A chest CT image segmentation and three-dimensional visualization method

Abstract

A method for segmenting and 3D visualization of chest CT images, belonging to the field of medical image processing and computer-aided diagnosis technology, comprises the following steps: S1, importing and standardizing preprocessing of medical image data; S2, segmenting of multiple tissue images of the thoracic cavity based on a hybrid strategy; S3, dual-mode 3D geometric reconstruction; and S4, interactive 3D visualization and rendering. Through these steps, this invention directly achieves accurate segmentation of multiple tissues and dual-mode 3D reconstruction of surfaces and volumes without relying on high-end hardware or complex manual operations, thus improving image analysis efficiency.

Description

Technical Field

[0001] This invention belongs to the field of medical image processing and computer-aided diagnosis technology, specifically relating to a method for segmenting and three-dimensional visualization of chest CT images. Background Technology

[0002] The rapid development of medical imaging technology, especially computed tomography (CT), has provided massive amounts of high-resolution volumetric data for the precise diagnosis and treatment planning of diseases. However, the stacked browsing method of two-dimensional tomographic images is no longer sufficient to meet the needs of modern clinical practice for understanding the spatial relationships of complex anatomical structures. Converting two-dimensional medical images into three-dimensional models has become a key technological direction for improving diagnostic efficiency, assisting surgical planning, and medical education.

[0003] In 3D visualization workflows, accurate tissue segmentation is a prerequisite for determining the quality of reconstructed models and their clinical application value. In recent years, deep learning technology has become the mainstream method for medical image segmentation. U-shaped network structures, especially 2D U-Net and its subsequent 3D variants, have shown particularly strong performance in this field due to their advantages in feature extraction and spatial information preservation. Studies have shown that by integrating different attention mechanisms and residual network configurations, the segmentation performance of models can be significantly improved. However, current segmentation techniques still face significant challenges. On the one hand, model performance heavily relies on large-scale, high-quality manually labeled data, but the high cost of acquiring such data limits the application of algorithms in primary healthcare institutions and for rare diseases. On the other hand, although many model variants have emerged, systematic benchmark studies for specific anatomical regions (such as the chest) are still lacking, leaving clinical researchers and engineers without clear guidelines for selecting and optimizing models. Furthermore, for tissues with low contrast and complex structures, such as blood vessels and small bronchi, existing algorithms still fall short in terms of segmentation completeness and accuracy, often resulting in fragmentation or missegmentation.

[0004] 3D reconstruction and visualization technologies, as a means of presenting segmented data, mainly fall into two major technical routes: surface rendering and volume rendering. Surface rendering algorithms, such as the classic moving cube algorithm, generate mesh models by extracting isosurfaces. They are computationally efficient, produce clear model structures, and are suitable for reconstructing hard tissues such as bones, but they lose internal detail information. Volume rendering algorithms, such as ray casting, directly sample and synthesize 3D volume data, completely preserving the grayscale information and internal structure of tissues. They have irreplaceable advantages in displaying soft tissues and vascular trees. However, volume rendering algorithms have high computational complexity and demanding requirements on hardware performance, especially graphics processors, resulting in slow rendering speeds and high interaction latency when processing large-scale data, making it difficult to achieve smooth real-time operation. Current technological optimizations, while improving rendering quality, often further exacerbate the contradiction between algorithm complexity and real-time performance.

[0005] Currently, there are several commercial and open-source medical image 3D reconstruction systems on the market, such as Mimics and 3DSlicer. These systems address the basic needs of converting images into 3D models to some extent, but they exhibit significant limitations in actual clinical deployment. First, semi-automatic software, such as Mimics, relies heavily on the operator's experience and patience for reconstruction quality and efficiency. The process of segmenting and differentiating different tissues (such as arteries and veins) is tedious and time-consuming, making it difficult to integrate into fast-paced clinical workflows. Second, most existing systems have relatively limited functionality, focusing either on traditional threshold-based segmentation and surface rendering or providing only a basic volume rendering viewer, lacking an integrated solution that deeply integrates cutting-edge deep learning automatic segmentation algorithms with high-quality, interactive volume rendering technology. Third, existing systems often have high requirements for computing resources (such as high-performance GPUs and large amounts of memory), and their software environment configuration is complex, making them difficult to promote in primary hospitals with limited hardware resources.

[0006] In summary, existing technologies face the following core bottlenecks in achieving intelligent, high-precision 3D visualization of chest CT images: In the reconstruction and visualization stages, it is difficult to simultaneously achieve high-quality volume rendering and efficient real-time interaction; at the system level, there is a lack of an integrated platform that combines advanced segmentation and rendering technologies, balances high precision and efficiency, and is easy to deploy and apply. Therefore, developing a chest CT image segmentation and 3D visualization method that can effectively address these issues is of great significance for promoting the clinical practice of precision medicine. Summary of the Invention

[0007] One objective of this invention is to propose a method for segmenting and visualizing chest CT images, which can directly achieve dual-mode three-dimensional reconstruction of accurate segmentation of multiple tissues and surface drawing without relying on high-end hardware and complex manual operation, thereby improving the efficiency of image analysis.

[0008] This invention is achieved through the following method: a method for segmenting and three-dimensional visualization of chest CT images, the steps of which are as follows: S1. Import and Standardized Preprocessing of Medical Imaging Data: Receive raw chest CT sequence data in DICOM or NIfTI format from imaging equipment; parse the file header of the data to extract slice image matrix, pixel spacing, slice thickness, and patient spatial orientation information; uniformly convert the extracted information to the system's internal coordinate system to construct three-dimensional volume data; perform Gaussian filtering on the three-dimensional volume data to suppress noise, and linearly transform the image grayscale contrast by adjusting the window width and window level to complete the standardized preprocessing of the data.

[0009] Gaussian filtering is achieved by convolving the image with a two-dimensional Gaussian convolution kernel. The formula for generating the Gaussian convolution kernel is: (1) Among them, G( x , y ) indicates that the generated Gaussian convolution kernel is located at coordinates ( x , y The value at () x , y ) represents the coordinates of each point within the convolution kernel relative to the kernel center, and σ is the standard deviation, which determines the smoothness of the filter. The original image is convolved using this kernel to obtain the smoothed image: I_smooth = I_original. G, thereby suppressing noise generated during image acquisition.

[0010] S2. Multi-tissue image segmentation of the thoracic cavity based on a hybrid strategy: Segmentation operations are performed on the preprocessed 3D volume data for cardiac tissue segmentation, vascular network segmentation, lung parenchyma segmentation, and skeletal structure segmentation.

[0011] S2.1) Heart tissue segmentation: Seed points are manually set on a 2D slice. Using a region growing algorithm, iterative growth is performed in 3D space based on a preset gray-level similarity threshold to extract the heart tissue region. The difference between the average gray-level value of the neighboring pixel p and the current region R is used to determine whether to merge. The criterion is: if |I(p) - μ_R| ≤ τ, then pixel p is merged into region R and μ_R is updated. τ is a preset threshold, where I(p) is the gray-level value of the neighboring pixel p, μ_R is the arithmetic mean of the gray levels of all pixels in the current growing region R, and τ is a preset gray-level similarity threshold, which is a non-negative integer used to control the strictness of region growth. The larger the value of τ, the more lenient the growth conditions. S2.2) Vascular network segmentation: Calculate the Hessian matrix for each point in the image. Construct a vascular enhancement filter by analyzing its eigenvalues ​​to highlight tubular structures. Then, perform threshold segmentation and morphological post-processing to extract the vascular tree. The Hessian matrix describes the second derivative in the neighborhood of a point in the image. For 3D images, its definition is as follows: (2) By analyzing the three eigenvalues ​​of the Hessian matrix |λ1|≤|λ2|≤|λ3|, a vascular enhancement response function is constructed. S2.3) Lung parenchyma segmentation: Two-dimensional slices or three-dimensional volume data blocks are input into a pre-trained U-Net convolutional neural network model. The model outputs an end-to-end lung parenchyma probability map through an encoder-decoder structure and skip connections. After thresholding, an accurate segmentation mask is obtained. The U-Net model learns the end-to-end mapping from the image to the segmentation mask through convolution, pooling, and upsampling operations. Its loss function usually adopts binary cross-entropy or Dice loss. S2.4) Skeletal structure segmentation: Initial segmentation is performed using a fixed high threshold (e.g., thresh = 226 HU, corresponding to the typical CT value of bones). That is, for any pixel p in the image, the segmentation result Seg(p) is defined as: (3) Seg(p) is the segmentation result, I(p) is the gray value of pixel p, and its unit is Henle unit (HU) in CT images; thresh is the preset bone segmentation threshold, and then morphological closing operation and hole filling are performed to obtain the complete bone structure.

[0012] S3. Dual-mode 3D geometric reconstruction: Based on the segmentation results obtained in step two, the user selects a reconstruction mode to perform 3D geometric reconstruction. The reconstruction modes include surface rendering reconstruction mode and volume rendering reconstruction mode.

[0013] S3.1) Surface Reconstruction: The Moving Cube algorithm is used, employing a binary segmentation mask of a specific tissue as a scalar field. By traversing voxels, searching the topological configuration table, and constructing triangular facets within the voxels, a triangular mesh model of the tissue surface is generated. The McLeod algorithm calculates the intersection of the isosurface and the voxel edge within each voxel cell using linear interpolation. If the scalar values ​​of the two endpoints v1 and v2 of the voxel edge are f(v1) and f(v2), respectively... 2) Given that the isosurface threshold is c, the intersection point coordinates P are obtained by linear interpolation using the following formula: (4) By connecting all the intersections within the same voxel, triangular facets are formed, which are eventually combined into a complete surface mesh. S3.2) Volume Reconstruction: A ray casting algorithm is used to resample the original or preprocessed 3D grayscale volume data by emitting a ray from a screen pixel. The sampled values ​​are then mapped to color and opacity using a transfer function, and optical integral synthesis is performed to generate a 2D projected image that preserves internal details. The ray casting algorithm is based on an emission and absorption model; the final color C and opacity A of the pixels on the screen are determined by the integral along the ray path. After discretization, a front-to-back or back-to-front synthesis formula is typically used. For example, the front-to-back synthesis formula is: (5) Among them, C i and α i C represents the color and opacity of the current sampling point, respectively. accum and A accum The accumulated color and opacity are initialized to 0.

[0014] S4. Interactive 3D Visualization and Rendering: Construct a graphical user interface and integrate a 3D rendering engine; synchronously display the 3D reconstructed model and 2D orthogonal slice views of cross sections, coronal planes, and sagittal planes in the interface, maintaining the linkage of all views; receive commands issued by the user through input devices to realize real-time rotation, translation, and scaling operations on the 3D model, as well as independent adjustment of the transparency and color attributes of different tissue models; provide controls for users to switch and select between the aforementioned two reconstruction modes of surface drawing and volume drawing.

[0015] The beneficial effects of this invention are as follows: This invention achieves precise segmentation of the heart, blood vessels, lung parenchyma, and bones through a hybrid segmentation strategy, combining traditional algorithms with U-Net deep learning, effectively improving the segmentation accuracy of low-contrast tissues. The integrated dual-mode reconstruction of surface rendering (MC algorithm) and volume rendering (ray casting method) balances reconstruction speed and detail restoration, allowing users to flexibly choose according to clinical needs. The system provides real-time interaction with the 3D model (rotation, scaling, and transparency adjustment) and interactive 2D three-view functionality, greatly optimizing diagnostic and surgical planning efficiency. Furthermore, the integrated workflow from data import to preprocessing, segmentation, reconstruction, and interaction is packaged into an independent executable file, reducing hardware dependence and deployment barriers, and possessing significant clinical application value. Attached Figure Description

[0016] Figure 1 This is a flowchart of the present invention.

[0017] Figure 2 The diagram shows the interface of the 3D visualization system; (a) is a description of the main interface and function areas of the system, and (b) is a 3D visualization effect of the system. Detailed Implementation

[0018] This embodiment uses the system of the present invention to process a set of real clinical chest CT image data. The specific implementation steps are as follows: S1: Data Import Users can use the "Import" function on the system interface to select a local folder containing chest CT scan sequences. The system supports the DICOM standard format. The system automatically reads all slice files within the folder, parses key information such as pixel data, image orientation, slice thickness, and pixel spacing, and correctly sorts and stacks them according to instance number, constructing a three-dimensional volumetric data in memory. Subsequently, the system's front-end interface updates synchronously, displaying views in three orthogonal planes—transverse, coronal, and sagittal—for initial browsing and positioning by the user.

[0019] S2: Image Preprocessing

[0020] The system preprocesses the raw data to optimize image quality. First, Gaussian filtering is automatically applied to smooth the image, suppressing noise while preserving key edges. Furthermore, the image's display contrast can be adjusted in real-time and interactively using window width and level controls on the interface to highlight the grayscale characteristics of different tissues, facilitating subsequent observation and segmentation operations.

[0021] S3: Multi-tissue image segmentation

[0022] (1) Blood vessel segmentation: The system automatically performs a blood vessel segmentation algorithm on the body data. The algorithm uses the Hessian matrix to filter and enhance the image to highlight the tubular structure features, and then combines threshold segmentation to extract the blood vessel network.

[0023] (2) Lung parenchyma segmentation: When the user clicks the “Lung parenchyma segmentation” button, the system automatically calls the pre-trained U-Net deep learning model. This model performs end-to-end semantic segmentation on the input CT slices and directly outputs a binary segmentation mask of the lung parenchyma.

[0024] (3) Skeletal segmentation: Based on the physical characteristics of CT values, the system applies threshold segmentation to the original data and optimizes the results through morphological operations to obtain a complete skeletal structure.

[0025] S4: Dual-mode 3D Reconstruction

[0026] After segmentation, the three-dimensional reconstruction mode can be selected according to different clinical observation needs: (1) Surface rendering: When the user selects the “Surface rendering” mode, the system applies the moving cube algorithm to the segmentation results of the specified tissue. The algorithm traverses the three-dimensional voxel mesh, extracts isosurfaces and constructs a triangular mesh model, which is suitable for displaying clear tissue surface structures.

[0027] (2) Volume rendering: When the user selects the "volume rendering" mode, the system applies a ray casting algorithm to the volume data. The algorithm simulates light penetrating the three-dimensional data field and synthesizes color and transparency to generate a projected image that retains continuous internal grayscale information and details, which is more suitable for observing the layers and spatial relationships of soft tissues.

[0028] S5: Interactive Visual Operation

[0029] The generated 3D model is rendered in real-time in the 3D view window of the system's main interface. The model can be rotated, scaled, and translated in real-time using the mouse. In the system's control panel, layered display and transparency adjustments can be selected. By dragging the sliders, the transparency of each component can be flexibly adjusted to achieve complex layered transparency effects. The virtual cutting plane in the 3D rendering window and the other 2D views are updated synchronously in real-time, enabling multi-view positioning.

[0030] S6: Output Results

[0031] All processing results can be exported for subsequent applications. Users can export the triangular mesh models generated by surface rendering in common 3D formats such as STL or OBJ for 3D printing or further analysis. Simultaneously, the tissue segmentation mask data generated by the system can also be saved in a standardized format.

Claims

1. A method for segmenting and three-dimensional visualization of chest CT images, characterized in that, The steps are as follows: S1. Import and Standardized Preprocessing of Medical Imaging Data: Receive raw chest CT sequence data in DICOM or NIfTI format from imaging equipment; parse the file header of the data, extract slice image matrix, pixel spacing, slice thickness and patient spatial orientation information; convert the extracted information to the system's internal coordinate system to construct three-dimensional volume data; Gaussian filtering is performed on the 3D volume data to suppress noise, and the image grayscale contrast is linearly transformed by adjusting the window width and window level to complete the data standardization preprocessing; S2. Multi-tissue image segmentation of the thoracic cavity based on a hybrid strategy: Segmentation operations are performed on the preprocessed 3D volume data for cardiac tissue segmentation, vascular network segmentation, lung parenchyma segmentation, and skeletal structure segmentation. S3. Dual-mode 3D geometric reconstruction: Based on the segmentation results obtained in step two, the user selects a reconstruction mode to perform 3D geometric reconstruction. The reconstruction modes include surface drawing reconstruction mode and volume drawing reconstruction mode. S4. Interactive 3D Visualization and Rendering: Constructing a graphical user interface and integrating a 3D rendering engine; The interface synchronously displays the 3D reconstructed model and 2D orthogonal slice views of cross sections, coronal planes, and sagittal planes, maintaining the linkage of all views; it receives commands from users through input devices to realize real-time rotation, translation, and scaling operations on the 3D model, as well as independent adjustment of the transparency and color attributes of different tissue models; it provides controls for users to switch and select between the aforementioned two reconstruction modes of surface drawing and volume drawing.

2. The method for segmenting and three-dimensional visualization of chest CT images according to claim 1, characterized in that, In S1, the Gaussian filtering operation is implemented by convolving the image with a two-dimensional Gaussian convolution kernel. The formula for generating the Gaussian convolution kernel is: （1） Among them, G( x , y ) indicates that the generated Gaussian convolution kernel is located at coordinates ( x , y The value at () x , y ) represents the coordinates of each point within the convolution kernel relative to the kernel center, and σ is the standard deviation, which determines the smoothness of the filter. The original image is convolved using this kernel to obtain the smoothed image: I_smooth = I_original. G, thereby suppressing noise generated during image acquisition.

3. The method for segmenting and three-dimensional visualization of chest CT images according to claim 1, characterized in that, In S2: S2.1) Heart tissue segmentation: Seed points are manually set on two-dimensional slices, and the heart tissue region is extracted by iterative growth in three-dimensional space using a region growing algorithm based on a preset gray-level similarity threshold. The decision to merge pixels is made by calculating the difference between the average gray value of neighboring pixel p and the current region R. The criterion is: if |I(p) - μ_R| ≤ τ, then pixel p is merged into region R and μ_R is updated. τ is a set threshold, where I(p) is the gray value of neighboring pixel p, μ_R is the arithmetic mean of the gray values ​​of all pixels in the current growth region R, and τ is a preset gray-level similarity threshold, which is a non-negative integer used to control the strictness of region growth. The larger the value of τ, the more lenient the growth conditions. S2.2) Vascular network segmentation: Calculate the Hessian matrix for each point in the image. Construct a vascular enhancement filter by analyzing its eigenvalues ​​to highlight tubular structures. Then, perform threshold segmentation and morphological post-processing to extract the vascular tree. The Hessian matrix describes the second derivative in the neighborhood of a point in the image. For 3D images, its definition is as follows: （2） By analyzing the three eigenvalues ​​of the Hessian matrix |λ1|≤|λ2|≤|λ3|, a vascular enhancement response function is constructed. S2.3) Lung parenchyma segmentation: Two-dimensional slices or three-dimensional volume data blocks are input into a pre-trained U-Net convolutional neural network model. The model outputs an end-to-end lung parenchyma probability map through an encoder-decoder structure and skip connections. After thresholding, an accurate segmentation mask is obtained. The U-Net model learns the end-to-end mapping from the image to the segmentation mask through convolution, pooling, and upsampling operations. Its loss function usually adopts binary cross-entropy or Dice loss. S2.4) Skeletal structure segmentation: Initial segmentation is performed using a fixed high threshold. That is, for any pixel p in the image, the segmentation result Seg(p) is defined as: （3） Where Seg(p) is the segmentation result, I(p) is the gray value of pixel p, and its unit is Henle unit (HU) in CT images; thresh is the preset bone segmentation threshold, and then morphological closing operation and hole filling are performed to obtain the complete bone structure.

4. The method for segmenting and three-dimensional visualization of chest CT images according to claim 1, characterized in that, In S3: S3.1) Surface Reconstruction: The Moving Cube algorithm is used, employing a binary segmentation mask of a specific tissue as a scalar field. By traversing voxels, searching the topological configuration table, and constructing triangular facets within the voxels, a triangular mesh model of the tissue surface is generated. The McLeod algorithm calculates the intersection of the isosurface and the voxel edge within each voxel cell using linear interpolation. If the scalar values ​​of the two endpoints v1 and v2 of the voxel edge are f(v1) and f(v2), respectively... 2) Given that the isosurface threshold is c, the intersection point coordinates P are obtained by linear interpolation using the following formula: （4） By connecting all the intersections within the same voxel, triangular facets are formed, which are eventually combined into a complete surface mesh. S3.2) Volume Reconstruction: A ray casting algorithm is used to resample the original or pre-processed 3D grayscale volume data by emitting a ray from a screen pixel. The sampled values ​​are then mapped to color and opacity using a transfer function, and optical integral synthesis is performed to generate a 2D projected image that preserves internal details. The ray casting algorithm is based on an emission and absorption model; the final color C and opacity A of the pixels on the screen are determined by the integral along the ray path. After discretization, a synthesis formula is used: （5） Among them, C i and α i C represents the color and opacity of the current sampling point, respectively. accum and A accum The accumulated color and opacity are initialized to 0.

Images