Marrow active area iconography positioning method and device based on deep learning, computer equipment and storage medium

Abstract

The marrow active area iconography positioning method based on deep learning comprises the following steps: acquiring a CT image of a patient, and preprocessing to obtain a plurality of basic images; inputting the basic image into a pre-trained iconography positioning model, wherein the iconography positioning model comprises a cyclic generative adversarial network module, a Transform module, a conditional diffusion model module and a segmentation network module; generating a synthetic PET image based on the basic image by using a cyclic generative adversarial network module; performing global optimization and local optimization on the synthesized PET image by using a Transform module to obtain a first optimized image; performing noise addition and noise removal on the first optimized image in sequence by using a conditional diffusion model module to obtain a second optimized image; and identifying the second optimized image by using a segmentation network module, extracting a bone marrow active region from the second optimized image, and determining the position of the bone marrow active region. The invention provides a PET image which can generate high-quality PET images under the condition that only CT images exist.

Description

Technical Field

[0001] The present invention relates to the technical field of imaging localization of active bone marrow regions, and more specifically to an imaging localization method, device, computer equipment and storage medium for active bone marrow regions based on deep learning. Background Art

[0002] Tumor radiotherapy is an important means of treating many cancers. Radiotherapy can effectively control tumor growth, improve tumor resection rate, and increase the probability of organ preservation. It is widely used in the treatment of local regional tumors such as rectal cancer and cervical cancer. However, the side effects of radiotherapy cannot be ignored, especially radiotherapy of tumors located in the pelvic area. Since a large amount of active bone marrow (ABM) in adults is concentrated in the pelvic area, these areas will inevitably be irradiated during radiotherapy, resulting in bone marrow suppression and causing severe hematological toxicity. Bone marrow suppression not only weakens the patient's immune system, but may also lead to treatment interruption or dose adjustment, seriously affecting the efficacy. Therefore, how to accurately locate and protect the active bone marrow area has become a key issue in reducing radiotherapy toxicity and improving treatment efficacy.

[0003] At present, 18F-fluorodeoxy-glucose positron emission tomography (18F-fluorodeoxy-glucose PET, 18F-FDG PET) is used clinically as a standard imaging method for active bone marrow areas. PET images can effectively locate bone marrow areas with active hematopoietic function by reflecting the active degree of glucose metabolism in the body. Since PET images can directly display metabolically active areas, they have high accuracy in identifying active bone marrow areas (ABM). However, PET imaging equipment is expensive and takes a long time to image. In many medical institutions, PET imaging is not a routine examination item in the diagnosis and treatment of localized tumors. In addition, radiotherapy often requires multiple treatments, and frequent PET imaging not only increases costs, but also brings additional burdens to patients. Therefore, how to achieve accurate positioning of active bone marrow areas by relying on other conventional imaging technologies such as CT images in the absence of PET functional imaging has become an important research direction in the development of radiotherapy technology.

[0004] In many existing technologies, atlas-based registration techniques are used to retrieve the most similar CT image and transplant its corresponding PET image into the treatment plan. However, this method requires a large database of CT image and PET image components with accurate matching relationships, which is very difficult to implement in practice. Summary of the invention

[0005] In order to address the deficiencies in the prior art, the present invention provides a deep learning-based imaging localization method, device, computer equipment and storage medium for bone marrow active area, which can generate high-quality PET images with only CT images, significantly improving the flexibility of clinical applications.

[0006] In order to achieve the above object, the specific scheme adopted by the present invention is:

[0007] The deep learning-based imaging localization method for bone marrow active areas includes the following steps:

[0008] Acquire CT images of the patient and perform preprocessing to obtain multiple basic images;

[0009] Input the base image into a pre-trained imaging localization model, which includes a recurrent generative adversarial network module, a Transformer module, a conditional diffusion model module, and a segmentation network module;

[0010] Generate synthetic PET (Positron Emission Computed Tomography) images based on the base images using a recurrent generative adversarial network module;

[0011] Using the Transformer module to perform global and local optimization on the synthetic PET image to obtain a first optimized image;

[0012] Using the conditional diffusion model module, adding noise and removing noise from the first optimized image in sequence to obtain a second optimized image;

[0013] The second optimized image is identified by using a segmentation network module, the bone marrow active area is extracted from the second optimized image, and the position of the bone marrow active area is determined.

[0014] Preferably, the method for preprocessing the CT image includes:

[0015] Traverse each voxel in the CT image and obtain the HU value of each voxel;

[0016] Normalize the HU value less than the preset first threshold value to the first threshold value, and normalize the HU value greater than the preset second threshold value to the second threshold value;

[0017] normalizing the first threshold value and the second threshold value to a first scale;

[0018] The CT image is normalized by using the bilinear interpolation method and the nearest neighbor interpolation method to generate multiple basic images with a size of n×n.

[0019] Preferably, the specific method of generating a synthetic PET image based on a basic image using a cyclic generative adversarial network module includes:

[0020] Generate preliminary PET images based on the base images using the positive generator of the recurrent generative adversarial network module;

[0021] Generate verification images based on preliminary PET images using the inverse generator of the recurrent generative adversarial network module;

[0022] Using two discriminators of the cyclic generative adversarial network module to respectively judge a first difference between the base image and the verification image, and a second difference between the preliminary PET image and the standard PET image;

[0023] When the first differences and the second differences corresponding to all the basic images satisfy a preset condition, all the preliminary PET images are combined to obtain a synthetic PET image.

[0024] Preferably, the specific method of using the Transformer module to perform global optimization and local optimization on the synthetic PET image includes:

[0025] Input the base image and standard PET image into the Transformer module;

[0026] The convolutional neural network is used to extract features from the basic image to obtain a high-dimensional feature vector, and the convolutional neural network is used to extract features from the standard PET image to obtain local feature information of the active bone marrow area;

[0027] Use the Transformer module to generate global features based on the high-dimensional feature vectors of all base images;

[0028] Based on the cross-attention mechanism, the global feature and local feature information are mapped into the synthetic PET image to obtain the first optimized image.

[0029] Preferably, the specific method of sequentially adding noise and removing noise from the first optimized image using the conditional diffusion model module includes:

[0030] Using a conditional diffusion model, adding predetermined noise to the first optimized image multiple times to obtain a noisy image;

[0031] The conditional diffusion model is used to denoise the noisy image multiple times to obtain a second optimized image.

[0032] Preferably, the specific method of using the segmentation network module to identify the second optimized image includes:

[0033] Using the convolution submodule of the segmentation network module to perform feature recognition on the second optimized image to obtain multiple candidate regions, and the recognition process is constrained by using the residual submodule of the segmentation network module;

[0034] The segmentation loss of the candidate region is evaluated, and if the evaluation is passed, the candidate region is recorded as a bone marrow active region;

[0035] The coordinates of the bone marrow active region in the second optimized image are determined, and the position of the bone marrow active region is generated based on the coordinates.

[0036] Preferably, the method for training the imaging localization model comprises:

[0037] Acquire multiple whole-body PET-CT image pairs, and divide them into a training atlas and a validation atlas based on a preset division ratio;

[0038] Assign an identifier to each whole-body PET-CT image pair;

[0039] Data screening and normalization of whole-body PET-CT image pairs;

[0040] The imaging localization model is trained using the training atlas, and the trained imaging localization model is verified using the verification atlas. If the verification passes, the imaging localization model is retained.

[0041] Bone marrow active area imaging positioning device, including:

[0042] A data acquisition module is used to collect CT images of patients and perform preprocessing to obtain multiple basic images;

[0043] A model implementation module is used to process the basic image using a pre-trained imaging localization model to determine the bone marrow active area and the location of the bone marrow active area;

[0044] The data output module is used to output the location of the bone marrow active area.

[0045] Computer equipment, including:

[0046] a memory storing a computer program;

[0047] A processor is used to read and execute the computer program to implement the above-mentioned deep learning-based imaging localization method for bone marrow active areas.

[0048] A storage medium is used to store a computer program, which, when executed, implements the above-mentioned deep learning-based imaging localization method for active bone marrow areas.

[0049] The present invention can generate high-quality PET images when only CT images are available, significantly improving the flexibility of clinical applications; the present invention adopts a technical means of combining the Transformer module with the conditional diffusion model module, so that the generated synthetic PET images have high precision in global and local feature preservation, and can accurately preserve the metabolic characteristics of active bone marrow areas in complex anatomical areas such as the pelvis, greatly improving the accuracy of positioning and the reliability of segmentation; the present invention realizes fully automatic recognition and labeling of active bone marrow areas, greatly improving efficiency, and reducing errors that may be caused by manual intervention, making the personalized design of radiotherapy plans more accurate and efficient. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0051] Figure 1 It is a structural schematic diagram of the imaging positioning model in the present invention;

[0052] Figure 2 It is a schematic diagram of the structure of the cyclic generative adversarial network module;

[0053] Figure 3 It is a schematic diagram of the structure of the Transformer module;

[0054] Figure 4 It is a structural diagram of the conditional diffusion model module;

[0055] Figure 5 It is a schematic diagram of the structure of the segmentation network module;

[0056] Figure 6 This is the processing result diagram of the first experiment;

[0057] Figure 7 This is the PSNR and SSIM result graph of the first experiment;

[0058] Figure 8 This is the DICE result diagram of the second experiment. DETAILED DESCRIPTION

[0059] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0060] like Figure 1 As shown, the imaging localization method of bone marrow active areas based on deep learning includes S1 to S6.

[0061] S1. Collect CT images of a patient and perform preprocessing to obtain a plurality of basic images. The method for preprocessing CT images includes S11 to S14.

[0062] S11, traverse each voxel in the CT image and obtain the HU value of each voxel. The voxel and HU value of the CT image belong to the conventional technology in the field and will not be described here.

[0063] S12, normalizing HU values ​​less than a preset first threshold value to a first threshold value, and normalizing HU values ​​greater than a preset second threshold value to a second threshold value. In this embodiment, the first threshold value of the HU value is set to 1000, and the second threshold value is set to -1000. Accordingly, HU values ​​greater than 1000 are normalized to 1000, and HU values ​​less than -1000 are normalized to -1000. By normalizing the HU values ​​of voxels in the CT image, the influence of abnormal values ​​can be reduced, and it is easier to process in subsequent processes, which can reduce the overall complexity of the method.

[0064] S13, normalizing the first limit value and the second limit value to a first scale. On the basis of setting the first limit value to 1000 and the second limit value to -1000, the range of the first scale is set to [0, 1]. By performing normalization again, the complexity of subsequent processing can be further reduced.

[0065] S14, using a bilinear interpolation method and a nearest neighbor interpolation method to normalize the size of the CT image to generate a plurality of basic images of size n×n. In this embodiment, the size of the CT image is normalized to 256×256. Both the bilinear interpolation method and the nearest neighbor interpolation method are mature existing technologies and will not be described in detail here.

[0066] S2, inputting the basic image into a pre-trained imaging localization model, wherein the imaging localization model includes a cyclic generative adversarial network module, a Transformer module, a conditional diffusion model module and a segmentation network module. The method for training the imaging localization model is as follows.

[0067] First, multiple whole-body PET-CT image pairs are acquired according to factors such as the patient's age, gender, and scanning time, and are divided into training atlases and verification atlases based on a preset division ratio. The specific division ratio of the training atlas and the verification atlas can be determined according to the number of whole-body PET-CT image pairs acquired. In this embodiment, 2000 whole-body PET-CT image pairs are acquired and divided according to a ratio of 8:2, that is, 1600 whole-body PET-CT image pairs are used as training atlases, and 400 whole-body PET-CT image pairs are used as verification atlases.

[0068] Second, an identifier is assigned to each whole-body PET-CT image pair. The identifier can be used to determine the unique correspondence between the two images in the whole-body PET-CT image pair to ensure that in the subsequent training process, the imaging localization model can accurately determine the association between the PET image and the CT image.

[0069] Third, data screening and normalization are performed on the whole-body PET-CT image pairs. The purpose of data screening is to delete obvious abnormal data to prevent these abnormal data from interfering with the training process of the imaging positioning model, resulting in slow convergence or even failure to converge. For PET images, when data screening is performed, the 2.5% with the highest scanning intensity and the 2.5% with the lowest scanning intensity are deleted, and then the remaining scanning intensity is normalized to the second scale. In this embodiment, the range of the second scale is also set to [0, 1]. On the other hand, the size of the PET image is also normalized to 256×256. The normalization method for CT images is as described in S11 to S14 and will not be repeated here.

[0070] Fourth, the imaging localization model is trained using the training atlas, and the trained imaging localization model is verified using the verification atlas. If the verification passes, the imaging localization model is retained. During the training process, the maximum number of iterations needs to be determined in advance to prevent the training process from being terminated when the imaging localization model cannot converge.

[0071] S3, using the recurrent generative adversarial network module to generate a synthetic PET (Positron Emission Computed Tomography) image based on the basic image. Figure 2As shown in Figure 1, the structure of the recurrent generative adversarial network mainly includes two generators and two discriminators. One of the generators is used as a positive generator to convert the input CT image into a synthetic PET image; the other generator is used as an inverse generator to remap the generated synthetic PET image back to the CT image to ensure the consistency between the generated image and the original image. Both generators use convolutional neural networks, which are constructed through multiple convolutional layers, activation functions and residual networks (ResNet Block). The core structural formula of the generator is G(X CT )=Conv(ResNet(X CT )), where G is the generator, X CT is the input CT image, and Resnet is the residual module for feature extraction. The two discriminators are used to judge the similarity between the generated synthetic PET image and the real PET image used as the reference, and the similarity between the CT image remapped by the inverse generator and the input CT image. The discriminator uses multiple convolution blocks to identify whether the generated PET image is close to the real PET image. The loss function of the discriminator uses the cross entropy loss function D loss = -log(D(Y PET ))-log(1-D(G(X CT ))), where D is the discriminator, Y PET is the real PET image, G(X CT ) is the generated PET image. In order to fully ensure that the generated synthetic PET images have high structural consistency and functional fidelity, the cycle generative adversarial network also introduces a cycle consistency loss function L Cycle The correlation between the CT image and the synthetic PET image is evaluated and constrained. The cycle consistency loss function is L Cycle (G,F)=‖X CT -F(G(X CT ))‖1+‖Y PET -F(G(Y PET ))‖1, where G is the positive generator, F is the inverse generator, and ‖.‖1 represents the L1 norm.

[0072] Based on the structure of the above-mentioned cyclic generative adversarial network model, a specific method of generating a synthetic PET image based on a basic image using a cyclic generative adversarial network module includes S31 to S34.

[0073] S31. Generate a preliminary PET image based on the base image using the positive generator of the cyclic generative adversarial network module.

[0074] S32. Generate a verification image based on the preliminary PET image using the inverse generator of the cyclic generative adversarial network module.

[0075] S33, using the two discriminators of the cyclic generative adversarial network module to respectively judge the first difference between the basic image and the verification image, and the second difference between the preliminary PET image and the standard PET image.

[0076] S34, when the first difference and the second difference corresponding to all the basic images meet the preset conditions, all the preliminary PET images are combined to obtain a synthetic PET image. The first difference, the second difference and the preset conditions can be determined according to actual needs.

[0077] In the cyclic generative adversarial network module of the present invention, a preliminary PET image is first generated by a generator, and then the generated preliminary PET image is verified by a discriminator. After the verification is successful, a synthetic PET image is finally merged, which can fully ensure the quality of the synthetic PET image.

[0078] S4. Use the Transformer module to perform global optimization and local optimization on the synthesized PET image to obtain a first optimized image. Figure 3 As shown in Figure 1, the Transformer module is mainly composed of a multi-head attention mechanism and a feedforward network. After the features of the CT image are processed by the attention mechanism, they are combined with the contextual information of other positions to form global features. The formula of the multi-head self-attention mechanism is Where Q is the query vector, K is the key vector, V is the value vector, and d k is the dimension of the key vector. Furthermore, the present invention also introduces a cross-attention mechanism in the Transformer module, which maps the global structural information of the CT image and the local detail information of the PET image to the generated sPET image, ensuring that the metabolic characteristics of the active bone marrow area are better captured and preserved in complex anatomical areas such as the pelvis.

[0079] The specific method of using the Transformer module to perform global optimization and local optimization on the synthetic PET image includes S41 to S44.

[0080] S41. Input the base image and the standard PET image into the Transformer module.

[0081] S42. Use a convolutional neural network to extract features from the basic image to obtain a high-dimensional feature vector, and use a convolutional neural network to extract features from the standard PET image to obtain local feature information of the bone marrow active area.

[0082] S43. Generate global features based on the high-dimensional feature vectors of all base images using the Transformer module.

[0083] S44. Map the global feature and local feature information to the synthetic PET image based on the cross-attention mechanism to obtain a first optimized image.

[0084] S5, using the conditional diffusion model module to sequentially add noise and remove noise from the first optimized image to obtain a second optimized image. Figure 4 As shown in the figure, the conditional diffusion model realizes multi-scale processing of images through a multi-layer convolutional neural network. Specifically, the input first optimized image will undergo preliminary feature extraction through the convolution block, and then enter the downsampling, intermediate processing and upsampling process, and gradually perform denoising and restoration of details. During the entire processing process, the conditional diffusion model ensures the retention of details and noise elimination of the generated image, so that the final output second optimized image has high clarity and structural fidelity, which can provide more accurate image support for radiotherapy planning. Through the upsampling, intermediate processing and downsampling process, images can be processed at different scales to ensure that the generated images have good structural consistency at different resolutions.

[0085] The specific method of sequentially adding noise and removing noise from the first optimized image using the conditional diffusion model module includes S51 to S52.

[0086] S51, using the conditional diffusion model to add predetermined noise to the first optimized image multiple times to obtain a noisy image. The specific method of adding noise is: where x t represents the image after noise injection, α t The intensity to use for the noise.

[0087] S52, denoising the noisy image multiple times using the conditional diffusion model to obtain a second optimized image. The specific method of denoising is:

[0088] In the conditional diffusion model, the present invention enhances the detail retention capability in the image by adding noise, retains key metabolic feature information during the denoising process, and ultimately ensures that the output second optimized image maintains high accuracy in the bone marrow active area.

[0089] S6. Use the segmentation network module to identify the second optimized image, extract the bone marrow active area from the second optimized image, and determine the location of the bone marrow active area. Figure 5As shown in the figure, the segmentation network is built based on a convolutional neural network, which mainly includes multiple convolution blocks and residual blocks. The convolution module is used to extract multi-level feature information from the second optimized image. Each convolution block is followed by batch normalization and activation function to ensure the accuracy of the segmentation result. The residual module introduces jump connections to ensure that high-resolution image detail information can still be retained in the deep convolution network. The residual module can be expressed as ResNet(X)=X+f(X), where f(X) is transformed by convolution, batch normalization and activation function.

[0090] Furthermore, the segmentation network uses a combination of cross entropy loss and Dice loss to ensure high-precision segmentation of the bone marrow active area. The Dice loss function is: Where P is the predicted segmentation result and G is the true segmentation label.

[0091] Based on the above network structure, the specific method of using the segmentation network module to identify the second optimized image includes S61 to S63.

[0092] S61. Use the convolution submodule of the segmentation network module to perform feature recognition on the second optimized image to obtain multiple candidate regions. The recognition process is constrained by the residual submodule of the segmentation network module.

[0093] S62. Evaluate the segmentation loss of the candidate region. If the evaluation is passed, record the candidate region as a bone marrow active region.

[0094] S63. Determine the coordinates of the bone marrow active area in the second optimized image, and generate the position of the bone marrow active area based on the coordinates.

[0095] Two experiments are conducted on the present invention using real CT images and PET images.

[0096] In the first experiment, two indicators, PSNR (Peak Signal-to-Noise Ratio) and SSIM (Structural Similarity Index Measure), were used for evaluation. The meanings and calculation methods of the two parameters are as follows.

[0097] PSNR is used to measure the reconstruction quality of an image, reflecting the similarity between the generated image and the original image. The larger the PSNR value, the higher the image quality. The calculation formula of PSNR is: in For an 8-bit image, Usually 255, MSE is the mean square error, and there are Among them, I(i,j) is the pixel value of the original image, K(i,j) is the pixel value of the generated image, and m and n are the width and height of the image.

[0098] SSIM is used to measure the similarity between two images in terms of brightness, contrast and structure. The value of SSIM is between [-1, 1]. The closer the value is to 1, the higher the similarity between the two images. The calculation formula of SSIM is: Among them, μ x and μ y are the means of images x and y, σ x 2 and σ y 2 are the means of images x and y, σ xy is the covariance between images x and y. C1 and C2 are constants introduced to stabilize the calculation, usually with the value of C1 = (k1L) 2 C2=(k2L) 2 , L is the maximum pixel value, k1 and k2 are usually 0.01 and 0.03 respectively.

[0099] The CT images, PET images and generated results are as follows: Figure 6 As shown, the calculation results of PSNR and SSIM parameters are as follows Figure 7 As shown, in Figure 7 In the figure, the present invention is marked as Ours, and UNet_128, UNet_256 and cGAN are three prior arts used as comparative examples.

[0100] Furthermore, in the second experiment, only CT images were used to generate synthetic PET images, and the DICE coefficient was used for evaluation. The DICE coefficient is an indicator used to measure the similarity between two sets. It is usually used to evaluate the accuracy of segmentation results. Its value is between 0 and 1. The closer it is to 1, the higher the overlap between the segmentation result and the true value. The calculation method of the DICE coefficient is Among them, A represents the predicted segmentation result, B represents the actual segmentation label, |A∩B| is the intersection of the predicted result and the actual label, |A| and |B| represent the area size of the predicted segmentation and the actual segmentation respectively. The methods based on Unet and Seg_GCN_50 in the prior art are used as comparative examples. The results are shown in Figure 8 shown.

[0101] It can be seen from the results of the above two experiments that the present invention can more accurately generate PET images based on CT images and mark the bone marrow active areas therein.

[0102] The present invention further provides a bone marrow active area imaging positioning device, comprising a data acquisition module, a model implementation module and a data output module.

[0103] The data acquisition module is used to collect CT images of patients and perform preprocessing to obtain multiple basic images.

[0104] The model implementation module is used to process the basic image using a pre-trained imaging localization model to determine the bone marrow active area and the location of the bone marrow active area.

[0105] The data output module is used to output the location of the bone marrow active area.

[0106] Those skilled in the art can clearly understand that, for the convenience and simplicity of description, the specific working process of the system and device described above can refer to the corresponding process in the aforementioned method embodiment, and will not be repeated here. In the several embodiments provided in the present disclosure, it should be understood that the disclosed system, device and method can be implemented in other ways. The device embodiments described above are merely schematic. For example, the division of the modules is only a logical function division. There may be other division methods in actual implementation. For example, multiple modules or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some communication interfaces, indirect coupling or communication connection of devices or modules, which can be electrical, mechanical or other forms.

[0107] The modules described as separate components may or may not be physically separated, and the components shown as modules may or may not be physical modules, that is, they may be located in one place or distributed on multiple network modules. Some or all of the modules may be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0108] The present invention also provides a computer device, including a memory and a processor.

[0109] A memory storing a computer program.

[0110] A processor is used to read and execute the computer program to implement the above-mentioned deep learning-based imaging localization method for bone marrow active areas.

[0111] The present invention finally provides a storage medium for storing a computer program, which, when executed, implements the above-mentioned deep learning-based imaging localization method for active bone marrow areas.

[0112] Based on the storage medium, the technical solution of the present disclosure, or the part that contributes to the prior art, or the part of the technical solution, can be embodied in the form of a software product, which is stored in a storage medium and includes several instructions for enabling a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the method described in each embodiment of the present disclosure. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM), random access memory (RAM), disk or optical disk, etc., various media that can store program codes.

[0113] The various embodiments in this specification are described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the various embodiments can be referenced to each other.

[0114] The above description of the disclosed embodiments enables one skilled in the art to implement or use the present invention. Various modifications to these embodiments will be apparent to one skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to the embodiments shown herein, but rather to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for imaging localization of active bone marrow areas based on deep learning, characterized in that: The steps include: Acquire CT images of the patient and perform preprocessing to obtain multiple basic images; Input the base image into a pre-trained imaging localization model, which includes a recurrent generative adversarial network module, a Transformer module, a conditional diffusion model module, and a segmentation network module; Generate synthetic PET (Positron Emission Computed Tomography) images based on the base images using a recurrent generative adversarial network module; Using the Transformer module to perform global and local optimization on the synthetic PET image to obtain a first optimized image; The conditional diffusion model module is used to add noise and remove noise from the first optimized image in turn to obtain a second optimized image; the segmentation network module is used to identify the second optimized image, extract the bone marrow active area from the second optimized image, and determine the position of the bone marrow active area.

2. The deep learning-based imaging localization method for active bone marrow regions according to claim 1, characterized in that: The method for preprocessing CT images includes: Traverse each voxel in the CT image and obtain the HU value of each voxel; Normalize the HU value less than the preset first threshold value to the first threshold value, and normalize the HU value greater than the preset second threshold value to the second threshold value; normalizing the first threshold value and the second threshold value to a first scale; The CT image is normalized by using the bilinear interpolation method and the nearest neighbor interpolation method to generate multiple basic images with a size of n×n.

3. The deep learning-based imaging localization method for bone marrow active areas according to claim 1, characterized in that: The specific method of using the recurrent generative adversarial network module to generate synthetic PET images based on the basic image includes: Generate preliminary PET images based on the base images using the positive generator of the recurrent generative adversarial network module; Generate verification images based on preliminary PET images using the inverse generator of the recurrent generative adversarial network module; Using two discriminators of the cyclic generative adversarial network module to respectively judge a first difference between the base image and the verification image, and a second difference between the preliminary PET image and the standard PET image; When the first differences and the second differences corresponding to all the basic images satisfy a preset condition, all the preliminary PET images are combined to obtain a synthetic PET image.

4. The method for imaging localization of active bone marrow regions based on deep learning according to claim 1, characterized in that: The specific methods of using the Transformer module to perform global and local optimization on synthetic PET images include: Input the base image and standard PET image into the Transformer module; The convolutional neural network is used to extract features from the basic image to obtain a high-dimensional feature vector, and the convolutional neural network is used to extract features from the standard PET image to obtain local feature information of the active bone marrow area; Use the Transformer module to generate global features based on the high-dimensional feature vectors of all base images; Based on the cross-attention mechanism, the global feature and local feature information are mapped into the synthetic PET image to obtain the first optimized image.

5. The deep learning-based imaging localization method for active bone marrow regions according to claim 1, characterized in that: The specific method of sequentially adding noise and removing noise from the first optimized image using the conditional diffusion model module includes: Using a conditional diffusion model, adding predetermined noise to the first optimized image multiple times to obtain a noisy image; The conditional diffusion model is used to denoise the noisy image multiple times to obtain a second optimized image.

6. The deep learning-based imaging localization method for bone marrow active areas according to claim 1, characterized in that: The specific method of using the segmentation network module to identify the second optimized image includes: Using the convolution submodule of the segmentation network module to perform feature recognition on the second optimized image to obtain multiple candidate regions, and the recognition process is constrained by using the residual submodule of the segmentation network module; The segmentation loss of the candidate region is evaluated, and if the evaluation is passed, the candidate region is recorded as a bone marrow active region; The coordinates of the bone marrow active region in the second optimized image are determined, and the position of the bone marrow active region is generated based on the coordinates.

7. The deep learning-based imaging localization method for active bone marrow regions according to claim 1, characterized in that: The method for training the imaging localization model includes: Acquire multiple whole-body PET-CT image pairs, and divide them into a training atlas and a validation atlas based on a preset division ratio; Assign an identifier to each whole-body PET-CT image pair; Data screening and normalization of whole-body PET-CT image pairs; The imaging localization model is trained using the training atlas, and the trained imaging localization model is verified using the verification atlas. If the verification passes, the imaging localization model is retained.

8. A bone marrow active area imaging positioning device, characterized in that: include: A data acquisition module is used to collect CT images of patients and perform preprocessing to obtain multiple basic images; A model implementation module is used to process the basic image using a pre-trained imaging localization model to determine the bone marrow active area and the location of the bone marrow active area; The data output module is used to output the location of the bone marrow active area.

9. Computer device, characterized in that include: a memory storing a computer program; A processor, configured to read and execute the computer program to implement the deep learning-based imaging localization method for active bone marrow areas as described in any one of claims 1 to 6.

10. A storage medium, characterized in that Used to store a computer program, which, when executed, implements the deep learning-based imaging localization method for bone marrow active areas as described in any one of claims 1-6.

Images