Three-dimensional quantitative analysis and report generation system for pulmonary nodules based on CT images

Abstract

This invention discloses a 3D quantitative analysis and report generation system for pulmonary nodules based on CT images, relating to the field of medical image processing and diagnostic analysis report technology. It includes a CT image access and processing module: accessing the target patient's chest thin-slice CT image sequence for standardized preprocessing to obtain standardized 3D image data; and a pulmonary nodule intelligent segmentation module: integrating a pre-trained deep learning model for pulmonary nodule segmentation, generating 3D segmentation mask data based on the input of the standardized 3D image data. This invention automatically identifies and segments nodules using a pre-trained deep learning model for pulmonary nodule segmentation, and performs precise volume measurement and 3D model reconstruction in parallel. It fuses the nodule model with multiple models of vascular and bronchial anatomy to generate a composite 3D scene model containing spatial relationships. Then, through non-rigid registration and intelligent matching algorithms, it achieves automatic comparison and growth quantification of nodules at different stages in the target patient.

Description

Technical Field

[0001] This invention relates to the field of medical image processing and diagnostic analysis report technology, specifically to a system for three-dimensional quantitative analysis and report generation of lung nodules based on CT images. Background Technology

[0002] In clinical lung cancer screening and pulmonary nodule management, patients need to undergo a lung CT scan, with thin-section chest CT being the core imaging tool. Radiologists need to identify and evaluate pulmonary nodules from CT sequences containing hundreds of cross-sectional images. To accurately assess pulmonary nodules, doctors typically rely on various tools. For example, currently available AI-assisted pulmonary nodule detection software can automatically outline suspicious nodules on CT images and provide preliminary two-dimensional measurement data. For complex nodules or cases requiring preoperative evaluation, doctors usually use a dedicated three-dimensional post-processing workstation. This workstation can reconstruct three-dimensional models of pulmonary nodules, pulmonary vessels, and bronchi based on CT data, allowing doctors to observe and analyze spatial anatomical relationships from multiple angles. Finally, doctors need to return to the medical image archive and communication system or a separate reporting system to synthesize all the above observations and write a pulmonary nodule characteristic diagnosis report in text form, manually extracting a few of the most representative two-dimensional images from the original images or the 3D workstation and attaching them to the report.

[0003] While existing CT analysis technology for lung nodules integrates advanced technologies such as AI detection and 3D reconstruction, it exhibits significant "fragmentation" at the workflow level. AI software, 3D workstations, and reporting systems operate independently, with data not interconnected. This directly leads to several drawbacks: First, the diagnostic process is fragmented, requiring doctors to repeatedly switch between different software programs and perform repetitive operations, resulting in a cumbersome, inefficient, and error-prone workflow. Second, crucial 3D spatial relationship information generated in the 3D workstation, such as the proximity and compression relationship between nodules and blood vessels, cannot be effectively transmitted to attending physicians or surgeons in clinical departments or to the final diagnostic report. The report remains limited to text and static 2D images, leading to the "dissipation" of 3D information. Third, for nodules requiring regular follow-up, doctors find it difficult to intuitively and accurately compare the 3D morphology and volume changes of the same nodule across different CT images.

[0004] In summary, existing technologies have the following shortcomings: key aspects such as intelligent detection of lung nodules, three-dimensional quantitative analysis, visualized three-dimensional reconstruction, and structured report generation are disconnected from each other, failing to form an automated, data-flow-connected closed loop. This results in low efficiency of clinical workflow, loss of three-dimensional spatial information, and unintuitive follow-up comparisons, ultimately affecting the accuracy and consistency of diagnosis and leading to biases in clinical decision-making.

[0005] The information disclosed in the background section is only intended to enhance the understanding of the background of this disclosure, and therefore may include information that does not constitute prior art known to those skilled in the art. Summary of the Invention

[0006] The purpose of this invention is to provide a three-dimensional quantitative analysis and report generation system for lung nodules based on CT images. This invention solves the problems in the background art by constructing an end-to-end automated system that integrates AI intelligent segmentation, three-dimensional quantitative analysis, anatomical structure fusion reconstruction, automatic temporal comparison and interactive report generation.

[0007] To achieve the above objectives, the present invention provides the following technical solution: a CT image-based three-dimensional quantitative analysis and report generation system for lung nodules, including a CT image access and processing module: accessing the chest thin-slice CT image sequence of the target patient and performing standardized preprocessing to obtain a standardized three-dimensional image data volume;

[0008] Lung nodule intelligent segmentation module: integrates a pre-trained deep learning model for lung nodule segmentation, and generates three-dimensional segmentation mask data based on the input of the standardized three-dimensional image data volume;

[0009] Quantitative analysis module: Calculates the set of quantitative feature parameters of nodules based on the three-dimensional segmentation mask data;

[0010] 3D Reconstruction and Fusion Engine: Based on the 3D image data volume and the 3D segmentation mask data, the engine performs parallel 3D model reconstruction of nodules and fusion of spatial relationship information to generate a lightweight and interactive 3D scene model.

[0011] Temporal comparison module: acquires the target patient's previous CT images and the corresponding three-dimensional segmentation mask of the nodules, and generates comparison results by performing spatiotemporal registration and matching alignment;

[0012] Automatic report generation engine: Creates a structured report template, receives the set of quantified feature parameters and the 3D scene model, as well as the comparison results, and automatically performs the filling, rendering and compositing operations of the structured report template to output an interactive electronic report.

[0013] Furthermore, the pre-trained lung nodule segmentation deep learning model is preferably a lung nodule segmentation model that is automatically configured and trained using the nnU-Net framework, and the lung nodule segmentation model includes an encoder-decoder.

[0014] Furthermore, the steps for generating the three-dimensional segmentation mask data are as follows:

[0015] The three-dimensional image data volume is input into the pre-trained lung nodule segmentation deep learning model. The input is downsampled through the encoder path to extract multi-scale deep semantic features. Then, it is upsampled through the decoder path and connected to the feature skip connection of the corresponding level of the encoder. At the end of the decoder, the Sigmoid activation function is applied to each voxel position to output a three-dimensional probability map. Each voxel position corresponds to a value between 0 and 1 in the output, which represents the probability that the voxel at the corresponding position belongs to the lung nodule category.

[0016] Threshold binarization is performed on the 3D probability map. A preset threshold of 0.5 is applied to convert each voxel in the 3D probability map into an initial binary foreground mask.

[0017] A connected component analysis algorithm is used to perform three-dimensional connected component analysis on the binary foreground mask to identify and label interconnected voxel sets. A unique temporary label is assigned to each connected voxel set to generate a label matrix.

[0018] Filtering noise regions requires calculating the volume of each labeled connected region. If the volume of a connected region is less than the preset minimum effective nodule volume threshold, the connected region is removed from the foreground and considered as noise.

[0019] For each connected region retained after filtering, a morphological opening operation is applied to smooth the boundaries of the connected regions and remove small noise points;

[0020] Each independent nodule region is assigned a unique identifier (ID) to generate the final 3D segmentation mask data.

[0021] Furthermore, the calculation steps for the quantized feature parameter set are as follows:

[0022] Receive 3D segmentation mask data, parse, traverse and extract coordinates of the independent lung nodules corresponding to each identifier ID, obtain voxel-nodule relationship, standard CT value of each voxel position, physical voxel size information of pixel spacing and slice thickness, and voxel coordinate set of each nodule.

[0023] The set of quantified feature parameters includes nodule volume, maximum three-dimensional diameter, average CT value, and semantic morphological descriptor;

[0024] The total number of voxels with a value of 1 in the 3D segmentation mask for each identifier ID is counted, and multiplied by the physical volume of a single voxel to obtain the nodule volume.

[0025] Extract the set of voxel coordinates of the 3D segmentation mask for each identifier ID, and calculate the maximum value of the Euclidean distance between any two voxel coordinates in the set as the maximum 3D diameter.

[0026] Based on the 3D segmentation mask coordinates of each identifier ID, backtrack to the normalized CT value array in the original CT image sequence, extract the gray value of the corresponding region, and calculate the average value;

[0027] Calculate the geometric descriptors of the sphericity, nodule surface area, and surface area-to-volume ratio of the 3D segmentation mask for each identifier ID;

[0028] The geometric descriptor is combined with the maximum three-dimensional diameter as a feature vector and input into a lightweight random forest classifier that has been trained on a nodule dataset labeled with semantic morphological descriptors of "lobulation sign" and "spiculation sign" by radiologists. The classifier outputs the probability values ​​of "lobulation sign" and "spiculation sign" and obtains the binary judgment result by setting a decision threshold of 0.5.

[0029] All parameters calculated for each nodule are integrated into a structured data record. After traversing all nodules, the structured data records corresponding to all nodules are integrated into the final set of quantized feature parameters.

[0030] Furthermore, the semantic morphological descriptor includes at least one of the lobed sign and the spiculated sign.

[0031] Furthermore, the steps for generating the 3D scene model are as follows:

[0032] Based on the 3D image data volume and the 3D segmentation mask data, the 3D reconstruction and fusion engine starts two parallel processing threads:

[0033] Thread 1: Use the moving cube algorithm to reconstruct the 3D segmentation mask of each nodule, generating a closed nodule surface mesh model composed of triangular facets.

[0034] Thread 2 uses a seed-based region growing blood vessel segmentation algorithm for segmentation and construction. Based on voxel-normalized CT values, a centerline model representing the topological structure of blood vessels is generated, and the surface model can be selectively obtained by moving cube surface reconstruction of the binary mask of blood vessels.

[0035] Meanwhile, a bronchus extraction algorithm based on connected component analysis threshold region growth is used for extraction and construction. From the same set of standardized CT value data, a bronchus centerline model is generated, and the surface model can be selectively obtained by moving cube surface reconstruction of the airway binary mask.

[0036] Import the nodule surface mesh model, the vascular topology centerline model, the vascular surface model, the bronchus centerline model, and the bronchus surface model into a unified 3D scene coordinate system.

[0037] By performing rigid or affine registration based on anatomical landmarks, spatial registration and consistent fusion of multiple models are ensured, and multiple models are assembled into a composite 3D scene model in a unified coordinate system after registration.

[0038] Furthermore, the structured report template includes a basic information area at the top of the GUI, a lung nodule analysis report area at the main body of the GUI, and a functional interaction area at the bottom of the GUI;

[0039] The basic information area at the top of the GUI includes the report title and basic patient information;

[0040] The main GUI lung nodule analysis report area includes a quantitative feature parameter set summary area, a two-dimensional multi-plane reconstructed view area, a three-dimensional scene model visualization area, a structured description and conclusion area, and a follow-up chapter comparison area;

[0041] The bottom functional interaction area of ​​the GUI includes icons for adjusting rotation / zoom / transparency in the 3D view area, as well as function buttons for exporting reports and sharing.

[0042] Furthermore, the steps for generating the interactive electronic report are as follows:

[0043] Load the structured report template, read the layout of each section in the structured report template, and determine the data positions and visualization interaction positions;

[0044] Receive the comparison results of the quantized feature parameter set, the composite 3D scene model file, and the time series, and fill in the text data placeholders through string formatting;

[0045] From the voxel-normalized CT value data, based on the center position of the nodule, PNG images of three orthogonal plane views (coronal, sagittal, and axial) are automatically extracted, generated, and saved, and then filled into the interactive visualization area.

[0046] The composite 3D scene model file and three orthogonal planar views are placed on a static resource directory accessible to a web server built using a framework such as Flask / Django. The text data is rendered into an HTML page, and the web server assigns a unique URL endpoint based on the patient ID and report timestamp for each generated report.

[0047] The glTF / glb file can be loaded directly using the Three.js library for interactive display, and can be accessed via a URL link.

[0048] Furthermore, the spatiotemporal registration and matching alignment are performed as follows:

[0049] A non-rigid image registration algorithm is used for temporal image registration, which globally aligns the CT image data of the current target patient with historical CT image data in terms of anatomical structure.

[0050] Within the registered space, the 3D mask overlap and centroid distance between each current nodule and all historical nodules are calculated. By setting a pre-defined threshold, individual nodule matching is performed to determine nodules that are aligned with each other.

[0051] For successfully matched nodule pairs, longitudinal time dimension analysis is performed by quantitatively calculating the volume change rate and morphological evolution index of the nodules, and a three-dimensional model comparison map of historical and current nodules is generated and the areas of volume growth are highlighted.

[0052] The nodule pairs that have undergone spatiotemporal registration and matching, the volume change rate and morphological evolution index of the quantified variables, and the generated historical-current 3D model comparison map are packaged into a structured comparison result data package, which is then used to fill the follow-up section of an interactive electronic report.

[0053] Furthermore, the method for three-dimensional quantitative analysis and report generation of pulmonary nodules based on CT images includes the following steps:

[0054] S1. Access the chest CT image sequence of the target patient, call the pre-trained lung nodule segmentation deep learning model to process the chest CT image sequence, automatically identify and segment at least one lung nodule region, and output the pixel-level three-dimensional segmentation mask corresponding to each lung nodule.

[0055] S2. Based on the three-dimensional segmentation mask, multi-dimensional feature extraction is performed on each lung nodule, and a set of quantitative feature parameters including nodule volume, maximum three-dimensional diameter, average CT value and semantic morphological descriptor is automatically calculated.

[0056] S3. Based on the 3D segmentation mask, the 3D model reconstruction and spatial relationship information fusion are initiated in parallel. This includes first generating a 3D surface mesh model of each lung nodule, and simultaneously segmenting and reconstructing a 3D model of the pulmonary vascular tree / bronchial tree from the same chest CT image sequence. Then, the 3D surface mesh model of the nodule is spatially registered and superimposed with the 3D model of the pulmonary vascular tree / bronchial tree to obtain a lightweight and interactive 3D scene model containing spatial relationship information.

[0057] S4. Pre-create a structured report template, automatically fill the corresponding data positions and visualization interaction positions of the structured report template with the quantitative feature parameter set and the three-dimensional scene model, and automatically generate a two-dimensional multi-plane reconstruction view containing at least one key morphology of the nodule, and synthesize an interactive electronic report that integrates quantitative data, two-dimensional multi-plane reconstruction view and embedded three-dimensional model.

[0058] S5. In response to the existence of previous CT images of the target patient, a time-series comparison process is executed in the quantitative analysis. The current three-dimensional segmentation mask of the lung nodule is automatically registered and matched with the corresponding nodule mask in the previous CT images in a time-space manner. The comparison results of volume change rate and morphological evolution index are calculated and the comparison results are synchronously updated to the follow-up section of the interactive electronic report.

[0059] The technical effects and advantages provided by the present invention in the above technical solution are as follows:

[0060] This invention employs a pre-trained deep learning model for lung nodule segmentation to automatically identify and segment nodules, while simultaneously performing precise volume measurement and 3D model reconstruction. It integrates the nodule model with multiple models of vascular and bronchial anatomy to generate a composite 3D scene model incorporating spatial relationships. Through non-rigid registration and intelligent matching algorithms, it achieves automatic comparison and growth quantification of nodules at different stages in the target patient. All analysis results are automatically integrated into a structured web report containing an embedded interactive 3D view model, quantitative data, and follow-up views. This enables fully automated, integrated generation of interactive reports simply by accessing CT images, improving diagnostic efficiency, the objectivity of follow-up assessments, the accuracy of follow-up, and the reliability of clinical decisions. Attached Figure Description

[0061] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0062] Figure 1 This is a block diagram of the CT image-based three-dimensional quantitative analysis and report generation system for lung nodules according to the present invention.

[0063] Figure 2 This is a flowchart of the method for three-dimensional quantitative analysis and report generation of lung nodules based on CT images according to the present invention.

[0064] Figure 3 This is a schematic diagram of the graphical user interface (GUI) of the interactive electronic report of the present invention. Detailed Implementation

[0065] Exemplary embodiments will now be described more fully with reference to the accompanying drawings. However, these exemplary embodiments can be implemented in many forms and should not be construed as limited to the examples set forth herein; rather, they are provided so that the description of this disclosure will be more complete and fully convey the concept of the exemplary embodiments to those skilled in the art.

[0066] Example 1

[0067] This invention provides, for example Figure 1 The CT image-based 3D quantitative analysis and report generation system for lung nodules shown includes a CT image access and processing module: a built-in DICOM parsing unit and a preprocessing unit. The DICOM parsing unit accesses the target patient's chest thin-slice CT image sequence through a standard DICOM interface, extracts metadata such as pixel array (CT value), slice thickness, pixel spacing, and patient orientation, and the preprocessing unit performs standardized preprocessing on the pixel array to obtain and parse the chest CT image sequence to obtain standardized 3D image data, which is then transmitted to the lung nodule intelligent segmentation module.

[0068] Specifically, the standardized preprocessing of CT image sequences includes, but is not limited to:

[0069] (1) The Heinz units HU of the original CT values ​​are mapped to the 0-255 standard range through linear transformation to ensure consistent image contrast;

[0070] (2) Use bilinear or cubic spline interpolation algorithms to perform isotropic resampling on CT image sequences and unify them to a voxel size of 1mm×1mm×1mm to ensure the accuracy of subsequent three-dimensional measurement and reconstruction.

[0071] (3) Output a normalized three-dimensional NumPy array or ITK image object as the standard data object for the entire system process.

[0072] The intelligent lung nodule segmentation module integrates a pre-trained deep learning model for lung nodule segmentation. Based on the input of standardized 3D image data, it automatically extracts multi-level features through an encoder-decoder structure. At the end of the decoder, it outputs a value between 0 and 1 for each voxel, representing the probability that the voxel belongs to the lung nodule category. A preset threshold of 0.5 is applied to binarize the probability to obtain an initial mask. Then, a connected component analysis algorithm is used to identify independent connected components, each of which represents an independent nodule. At the same time, post-processing such as morphological opening operations are applied to eliminate small noise points. A 3D segmentation mask matrix with a unique identifier (ID) is generated for each independent nodule region. The generated 3D segmentation mask data is then transmitted to the quantization analysis module.

[0073] Specifically, the pre-trained deep learning model for lung nodule segmentation is preferably a lung nodule segmentation model that is automatically configured and trained using the nnU-Net framework, and the lung nodule segmentation model includes an encoder-decoder, exhibiting excellent robustness and accuracy.

[0074] Specifically, the steps for generating 3D segmentation mask data are as follows:

[0075] The three-dimensional image data volume is input into a pre-trained deep learning model for lung nodule segmentation. The input is downsampled through the encoder path to extract multi-scale deep semantic features, and then upsampled through the decoder path and connected to the features of the corresponding level of the encoder to restore spatial details. At the end of the decoder, the Sigmoid activation function is applied to each voxel position to output a three-dimensional probability map. Each voxel position corresponds to a value between 0 and 1 in the output, which represents the probability that the voxel at the corresponding position belongs to the lung nodule category.

[0076] The 3D probability map is subjected to threshold binarization. A preset threshold of 0.5 is applied. For each voxel in the 3D probability map, when the output probability value is not less than the preset threshold, the corresponding voxel is identified as the foreground (nodule) and assigned a value of 1; when the output probability value is less than the preset threshold, the corresponding voxel is identified as the background and assigned a value of 0. In this way, the 3D probability map is transformed into an initial binary foreground mask.

[0077] A connected component analysis algorithm is used to perform three-dimensional connected component analysis on the binary foreground mask to identify and label interconnected sets of voxels for identifying potential independent nodules. The analysis adopts the 26-neighborhood connectivity criterion, that is, voxels are considered to be adjacent in the front-back, left-right, up-down and all diagonal directions in three-dimensional space. By scanning the entire binary foreground mask, a unique temporary label is assigned to each connected set of voxels to generate a label matrix.

[0078] Based on prior clinical knowledge, connected regions with excessively small volumes are very likely to be image noise or irrelevant tissue. Therefore, filtering noise regions requires calculating the volume of each labeled connected region, i.e., counting the total number of voxels in the connected region. When the volume of a connected region is less than the preset minimum effective nodule volume threshold, the connected region is removed from the foreground and regarded as noise.

[0079] For each connected region retained after filtering, a morphological opening operation is applied to smooth the boundaries of the connected regions and remove any possible small burrs or raised micro noise points. The morphological opening operation is a composite operation of erosion followed by dilation, performed using a preset three-dimensional structuring element. The preset three-dimensional structuring element is preferably a small-sized sphere or cube core. This composite operation can eliminate unstable boundary noise while maintaining the main shape of the region.

[0080] Each independent nodule region that has undergone the above refinement process is assigned a unique identifier (ID) to generate the final 3D segmentation mask data.

[0081] Quantitative Analysis Module: Calculates a set of quantitative feature parameters of nodules based on 3D segmentation mask data, including nodule volume, maximum 3D diameter, average CT value and semantic morphological descriptor, and sends them to the report automatic generation engine;

[0082] Specifically, the calculation steps for the quantization feature parameter set are as follows:

[0083] The system receives 3D segmentation mask data, parses the individual lung nodules corresponding to each identifier ID, traverses the nodule region, and extracts coordinates to obtain the voxel-nodule relationship, the standard CT value of each voxel location, the physical voxel size information of pixel spacing and slice thickness, and the voxel coordinate set of each nodule. The expression for the voxel-nodule relationship is as follows: In the formula, Represented as size Integer matrix, Represented as voxel positions, Represented as the first A non-zero node identifier, and Includes One nodule and ;

[0084] Expression for the standard CT value at each voxel location In the formula, Standard CT values, representing voxel locations, range from 0 to 255.

[0085] The expression for physical voxel size information is as follows In the formula, Represented as pixel spacing, This is expressed as slice thickness;

[0086] No. The expression for the set of voxel coordinates of each nodule is: In the formula, Represented as the first A set of voxel coordinates for each nodule. Represented as the first The first nodule Individual location, Represented as the first The total number of voxels contained in each nodule. Represented as the first Voxel index in each nodule;

[0087] The set of quantified feature parameters includes nodule volume, maximum three-dimensional diameter, average CT value, and semantic morphological descriptor;

[0088] Nodule volume calculation process: The total number of voxels with a value of 1 in the 3D segmentation mask for each identifier ID is counted, and multiplied by the physical volume of a single voxel to obtain the nodule volume. The formula for calculating the nodule volume is as follows: In the formula, Represented as the first Volume of each nodule;

[0089] The maximum 3D diameter calculation process is as follows: Extract the voxel coordinate set of the 3D segmentation mask for each identifier ID, and calculate the maximum Euclidean distance between any two voxel coordinates in this set. This maximum distance is taken as the maximum 3D diameter. The formula for calculating the maximum 3D diameter is as follows: In the formula, Represented as the first The maximum three-dimensional diameter of each nodule , Represented as the first Any two voxel coordinates of the set of voxel coordinates of a node, Represented as the Euclidean distance symbol;

[0090] The average CT value calculation process is as follows: Based on the 3D segmentation mask coordinates for each identifier ID, the normalized CT value array in the original CT image sequence is traced back to extract the grayscale values ​​of the corresponding regions, and the average value is calculated. The formula for calculating the average CT value is as follows: In the formula, Expressed as average CT value, Represented as voxel position Standard CT values;

[0091] Semantic morphological descriptor calculation process: Calculate the geometric descriptor of the sphericity, nodule surface area, and surface area-to-volume ratio of the 3D segmentation mask for each identifier ID. The formula for calculating sphericity is as follows: In the formula, Represented as the first The sphericity of each nodule, Represented as the first The three-dimensional mesh surface area of ​​each nodule;

[0092] The formula for calculating the surface area to volume ratio is: In the formula, Represented as the first The surface area-to-volume ratio of each nodule;

[0093] The geometric descriptor, combined with the maximum 3D diameter, is used as a feature vector. This vector is input into a lightweight random forest classifier trained on a nodule dataset labeled with semantic morphological descriptors "lobulation sign" and "spiculation sign" by radiologists. The classifier outputs probability values ​​for "lobulation sign" and "spiculation sign," and a binary judgment result is obtained by comparing these probabilities using a decision threshold of 0.5.

[0094] The expression for the probability values ​​of the "lobed sign" and "spiculated sign" output by the lightweight random forest classifier is as follows: ,and In the formula, This represents the probability value of the lobulation feature output by the lightweight random forest classifier. This represents the probability value of the corresponding spurious feature output by the lightweight random forest classifier. Represented as the first The feature vectors of each nodule Represented as a lightweight random forest classifier, This is indicated by the lobulation sign. This is indicated by spiculation;

[0095] The expression for the result of a binary judgment is: In the formula, This is represented as the binarized determination result of the lobulation sign. This is an indicator function that takes the value 1 when the condition is true and 0 otherwise. This represents setting a decision threshold. This is represented as the binarized determination result of the spiculation feature;

[0096] All calculated parameters for each node are integrated into a structured data record. After traversing all nodes, the structured data records corresponding to all nodes are integrated into the final set of quantized feature parameters. The expression for the structured data record of each node is: In the formula, Represented as the first Structured data records for each nodule;

[0097] The expression for the quantized feature parameter set is: .

[0098] Specifically, semantic morphological descriptors include at least one of the lobed sign and the spiculated sign.

[0099] 3D Reconstruction and Fusion Engine: Based on the 3D image data volume and 3D segmentation mask data, the engine performs parallel 3D model reconstruction of nodules and fusion of spatial relationship information to generate a lightweight, interactive 3D scene model, which is then sent to the report generation engine.

[0100] Specifically, the steps for generating a 3D scene model are as follows:

[0101] Based on the 3D image data volume and the 3D segmentation mask data, the 3D reconstruction and fusion engine starts two parallel processing threads:

[0102] Thread 1: Use the moving cube algorithm to reconstruct the 3D segmentation mask of each nodule, generating a closed nodule surface mesh model composed of triangular facets.

[0103] Thread 2 uses a seed-based region growing blood vessel segmentation algorithm for segmentation and construction. Based on voxel-normalized CT values, a centerline model representing the topological structure of blood vessels is generated, and the surface model can be selectively obtained by moving cube surface reconstruction of the binary mask of blood vessels.

[0104] Meanwhile, a bronchus extraction algorithm based on connected component analysis threshold region growth is used for extraction and construction. From the same set of standardized CT value data, a bronchus centerline model is generated, and the surface model can be selectively obtained by moving cube surface reconstruction of the airway binary mask.

[0105] Import the nodule surface mesh model, the vascular topology centerline model, the vascular surface model, the bronchus centerline model, and the bronchus surface model into a unified 3D scene coordinate system that can be managed using PyVista or VTK libraries.

[0106] By performing rigid or affine registration based on anatomical landmarks, spatial registration and consistent fusion of multiple models are ensured. In the unified coordinate system after registration, multiple models are assembled into a composite 3D scene model. The 3D scene model contains multiple objects such as nodules, blood vessels, and bronchi, and can be exported as a composite 3D scene model file in glTF / glb format.

[0107] Automatic report generation engine: Creates a structured report template, receives a set of quantitative feature parameters and a 3D scene model, as well as the comparison results from the time series comparison module, and automatically performs the filling, rendering and compositing operations of the structured report template, outputting an interactive electronic report in the form of a URL link;

[0108] Specifically, the structured report template includes a basic information area at the top of the GUI, a lung nodule analysis report area in the main body of the GUI, and a functional interaction area at the bottom of the GUI;

[0109] The basic information area at the top of the GUI includes the report title and basic patient information;

[0110] The main GUI lung nodule analysis report area includes a quantitative feature parameter set summary area, a two-dimensional multi-plane reconstructed view area, a three-dimensional scene model visualization area, a structured description and conclusion area, and a follow-up chapter comparison area;

[0111] The bottom functional interaction area of ​​the GUI includes icons for adjusting rotation / zoom / transparency in the 3D view area, as well as function buttons for exporting reports and sharing.

[0112] Specifically, the steps for generating an interactive electronic report are as follows:

[0113] Load the structured report template, read the layout of each section in the structured report template, and determine the data positions and visualization interaction positions;

[0114] Receive the comparison results of the quantized feature parameter set, the composite 3D scene model file, and the time series, and fill in the text data placeholders through string formatting;

[0115] From the voxel-normalized CT value data, based on the center position of the nodule, PNG images of three orthogonal plane views (coronal, sagittal, and axial) are automatically extracted, generated, and saved, and then filled into the interactive visualization area.

[0116] The composite 3D scene model file and three orthogonal planar views are placed on a static resource directory accessible to a web server built using a framework such as Flask / Django. The text data is rendered into an HTML page, and the web server assigns a unique URL endpoint based on the patient ID and report timestamp for each generated report.

[0117] The glTF / glb file can be loaded directly using the Three.js library for interactive display, and can be accessed via a URL link;

[0118] Doctors can access a fully functional, interactive electronic report by visiting the URL link on any electronic device with a modern web browser, without needing to install any additional plugins.

[0119] Temporal comparison module: When a target patient has a history of CT studies, the temporal comparison process is triggered to obtain the target patient's previous CT images and the corresponding three-dimensional segmentation mask of the nodules. By performing spatiotemporal registration and matching alignment, the comparison results are generated and sent to the report automatic generation engine.

[0120] Specifically, the steps for spatiotemporal registration and matching alignment are as follows:

[0121] Using a non-rigid image registration algorithm for temporal image registration, the CT image data of the current target patient is globally aligned with historical CT image data in terms of anatomical structure to eliminate the influence of differences in respiratory status and body position.

[0122] Within the registered space, the 3D mask overlap and centroid distance between each current nodule and all historical nodules are calculated. By setting a pre-defined threshold, individual nodule matching is performed to determine nodules that are aligned with each other.

[0123] For successfully matched nodule pairs, longitudinal time dimension analysis is performed by quantitatively calculating the volume change rate and morphological evolution index of the nodules, and a three-dimensional model comparison map of historical and current nodules is generated and the areas of volume growth are highlighted.

[0124] The nodule pairs that have undergone spatiotemporal registration and matching, the volume change rate and morphological evolution index of the quantified variables, and the generated historical-current 3D model comparison map are packaged into a structured comparison result data package. This data package is then used as a follow-up section to populate an interactive electronic report and integrated into the automated process of horizontal single analysis to improve the objectivity, time management, and efficiency of follow-up assessment. The morphological evolution index includes changes in sphericity, surface area to volume ratio, etc.

[0125] Specifically, the method for three-dimensional quantitative analysis and report generation of lung nodules based on CT images includes the following steps:

[0126] S1. Access the chest CT image sequence of the target patient, call the pre-trained lung nodule segmentation deep learning model to process the chest CT image sequence, automatically identify and segment at least one lung nodule region, and output the pixel-level three-dimensional segmentation mask corresponding to each lung nodule.

[0127] S2. Based on the three-dimensional segmentation mask, multi-dimensional feature extraction is performed on each lung nodule, and a set of quantitative feature parameters including nodule volume, maximum three-dimensional diameter, average CT value and semantic morphological descriptor is automatically calculated.

[0128] S3. Based on the 3D segmentation mask, the 3D model reconstruction and spatial relationship information fusion are initiated in parallel. This includes first generating a 3D surface mesh model of each lung nodule, and simultaneously segmenting and reconstructing a 3D model of the pulmonary vascular tree / bronchial tree from the same chest CT image sequence. Then, the 3D surface mesh model of the nodule is spatially registered and superimposed with the 3D model of the pulmonary vascular tree / bronchial tree to obtain a lightweight and interactive 3D scene model containing spatial relationship information.

[0129] S4. Pre-create a structured report template, automatically fill the corresponding data positions and visualization interaction positions of the structured report template with the quantitative feature parameter set and the three-dimensional scene model, and automatically generate a two-dimensional multi-plane reconstruction view containing at least one key morphology of the nodule, and synthesize an interactive electronic report that integrates quantitative data, two-dimensional multi-plane reconstruction view and embedded three-dimensional model.

[0130] S5. In response to the existence of previous CT images of the target patient, a time-series comparison process is executed in the quantitative analysis. The current three-dimensional segmentation mask of the lung nodule is automatically registered and matched with the corresponding nodule mask in the previous CT images in a time-space manner. The comparison results of volume change rate and morphological evolution index are calculated and the comparison results are synchronously updated to the follow-up section of the interactive electronic report.

[0131] Specifically, the 3D scene model is 3D data that has been simplified through mesh processing and converted into a WebGL-compatible format.

[0132] Specifically, the interactive electronic report is in HTML5 web page format or a PDF document with an embedded 3D model access link.

[0133] The above formulas are all dimensionless calculations. The formulas are derived from software simulations based on a large amount of collected data to obtain the most recent real-world results. The preset parameters in the formulas are set by those skilled in the art according to the actual situation.

[0134] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, as a computer program product. The computer program product includes one or more computer instructions or computer programs. When the computer instructions or computer programs are loaded or executed on a computer, all or part of the processes or functions described in the embodiments of this application are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that includes one or more sets of available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium. A semiconductor medium can be a solid-state drive.

[0135] It should be understood that, in the various embodiments of this application, the order of the above-mentioned processes does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.

[0136] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0137] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A system for three-dimensional quantitative analysis and report generation of pulmonary nodules based on CT images, characterized in that, Includes a CT image access and processing module: it accesses the target patient's chest thin-section CT image sequence, performs standardized preprocessing, and obtains standardized three-dimensional image data. Lung nodule intelligent segmentation module: integrates a pre-trained deep learning model for lung nodule segmentation, and generates three-dimensional segmentation mask data based on the input of the standardized three-dimensional image data volume. The pre-trained deep learning model for lung nodule segmentation is an automatically configured and trained lung nodule segmentation model using the nnU-Net framework, and the lung nodule segmentation model includes an encoder-decoder. Quantitative analysis module: Calculates the set of quantitative feature parameters for nodules based on the 3D segmentation mask data, wherein the calculation steps for the set of quantitative feature parameters are as follows: Receive 3D segmentation mask data, parse, traverse and extract coordinates of the independent lung nodules corresponding to each identifier ID, obtain voxel-nodule relationship, standard CT value of each voxel position, physical voxel size information of pixel spacing and slice thickness, and voxel coordinate set of each nodule. The set of quantified feature parameters includes nodule volume, maximum three-dimensional diameter, average CT value, and semantic morphological descriptor; The total number of voxels with a value of 1 in the 3D segmentation mask for each identifier ID is counted, and multiplied by the physical volume of a single voxel to obtain the nodule volume. Extract the set of voxel coordinates of the 3D segmentation mask for each identifier ID, and calculate the maximum value of the Euclidean distance between any two voxel coordinates in the set as the maximum 3D diameter. Based on the 3D segmentation mask coordinates of each identifier ID, backtrack to the normalized CT value array in the original CT image sequence, extract the gray value of the corresponding region, and calculate the average value; Calculate the geometric descriptors of the sphericity, nodule surface area, and surface area-to-volume ratio of the 3D segmentation mask for each identifier ID; The geometric descriptor is combined with the maximum three-dimensional diameter as a feature vector and input into a lightweight random forest classifier that has been trained on a nodule dataset labeled with semantic morphological descriptors "lobulation sign" and "spiculation sign" by radiologists. The classifier outputs the probability values ​​corresponding to "lobulation sign" and "spiculation sign" and obtains the binary judgment result by setting a decision threshold of 0.

5. All parameters calculated for each node are integrated into a structured data record, and after traversing all nodes, the structured data records corresponding to all nodes are integrated into the final set of quantized feature parameters. 3D Reconstruction and Fusion Engine: Based on the 3D image data volume and the 3D segmentation mask data, the engine performs parallel 3D model reconstruction of nodules and fusion of spatial relationship information to generate a lightweight and interactive 3D scene model. Temporal comparison module: acquires the target patient's previous CT images and the corresponding three-dimensional segmentation mask of the nodules, and generates comparison results by performing spatiotemporal registration and matching alignment; Automatic report generation engine: Creates a structured report template, receives the set of quantified feature parameters and the 3D scene model, as well as the comparison results, and automatically performs the filling, rendering and compositing operations of the structured report template to output an interactive electronic report.

2. The CT image-based three-dimensional quantitative analysis and report generation system for lung nodules according to claim 1, characterized in that, The steps for generating the 3D segmentation mask data are as follows: The three-dimensional image data volume is input into the pre-trained lung nodule segmentation deep learning model. The input is downsampled through the encoder path to extract multi-scale deep semantic features. Then, it is upsampled through the decoder path and connected to the feature skip connection of the corresponding level of the encoder. At the end of the decoder, the Sigmoid activation function is applied to each voxel position to output a three-dimensional probability map. Each voxel position corresponds to a value between 0 and 1 in the output, which represents the probability that the voxel at the corresponding position belongs to the lung nodule category. Threshold binarization is performed on the 3D probability map. A preset threshold of 0.5 is applied to convert each voxel in the 3D probability map into an initial binary foreground mask. A connected component analysis algorithm is used to perform three-dimensional connected component analysis on the binary foreground mask to identify and label interconnected voxel sets. A unique temporary label is assigned to each connected voxel set to generate a label matrix. Filtering noise regions requires calculating the volume of each labeled connected region. If the volume of a connected region is less than the preset minimum effective nodule volume threshold, the connected region is removed from the foreground and considered as noise. For each connected region retained after filtering, a morphological opening operation is applied to smooth the boundaries of the connected regions and remove small noise points; Each independent nodule region is assigned a unique identifier (ID) to generate the final 3D segmentation mask data.

3. The CT image-based three-dimensional quantitative analysis and report generation system for lung nodules according to claim 2, characterized in that, The semantic morphological descriptor includes at least one of the lobed sign and the spiculated sign.

4. The CT image-based three-dimensional quantitative analysis and report generation system for lung nodules according to claim 3, characterized in that, The steps for generating the 3D scene model are as follows: Based on the 3D image data volume and the 3D segmentation mask data, the 3D reconstruction and fusion engine starts two parallel processing threads: Thread 1: Use the moving cube algorithm to reconstruct the 3D segmentation mask of each nodule, generating a closed nodule surface mesh model composed of triangular facets. Thread 2 uses a seed-based region growing blood vessel segmentation algorithm for segmentation and construction. Based on voxel-normalized CT values, a centerline model representing the topological structure of blood vessels is generated, and the surface model can be selectively obtained by moving cube surface reconstruction of the binary mask of blood vessels. Meanwhile, a bronchus extraction algorithm based on connected component analysis threshold region growth is used for extraction and construction. From the same set of standardized CT value data, a bronchus centerline model is generated, and the surface model can be selectively obtained by moving cube surface reconstruction of the airway binary mask. Import the nodule surface mesh model, the vascular topology centerline model, the vascular surface model, the bronchus centerline model, and the bronchus surface model into a unified 3D scene coordinate system. By performing rigid or affine registration based on anatomical landmarks, spatial registration and consistent fusion of multiple models are ensured, and multiple models are assembled into a composite 3D scene model in a unified coordinate system after registration.

5. The CT image-based three-dimensional quantitative analysis and report generation system for lung nodules according to claim 1, characterized in that, The structured report template includes a basic information area at the top of the GUI, a lung nodule analysis report area in the main body of the GUI, and a functional interaction area at the bottom of the GUI; The basic information area at the top of the GUI includes the report title and basic patient information; The main GUI lung nodule analysis report area includes a quantitative feature parameter set summary area, a two-dimensional multi-plane reconstructed view area, a three-dimensional scene model visualization area, a structured description and conclusion area, and a follow-up chapter comparison area; The bottom functional interaction area of ​​the GUI includes icons for adjusting rotation / zoom / transparency in the 3D view area, as well as function buttons for exporting reports and sharing.

6. The system for three-dimensional quantitative analysis and report generation of pulmonary nodules based on CT images according to claim 1, characterized in that, The steps for generating the interactive electronic report are as follows: Load the structured report template, read the layout of each section in the structured report template, and determine the data positions and visualization interaction positions; Receive the comparison results of the quantized feature parameter set, the composite 3D scene model file, and the time series, and fill in the text data placeholders through string formatting; From the voxel-normalized CT value data, based on the center position of the nodule, PNG images of three orthogonal plane views (coronal, sagittal, and axial) are automatically extracted, generated, and saved, and then filled into the interactive visualization area. The composite 3D scene model file and three orthogonal planar views are placed on a static resource directory accessible to a web server built using a framework such as Flask / Django. The text data is rendered into an HTML page, and the web server assigns a unique URL endpoint based on the patient ID and report timestamp for each generated report. The glTF / glb file can be loaded directly using the Three.js library for interactive display, and can be accessed via a URL link.

7. The system for three-dimensional quantitative analysis and report generation of pulmonary nodules based on CT images according to claim 1, characterized in that, The spatiotemporal registration and matching alignment steps are as follows: A non-rigid image registration algorithm is used for temporal image registration, which globally aligns the CT image data of the current target patient with historical CT image data in terms of anatomical structure. Within the registered space, the 3D mask overlap and centroid distance between each current nodule and all historical nodules are calculated. By setting a pre-defined threshold, individual nodule matching is performed to determine nodules that are aligned with each other. For successfully matched nodule pairs, longitudinal time dimension analysis is performed by quantitatively calculating the volume change rate and morphological evolution index of the nodules, and a three-dimensional model comparison map of historical and current nodules is generated and the areas of volume growth are highlighted. The nodule pairs that have undergone spatiotemporal registration and matching, the volume change rate and morphological evolution index of the quantified variables, and the generated historical-current 3D model comparison map are packaged into a structured comparison result data package, which is then used to fill the follow-up section of an interactive electronic report.

8. The CT image-based three-dimensional quantitative analysis and report generation system for lung nodules according to claim 1, characterized in that, The generation method steps of the CT image-based 3D quantitative analysis and report generation system for lung nodules are as follows: S1. Access the chest CT image sequence of the target patient, call the pre-trained lung nodule segmentation deep learning model to process the chest CT image sequence, automatically identify and segment at least one lung nodule region, and output the pixel-level three-dimensional segmentation mask corresponding to each lung nodule. S2. Based on the three-dimensional segmentation mask, multi-dimensional feature extraction is performed on each lung nodule, and a set of quantitative feature parameters including nodule volume, maximum three-dimensional diameter, average CT value and semantic morphological descriptor is automatically calculated. S3. Based on the 3D segmentation mask, the 3D model reconstruction and spatial relationship information fusion are initiated in parallel. This includes first generating a 3D surface mesh model of each lung nodule, and simultaneously segmenting and reconstructing a 3D model of the pulmonary vascular tree / bronchial tree from the same chest CT image sequence. Then, the 3D surface mesh model of the nodule is spatially registered and superimposed with the 3D model of the pulmonary vascular tree / bronchial tree to obtain a lightweight and interactive 3D scene model containing spatial relationship information. S4. Pre-create a structured report template, automatically fill the corresponding data positions and visualization interaction positions of the structured report template with the quantitative feature parameter set and the three-dimensional scene model, and automatically generate a two-dimensional multi-plane reconstruction view containing at least one key morphology of the nodule, and synthesize an interactive electronic report that integrates quantitative data, two-dimensional multi-plane reconstruction view and embedded three-dimensional model. S5. In response to the existence of previous CT images of the target patient, a time-series comparison process is executed in the quantitative analysis. The current three-dimensional segmentation mask of the lung nodule is automatically registered and matched with the corresponding nodule mask in the previous CT images in a time-space manner. The comparison results of volume change rate and morphological evolution index are calculated and the comparison results are synchronously updated to the follow-up section of the interactive electronic report.

Images