrity analysis, and storing the pulmonary fissure integrity analysis results in the application server for data aggregation; Step (2), lung lobe integrity judgment: Based on the lung lobe segmentation results of step (1) and the lung fissure integrity analysis results, confirm whether the lung lobe segmentation is complete and valid; if the lung lobe segmentation is complete and valid, enter the texture recognition operation; if the lung lobe segmentation is incomplete, end the current texture recognition process; Step (3), texture recognition: performing thick-slice image reconstruction on the lung lobe segmentation result of step (1) to obtain a reconstructed CT value array, overlapping the CT value arrays of multiple cross-sectional layers of adjacent slice thicknesses to obtain a multi-channel 3D input array, and then inputting the multi-channel 3D input array layer by layer into the recognition model to obtain lung texture inference results, and storing the lung texture inference results in the application server for data aggregation; Step (4), volume of interest post-processing: obtain the volume of interest (VOI) of normal area, ground-glass shadow area, grid shadow area, consolidation shadow area, honeycomb sign area, and emphysema area through layer-by-layer fusion, discrete point suppression, and connected cavity analysis, and store the above volume of interest VOI in the application server for data aggregation; Step (5), generating a structured report: aggregating and storing the data stored in the application server in steps (1)-(4), and running the recognition model to generate a structured report on interstitial lung disease.

8. A diffuse interstitial lung disease texture recognition model according to claim 7, characterized in that: In step (1), the pulmonary fissure integrity analysis includes the following steps: (1.1) Surface mesh generation (1.2) Calculate the Euler characteristic: χ = V - E + F; (1.3) Fissure continuity detection: geodetic distance analysis is performed; (1.4) Verification of hilar regional connectivity; (1.5) Volume rationality test: Single lobe volume > 5% of total lung volume.

9. The texture recognition model for diffuse interstitial lung disease based on CT images and deep learning according to claim 7, characterized in that: In step (2), the lung lobe segmentation is complete and effective as follows: (2.1) Pulmonary fissure detection completeness score > 0.85; (2.2) The left and right lungs are completely separated without adhesion; (2.3) Main bronchial connectivity is normal; (2.4) The anatomical position has been verified and the positions of each lobe are correct.

10. The texture recognition model for diffuse interstitial lung disease according to claim 7, characterized in that: In step (4), layer-by-layer fusion is performed: the inference results of each layer are stacked into a VOI array of the same size as the original CT; discrete point suppression is performed: isolated lesions with a volume smaller than 4 mm3 are removed through 6-domain three-dimensional connected domain analysis.

Images