CT image artifact removing method and system for pancreatic cancer focus recognition
By extracting the foreground region from CT images of pancreatic cancer lesions and constructing a feature importance map, combined with U-Net and GAN models, the problem of artifact interference in CT images was solved, achieving high-precision segmentation of intratumoral and peritumoral tissues, thus improving diagnostic accuracy and visual effects.
Patent Information
- Application Number
- CN202511383917.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-09-26
- Publication Date
- 2026-02-06
AI Technical Summary
In existing technologies, CT images are difficult to effectively remove artifacts when segmenting pancreatic cancer lesions, affecting the segmentation accuracy of intratumoral and peritumoral tissues.
By extracting the foreground region from pancreatic CT images, a feature importance map is constructed. An artifact removal model is built using U-Net and Generative Adversarial Network (GAN). Artifact removal is performed based on the feature importance map. The model is then optimized by combining the feature importance map and the loss function to achieve accurate artifact removal.
It improves the segmentation accuracy of intratumoral and peritumoral tissues, ensures that the artifact-free CT images conform to the doctor's reading habits, avoids visual interference, and improves the accuracy of diagnosis.
Smart Images
Figure CN121482216A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of medical image processing technology, specifically to a method and system for artifact removal from CT images used for pancreatic cancer lesion identification. Background Technology
[0002] Pancreatic ductal adenocarcinoma (PDAC) is one of the most destructive malignancies, and computed tomography (CT) is the preferred imaging modality for its diagnosis and evaluation. In recent years, radiomics, as a technique for analyzing imaging features of regions of interest, has demonstrated excellent performance in predicting tumor biological complexity and intratumoral heterogeneity. Through high-throughput analysis of intratumoral CT images, radiomics has also been used to predict the biological behavior of PDAC tumors. While most research focuses on the primary tumor, recent studies have highlighted the important auxiliary role of the peritumor region in assessing cancer heterogeneity and its importance in prognostic prediction. Therefore, current radiomic analyses of pancreatic ductal adenocarcinoma often combine the analysis of the tumor region and the peritumor region.
[0003] Existing technologies typically involve directly segmenting the intratumoral and peritumoral tissues on CT images to obtain the tumor region and the surrounding area. This segmentation method is difficult to eliminate artifact interference, which affects the segmentation accuracy of the intratumoral and peritumoral tissues. Summary of the Invention
[0004] The purpose of this invention is to provide a method and system for removing artifacts from CT images for pancreatic cancer lesion identification, in order to solve the technical problem in the prior art that it is difficult to remove the interference of artifacts, which affects the segmentation accuracy of intratumoral and peritumoral tissues.
[0005] To solve the above-mentioned technical problems, the present invention specifically provides the following technical solution: A method for artifact removal from CT images for pancreatic cancer lesion identification includes the following steps: Obtain CT images of the pancreas; Extract the foreground region from pancreatic CT images and create a feature importance map for the foreground region; Using a pre-established artifact removal model, the foreground region is processed based on the feature importance map to obtain an artifact-free foreground region; The artifact-free foreground region is fused with the pancreatic CT image to obtain an artifact-free pancreatic CT image.
[0006] As a preferred embodiment of the present invention, the method for extracting the foreground region includes: For each pixel (x,y) in a pancreatic CT image, define a local window of size N*N centered on it as its neighborhood space N(x,y); Calculate the gradient magnitude of all pixels in the neighborhood space N(x,y), and use it as the local gradient magnitude of pixel (x,y). ,in, , These are the gradient components of the neighborhood space N(x,y) calculated by the Sobel operator in the x and y directions, respectively. Calculate the variance of the intensity of all pixels in the neighborhood space N(x,y), and use it as the local intensity variance of pixel (x,y). , ,in, Let N(x,y) be the mean intensity of all pixels in the neighborhood space N(x,y). For pixels in the neighborhood space N(x,y) The pixel intensity at a given location, where N is the total number of pixels in each single dimension of the neighborhood space N(x,y) in the x,y dimension; Local intensity variance and local gradient magnitude Normalization and weighted combination are performed to determine the feature importance of pixel (x,y). ,in, After normalization , After normalization , for The weight, for The weights; Feature importance The feature importance is compared with a preset threshold, and features exceeding the preset threshold are considered more important. The corresponding pixel (x,y) is labeled as the foreground region K(x,y).
[0007] As a preferred embodiment of the present invention, the feature importance of each pixel in the foreground region K(x,y) is obtained. The feature importance map of the foreground region is then normalized and used as a feature importance map. .
[0008] As a preferred embodiment of the present invention, the method for constructing the artifact removal model includes: Multiple pancreatic CT images were selected, and the foreground region was extracted from the pancreatic CT images. The feature importance map of the foreground region was calculated, and a pre-established U-Net network was used to segment intratumoral and peritumoral tissues in the foreground region to obtain intratumoral and peritumoral segmentation images, which were used as ground-value images for intratumoral segmentation. and peritumor segmentation ground truth image ; The foreground region and a weighted combination of the feature importance map are used as input to the neural network to remove artifacts from the foreground region. This is the output term of the neural network; The foreground region output by the neural network is used as the input to the U-Net network, which then outputs the intratumoral tissue segmentation result. Results of peritumoral tissue segmentation ; Based on the results of intratumoral tissue segmentation Results of peritumoral tissue segmentation Compared with the true value image of intratumoral segmentation and peritumor segmentation ground truth image The differences between them, and the foreground region for artifact removal With foreground area The differences between them are combined into a loss function. The neural network is trained to obtain the artifact removal model. ; in, This refers to the cross-entropy operator; GAN stands for Gyrelian Network. , is a hyperparameter, and T is the training period of the GAN neural network.
[0009] As a preferred embodiment of the present invention, the method for obtaining the artifact-free foreground region includes: The foreground regions are weighted and combined using a feature importance map to obtain a weighted foreground region; The weighted foreground region is input into the artifact removal model, and the artifact removal model outputs the artifact-removed foreground region.
[0010] As a preferred embodiment of the present invention, the U-Net network includes a U-Net network for intratumoral tissue segmentation. and the U-Net network for peritumoral tissue segmentation ; in, Outputting the foreground region of the U-Net network for intratumoral tissue segmentation The results of intratumoral tissue segmentation, Outputting the foreground region of the U-Net network for intratumoral tissue segmentation The results of intratumoral tissue segmentation, Foreground region, UNet is a U-Net network.
[0011] As a preferred embodiment of the present invention, the artifact-free foreground region is replaced with the corresponding region in the pancreatic CT image, and the junction is smoothed to obtain an artifact-free pancreatic CT image.
[0012] As a preferred embodiment of the present invention, all pancreatic CT images are of uniform size.
[0013] As a preferred embodiment of the present invention, the present invention provides a CT image artifact removal system for pancreatic cancer lesion identification, applied to the aforementioned CT image artifact removal method for pancreatic cancer lesion identification, the system comprising: The data acquisition unit is used to acquire CT images of the pancreas; The feature processing unit is used to extract the foreground region in pancreatic CT images and to create a feature importance map for the foreground region; The artifact removal unit is used to perform artifact removal processing on the foreground region based on the feature importance map using a pre-established artifact removal model to obtain an artifact-removed foreground region, and then fuse the artifact-removed foreground region with the pancreatic CT image to obtain an artifact-removed pancreatic CT image.
[0014] As a preferred embodiment of the present invention, the artifact removal model ; in, To remove artifacts in the foreground area, Foreground area GAN stands for neural network, referring to the importance of features.
[0015] Compared with the prior art, the present invention has the following advantages: This invention focuses on the precise and adaptive processing of the most critical diagnostic region (foreground region). It uses a feature importance map as a guide and a generative model (GAN) with downstream segmentation task performance as the ultimate optimization goal to remove artifacts, ultimately achieving artifact removal effects that serve diagnostic purposes. Attached Figure Description
[0016] To more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings in the following description are merely exemplary, and those skilled in the art can derive other embodiments based on the provided drawings without creative effort.
[0017] Figure 1 This is a flowchart of a CT image artifact removal method for pancreatic cancer lesion identification provided in an embodiment of the present invention; Figure 2 This is a block diagram of a CT image artifact removal system for pancreatic cancer lesion identification provided in an embodiment of the present invention. Detailed Implementation
[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0019] like Figure 1 As shown, this invention provides a CT image artifact removal method for pancreatic cancer lesion identification, comprising the following steps: Obtain CT images of the pancreas; Extract the foreground region from pancreatic CT images and create a feature importance map for the foreground region; Using a pre-established artifact removal model, the foreground region is processed based on the feature importance map to obtain an artifact-free foreground region; The artifact-free foreground region is fused with the pancreatic CT image to obtain an artifact-free pancreatic CT image.
[0020] In removing artifacts, this invention focuses on removing the effective pixel area, or the peritumoral tissue area and intratumoral tissue area, concentrating on the important areas for removal, thus avoiding ineffective removal of the background area and the waste of resources and efficiency.
[0021] Methods for extracting foreground regions include: For each pixel (x,y) in a pancreatic CT image, define a local window of size N*N centered on it as its neighborhood space N(x,y); Calculate the gradient magnitude of all pixels in the neighborhood space N(x,y), and use it as the local gradient magnitude of pixel (x,y). ,in, , These are the gradient components of the neighborhood space N(x,y) calculated by the Sobel operator in the x and y directions, respectively. Calculate the variance of the intensity of all pixels in the neighborhood space N(x,y), and use it as the local intensity variance of pixel (x,y). , ,in, Let N(x,y) be the mean intensity of all pixels in the neighborhood space N(x,y). For pixels in the neighborhood space N(x,y) The pixel intensity at a given location, where N is the total number of pixels in each single dimension of the neighborhood space N(x,y) in the x,y dimension; Local intensity variance and local gradient magnitude Normalization and weighted combination are performed to determine the feature importance of pixel (x,y). ,in, After normalization , After normalization , for The weight, for The weights; Feature importance The feature importance is compared with a preset threshold, and features exceeding the preset threshold are considered more important. The corresponding pixel (x,y) is labeled as the foreground region K(x,y), which includes the area of intratumoral tissue and peritumoral tissue.
[0022] The feature importance of each pixel in the foreground region K(x,y) will be obtained. The feature importance map of the foreground region is then normalized and used as a feature importance map. .
[0023] This invention utilizes gradient magnitude to assess sensitivity to edges and structures (such as tumor boundaries and blood vessels), and intensity variance to assess sensitivity to texture complexity (such as heterogeneity within the tumor and peritumoral invasive features). These are precisely the visual features that radiologists and segmentation networks are most interested in, enabling the automatic and quantitative identification of regions in an image rich in diagnostic information.
[0024] This invention effectively excludes uniform, information-free background areas (such as uniform fat or blank areas) from subsequent calculations through threshold processing, significantly reducing unnecessary computation and avoiding meaningless processing of these areas that may introduce noise.
[0025] Feature importance map generated by this invention This method quantifies the importance of pixel features in an image, providing pixel-level guidance for subsequent artifact removal models. It can provide the model with prior knowledge of the importance of pixel features, allowing the model to refine the processing of highly important regions based on the importance of pixel features.
[0026] Methods for constructing artifact removal models include: Multiple pancreatic CT images were selected, and the foreground region was extracted from the pancreatic CT images. The feature importance map of the foreground region was calculated, and a pre-established U-Net network was used to segment intratumoral and peritumoral tissues in the foreground region to obtain intratumoral and peritumoral segmentation images, which were used as ground-value images for intratumoral segmentation. and peritumor segmentation ground truth image ; The foreground region and a weighted combination of feature importance maps are used as inputs to the neural network to remove artifacts from the foreground region. This is the output term of the neural network; The foreground region output by the neural network is used as the input to the U-Net network, which then outputs the intratumoral tissue segmentation result. Results of peritumoral tissue segmentation ; Based on the results of intratumoral tissue segmentation Results of peritumoral tissue segmentation Compared with the true value image of intratumoral segmentation and peritumor segmentation ground truth image The differences between them, and the foreground region for artifact removal With foreground area The differences between them are combined into a loss function. Train the neural network to obtain the artifact removal model. ; in, This refers to the cross-entropy operator; GAN stands for Gyrelian Network. , is a hyperparameter, and T is the training period of the GAN neural network.
[0027] In this invention, the effect of artifact removal is used to indirectly quantify the effectiveness of artifact removal, which is the loss function. and These two aspects ensure that the artifact removal model outputs high-precision segmentation performance for intratumoral and peritumoral tissues, indirectly guaranteeing the optimal artifact removal effect of the model and improving the artifact removal effect at the tissue segmentation level.
[0028] The present invention also includes a loss term. To maintain maximum consistency between the foreground region before and after artifact removal, prevent excessive random generation by the generator, ensure that the artifact-removed image does not deviate excessively from the original anatomical structure, and constrain the artifact removal results at the tissue segmentation level.
[0029] in, During training, α exhibits a dynamic change from small to large. In the early stages of training, α is relatively small, resulting in weak constraints on image fidelity. This allows the generator a larger exploration space to generate images that optimize segmentation results, enhancing the diversity of solutions and making it easier to escape local optima. In the later stages of training, α increases, strengthening the constraints on image realism, stabilizing the optimal solutions from the early stages, ensuring their anatomical rationality, and enabling the model to converge well.
[0030] This invention employs a Generative Adversarial Network (GAN) architecture to construct the artifact removal model. This architecture is well-suited for learning the complex, non-linear mapping from artifact presence to artifact removal, and can effectively handle artifacts with various complex structures. Other neural networks with similar functionalities can also be used as needed.
[0031] Methods for obtaining the foreground region for artifact removal include: The foreground regions are weighted and combined using a feature importance map to obtain a weighted foreground region; The weighted foreground region is input into the artifact removal model, and the artifact removal model outputs the artifact-removed foreground region.
[0032] This invention utilizes a feature importance map to weight the foreground region before inputting it into the artifact removal model. This explicitly guides the model to focus on high-importance regions during the input stage. This significantly simplifies the model's learning process, eliminating the need to learn from scratch which regions are important. Instead, it can concentrate all its modeling capabilities on learning how to perform optimal artifact removal transformations on these key regions.
[0033] The U-Net network includes a U-Net network for intratumoral tissue segmentation. and the U-Net network for peritumoral tissue segmentation ; in, Outputting the foreground region of the U-Net network for intratumoral tissue segmentation The results of intratumoral tissue segmentation, Outputting the foreground region of the U-Net network for intratumoral tissue segmentation The results of intratumoral tissue segmentation, Foreground region, UNet is a U-Net network.
[0034] Because intratumoral and peritumoral tissues differ significantly in texture, boundaries, and physiological characteristics, this invention uses two independent U-Nets to process intratumoral and peritumoral tissues respectively. This allows each network depth to focus on learning subtle features of a specific category, resulting in higher accuracy than using a single general network to segment both simultaneously. Furthermore, these two high-performance segmentation networks provide extremely reliable and accurate gradient feedback (through segmentation loss) for training the artifact removal model. For example, if the segmentation network itself is poorly performing, its feedback gradients will be noisy and unable to effectively guide the optimization of the artifact removal model.
[0035] The artifact-free foreground region is replaced with the corresponding region in the pancreatic CT image, and the junction is smoothed to obtain the artifact-free pancreatic CT image.
[0036] This invention replaces the foreground area after artifact removal with the original image and smooths the junction, so that the final output is a complete and globally consistent CT image, which conforms to the doctor's reading habits. It can also avoid sharp breaks at the boundary of the replacement area, so that the artifact removal area blends naturally with the surrounding unprocessed background, making it more visually natural and not introducing new visual interference.
[0037] All pancreatic CT images are of a uniform size.
[0038] like Figure 2 As shown, this invention provides a CT image artifact removal system for pancreatic cancer lesion identification, applied to a CT image artifact removal method for pancreatic cancer lesion identification. The system includes: The data acquisition unit is used to acquire CT images of the pancreas; The feature processing unit is used to extract the foreground region in pancreatic CT images and to create a feature importance map for the foreground region; The artifact removal unit is used to perform artifact removal processing on the foreground region based on the feature importance map using a pre-established artifact removal model to obtain an artifact-removed foreground region, and then fuse the artifact-removed foreground region with the pancreatic CT image to obtain an artifact-removed pancreatic CT image.
[0039] Artifact Removal Model ; in, To remove artifacts in the foreground area, Foreground area GAN stands for neural network, referring to the importance of features.
[0040] This invention focuses on the precise and adaptive processing of the most critical diagnostic region (foreground region). It uses a feature importance map as a guide and a generative model (GAN) with downstream segmentation task performance as the ultimate optimization goal to remove artifacts, ultimately achieving artifact removal effects that serve diagnostic purposes.
[0041] The above embodiments are merely exemplary embodiments of this application and are not intended to limit this application. The scope of protection of this application is defined by the claims. Those skilled in the art can make various modifications or equivalent substitutions to this application within its substance and scope of protection, and such modifications or equivalent substitutions should also be considered to fall within the scope of protection of this application.
Claims
1. A method for artifact removal from CT images for pancreatic cancer lesion identification, characterized in that, Includes the following steps: Obtain CT images of the pancreas; Extract the foreground region from pancreatic CT images and create a feature importance map for the foreground region; Using a pre-established artifact removal model, the foreground region is processed based on the feature importance map to obtain an artifact-free foreground region; The artifact-free foreground region is fused with the pancreatic CT image to obtain an artifact-free pancreatic CT image.
2. The CT image artifact removal method for pancreatic cancer lesion identification according to claim 1, characterized in that: The method for extracting the foreground region includes: For each pixel (x,y) in a pancreatic CT image, define a local window of size N*N centered on it as its neighborhood space N(x,y); Calculate the gradient magnitude of all pixels in the neighborhood space N(x,y), and use it as the local gradient magnitude of pixel (x,y). ,in, , These are the gradient components of the neighborhood space N(x,y) calculated by the Sobel operator in the x and y directions, respectively. Calculate the variance of the intensity of all pixels in the neighborhood space N(x,y), and use it as the local intensity variance of pixel (x,y). , ,in, Let N(x,y) be the mean intensity of all pixels in the neighborhood space N(x,y). For pixels in the neighborhood space N(x,y) The pixel intensity at a given location, where N is the total number of pixels in each single dimension of the neighborhood space N(x,y) in the x,y dimension; Local intensity variance and local gradient magnitude Normalization and weighted combination are performed to determine the feature importance of pixel (x,y). ,in, After normalization , After normalization , for The weight, for The weights; Feature importance The feature importance is compared with a preset threshold, and features exceeding the preset threshold are considered more important. The corresponding pixel (x,y) is labeled as the foreground region K(x,y).
3. The CT image artifact removal method for pancreatic cancer lesion identification according to claim 2, characterized in that: The feature importance of each pixel in the foreground region K(x,y) will be obtained. The feature importance map of the foreground region is then normalized and used as a feature importance map. .
4. The CT image artifact removal method for pancreatic cancer lesion identification according to claim 3, characterized in that: The method for constructing the artifact removal model includes: Multiple pancreatic CT images were selected, and the foreground region was extracted from the pancreatic CT images. The feature importance map of the foreground region was calculated, and a pre-established U-Net network was used to segment intratumoral and peritumoral tissues in the foreground region to obtain intratumoral and peritumoral segmentation images, which were used as ground-value images for intratumoral segmentation. and peritumor segmentation ground truth image ; The foreground region and a weighted combination of the feature importance map are used as input to the neural network to remove artifacts from the foreground region. This is the output term of the neural network; The foreground region output by the neural network is used as the input to the U-Net network, which then outputs the intratumoral tissue segmentation result. Results of peritumoral tissue segmentation ; Based on the results of intratumoral tissue segmentation Results of peritumoral tissue segmentation Compared with the true value image of intratumoral segmentation and peritumor segmentation ground truth image The differences between them, and the foreground region for artifact removal With foreground area The differences between them are combined into a loss function. The neural network is trained to obtain the artifact removal model. ; in, This refers to the cross-entropy operator; GAN stands for Gyrelian Network. , is a hyperparameter, and T is the training period of the GAN neural network.
5. The CT image artifact removal method for pancreatic cancer lesion identification according to claim 4, characterized in that: The method for obtaining the artifact-removed foreground region includes: The foreground regions are weighted and combined using a feature importance map to obtain a weighted foreground region; The weighted foreground region is input into the artifact removal model, and the artifact removal model outputs the artifact-removed foreground region.
6. The CT image artifact removal method for pancreatic cancer lesion identification according to claim 5, characterized in that: The U-Net network includes a U-Net network for intratumoral tissue segmentation. and the U-Net network for peritumoral tissue segmentation ; in, Outputting the foreground region of the U-Net network for intratumoral tissue segmentation The results of intratumoral tissue segmentation, Outputting the foreground region of the U-Net network for intratumoral tissue segmentation The results of intratumoral tissue segmentation, Foreground region, UNet is a U-Net network.
7. The CT image artifact removal method for pancreatic cancer lesion identification according to claim 6, characterized in that: The artifact-free foreground region is replaced with the corresponding region in the pancreatic CT image, and the junction is smoothed to obtain the artifact-free pancreatic CT image.
8. The CT image artifact removal method for pancreatic cancer lesion identification according to claim 1, characterized in that: All pancreatic CT images were of a uniform size.
9. A CT image artifact removal system for pancreatic cancer lesion identification, characterized in that, A method for CT image artifact removal for pancreatic cancer lesion identification according to any one of claims 1-8, the system comprising: The data acquisition unit is used to acquire CT images of the pancreas; The feature processing unit is used to extract the foreground region in pancreatic CT images and to create a feature importance map for the foreground region; The artifact removal unit is used to perform artifact removal processing on the foreground region based on the feature importance map using a pre-established artifact removal model to obtain an artifact-removed foreground region, and then fuse the artifact-removed foreground region with the pancreatic CT image to obtain an artifact-removed pancreatic CT image.
10. A method for CT image artifact removal for pancreatic cancer lesion identification according to claim 9, characterized in that: The artifact removal model ; in, To remove artifacts in the foreground area, Foreground area GAN stands for neural network, referring to the importance of features.