A PET image generation method, system and device based on 3D wavelet transform Brown bridge diffusion and a storage medium

Abstract

This invention relates to the field of hospital image analysis technology, providing a method, system, device, and storage medium for generating PET images based on 3D wavelet transform Brownian bridge diffusion. It addresses the technical problem of improving the accuracy and reliability of PET images in existing nasopharyngeal carcinoma medical image analysis processes. The method divides MRI and PET images into low-frequency structural information and high-frequency detail information through wavelet decomposition, then uses bi-branch Brownian bridge diffusion to model global semantics and local details respectively, and finally performs fusion through inverse wavelet transform to obtain a more stable and accurate final three-dimensional synthesized PET image. This invention is applicable to nasopharyngeal carcinoma or other tumors, or other medical image analysis processes.

Description

Technical Field

[0001] This invention relates to the field of hospital image analysis technology, and more specifically, to a method, system, device, and storage medium for generating PET images based on 3D wavelet transform Brownian bridge diffusion. Background Technology

[0002] Nasopharyngeal carcinoma is a head and neck malignant tumor with a significant geographical prevalence, particularly in southern China and Southeast Asia. Due to the complex anatomical structures of the nasopharynx, including the skull base, cranial nerve foramina, parapharyngeal space, and major blood vessels, tumors often spread in a stepwise manner. Therefore, accurate assessment of the extent of invasion is crucial for treatment planning, radiotherapy target delineation, and prognosis. Currently, prognostic prediction primarily relies on magnetic resonance imaging (MRI) combined with patient clinical indicators for comprehensive evaluation. Positron emission tomography (PET) can provide metabolic and functional information; therefore, combining PET images with MRI images has a significant role in predicting the prognosis of nasopharyngeal carcinoma.

[0003] However, the acquisition and fusion of PET images still face many limitations. On the one hand, because the PET imaging process requires the injection of radioactive isotopes, this imposes an ionizing radiation burden on patients, posing a potential long-term risk for nasopharyngeal carcinoma patients who require multiple follow-up assessments; on the other hand, the cost of acquiring PET images is high.

[0004] In addition, PET images have problems such as oversmoothing, blurred boundaries, distorted boundaries, drifting of details, loss of details, and local artifacts, such as failing to reproduce details such as lymph nodes well.

[0005] Therefore, improving the accuracy and reliability of PET images is a technical problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0006] This application aims to address the technical problem of improving the accuracy and reliability of PET images in the existing nasopharyngeal carcinoma medical image analysis process, and provides a PET image generation method, system, device, and storage medium based on 3D wavelet transform Brownian bridge diffusion.

[0007] A first aspect of this application provides a PET image generation method based on 3D wavelet transform Brownian bridge diffusion, comprising the following steps:

[0008] Step (1): Preprocess the acquired MRI and PET image data to obtain preprocessed MRI three-dimensional images. and preprocessed PET 3D images ;

[0009] Step (2): The 3D Haar wavelet transform algorithm is used to process the preprocessed MRI three-dimensional image. Decomposed into low-frequency components and high frequency components The preprocessed PET 3D image Decomposed into low-frequency components and high frequency components ;

[0010] Step (3), the low-frequency component Low-frequency components Defined as low-frequency components Low-frequency components , to high frequency components High-frequency components Defined as high-frequency components High-frequency components ;

[0011] Low frequency components As a bridging starting point, the low-frequency components As the bridging endpoint, low-frequency Brownian bridge diffusion processing is performed to obtain the low-frequency bridge branch. ;

[0012] Calculate high frequency residuals : ; to high frequency components As a bridging starting point, the high-frequency residual As the bridging endpoint, high-frequency Brownian bridge diffusion processing is performed to obtain residual information. Calculate the high-frequency bridge branch : ;

[0013] Step (4): Use the inverse wavelet transform algorithm to divide the low-frequency bridge branch. and high frequency bridge branch The final 3D synthetic PET image is obtained by fusing the images. .

[0014] Preferably, the preprocessing step (1) is as follows:

[0015] First, the MRI and PET image data were registered.

[0016] Secondly, the MRI and PET image data were resampled at a consistent resolution.

[0017] Finally, the resampled MRI and PET image data were normalized to obtain the preprocessed 3D MRI images. and preprocessed PET 3D images .

[0018] Preferably, the registration process uses CT images as a registration reference to register MRI image data and PET image data; the CT images are acquired simultaneously with the PET image data.

[0019] Preferably, a joint loss function is used to optimize the diffusion process. The joint loss function is:

[0020] ;

[0021] In the formula, Indicates weight, Indicates weight, Indicates weight, For low-frequency branch loss, For high-frequency branch loss, For image domain reconstruction loss, Used to enhance high-frequency boundary and texture representation. Used to constrain the final 3D synthesized PET image Consistency with the PET image data in step (1) in the image space.

[0022] A second aspect of this application provides a PET image generation system based on 3D wavelet transform Brownian bridge diffusion, including a preprocessing module, a data decomposition module, a dual-bridge diffusion generation module, and a joint reconstruction optimization module;

[0023] The preprocessing module is configured to preprocess the acquired MRI and PET image data to obtain preprocessed MRI three-dimensional images. and preprocessed PET 3D images ;

[0024] The data decomposition module is configured to use the 3D Haar wavelet transform algorithm to decompose the preprocessed MRI three-dimensional images. Decomposed into low-frequency components and high frequency components The preprocessed PET 3D image Decomposed into low-frequency components and high frequency components ;

[0025] The dual-bridge diffusion generation module is configured to generate low-frequency components. Low-frequency components Defined as low-frequency components Low-frequency components , to high frequency components High-frequency components Defined as high-frequency components High-frequency components ;

[0026] Low frequency components As a bridging starting point, the low-frequency components As the bridging endpoint, low-frequency Brownian bridge diffusion processing is performed to obtain the low-frequency bridge branch. ;

[0027] Calculate high frequency residuals : ; to high frequency components As a bridging starting point, the high-frequency residual As the bridging endpoint, high-frequency Brownian bridge diffusion processing is performed to obtain residual information. Calculate the high-frequency bridge branch : ;

[0028] The joint reconstruction optimization module is configured to use an inverse wavelet transform algorithm to reconstruct the low-frequency bridge branch. and high frequency bridge branch The final 3D synthetic PET image is obtained by fusing the images. .

[0029] Preferably, the preprocessing module performs the preprocessing process as follows:

[0030] First, the MRI and PET image data were registered.

[0031] Secondly, the MRI and PET image data were resampled at a consistent resolution.

[0032] Finally, the resampled MRI and PET image data were normalized to obtain the preprocessed 3D MRI images. and preprocessed PET 3D images .

[0033] Preferably, the registration process uses CT images as a registration reference to register MRI image data with PET image data; the CT images are acquired simultaneously with the PET image data.

[0034] A third aspect of this application provides a storage medium storing a computer program thereon, which, when executed by a processor, implements each step of any of the above-described PET image generation methods based on 3D wavelet transform Brownian bridge diffusion.

[0035] A fourth aspect of this application provides a PET image generation apparatus based on 3D wavelet transform Brownian bridge diffusion, comprising a memory and a processor; the memory is used to store a program; the processor is used to execute the program to implement each step of any of the above-described PET image generation methods based on 3D wavelet transform Brownian bridge diffusion.

[0036] The beneficial effects of this application are that it provides more accurate, reliable, and visually realistic PET image information, providing more reliable data for the detection, precise localization, auxiliary diagnosis, radiotherapy planning, efficacy evaluation, and prognostic analysis of nasopharyngeal carcinoma.

[0037] The obtained PET image information better takes into account the overall structure, local texture and local detail expression, and reduces common problems in existing technologies such as oversmoothing, blurred boundaries, boundary distortion, detail drift, missing details and local artifacts.

[0038] The low-frequency-high-frequency decoupled dual-bridge diffusion mechanism separates global structure restoration from local detail compensation, which helps reduce oversmoothing, detail drift and boundary distortion in complex regions.

[0039] Combining MRI with PET scans can alleviate limitations in terms of radioactive tracers, equipment costs, and clinical accessibility, providing more convenient supplementary metabolic information for follow-up and auxiliary assessment. It avoids multiple PET imaging procedures, minimizing the ionizing radiation burden on patients from radioactive isotopes during the PET imaging process. Furthermore, reducing the number of PET imaging procedures lowers testing costs.

[0040] This invention is not limited to nasopharyngeal carcinoma, but can also be applied to other tumors or other medical imaging analysis processes.

[0041] Further features and aspects of this application will be clearly described in the following detailed description with reference to the accompanying drawings. Attached Figure Description

[0042] Figure 1 This is a diagram of the PET image generation system architecture based on 3D wavelet transform Brownian bridge diffusion according to the present invention. Detailed Implementation

[0043] The specific embodiments described below are merely preferred embodiments of this application, and the scope of protection of this application is not limited thereto. Those skilled in the art can make modifications or variations based on the principles, concepts, and spirit of this application, and the resulting technical solutions should all be covered within the scope of protection of this application.

[0044] refer to Figure 1The PET image generation method based on 3D wavelet transform Brownian bridge diffusion includes the following steps:

[0045] Step (1): Preprocess the MRI and PET image data.

[0046] For nasopharyngeal carcinoma patients, MRI 3D image data and PET 3D image data were acquired using MRI image acquisition equipment and a PET / CT integrated machine. The PET / CT integrated machine, as a positron emission tomography (PET) and X-ray computed tomography (CT) system, simultaneously acquires CT images during the acquisition of PET 3D image data; the PET 3D image data and CT images are strictly registered. Since MRI and PET images are acquired separately, patient positioning will inevitably differ under different acquisition conditions, and the field of view of the two images is not the same. Therefore, it is necessary to unify these two modalities into a common physical space. However, the PET modality has a thicker scan slice count, and compared to the CT modality, its rigid registration anchor points, such as the skull, are not as clear. CT images have a clearer bone structure and a thinner scan slice count. Therefore, using the CT image as a reference, the MRI and PET modalities are unified into the same physical space, that is, the MRI and PET modalities are unified into the physical space of the CT image. Therefore, using the CT image as a fixed registration target, the MRI 3D image data is registered to obtain strictly registered MRI and PET modalities.

[0047] Next, the MRI 3D image data and the PET 3D image data are resampled to ensure consistent resolution and field of view. This results in resampled MRI 3D image data and resampled PET 3D image data with consistent resolution and field of view. The two data form a voxel-level correspondence in 3D space, achieving spatial alignment between the two modalities.

[0048] Subsequently, the resampled MRI 3D image data and the resampled PET 3D image data were normalized to obtain the preprocessed MRI 3D images. and preprocessed PET 3D images .

[0049] Step (2) involves processing the preprocessed MRI three-dimensional images. and preprocessed PET 3D images The data is input into the data decomposition module, which uses the 3D Haar wavelet transform algorithm to decompose the preprocessed MRI 3D images. Decomposed into low-frequency components and high frequency components The preprocessed PET 3D image Decomposed into low-frequency components and high frequency components Among them, the low-frequency component , High-frequency components are used to characterize organ contours, main anatomical morphology, and coarse-scale intensity distribution. , It is used to characterize boundaries, textures, local contrast, and subtle structural changes. Through this explicit decomposition method, the global semantic recovery problem and the local detail compensation problem, which were originally coupled in the same space, are separated into two relatively independent but mutually cooperative sub-tasks.

[0050] The 3D Haar wavelet transform algorithm yielded a low-frequency subband and seven high-frequency subbands corresponding to three dimensions. These high-frequency subbands contain detailed information in each direction.

[0051] Step (3), Brownian bridge diffusion treatment.

[0052] The low-frequency component input to the dual-bridge diffusion generation module Low-frequency components Defined as low-frequency components Low-frequency components The high-frequency components input to the dual-bridge diffusion generation module High-frequency components Defined as high-frequency components High-frequency components Among them, high-frequency components and It is composed of seven high-frequency subbands of each mode spliced ​​together.

[0053] The dual-bridge diffusion generation module will generate low-frequency components. As a bridging starting point, the low-frequency components As the bridging endpoint, a low-frequency Brownian bridge diffusion process is established within the Brownian bridge diffusion framework:

[0054]

[0055] In the formula, The time step scheduling coefficient, For noise intensity, It is standard Gaussian noise.

[0056] The low-frequency bridge branch is then obtained through the following formula. :

[0057]

[0058] In the formula, Represents low-frequency components with low-frequency components The difference (i.e., the low-frequency difference between the MRI modality and the PET modality). This represents the noise in the network prediction.

[0059] Constructing low-frequency bridging branches allows for the priority establishment of a stable structural framework, reducing the difficulty of subsequent high-frequency generation. The role of low-frequency bridging branches is primarily to recover the low-frequency semantic framework related to overall anatomical structure and metabolic distribution in PET, providing stable and reliable global conditions for the generation of subsequent high-frequency details. This low-frequency-first approach reduces the training difficulty in complex 3D cross-modal mapping and enhances the stability of the generation process.

[0060] The dual-bridge diffusion generation module will also generate high-frequency components. As a bridging starting point, the high-frequency residual As the bridging endpoint, a high-frequency Brownian bridge diffusion process is established within the Brownian bridge diffusion framework:

[0061]

[0062] In the formula, the high-frequency residual for: ; The time step scheduling coefficient, For noise intensity, It is standard Gaussian noise.

[0063] Furthermore, under the condition of low-frequency bridge branch prediction results, the residual information is calculated using the following formula. :

[0064]

[0065] Furthermore, the high-frequency bridge branch is calculated using the following formula. :

[0066]

[0067] This design enables the generation of high-frequency details to be based on the correct global structure, which is more conducive to restoring local contrast, texture information and boundary features, reducing the oversmoothing and detail drift problems in traditional methods, and reducing the boundary misalignment, texture drift and artifact problems common in single-branch models.

[0068] Constructing a high-frequency bridge branch The process does not directly generate all high-frequency information from scratch, but focuses on learning the high-frequency changes that are actually needed during the MRI to PET conversion, and uses the low-frequency bridge branch as a conditional prior to guide the high-frequency bridge to restore boundaries, textures and local contrasts in the correct structural locations.

[0069] The Brownian bridge diffusion model used in the study of the Brownian bridge diffusion process was trained using a paired MRI-PET dataset for nasopharyngeal carcinoma provided by the hospital. The data acquisition equipment was provided by Shanghai United Imaging Healthcare Co., Ltd. The dataset contained 267 paired 3D volumetric data points. After removing missing sequences, cases with acquisition failures, and severe motion artifacts, the data was divided into training, validation, and test sets, with 86, 10, and 39 cases respectively. The Adam optimizer was used during training, with an initial learning rate of 1×10⁻⁶. -4 And use an exponential moving average strategy for stable training.

[0070] The Brownian bridge diffusion model was deployed on a single NVIDIA V100 (32GB) GPU, implemented using PyTorch 2.4.1 and CUDA 11.8. The diffusion steps were set to 1000, and a linear strategy was used for bridging time scheduling, with an initial time coefficient of 0.001 and a final time coefficient of 0.999. The noise variance scaling factor was set to 2.0, and a deterministic sampling strategy was used in the sampling phase, combined with skip-step linear sampling to improve inference efficiency.

[0071] Step (4), branch the low-frequency bridge and high frequency bridge branch The common input joint reconstruction optimization module uses the inverse wavelet transform algorithm to divide the low-frequency bridge branch. and high frequency bridge branch The final 3D synthetic PET image is obtained by fusing the images. The fusion is accomplished using the following formula:

[0072]

[0073] Through inverse transformation, low-frequency global structural information and high-frequency local detail information are fused into the same image space. Since low-frequency information has already determined the global morphology and metabolic framework, while high-frequency information further supplements edges and details, the result after inverse transformation fusion can simultaneously maintain global structural fidelity and local texture representation. Cross-modal generation of nasopharyngeal carcinoma MRI and PET images, combining wavelet domain modeling with image domain joint optimization, improves generation efficiency while enhancing the structural fidelity, detail quality, and quantitative consistency of the synthesized PET images.

[0074] To ensure the fusion result remains consistent in both the frequency and image domains, a joint loss function can be used to optimize the diffusion process. The joint loss function is:

[0075]

[0076] In the formula, Indicates weight, Indicates weight, Indicates weight, For low-frequency branch loss, For high-frequency branch loss, For image domain reconstruction loss, Used to enhance high-frequency boundary and texture representation. Used to constrain the final 3D synthesized PET image Consistency with the PET image data in step (1) in the image space.

[0077] Through the above joint optimization, structural fidelity, detail quality, and metabolic quantification stability can be improved simultaneously.

[0078] This invention better balances the representation of overall structure, local texture, and local details, reducing problems common in traditional methods such as over-smoothing, blurred boundaries, boundary distortion, detail drift, missing details, and local artifacts. It also maintains good three-dimensional spatial continuity and improves the stability and consistency of the generated results at the metabolic quantification level. This enhances the accuracy and visual realism of synthesized PET images.

[0079] To verify the effectiveness of the method of the present invention, three representative cross-modal generation methods were selected as comparison objects. The first category is PASTA and CoCoLIT, which are based on 2.5D conditional diffusion. The second category is Vox2Vox and HA-GAN, which are based on 3D GAN. The third category is 3D-Med, 3D-WLDM, cWDM and FICD, which are based on 3D diffusion models. Evaluation metrics fall into two categories. The first category comprises image quality assessment metrics, including the following three: ① Structural / morphological level: Multiscale structural similarity (MS-SSIM) is used to demonstrate whether the model can correctly recover organ structures, boundaries, and multiscale details, avoiding structural mismatches or local texture anomalies; ② Intensity error level: Peak signal-to-noise ratio (PSNR) is used to demonstrate the accuracy of voxel-level intensity reconstruction and its ability to reduce overall error and noise. This metric is sensitive to intensity shifts and can serve as a "lower bound" for error; ③ Distribution level: Maximum mean difference (MMD) is used to demonstrate whether the model-generated results are statistically aligned with real PET images, avoiding situations where only the test samples are "fitted" individually, but the overall distribution still deviates from the real data (e.g., smoother overall, lower contrast, unrealistic noise and texture). The second category consists of clinically relevant quantitative metrics. SUV (standard uptake value) is an important standard value for judging tumor and lymph node characteristics in PET images. SUVmean (mean standard uptake value) reflects whether the global metabolic level has shifted, and SUVmax (maximum standard uptake value) reflects whether the peak values ​​of the examination hotspots have been weakened (over-smoothed) or exaggerated (artifacts / noise). SUVmean and SUVmax are particularly crucial for downstream clinical analysis. The experimental comparison results are shown in Table 1 below.

[0080] Table 1 Comparison of Experimental Results

[0081] Method Name MS-SSIM PSNR MMD SUVmean SUVmax PASTA 0.78 27.00 78.57 0.8302 0.8567 CoCoLIT 0.78 16.87 84.62 0.5500 0.6734 Vox2Vox 0.86 17.44 82.14 0.8846 0.9116 HA-GAN 0.57 17.26 0.13 18.03 29.36 3D-Med 0.68 17.51 0.15 421.24 25.47 3D-WLDM 0.66 14.65 0.16 28.59 42.70 cWDM 0.79 20.75 0.03 34.18 26.20 FICD 0.59 15.39 0.11 45.54 26.88 This invention 0.87 20.05 0.12 12.11 24.13

[0082] Experimental results show that the method of this invention achieves best or tied-best performance on most evaluation metrics. Specifically, on the test set, this method achieved an MS-SSIM of 0.87 and reduced SUVmean and SUVmax to 12.11 and 24.13, respectively, indicating significant advantages in structural fidelity and consistency of metabolic quantification. Compared with existing methods, the PET images generated by this invention have clearer boundaries and more reasonable local contrast in nasopharyngeal carcinoma lesions and adjacent complex anatomical structures, and can better preserve metabolic hotspots and structural continuity.

[0083] The method of this invention has achieved superior performance in multiple indicators such as structural similarity, detail quality, and SUV value, indicating that it can provide more reliable synthetic PET information for the detection, precise localization, auxiliary diagnosis, radiotherapy planning, efficacy evaluation, and prognostic analysis of nasopharyngeal carcinoma, and has good prospects for clinical application.

Claims

1. A method for PET image generation based on 3D wavelet transform Brownian bridge diffusion, characterized in that, Includes the following steps: Step (1), pre-processing MRI image data and PET image data collected respectively to obtain pre-processed MRI three-dimensional image and pre-processed PET three-dimensional image ; Step (2): The 3D Haar wavelet transform algorithm is used to process the preprocessed MRI three-dimensional image. Decomposed into low-frequency components and high frequency components The preprocessed PET 3D image Decomposed into low-frequency components and high frequency components ; Step (3), the low-frequency component Low-frequency components Defined as low-frequency components Low-frequency components , to high frequency components High-frequency components Defined as high-frequency components High-frequency components ; Low frequency components As a bridging starting point, the low-frequency components As the bridging endpoint, low-frequency Brownian bridge diffusion processing is performed to obtain the low-frequency bridge branch. ; Calculate high-frequency residuals : ; to high frequency components As a bridging starting point, the high-frequency residual As the bridging endpoint, high-frequency Brownian bridge diffusion processing is performed to obtain residual information. Calculate the high-frequency bridge branch : ; Step (4): Use the inverse wavelet transform algorithm to divide the low-frequency bridge branch. and high frequency bridge branch The final 3D synthetic PET image is obtained by fusing the images. .

2. The PET image generation method based on 3D wavelet transform Brownian bridge diffusion according to claim 1, characterized in that, The preprocessing process in step (1) is as follows: First, the MRI and PET image data were registered. Secondly, the MRI and PET image data were resampled at a consistent resolution. Finally, the resampled MRI and PET image data were normalized to obtain the preprocessed 3D MRI images. and preprocessed PET 3D images .

3. The PET image generation method based on 3D wavelet transform Brownian bridge diffusion according to claim 2, characterized in that, The registration process uses CT images as a registration reference to register MRI image data and PET image data; the CT images are acquired simultaneously with the PET image data.

4. The PET image generation method based on 3D wavelet transform Brownian bridge diffusion according to claim 2 or 3, characterized in that, The diffusion process is optimized using a joint loss function, which is: ； In the formula, Indicates weight, Indicates weight, Indicates weight, For low-frequency branch loss, For high-frequency branch loss, For image domain reconstruction loss, Used to enhance high-frequency boundary and texture representation. Used to constrain the final 3D synthesized PET image Consistency with the PET image data in step (1) in the image space.

5. A PET image generation system based on 3D wavelet transform Brownian bridge diffusion, characterized in that, It includes a preprocessing module, a data decomposition module, a dual-bridge diffusion generation module, and a joint reconstruction optimization module; The preprocessing module is configured to preprocess the acquired MRI and PET image data respectively to obtain preprocessed MRI three-dimensional images. and preprocessed PET 3D images ; The data decomposition module is configured to use a 3D Haar wavelet transform algorithm to decompose the preprocessed MRI three-dimensional images. Decomposed into low-frequency components and high frequency components The preprocessed PET 3D image Decomposed into low-frequency components and high frequency components ; The dual-bridge diffusion generation module is configured to generate low-frequency components. Low-frequency components Defined as low-frequency components Low-frequency components , to high frequency components High-frequency components Defined as high-frequency components High-frequency components ; Low frequency components As a bridging starting point, the low-frequency components As the bridging endpoint, low-frequency Brownian bridge diffusion processing is performed to obtain the low-frequency bridge branch. ; Calculate high-frequency residuals : ; to high frequency components As a bridging starting point, the high-frequency residual As the bridging endpoint, high-frequency Brownian bridge diffusion processing is performed to obtain residual information. Calculate the high-frequency bridge branch : ; The joint reconstruction optimization module is configured to use an inverse wavelet transform algorithm to reconstruct the low-frequency bridge branch. and high frequency bridge branch The final 3D synthetic PET image is obtained by fusing the images. .

6. The PET image generation system based on 3D wavelet transform Brownian bridge diffusion according to claim 5, characterized in that, The preprocessing module performs the following preprocessing process: First, the MRI and PET image data were registered. Secondly, the MRI and PET image data were resampled at a consistent resolution. Finally, the resampled MRI and PET image data were normalized to obtain the preprocessed 3D MRI images. and preprocessed PET 3D images .

7. The PET image generation system based on 3D wavelet transform Brownian bridge diffusion according to claim 6, characterized in that, The registration process uses CT images as a registration reference to register MRI image data with PET image data; the CT images are acquired simultaneously with the PET image data.

8. A storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method as described in any one of claims 1-4.

9. A PET image generation device based on 3D wavelet transform Brownian bridge diffusion, characterized in that, Including memory and processor; The memory is used to store programs; The processor is configured to execute the program to implement the steps of the method as described in any one of claims 1-4.

Images