Method for ct image three-dimensional visualization reconstruction of high-density component composite material based on deep learning

Abstract

The application discloses a CT image three-dimensional visualization reconstruction method of a composite material containing high-density components based on deep learning, and comprises the following steps: acquiring two-dimensional CT slice image data and inputting the same as initial data; pre-processing the input initial data; marking a plurality of regions of interest and merging the same into multiple regions of interest; selecting a U-net 3D model, setting corresponding training parameters and performing deep learning on the multiple regions of interest; and reconstructing a three-dimensional visualization image and performing quantitative analysis. The method combines artificial marking and deep learning, realizes reconstruction of the internal microstructure of the composite material containing high-density components, overcomes the difficulty that the high-density components cause artifacts, the other components have low contrast and are difficult to identify. The reconstruction result of the application is close to the real situation and has a certain universality, and provides a good foundation for visualization reconstruction, microstructure evolution and preparation process improvement of the composite material containing high-density components.

Description

Technical Field

[0001] This invention relates to the field of industrial CT image processing technology, specifically to a method for three-dimensional visualization reconstruction of CT images based on deep learning and high-density component composite materials. Background Technology

[0002] X-ray computed tomography (CT) technology is widely used in industry to accurately acquire the three-dimensional structural features of materials, thereby effectively performing non-destructive testing on the interiors of various types of materials. In recent years, industrial CT has gradually become a research hotspot in the analysis and characterization of composite materials, serving as an auxiliary tool for research on damage evolution and failure analysis of composite materials. Industrial CT technology is currently one of the most intuitive and reliable methods for non-destructive testing of the internal microstructure of composite materials.

[0003] The detection accuracy of high-precision industrial CT is ultimately directly determined by the accuracy of image segmentation. Efficient and accurate image segmentation methods are crucial prerequisites for achieving precise visualization of materials and qualitative or quantitative analysis of the internal physical characteristics of the inspected object. Currently, while image segmentation based on thresholds, edges, and regions is convenient and efficient, it is severely affected by noise and artifacts, resulting in insufficient segmentation accuracy and poor versatility. Relying solely on such segmentation methods is far from meeting the characterization needs of complex materials. Furthermore, segmentation methods based on manually extracted target features are costly, have low accuracy, and are highly individualized, making further development and widespread application difficult.

[0004] However, current image analysis tools for CT data are mostly developed for medical applications and cannot automatically identify special regions in the field of materials engineering. As multi-component materials, composite materials have complex and diverse internal structures, and traditional CT image segmentation methods cannot accurately obtain their internal structural features such as components, defects, and pores. Therefore, efficiently and accurately acquiring CT image data information of composite materials has become a difficult point and challenge for CT image segmentation technology.

[0005] With the development of artificial intelligence technology and algorithms, deep learning-based image segmentation has demonstrated its potential. Deep convolutional neural networks (CNNs) are characterized by data-driven approaches, strong applicability, and automatic extraction of data features. They provide opportunities to learn how to segment an entire dataset or other similar datasets from limited samples, greatly improving detection accuracy and image recognition robustness, and effectively solving the problems caused by traditional segmentation methods. Patent CN112330682B discloses an industrial CT segmentation network model based on an encoder-decoder structure using deep convolutional neural networks. This segmentation model has advantages such as high efficiency, high accuracy, certain versatility, and avoidance of subjective interference. Patent CN115115610A discloses a method for identifying internal defects in composite materials using an improved convolutional neural network, which improves recognition accuracy and speed. The above methods have all improved the algorithms and demonstrated the effectiveness of deep learning in industrial CT image segmentation. However, due to the lack of objective solutions for the number of reference samples and microstructure required for the initial evaluation and recognition of such materials, as well as the lack of evaluation of the accuracy of the training parameters selected for the deep learning model, deep learning-based image segmentation technology still needs breakthroughs in the field of composite materials. Currently, research on CT image segmentation schemes for composite materials is scarce and lacks a systematic approach. There is also a lack of relevant data on CT image segmentation technology for composite materials containing high-density components. However, the detection of the evolution of the internal microstructure of composite materials is of great significance for the preparation process and material densification. Therefore, proposing a three-dimensional visualization reconstruction method for CT images of composite materials containing high-density components based on deep learning is of great significance. Summary of the Invention

[0006] To address the shortcomings of CT image segmentation technology for composite materials with added high-density components, this invention proposes a deep learning-based method for 3D visualization reconstruction of CT images of composite materials containing high-density components. By combining machine labeling, manual labeling, and deep learning, accurate 3D reconstruction results of composite materials can be obtained efficiently, thus solving the problem of difficult segmentation between components of composite materials after adding high-density components.

[0007] To achieve the above objectives, the present invention adopts the following technical solution: A deep learning-based method for 3D visualization reconstruction of CT images containing high-density component composite materials includes the following steps: (1) For SiC containing high-density components f Two-dimensional CT slice images of SiC composite materials were obtained through CT scanning and used as initial data for three-dimensional visualization reconstruction. The high-density component consisted of tungsten powder and zirconium boride powder. CT scans were performed on two SiC composites with different amounts of this high-density component. f / SiC composite material, wherein the mass fraction of high-density component added to sample I is 3%-5%, and the mass fraction of high-density component added to sample II is 15%-20%.

[0008] (2) Preprocess the initial data input in step (1) to obtain preprocessed two-dimensional CT slice image data; (3) First, perform threshold segmentation on the preprocessed 2D CT slice image data in step (2), select an appropriate gray value range, distinguish the region of high-density components in the preprocessed 2D CT slice image data, and mark this region as a region of interest, naming this region of interest as metal; then select the complementary range of the segmented gray value range of high-density components, and establish another region of interest, named the whole; then uniformly select a preset number of 2D CT slice images in the 2D CT slice image data, and establish different regions of interest for the whole on the selected 2D CT slice images by manual marking, named pores and SiC. f / SiC, and finally merges the different regions of interest into a single multi-region of interest; (4) Based on the 3D U-net model, set the corresponding training parameters, perform deep learning on the multiple regions of interest obtained in step (3), and obtain the trained 3D U-net model after learning; (5) Apply the model trained in step (4) to the two-dimensional CT slice image data to reconstruct SiC containing high-density components. f Three-dimensional visualization images of SiC composite materials and their quantitative analysis.

[0009] Furthermore, in step (1), the SiC containing high-density components... f The specific details of the CT scan of the SiC composite material are as follows: the resolution used for the scan is 1.0047 μm / pixel, and the data is input in TIFF format.

[0010] Furthermore, the specific content of step (2) is as follows: preprocessing includes first cropping the non-material structure of the outer solid area of ​​the two-dimensional CT slice image data, retaining the area that needs to be analyzed and reconstructed, and reducing the amount of data that needs to be processed later; secondly, optimizing the two-dimensional CT slice image data so that the gray values ​​of each component remain different, which facilitates the subsequent manual labeling work.

[0011] Furthermore, the specific content of the manual marking in step (3) is: distinguishing between pores and SiC. f When using SiC, Boolean operations are used to select pores within the whole; when selecting manually, the metal segmented by a threshold is used as the boundary; pores and SiC f / SiC identification was performed using rolling two-dimensional CT slice image data; ultimately, the whole was divided into pores and SiC. f / SiC; The number of 2D CT slice images is selected as 10-20, and the manual labeling accounts for 1-1.5% of the 2D CT slice image data; In order to correct the manually labeled area and shorten the labeling time, local Otsu and morphological operations are used to refine the pores.

[0012] Furthermore, the specific content of the morphological operation in step (3) is as follows: when performing morphological operations, the structural element of the operation is selected to be spherical in shape, with a core size of 3-5, and 1-3 erosions and 1-3 expansions are performed to remove the diffuse point-like areas caused by the increase of high-density components. Further, the specific content of the training parameters in step (4) is as follows: the block size is set to 32, the step size is 0.5-1, the batch size is 64, 128 or 256, the number of times the image segmentation algorithm returns is 100, the loss function Orsdiceloss and the optimization algorithm Adadelta are selected to train the labeled voxels, and training is stopped when the number of times is 40-80; 20%-30% of the training data is used as validation data.

[0013] Further, the specific content of step (5) is as follows: the quantitative analysis adopts connectivity analysis, creates a six-connected multiple region of interest for the pores and selects the volume as the analysis object to obtain the proportion of pores in different volume ranges and the three-dimensional reconstruction distribution map.

[0014] The beneficial effects and advantages of this invention are: 1. This invention realizes the three-dimensional reconstruction and quantitative analysis of composite materials with added high-density components, forms a complete processing scheme and process, demonstrates the operation steps of manual labeling and the training parameters required in deep learning, provides a corresponding reference for the segmentation and reconstruction processing of CT images of this type of material, and provides a good auxiliary means and analytical basis for the optimization of composite material preparation, monitoring of densification process, damage evolution and failure analysis of composite materials with added high-density components; 2. By combining threshold segmentation, morphological operations, and Boolean operations in an orderly manner, high-density components are accurately labeled using threshold segmentation. Boolean operations are then used to distinguish components with similar gray values. Morphological operations are used to refine the process. This reasonable labeling method achieves accurate and efficient segmentation of different components in composite materials with added high-density components, greatly reducing image processing costs.

[0015] 3. Without altering the algorithm, this invention utilizes a 3D U-net model to efficiently and accurately segment and visualize the three-dimensional reconstruction of components in composite materials with added high-density components, thus improving the method's versatility. This method is simple to operate and requires no advanced technical expertise from the user. Attached Figure Description

[0016] Figure 1 The image processing flowchart provided by this invention; Figure 2 Two-dimensional CT slice images of sample I and sample II in this invention; (a) Two-dimensional CT slice image of sample I; (b) Two-dimensional CT slice image of sample II; Figure 3 The images shown are: (a) Threshold segmentation effect; (b) Manual labeling effect; (c) Deep learning effect; (d) 3D visualization reconstruction of CT images. Detailed Implementation

[0017] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments: Example 1: like Figure 1 As shown in the figure, this embodiment of the method for three-dimensional visualization reconstruction of CT images containing high-density component composite materials based on deep learning includes the following steps: (1) For SiC containing high-density components f Two-dimensional CT slice images of SiC composite materials were acquired using CT scanning at a resolution of 1.0047 μm / pixel. The acquired 2D CT slice images were input in TIFF format as the initial data for 3D visualization reconstruction. The high-density component consisted of tungsten powder and zirconium boride powder. CT scans were performed on SiC composite materials with the high-density component added. f / SiC composite material sample I, wherein the mass fraction of the high-density component added to sample I is 4.66%; (2) Preprocess the initial data input in step (1). The preprocessing includes first cropping the non-material structure peripheral solid areas of the two-dimensional CT slice image data to retain the areas that need to be analyzed and reconstructed, thereby reducing the amount of data to be processed later; secondly, optimizing the two-dimensional CT slice image data, with the window level of sample I selected in the range of 9685.98-27262.87, so that the gray values ​​of each component remain different, which facilitates the subsequent manual labeling work. The preprocessed two-dimensional CT slice image data is obtained, as follows: Figure 2 As shown in (a), the two-dimensional CT slice image data of sample I after preprocessing can clearly distinguish the metal, pores and SiC. f The three regions of interest in SiC will help in the marking of the regions of interest in the subsequent step (3).

[0018] (3) Threshold segmentation is first performed on the preprocessed two-dimensional CT slice image data in step (2). The threshold segmentation range for sample I is selected as 17750.20-59035.65. The region of high-density components in the preprocessed two-dimensional CT slice image data is distinguished, and this region is marked as a region of interest. This region of interest is named metal, such as Figure 3 As shown in (a)(I), the metal has been accurately marked after thresholding, and the marked areas correspond to the color block annotations below. Then, the complementary range of the gray value range of the segmented high-density components is selected to establish another region of interest, named the whole. Next, 12 two-dimensional CT slice images are evenly selected from the two-dimensional CT slice image data. Different regions of interest are established on the whole through manual marking on the selected 12 two-dimensional CT slice images, named pores and SiC. f / SiC, such as Figure 3 As shown in (b)(I), after thresholding and artificial labeling, metals, pores, and SiC... f The three regions of interest (ROIs) in the SiC image were completely distinguished, and each ROI corresponds to a color block annotation at the bottom of the image. Manually labeled voxels account for 1.10% of the total voxels in the 2D CT slice image data. Finally, the established ROIs were merged into a single multi-ROI. To correct the manually labeled areas and shorten the labeling time, local Otsu and morphological operations were used to refine the porosity. During the morphological operations, the structural element was selected as a spherical shape with a kernel size of 3. One erosion and one expansion were performed to remove diffuse point-like regions caused by the increase in high-density components. (4) Based on the 3D U-net model, set the corresponding training parameters, set the block size to 32, the step size to 0.5, the batch size to 256, and the number of times the image segmentation algorithm returns to 100. Select the loss function Orsdiceloss and the optimization algorithm Adadelta to train the labeled voxels. Stop training at the 50th iteration. Use 20% of the training data as validation data. Perform deep learning on the multiple regions of interest obtained in step (3) to obtain the trained 3D U-net model. (5) Apply the model trained in step (4) to the two-dimensional CT slice image data to reconstruct SiC containing high-density components. f Three-dimensional visualization images of SiC composite materials and their quantitative analysis, such as Figure 3 As shown in (c)(I) and (d)(I), after deep learning, the three regions of interest in the two-dimensional CT slice image data of sample I are reasonably distinguished, and the reconstructed three-dimensional visualization image does not show distortion. However, there are still recognition errors between different regions of interest. Among them, SiC fThe error for SiC is 9.42%, and the error for metals is 9.82%. The quantitative analysis employs connectivity analysis, creating six-connected multiple regions of interest for the pores and selecting volume as the analysis object to obtain the proportion of pores within different volume ranges and a three-dimensional reconstructed distribution map.

[0019] Example 2: like Figure 1 As shown in the figure, this embodiment of the method for three-dimensional visualization reconstruction of CT images containing high-density component composite materials based on deep learning includes the following steps: (1) For SiC containing high-density components f Two-dimensional CT slice images of SiC composite materials were acquired using CT scanning at a resolution of 1.0047 μm / pixel. The acquired 2D CT slice images were input in TIFF format as the initial data for 3D visualization reconstruction. The high-density component consisted of tungsten powder and zirconium boride powder. CT scans were performed on SiC composite materials with the high-density component added. f / SiC composite material sample II, wherein the mass fraction of the high-density component added to sample II is 16.87%; (2) Preprocess the initial data input in step (1). The preprocessing includes first cropping the non-material structure peripheral solid areas of the two-dimensional CT slice image data to retain the areas that need to be analyzed and reconstructed, thereby reducing the amount of data to be processed later; secondly, optimizing the two-dimensional CT slice image data, with the window level of sample I selected in the range of 3978.98-68162.48, so that the gray values ​​of each component remain different, which facilitates the subsequent manual labeling work. The preprocessed two-dimensional CT slice image data is obtained, as follows: Figure 2 As shown in (b), the two-dimensional CT slice image data of sample II after preprocessing can clearly distinguish the metal, pores, and SiC. f The three regions of interest in SiC will help in the marking of the regions of interest in the subsequent step (3).

[0020] (3) Threshold segmentation is first performed on the preprocessed two-dimensional CT slice image data in step (2). The threshold segmentation range for sample I is selected as 19698.99-59710.86. The region of high-density components in the preprocessed two-dimensional CT slice image data is distinguished, and this region is marked as a region of interest. This region of interest is named metal, such as Figure 3(a)(II); As can be seen from the figure: after thresholding, the metal has been accurately marked, and the marked area corresponds to the color block annotation below; then, the complementary range of the gray value range of the segmented high-density component is selected to establish another region of interest, named the whole; then, 15 two-dimensional CT slice images are evenly selected from the two-dimensional CT slice image data, and different regions of interest are established on the whole by manual marking on the selected 15 two-dimensional CT slice images, named pores and SiC. f / SiC, such as Figure 3 In Figure (b)(II), it can be seen that after thresholding and artificial labeling, the metal, pores, and SiC... f The three regions of interest (ROIs) in the SiC image were completely distinguished, and each ROI corresponds to a color block annotation at the bottom of the image. Manually labeled voxels account for 1.26% of the total voxels in the 2D CT slice image data. Finally, the established ROIs were merged into a single multi-ROI. To correct the manually labeled areas and shorten the labeling time, local Otsu and morphological operations were used to refine the porosity. During morphological operations, the structural element was selected as a spherical shape with a kernel size of 3. One erosion and two expansion operations were performed to remove diffuse point-like regions caused by the increase in high-density components. (4) Based on the 3D U-net model, set the corresponding training parameters, set the block size to 32, the step size to 0.5, the batch size to 128, and the number of times the image segmentation algorithm returns to 100. Select the loss function Orsdiceloss and the optimization algorithm Adadelta to train the labeled voxels. Stop training at the 76th iteration. Use 25% of the training data as validation data. Perform deep learning on the multiple regions of interest obtained in step (3) to obtain the trained 3D U-net model. (5) Apply the model trained in step (4) to the two-dimensional CT slice image data to reconstruct SiC containing high-density components. f Three-dimensional visualization images of SiC composite materials and their quantitative analysis, such as Figure 3 Figures (c)(II) and (d)(II) show that after deep learning, the three regions of interest in the two-dimensional CT slice image data of sample II are reasonably distinguished, and the reconstructed three-dimensional visualization image does not show distortion. However, there are still recognition errors between different regions of interest. Among them, SiC f The error for SiC is 2.58%, and the error for metals is 10.49%. The quantitative analysis employs connectivity analysis, creating six-connected multiple regions of interest for the pores and selecting volume as the analysis object to obtain the proportion of pores within different volume ranges and a three-dimensional reconstructed distribution map.

Claims

1. A method for three-dimensional visualization reconstruction of CT images containing high-density component composite materials based on deep learning, characterized in that, Includes the following steps: Step 1: SiC containing high-density components f Two-dimensional CT slice image data of / SiC composite material is obtained by CT scanning, and the obtained two-dimensional CT slice image data is used as the initial data for three-dimensional visualization reconstruction. The high-density component consisted of tungsten powder and zirconium boride powder. CT scans were performed on two SiC substrates with different amounts of the high-density component. f / SiC composite material, wherein the mass fraction of high-density component added to sample I is 3%-5%, and the mass fraction of high-density component added to sample II is 15%-20%; Step 2: Preprocess the initial data input in Step 1 to obtain preprocessed 2D CT slice image data; Step 3: First, perform threshold segmentation on the preprocessed 2D CT slice image data from Step 2, select an appropriate grayscale range, distinguish the regions of high-density components in the preprocessed 2D CT slice image data, and mark this region as a region of interest, named "metal". Then, select the complementary range of the segmented high-density component grayscale range to establish another region of interest, named "the whole". Next, uniformly select a preset number of 2D CT slice images from the 2D CT slice image data, and manually mark different regions of interest on the "whole" on the selected 2D CT slice images, named "pores" and "SiC". f / SiC, and finally merges the different regions of interest into a single multi-region of interest; Step 4: Based on the 3D U-net model, set the corresponding training parameters, perform deep learning on the multiple regions of interest obtained in Step 3, and obtain the trained 3D U-net model after learning; Step 5: Apply the model trained in Step 4 to the 2D CT slice image data to reconstruct SiC containing high-density components. f Three-dimensional visualization images of SiC composite materials and their quantitative analysis; Step 2 specifically involves: preprocessing, which first involves cropping the non-material structure peripheral solid regions of the 2D CT slice image data to retain the areas that need to be analyzed and reconstructed, thereby reducing the amount of data that needs to be processed later; secondly, optimizing the 2D CT slice image data to maintain the difference in grayscale values ​​between each component, which facilitates the subsequent manual labeling work. The specific content of the manual marking in step 3 is to distinguish between pores and SiC. f When using SiC, Boolean operations are used to select pores within the whole; when selecting manually, the metal segmented by a threshold is used as the boundary; pores and SiC f / SiC identification was performed using rolling two-dimensional CT slice image data; ultimately, the whole was divided into pores and SiC. f / SiC; The number of 2D CT slice images is selected as 10-20, and the manual labeling accounts for 1-1.5% of the 2D CT slice image data; In order to correct the manually labeled areas and shorten the labeling time, local Otsu and morphological operations are used to refine the pores; The specific content of morphological operations is as follows: When performing morphological operations, the structural element of the operation is selected to be spherical in shape, with a core size of 3-5, and 1-3 erosions and 1-3 expansions are performed to remove the diffuse point-like areas caused by the increase of high-density components.

2. The method for three-dimensional visualization reconstruction of CT images containing high-density component composite materials based on deep learning according to claim 1, characterized in that, Step 1: SiC containing high-density components f The SiC composite material underwent CT scanning, as detailed below: The scanning resolution was 1.0047 μm / pixel, and the data was input in TIFF format.

3. The method for three-dimensional visualization reconstruction of CT images containing high-density component composite materials based on deep learning according to claim 1, characterized in that, In step 4, the training parameters are set with a block size of 32, a step size of 0.5-1, a batch size of 64, 128, or 256, and the number of times the image segmentation algorithm returns is 100. The loss function Orsdiceloss and the optimization algorithm Adadelta are selected to train the labeled voxels. Training is stopped when the number of iterations is 40-80. 20%-30% of the training data is used as validation data.

4. The method for three-dimensional visualization reconstruction of CT images containing high-density component composite materials based on deep learning according to claim 1, characterized in that, The quantitative analysis described in step 5 uses connectivity analysis to create six-connected multiple regions of interest for the pores and selects the volume as the analysis object to obtain the proportion of pores within different volume ranges and the three-dimensional reconstructed distribution map.

Images