Method for solving PSMA PET image dispersion problem based on CT image
By combining high-resolution anatomical information and deep learning methods of CT images, the diffusion effect problem of PSMA PET images is solved, precise positioning and high-precision segmentation of lesions are achieved, and image support for prostate cancer diagnosis and treatment is improved.
Patent Information
- Application Number
- CN202510377505.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-03-28
- Publication Date
- 2025-08-01
AI Technical Summary
The existing PSMA PET images are affected by radiation attenuation, noise and other factors in the diagnosis and treatment of prostate cancer, resulting in image diffusion effects, affecting the precise positioning and quantitative analysis of lesions. The existing segmentation methods cannot effectively solve the problem of image blur.
The organ and lesion profile are extracted by U-Net network based on CT images, combined with PSMA PET images for registration and fusion, and the high-resolution anatomical structure information of CT images are used for diffusion area compensation, and combined with deep learning and GAN network optimization segmentation results, diffusion artifacts are removed.
The segmentation accuracy of PSMA PET images was improved, the true boundaries of the lesion area were clarified, the segmentation accuracy and recall rate reached more than 95%, the lesion structure similarity index reached 91.5%, the adaptability and stability were high, and the cross-verification accuracy reached 86%.
Smart Images
Figure CN120411271A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of nuclear medicine imaging, and particularly to a method for solving the diffusion problem of PSMA PET images based on CT images. Background Art
[0002] PSMA (Prostate Specific Membrane Antigen) tracer PET (Positron Emission Tomography) technology has become an important means in the diagnosis and treatment effect evaluation of prostate cancer. However, in clinical applications, PET images are often affected by factors such as radiation attenuation, noise, and resolution, resulting in a certain diffusion effect in the images, which affects the precise localization and quantitative analysis of lesions. Especially in the early diagnosis and treatment of prostate cancer, the improvement of image quality is of crucial significance.
[0003] In PET / CT images, CT images can provide high-resolution anatomical structure information, including organ contours, tissue densities, anatomical regions, etc., while PET images reflect the metabolic activities and biological characteristics of tumors. Due to the diffusion effect of PET tracers ( 18 F-FDG and 68 Ga-PSMA), PET images usually appear blurred in the lesion area, resulting in inaccurate localization and segmentation of tumors. Most of the existing image segmentation methods focus on improving the segmentation accuracy of images through traditional algorithms (such as thresholding method, region growing method, etc.), but most methods cannot effectively solve the image blurring problem caused by the diffusion effect. Therefore, there is an urgent need for a new image segmentation method that can effectively reduce the diffusion effect in PSMA PET images and improve the segmentation accuracy of the lesion area. Summary of the Invention
[0004] Technical Problems to be Solved
[0005] In view of the above-mentioned disadvantages and deficiencies of the prior art, the present invention provides a method for solving the diffusion problem of PSMA PET images based on CT images, which solves the technical problem of how to reduce the diffusion effect in PSMA PET images and improve the segmentation accuracy of the lesion area.
[0006] Technical Solutions
[0007] To achieve the above object, the main technical solutions adopted by the present invention include:
[0008] In a first aspect, the present invention provides a method for solving the diffusion problem of PSMA PET images based on CT images, including:
[0009] Extracting the contours of organs and lesions in the CT image based on the U-Net network to obtain contour labels;
[0010] Fuse the PSMA PET image with the CT image after registration;
[0011] Determine the diffusion region in the PSMA PET image based on the contour label in the fused image;
[0012] Compensate for the diffusion region in the PSMA PET image;
[0013] Perform organ segmentation on the compensated PSMA PET image based on the U-Net network combined with the contour label to obtain the segmentation result;
[0014] Post-process the segmentation result to obtain the final segmentation result of the PSMA PET image.
[0015] Optionally, the U-Net network includes an encoder and a decoder, and multiple convolutional layers are included in the encoder and decoder of the U-Net network.
[0016] Optionally, the calculation formula of a single convolutional layer is as follows:
[0017]
[0018] Among them, is the j-th feature map of the l-th layer, is the i-th feature map of the (l - 1)-th layer, * represents the convolution operation, is the convolution kernel from the i-th feature map to the j-th feature map of the l-th layer, is the bias of the j-th feature map in the l-th layer, f(·) is the activation function, M j is the set of the upper-layer feature maps connected to the j-th feature map.
[0019] Optionally, max pooling is used for downsampling in the encoder, and the formula is as follows:
[0020]
[0021] Among them, downsample(·) is a 2×2 max pooling operation;
[0022] Transposed convolution is used to upsample the feature map in the decoder, and its calculation formula is as follows:
[0023]
[0024] Optionally, the loss function of the U-Net network is the cross-entropy loss function. For a binary classification problem, its formula is as follows:
[0025]
[0026] Among them, N is the number of samples, and y i is the true label of the i-th sample, and
[0027]
[0028] is the predicted probability of the i-th sample. Optionally, compensating for the diffusion region in the PSMA PET image includes: Using the deconvolution method of the physical model, combining contour labels based on the deep learning model, and / or compensating for the diffusion region in the PSMA PET image based on the GAN network.
[0029] Optionally, the discriminator loss function L D of the GAN network is:
[0030] L D = -E x [logD(x)] - E z [log(1 - D(G(z)))]), where x is the true data distribution, z is the prior noise distribution, D(x) is the judgment probability of the discriminator for the true data x, D(G(z)) is the judgment probability of the discriminator for the generated data G(z), where G is the generator, and E x is the expectation of the true data distribution, and E z is the expectation of the generated data distribution;
[0031] The generator loss function L G of the GAN network is:
[0032] L G = -E z [logD(G(z))];
[0033] During the adversarial training process, the discriminator is updated as:
[0034] , where θ D is the parameter of the discriminator;
[0035] During the adversarial training process, the generator is updated as:
[0036] , where θ G is the parameter of the generator.
[0037] Optionally, post-processing the segmentation result includes:
[0038] Sharpening the edge of the segmentation result, removing artifacts, and smoothing the segmentation boundary.
[0039] Second aspect, the present invention provides a computer-readable storage medium, on which a computer program is stored, and when the program is executed, it implements the method for solving the diffusion problem of PSMA PET images based on CT images described in any one of the above first aspects.
[0040] Third aspect, the present invention provides a storage device, including a storage medium and a processor, the storage medium stores a computer program, and when the program is executed by the processor, it implements the method for solving the diffusion problem of PSMA PET images based on CT images described in any one of the above first aspects.
[0041] Beneficial effects
[0042] The beneficial effects of the present invention are as follows: The method for solving the diffusion problem of PSMA PET images based on CT images of the present invention uses the anatomical information of CT images to guide the diffusion segmentation of PET images, removes diffusion artifacts, and improves the segmentation accuracy. Compensating for and removing the diffusion effect requires careful correction of the blurred areas in the PET images, and the high-resolution anatomical structure information provided by CT images can provide an effective reference for removing the diffusion effect. Specifically, by extracting organ contour labels on the CT images and registering and fusing them with the PET images, it can help distinguish the true organ contours in the PET images from the artifacts caused by the diffusion effect. Description of the drawings
[0043] Figure 1 It is a schematic diagram of the fused display of PET / CT images before diffusion correction provided by an embodiment of the present invention;
[0044] Figure 2 It is a schematic diagram of the fused display of PET / CT images after diffusion correction provided by an embodiment of the present invention;
[0045] Figure 3 It is a schematic diagram of the organ segmentation effect of CT images provided by an embodiment of the present invention;
[0046] Figure 4 It is a schematic diagram of the display of the PET image before diffusion correction and the true label of the primary prostate lesion provided by an embodiment of the present invention;
[0047] Figure 5 It is a schematic diagram of the display of the PET image after diffusion correction and the true label of the primary prostate lesion provided by an embodiment of the present invention. Detailed implementation manners
[0048] For a better explanation and understanding of the present invention, the following will describe the present invention in detail with reference to the accompanying drawings and through specific embodiments. Although exemplary embodiments of the present invention are shown in the drawings, it should be understood that the present invention can be implemented in various forms and should not be limited by the embodiments set forth herein. On the contrary, these embodiments are provided to enable a clearer and more thorough understanding of the present invention and to fully convey the scope of the present invention to those skilled in the art.
[0049] In a first aspect, this embodiment provides a method for solving the diffusion problem of PSMA PET images based on CT images, including:
[0050] S1, extracting the contours of organs and lesions in the CT image based on the U-Net network to obtain contour labels.
[0051] First, using the high-resolution anatomical structure information of the CT image, extract the contours of the target whole body organs and tumor lesions. This step adopts an automatic segmentation algorithm, a U-Net network segmentation method based on deep learning, to extract the contour labels of organs and lesions. By this method, accurate organ contours can be obtained as a reference in PET image analysis. These organ contour labels can not only clarify the positions of each anatomical structure in the PET image but also serve as a "guidance" in the subsequent diffusion effect compensation process to help distinguish which regions are affected by the diffusion effect and which regions are actual high-metabolism foci.
[0052] U-Net is a successful medical image segmentation network, and its core is an encoder-decoder structure. The following is its main formulaic expression:
[0053] First, in the encoder and decoder of U-Net, convolutional layers are widely used to extract image features. For a convolutional layer, its calculation formula is as follows:
[0054]
[0055] Among them, is the j-th feature map of the l-th layer, is the i-th feature map of the (l - 1)-th layer, * represents the convolution operation, is the convolution kernel from the i-th feature map to the j-th feature map in the l-th layer, is the bias of the j-th feature map in the l-th layer, f(·) is the activation function, M j is the set of feature maps of the upper layer connected to the j-th feature map.
[0056] Secondly, for the downsampling layer, max pooling is usually used for downsampling in the encoder, and the formula is as follows:
[0057]
[0058] Among them, downsample(·) is usually a 2×2 max pooling operation, that is, taking the maximum value within a 2×2 window with a stride of 2. It can reduce the size of the feature map while increasing the receptive field, which helps to extract more advanced features.
[0059] The upsampling layer in the decoder needs to upsample the feature map to restore it to a size close to the input image. The commonly used upsampling method is transposed convolution (also called deconvolution), and its calculation formula is as follows:
[0060]
[0061] The operation here is similar to that of the convolutional layer, but the setting of the convolutional kernel enables it to achieve the effect of upsampling, that is, increasing the size of the feature map. Usually, the stride of the transposed convolution will be set to 2 or other appropriate values to gradually restore the feature map to the size of the original image.
[0062] When training the U-Net, the commonly used loss function is the cross-entropy loss function. For binary classification problems, its formula is as follows:
[0063]
[0064] Among them, N is the number of samples, y i is the true label of the i-th sample, is the predicted probability of the i-th sample. When performing multi-classification tasks, the cross-entropy loss function will have a more complex form, such as the Softmax cross-entropy loss function, etc. In practical applications, the specific implementation of the U-Net may be adjusted and optimized according to different tasks and data, such as adding a regularization term to the loss function, etc.
[0065] S2, fuse the PSMA PET image and the CT image after registration.
[0066] After obtaining the organ contour labels of the CT image, the next step is to accurately register the PET image with the CT image. The purpose of registration is to ensure that the anatomical information of the CT image and the metabolic information of the PET image can be completely aligned in space. The registration method includes the mutual information (MI) method based on images.
[0067] S3, determine the diffusion region in the PSMA PET image in the fused image based on the contour labels.
[0068] After registration is completed, the PET image is fused with the CT image to ensure that the lesion area in the PET image can be accurately mapped to the anatomical structure of the CT image. At this time, the lesion area in the PET image is docked with the organ contour label of the CT image, which can effectively identify and locate the lesion, and further identify the diffusion blur effect area in the PET image.
[0069] S4. Compensate for the diffusion area in the PSMA PET image.
[0070] During the registration and fusion process of the PET image, the organ contour label provided by the CT image is used to guide the compensation of the diffusion effect in the PET image. The diffusion effect usually manifests as blurred image edges, especially in the edge area of tumor lesions. Through the contour label of the CT image, these areas can be accurately located, and the following steps are used to compensate for the diffusion effect: 1. Local area identification and correction: Through the anatomical area determined by the CT label, the blurred part of the corresponding area in the PET image is identified. For these areas, the deconvolution method of the physical model is used to correct the image to remove the blur effect caused by scattering and attenuation. At the same time, deep learning methods (such as convolutional neural networks) can be combined to further restore the details of these areas and reduce the interference of artifacts. 2. Segmentation correction based on contour labels: Use the organ contour label in the CT image to guide the segmentation process in the PET image. Through the label information, the diffusion effect mainly affects the edges of the lesion area, rather than the whole lesion. Therefore, combined with the label information, the deep learning model can better identify and remove the artifacts in these edge areas and improve the segmentation accuracy. 3. Adaptive optimization of the deep learning GAN network: For different patients and image qualities, the diffusion effect compensation process needs to use the deep learning GAN neural network for adaptive adjustment. The present invention adopts the deep learning GAN neural network adaptive filtering method and dynamic adjustment strategy to strengthen the compensation in the areas with strong diffusion effect, while maintaining the original image details in the areas with weak influence.
[0071] The GAN network formula is as follows:
[0072] (1) Discriminator loss function:
[0073] L D =-E x [logD(x)]-E z [log(1-D(G(z)))]
[0074] Here, the coefficient 1 / 2 in front is omitted because in the optimization process, the coefficient does not affect the direction of minimizing the loss function and can be absorbed by parameters such as the learning rate in actual calculations. At the same time, for the sake of concise representation, the subscripts of the expectation symbols sampled from different distributions are omitted, and it is defaulted that x comes from the real data distribution and z comes from the prior noise distribution.
[0075] (2) Generator loss function:
[0076] L G = -E z [logD(G(z))]
[0077] The coefficient 1 / 2 is also omitted.
[0078] (3) Adversarial training process, discriminator update:
[0079]
[0080] (4) Adversarial training process, generator update:
[0081]
[0082] Among them, α is usually not directly used to represent a specific quantity in the GAN training process, but in some cases, α may represent the learning rate or other hyperparameters related to the training process.
[0083] The above are the basic formulas and training processes of the GAN network. In order to improve the training stability and generation effect, some improvements and adjustments will be made to the loss function, such as using additional regularization terms.
[0084] S5. Based on the U-Net network and combined with the contour labels, perform organ segmentation on the compensated PSMA PET image to obtain the segmentation result.
[0085] After the above-mentioned diffusion effect compensation, the quality of the PET image is significantly improved, and the boundaries of the tumor lesions are clearer. Next, use the deep learning segmentation network U-Net for accurate organ segmentation. The deep learning formula is introduced above. This process uses the aforementioned organ contour labels as prior knowledge to optimize the output of the segmentation network and reduce segmentation errors caused by image blurring. Specifically, during the segmentation process, combined with the organ label information of the CT image, the deep learning model can clearly distinguish the tumor lesion area from other background tissue areas, thereby effectively removing artifacts. Using the fused multi-modal image, the segmentation model can perform tumor localization and segmentation with higher accuracy, especially in low-contrast areas, where the accuracy improvement is more obvious.
[0086] S6. Post-process the segmentation result to obtain the final segmentation result of the PSMA PET image.
[0087] Finally, morphological optimization is performed on the segmentation results, including operations such as edge sharpening, removing small artifacts, and smoothing the segmentation boundaries, to further improve the accuracy and robustness of the segmentation results. For some small artifacts that may still exist, the post-processing algorithm can ensure the clarity and precision of the final segmentation results by repairing and removing them.
[0088] A method for solving the diffusion problem of PSMA PET images based on CT images provided in this embodiment. The high-resolution anatomical information provided by CT images, especially the contour labels of organs, can help accurately locate and distinguish real lesions and diffusion artifacts in PET images. In this way, the blurred areas in PET images can be effectively compensated by precise anatomical guidance. The precision and recall rates of organ boundaries can both reach over 95%.
[0089] By combining CT contour labels, the present invention can dynamically adjust the intensity of diffusion effect compensation in different regions, avoid over-correction or incomplete correction, and improve the accuracy of lesion segmentation. The structural similarity index (SSIM) at the lesion site can reach 91.5%.
[0090] This method can process PET / CT images of different qualities and types through an adaptive compensation strategy and a deep learning model, has high adaptability and stability, and has a high degree of automation in the processing process, reducing manual intervention. The cross-validation accuracy rate can reach 86%.
[0091] The present invention combines the organ contour labels of CT images to assist in compensating for the diffusion effect and removing artifacts of PET images. This mechanism can not only improve the quality of PET images, but also accurately locate tumor lesions through precise segmentation algorithms, especially in regions with low contrast and high noise, with a significant improvement in accuracy. Through the deep fusion of multi-modal information, the present invention can provide strong imaging support for the early diagnosis and treatment of tumors such as prostate cancer.
[0092] Figure 1 For the fused display of PET / CT images before diffusion correction, Figure 2 For the fused display of PET / CT images after diffusion correction, Figure 3 For the AI organ segmentation effect of CT images, Figure 4 For the display of the PET image before diffusion correction and the true label of the primary prostate lesion, Figure 5 For the display of the PET image after diffusion correction and the true label of the primary prostate lesion.
[0093] In a second aspect, an embodiment of the present invention provides a computer-readable storage medium, on which a computer program is stored, and when the program is executed, it implements the method for solving the diffusion problem of PSMA PET images based on CT images according to any one of the above first aspects.
[0094] In a third aspect, an embodiment of the present invention provides a storage device, including a storage medium and a processor, where the storage medium stores a computer program, and when the program is executed by the processor, it implements the method for solving the diffusion problem of PSMA PET images based on CT images according to any one of the above first aspects.
[0095] Those skilled in the art should understand that the embodiments of the present invention can be provided as a method, a system, or a computer program product. Therefore, the present invention can adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware aspects. Moreover, the present invention can adopt the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.
[0096] Obviously, those skilled in the art can make various modifications and variations to the present invention without departing from the spirit and scope of the present invention. Thus, if these modifications and variations of the present invention fall within the scope of the claims of the present invention and their equivalent technologies, the present invention should also include these modifications and variations.
[0097] Although the embodiments of the present invention have been shown and described above, it can be understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those of ordinary skill in the art can make changes, modifications, substitutions, and variations to the above embodiments within the scope of the present invention.
Claims
1. A method for solving the diffusion problem of PSMA PET images based on CT images, characterized in that, Including: Extract the contours of organs and lesions in the CT image based on the U-Net network to obtain contour labels; Register and fuse the PSMA PET image with the CT image; Determine the diffusion region in the PSMA PET image in the fused image based on the contour labels; Compensate the diffusion region in the PSMA PET image; Perform organ segmentation on the compensated PSMA PET image based on the U-Net network combined with the contour labels to obtain a segmentation result; Post-process the segmentation result to obtain the final segmentation result of the PSMA PET image.
2. The method for solving the diffusion problem of PSMA PET images based on CT images according to claim 1, characterized in that: The U-Net network includes an encoder and a decoder, and multiple convolutional layers are included in the encoder and decoder of the U-Net network.
3. The method for solving the diffusion problem of PSMA PET images based on CT images according to claim 2, characterized in that: The calculation formula of a single convolutional layer is as follows: Among them, is the j-th feature map of the l-th layer, is the i-th feature map of the (l-1)-th layer, * represents the convolution operation, is the convolution kernel from the i-th feature map to the j-th feature map of the l-th layer, is the bias of the j-th feature map in the l-th layer, f(·) is the activation function, M j is the set of feature maps of the previous layer connected to the j-th feature map.
4. A method for solving the diffusion problem of PSMA PET images based on CT images according to claim 3, characterized in that Max-pooling is used for downsampling in the encoder, and the formula is as follows: where downsample(·) is a 2×2 max-pooling operation; Transposed convolution is used in the decoder to upsample the feature map, and its calculation formula is as follows:
5. A method for solving the diffusion problem of PSMA PET images based on CT images according to claim 4, the loss function of the U-Net network is a cross-entropy loss function, and for a binary classification problem, its formula is as follows: Among them, N is the number of samples, and y i is the true label of the i-th sample, and is the predicted probability of the i-th sample.
6. The method for solving the diffusion problem of PSMA PET images based on CT images according to claim 5, wherein Compensating the diffusion region in the PSMA PET image includes: Compensate the diffusion region in the PSMA PET image by using the deconvolution method of the physical model, based on the deep learning model combined with the contour labels, and / or based on the GAN network.
7. A method for solving the diffusion problem of PSMA PET images based on CT images according to claim 6, characterized in that, The discriminator loss function L of the GAN network D is as follows: L D = -E x [logD(x)] - E z [log(1 - D(G(z)))],where x is the real data distribution, z is the prior noise distribution, D(x) is the discrimination probability of the discriminator for the real data x, D(G(z)) is the discrimination probability of the discriminator for the generated data G(z), where G is the generator, and E x is the expectation of the real data distribution, and E z is the expectation of the generated data distribution; The generator loss function L of the GAN network G for: L G = -E z [logD(G(z))]; During the adversarial training process, the discriminator is updated as: Among them, θ D is the parameter of the discriminator; During the adversarial training process, the generator is updated as: Among them, θ G is a parameter of the generator.
8. A method for solving the diffusion problem of PSMA PET images based on CT images according to claim 7, characterized in that, Post-processing the segmentation result includes: Sharpen the edges, remove artifacts, and smooth the segmentation boundary of the segmentation result.
9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the method for solving the diffusion problem of PSMA PET images based on CT images according to any one of claims 1 to 8.
10. A storage device, comprising a storage medium and a processor, wherein the storage medium stores a computer program, characterized in that, When the processor executes the computer program, it implements the method for solving the diffusion problem of PSMA PET images based on CT images according to any one of claims 1 to 8.