A CT-to-PET image generation method and system based on multi-scale conditional diffusion

Abstract

This invention discloses a method and system for generating CT-to-PET images based on multi-scale conditional diffusion. The method includes: CT image preprocessing; extracting multi-scale CT feature maps and performing edge feature extraction and fusion to obtain an edge enhancement condition map; performing multi-level pooling processing on the multi-scale CT feature maps and calculating an attenuation map to generate semantic enhancement condition features; stitching together the edge enhancement condition map, semantic enhancement condition features, and CT attenuation map to construct conditional information and inputting it into a diffusion generation model to gradually generate the target PET image. This invention introduces multi-scale edge enhancement conditions and an attenuation map calculated based on CT images as semantic guidance information, overcoming the problems of insufficient anatomical structure preservation and blurred boundary regions in existing technologies. It fully utilizes the physical priors of PET imaging and accurate metabolic distribution prediction, accurately capturing the intrinsic relationship between structure and metabolism at different spatial scales, and enhancing the semantic discriminativeness of the generated PET image.

Description

Technical Field

[0001] This invention belongs to the technical field of image processing, specifically relating to a method and system for generating CT to PET images based on multi-scale conditional diffusion. Background Technology

[0002] Positron emission tomography (PET) is an important medical imaging technique that can reflect the metabolic activity of human tissues, playing a significant role in tumor diagnosis and treatment evaluation. However, PET imaging requires the injection of radioactive tracers, posing challenges such as high radiation risks, high costs, and limited equipment resources. In contrast, computed tomography (CT) offers faster imaging speeds, lower costs, and wider applications, but it only provides anatomical structural information and cannot directly reflect metabolic function. Therefore, generating corresponding PET images from CT images has become an important research direction in the field of medical image processing.

[0003] Existing methods for generating CT to PET images mainly include those based on generative adversarial networks (GANs) and those based on diffusion models. GAN-based methods typically achieve image conversion by learning the mapping relationship between CT and PET images. However, these methods are prone to instability during training, resulting in blurred images and distorted anatomical structures, making it difficult to guarantee the reliability of the generated PET images. In recent years, diffusion models have been introduced into medical image conversion tasks due to their higher generation stability. However, existing diffusion generation methods often employ simple conditional fusion, typically using only CT or prior images as single conditional inputs, making it difficult to fully utilize the multi-scale structural information in CT images. Furthermore, some methods rely on static attention maps or attenuation maps as prior information. This prior information cannot be adaptively adjusted according to the generation process, resulting in limited guidance for key metabolic regions. In addition, existing methods often lack effective constraints on image edges and anatomical boundaries during generation, easily leading to blurred or inconsistent structural details and metabolic distributions in the generated PET images. Summary of the Invention

[0004] The main objective of this invention is to overcome the shortcomings and deficiencies of the prior art and provide a CT-to-PET image generation method and system based on multi-scale conditional diffusion. By constructing a multi-scale conditional mechanism, structural edge information and attenuation semantic information are jointly introduced during the diffusion generation process to achieve refined guidance of the PET image generation process, thereby improving the anatomical consistency and metabolic accuracy of the generated results.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: One aspect of the present invention provides a method for generating CT-to-PET images based on multi-scale conditional diffusion, comprising the following steps: Acquire the CT images to be processed and perform intensity normalization, size unification, and format conversion; The processed CT images are input into a feature extraction network to generate several CT feature maps at different spatial scales; Edge features are extracted and fused from CT feature maps at different spatial scales to obtain an edge enhancement condition map; the edge feature extraction includes edge feature extraction along the horizontal and vertical directions; CT attenuation maps are calculated based on the processed CT images; multi-level pooling is performed on the CT feature maps at different spatial scales to obtain fused features; using the CT attenuation maps as query information, feature association is performed with the fused features to generate semantically enhanced conditional features. The edge enhancement condition map, semantic enhancement condition features, and CT attenuation map are stitched together to construct condition information at different spatial scales; Conditional information at different spatial scales is input into the diffusion generation model to gradually generate target PET images.

[0006] As a preferred technical solution, the feature extraction network has several feature extraction branches with different receptive fields or different spatial resolutions, which are used to generate several CT feature maps with different spatial scales.

[0007] As a preferred technical solution, the step of extracting and fusing edge features from CT feature maps at different spatial scales to obtain an edge enhancement condition map is specifically as follows: For each CT feature map at a different spatial scale, gradient features are calculated along the horizontal and vertical directions respectively; The gradient features calculated in the horizontal and vertical directions are combined to generate an edge feature map that characterizes the anatomical boundary of the corresponding spatial scale. Edge feature maps obtained at different spatial scales are overlaid to generate a unified edge enhancement condition map.

[0008] As a preferred technical solution, the step of calculating the CT attenuation map based on the processed CT image specifically involves: calculating the attenuation map corresponding to the CT image based on the processed CT image according to a preset physical mapping relationship. The process of using CT attenuation maps as query information and associating them with fused features to generate semantic enhancement conditional features involves: associating features through a cross-attention mechanism to obtain attenuation semantic enhancement features; and fusing attenuation semantic enhancement features obtained at different spatial scales to generate unified semantic enhancement conditional features.

[0009] As a preferred technical solution, the step of inputting conditional information at different spatial scales into the diffusion generation model to gradually generate the target PET image specifically involves: Conditional information at different spatial scales is input into the Brownian bridge diffusion generation model. The Brownian bridge diffusion generation model is equipped with an encoder and a decoder. It takes the current diffusion state, time step information and conditional information as input, and gradually removes noise so that the samples converge from the random state to the target image distribution along the Brownian bridge path. The Brownian bridge diffusion generation model is also equipped with a conditional constraint channel for incorporating the information of the input image into the reverse diffusion process. In the reverse diffusion process of the diffusion generation model, the noise removal process is dynamically modulated according to the condition information of different spatial scales. The target PET image is gradually generated through multiple backdiffusion iterations.

[0010] Another aspect of the present invention provides a CT-to-PET image generation system based on multi-scale conditional diffusion, including a data acquisition and preprocessing module, a multi-scale feature extraction module, a multi-scale edge information enhancement module, a CT attenuation map calculation module, a multi-scale semantic information extraction module, a feature fusion module, and a multi-scale conditional diffusion module. The data acquisition and preprocessing module is used to acquire the CT image to be processed and perform intensity normalization, size unification and format conversion; The multi-scale feature extraction module is used to input the processed CT image into the feature extraction network to generate several CT feature maps at different spatial scales; The multi-scale edge information enhancement module is used to extract and fuse edge features from CT feature maps at different spatial scales to obtain an edge enhancement condition map; the edge feature extraction includes edge feature extraction along the horizontal and vertical directions; The CT attenuation map calculation module is used to calculate the CT attenuation map based on the processed CT image; and to perform multi-level pooling processing on the CT feature maps of different spatial scales to obtain fused features; The multi-scale semantic information extraction module is used to use CT attenuation map as query information, perform feature association with fused features, and generate semantic enhancement conditional features. The feature fusion module is used to stitch together the edge enhancement condition map, semantic enhancement condition features and CT attenuation map to construct condition information at different spatial scales; The multi-scale conditional diffusion module is used to input conditional information at different spatial scales into the diffusion generation model to gradually generate the target PET image.

[0011] As a preferred technical solution, the feature extraction network has several feature extraction branches with different receptive fields or different spatial resolutions, which are used to generate several CT feature maps with different spatial scales.

[0012] As a preferred technical solution, the diffusion generation model adopts the Brownian bridge diffusion generation model; the Brownian bridge diffusion generation model is equipped with an encoder and a decoder, which take the current diffusion state, time step information and condition information as input, and gradually removes noise so that the samples converge from the random state to the target image distribution along the Brownian bridge path; the Brownian bridge diffusion generation model is also equipped with a condition constraint channel for incorporating the information of the input image into the reverse diffusion process.

[0013] Compared with the prior art, the present invention has the following advantages and beneficial effects: (1) This invention addresses the problems of insufficient anatomical structure preservation and blurred boundary regions in existing CT-to-PET image conversion methods by introducing multi-scale edge enhancement conditions to explicitly model the structural boundary information of CT images at different spatial scales. Specifically, based on multi-scale feature extraction of CT images, this invention performs gradient calculations on features at each scale along the horizontal and vertical directions, and fuses the obtained edge features to generate an edge enhancement condition map that characterizes anatomical boundary information. By introducing this edge enhancement condition map into the diffusion generation process, the model can continuously focus on organ contours and tissue boundary regions when generating PET images, thereby effectively reducing structural misalignment and edge blurring, and significantly improving the anatomical consistency between the synthesized PET image and the original CT image.

[0014] (2) This invention addresses the problem that existing generative models do not fully utilize the physical priors of PET imaging and have inaccurate predictions of metabolic distribution by introducing attenuation maps calculated from CT images as semantic guidance information. Attenuation maps reflect the absorption characteristics of different tissues to radiation signals and are important physical prior information in PET imaging. This invention utilizes the cross-attention mechanism between attenuation maps and multi-scale CT features to establish a correlation between structural information and radiation absorption characteristics, enabling the generative model to fully consider the physical differences of different tissue regions when predicting PET metabolic distribution, thereby improving the rationality and accuracy of the synthetic PET images in the expression of high-metabolic and low-metabolic regions.

[0015] (3) This invention effectively enhances the model's ability to express complex metabolic patterns by constructing multi-scale semantically enhanced conditional features. Specifically, this invention performs multi-level pooling fusion on multi-scale CT features to obtain fused features containing multi-scale contextual information. Using attenuation maps as query information, a cross-attention mechanism is used to weight the fused features, thereby suppressing irrelevant or redundant information and highlighting key features related to metabolic distribution. This technical solution enables the model to accurately capture the intrinsic relationship between structure and metabolism at different spatial scales, enhancing the semantic discriminativeness of the generated PET images.

[0016] (4) This invention constructs multi-scale conditional information by fusing edge enhancement conditional maps, semantic enhancement conditional features, and attenuation maps, and introduces this information into the back diffusion process of the diffusion generation model to dynamically modulate the noise removal process. Compared with existing diffusion methods that rely solely on single or static conditions, the multi-scale conditional fusion mechanism of this invention is illustrated in the following experiment. The experiment uses actual acquired CT and PET paired medical image data as the experimental dataset, with CT images as input data and real PET images as reference standard images. The dataset contains samples from various locations and with different metabolic distribution characteristics to verify the applicability and stability of the method in complex medical scenarios. Attached Figure Description

[0017] Figure 1 This is a schematic flowchart of a CT-to-PET image generation method based on multi-scale conditional diffusion according to an embodiment of the present invention. Figure 2 This is a principle block diagram of a CT-to-PET image generation method based on multi-scale conditional diffusion according to an embodiment of the present invention. Detailed Implementation

[0018] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are merely some embodiments of the present application, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present application without creative effort are within the scope of protection of the present application.

[0019] Example 1: like Figure 1 , Figure 2 As shown, this embodiment provides a CT-to-PET image generation method based on multi-scale conditional diffusion, including the following steps: S1: Obtain input data: Obtain the CT image to be processed as input data, and perform normalization and preprocessing operations on the CT image.

[0020] Furthermore, step S1 includes the following steps: S11: Acquire CT image data to be processed, wherein the CT image is two-dimensional or three-dimensional medical image data; S12: Perform intensity normalization processing on the CT image to eliminate intensity differences caused by different scanning conditions; S13: Perform size unification and format conversion on the normalized CT images to meet the input requirements of the subsequent feature extraction network.

[0021] S2: Multi-scale feature extraction: The preprocessed CT image is input into the feature extraction network to generate multiple CT feature maps at different spatial scales, which are used to characterize the structural information of the CT image at different resolutions.

[0022] Furthermore, step S2 includes the following steps: S21: Input the preprocessed CT image into the feature extraction network and extract the initial structural features of the image through convolution operation; S22: During the feature extraction process, multiple feature extraction branches with different receptive fields or different spatial resolutions are set up to generate multiple CT feature maps with different spatial scales. S23: Output the CT feature maps of different spatial scales respectively to generate subsequent edge enhancement condition maps and attenuation semantic enhancement features / semantic enhancement condition features.

[0023] S3: Edge enhancement condition generation: Edge feature extraction is performed on the multiple CT feature maps of different spatial scales obtained in step S2. The edge feature extraction includes gradient operations along the horizontal and vertical directions. Edge features obtained at various scales are fused to generate an edge enhancement condition map that represents anatomical boundary information.

[0024] Furthermore, step S3 includes the following steps: S31: For each spatial scale CT feature map obtained in step S2, calculate the gradient features along the horizontal and vertical directions respectively to obtain the edge response information at the corresponding spatial scale. S32: Combine the horizontal gradient features with the vertical gradient features to generate an edge feature map that characterizes the anatomical boundary of the corresponding spatial scale. S33: Overlay edge feature maps obtained at different spatial scales to generate a unified edge enhancement condition map.

[0025] S4: Attenuation semantic condition generation: Attenuation map is calculated based on CT images to represent the absorption characteristics of different tissues to radiation signals; Multi-scale CT feature maps are subjected to multi-level pooling to obtain fused features; Using CT image attenuation maps as query information, and associating them with the fused features through a cross-attention mechanism, multi-scale semantic enhancement conditional features are generated.

[0026] Furthermore, step S4 includes the following steps: S41: Based on the preprocessed CT image obtained in step S1, the CT attenuation map corresponding to the CT image is calculated according to the preset physical mapping relationship. The attenuation map is used to characterize the absorption characteristics of different tissue regions to radiation signals. S42: Perform multi-level pooling on the CT feature maps at different scales obtained in step S2 to obtain fused features containing multi-scale contextual information; S43: Using the CT attenuation map as guiding information, associate it with the fused features through a cross-attention mechanism to generate attenuation semantic enhancement features; S44: Fuse the attenuation semantic enhancement features obtained at different spatial scales to generate a unified semantic enhancement conditional feature.

[0027] S5: Multi-scale conditional fusion: The edge enhancement conditional map obtained in step S3, the semantic enhancement conditional features obtained in step S4, and the CT attenuation map are stitched together to construct conditional information at different spatial scales, which is used as the conditional input for the subsequent diffusion generation process.

[0028] Furthermore, step S5 includes the following steps: S51: The edge enhancement condition map obtained in step S3, the semantic enhancement condition features obtained in step S4, and the CT attenuation map are stitched together to construct condition information at different spatial scales.

[0029] S6: Conditional diffusion to generate PET images: Input the multi-scale conditional information constructed in step S5 into the diffusion generation model, dynamically modulate the noise removal process during the back diffusion process, and gradually generate the target PET image.

[0030] Furthermore, step S6 includes the following steps: S61: Input the multi-scale conditional information constructed in step S5 into the diffusion generation model; S62: During the reverse diffusion process of the diffusion model, the noise removal process is dynamically modulated according to the condition information of different spatial scales; S63: The target PET image is generated step by step through multiple back diffusion iterations.

[0031] Furthermore, in step S61, the Brownian Bridge Diffusion Model (BBDM) is adopted.

[0032] The Brownian bridge diffusion generation model consists of three parts: a forward diffusion process, a reverse diffusion generation network, and a conditional constraint mechanism. Its core structure revolves around the "constrained diffusion-denoising" process.

[0033] In terms of model structure, the Brownian bridge diffusion generation model uses the input image and the target image as the two endpoints of the diffusion process, and the UNet network as the denoising network. First, in the forward diffusion stage, the model starts with the target image and gradually introduces Gaussian noise to map the target image to an intermediate random state space. This process is not a completely unconstrained random diffusion, but is controlled by the Brownian bridge process, so that the noise evolution path is constrained by the distribution of both the input image and the target image.

[0034] In the back-diffusion stage, the Brownian bridge diffusion generation model employs a parameterized Unet denoising network as its core generation module. This network is typically based on an encoder-decoder structure and is used to predict noise components or intermediate states at different diffusion time steps. The denoising network receives the current diffusion state, time step information, and conditional information as input and progressively removes noise, causing the samples to converge from random states to the target image distribution along the Brownian bridge path.

[0035] Furthermore, the Brownian bridge diffusion generation model explicitly introduces a conditional constraint channel into its structure to incorporate information from the input image into the backdiffusion process, thereby establishing a direct correspondence between the input domain and the target domain during the generation process.

[0036] Through this structural design, the Brownian bridge diffusion generation model significantly reduces the problems of excessive randomness and structural inconsistency commonly found in traditional generation models, while maintaining the generation stability of diffusion models.

[0037] Example 2: In this embodiment, a CT-to-PET image generation method based on multi-scale conditional diffusion described in Example 1 was compared with several existing CT-to-PET image generation algorithms. The existing technologies used included image conversion methods based on generative adversarial networks and generation methods based on traditional conditional diffusion models. The comparative experimental results were evaluated using the absolute error between synthetic and real PET images (MAE, used to measure pixel-level deviation, particularly important in PET imaging as it directly relates to the quantitative accuracy of radioactive tracer uptake (SUV). Lower values ​​indicate better performance), structural similarity index (SSIM, used to evaluate the structural consistency between synthetic and real PET images; higher values ​​indicate better performance), peak signal-to-noise ratio (PSNR, used to evaluate the overall reconstruction quality between synthetic and real PET images; higher values ​​indicate better performance), and perceptual image loss (LPIPS, based on depth features to evaluate the perceptual distance between images, capturing fine-grained differences that traditional indices fail to reflect, making the evaluation of metabolic pattern reconstruction and visual realism more comprehensive. Lower values ​​indicate better performance). The results are shown in Table 1.

[0038] ; Table 1. Comparative experimental results.

[0039] Based on the comparative experimental results in Table 1, it can be seen that this method achieves better results in multiple evaluation indicators. It is particularly outstanding in terms of overall reconstruction quality and structural consistency between synthetic PET images and real PET images, capturing fine-grained differences, and pixel-level deviations. It can generate PET images that are more in line with the characteristics of medical imaging and has high practical value.

[0040] Example 3: In this embodiment, a CT-to-PET image generation system based on multi-scale conditional diffusion is provided. The system includes a data acquisition and preprocessing module, a multi-scale feature extraction module, a multi-scale edge information enhancement module, a CT attenuation map calculation module, a multi-scale semantic information extraction module, a feature fusion module, and a multi-scale conditional diffusion module. The data acquisition and preprocessing module is used to acquire the CT image to be processed and perform intensity normalization, size unification and format conversion; The multi-scale feature extraction module is used to input the processed CT image into the feature extraction network to generate several CT feature maps at different spatial scales; The multi-scale edge information enhancement module is used to extract and fuse edge features from CT feature maps at different spatial scales to obtain an edge enhancement condition map; the edge feature extraction includes edge feature extraction along the horizontal and vertical directions; The CT attenuation map calculation module is used to calculate the CT attenuation map based on the processed CT image; and to perform multi-level pooling processing on the CT feature maps of different spatial scales to obtain fused features; The multi-scale semantic information extraction module is used to use CT attenuation map as query information, perform feature association with fused features, and generate semantic enhancement conditional features. The feature fusion module is used to stitch together the edge enhancement condition map, semantic enhancement condition features and CT attenuation map to construct condition information at different spatial scales; The multi-scale conditional diffusion module is used to input conditional information at different spatial scales into the diffusion generation model to gradually generate the target PET image.

[0041] Furthermore, the feature extraction network has several feature extraction branches with different receptive fields or different spatial resolutions, used to generate several CT feature maps with different spatial scales.

[0042] Furthermore, the diffusion generation model adopts the Brownian bridge diffusion generation model; the Brownian bridge diffusion generation model is equipped with an encoder and a decoder, which takes the current diffusion state, time step information and condition information as input, and gradually removes noise so that the samples converge from the random state to the target image distribution along the Brownian bridge path; the Brownian bridge diffusion generation model is also equipped with a condition constraint channel for incorporating the information of the input image into the reverse diffusion process.

[0043] It should be noted that the system provided in the above embodiments is only an example of the division of the above functional modules. In practical applications, the above functions can be assigned to different functional modules as needed, that is, the internal structure can be divided into different functional modules to complete all or part of the functions described above. The system can be applied to a CT to PET image generation method based on multi-scale conditional diffusion in the above embodiments.

[0044] It should be understood that various parts of this application can be implemented using hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented using software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware, as in another embodiment, it can be implemented using any one or a combination of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.

[0045] The above embodiments are preferred embodiments of the present invention, but the embodiments of the present invention are not limited to the above embodiments. Any changes, modifications, substitutions, combinations, or simplifications made without departing from the spirit and principle of the present invention shall be considered equivalent substitutions and shall be included within the protection scope of the present invention.

Claims

1. A method for generating CT-to-PET images based on multi-scale conditional diffusion, characterized in that, Includes the following steps: Acquire the CT images to be processed and perform intensity normalization, size unification, and format conversion; The processed CT images are input into a feature extraction network to generate several CT feature maps at different spatial scales; Edge features are extracted and fused from CT feature maps at different spatial scales to obtain an edge enhancement condition map; the edge feature extraction includes edge feature extraction along the horizontal and vertical directions; CT attenuation maps are calculated based on the processed CT images; multi-level pooling is performed on the CT feature maps at different spatial scales to obtain fused features; Using CT attenuation maps as query information, feature association is performed with fused features to generate semantically enhanced conditional features; The edge enhancement condition map, semantic enhancement condition features, and CT attenuation map are stitched together to construct condition information at different spatial scales; Conditional information at different spatial scales is input into the diffusion generation model to gradually generate target PET images.

2. The CT-to-PET image generation method based on multi-scale conditional diffusion according to claim 1, characterized in that, The feature extraction network has several feature extraction branches with different receptive fields or different spatial resolutions, which are used to generate several CT feature maps with different spatial scales.

3. The CT-to-PET image generation method based on multi-scale conditional diffusion according to claim 1, characterized in that, The process involves extracting and fusing edge features from CT feature maps at different spatial scales to obtain an edge enhancement condition map, specifically as follows: For each CT feature map at a different spatial scale, gradient features are calculated along the horizontal and vertical directions respectively; The gradient features calculated in the horizontal and vertical directions are combined to generate an edge feature map that characterizes the anatomical boundary of the corresponding spatial scale. Edge feature maps obtained at different spatial scales are overlaid to generate a unified edge enhancement condition map.

4. The CT-to-PET image generation method based on multi-scale conditional diffusion according to claim 1, characterized in that, The step of calculating the CT attenuation map based on the processed CT image specifically involves: calculating the attenuation map corresponding to the CT image based on the processed CT image according to a preset physical mapping relationship. The process of using CT attenuation maps as query information and associating them with fused features to generate semantic enhancement conditional features involves: associating features through a cross-attention mechanism to obtain attenuation semantic enhancement features; and fusing attenuation semantic enhancement features obtained at different spatial scales to generate unified semantic enhancement conditional features.

5. The CT-to-PET image generation method based on multi-scale conditional diffusion according to claim 1, characterized in that, The step of inputting conditional information at different spatial scales into the diffusion generation model to gradually generate target PET images specifically involves: Conditional information at different spatial scales is input into the Brownian bridge diffusion generation model. The Brownian bridge diffusion generation model is equipped with an encoder and a decoder. It takes the current diffusion state, time step information and conditional information as input, and gradually removes noise so that the samples converge from the random state to the target image distribution along the Brownian bridge path. The Brownian bridge diffusion generation model is also equipped with a conditional constraint channel for incorporating the information of the input image into the reverse diffusion process. In the reverse diffusion process of the diffusion generation model, the noise removal process is dynamically modulated according to the condition information of different spatial scales. The target PET image is gradually generated through multiple backdiffusion iterations.

6. A CT-to-PET image generation system based on multi-scale conditional diffusion, characterized in that, It includes a data acquisition and preprocessing module, a multi-scale feature extraction module, a multi-scale edge information enhancement module, a CT attenuation map calculation module, a multi-scale semantic information extraction module, a feature fusion module, and a multi-scale conditional diffusion module; The data acquisition and preprocessing module is used to acquire the CT image to be processed and perform intensity normalization, size unification and format conversion; The multi-scale feature extraction module is used to input the processed CT image into the feature extraction network to generate several CT feature maps at different spatial scales; The multi-scale edge information enhancement module is used to extract and fuse edge features from CT feature maps at different spatial scales to obtain an edge enhancement condition map; the edge feature extraction includes edge feature extraction along the horizontal and vertical directions; The CT attenuation map calculation module is used to calculate the CT attenuation map based on the processed CT image; and to perform multi-level pooling processing on the CT feature maps of different spatial scales to obtain fused features; The multi-scale semantic information extraction module is used to use CT attenuation map as query information, perform feature association with fused features, and generate semantic enhancement conditional features. The feature fusion module is used to stitch together the edge enhancement condition map, semantic enhancement condition features and CT attenuation map to construct condition information at different spatial scales; The multi-scale conditional diffusion module is used to input conditional information at different spatial scales into the diffusion generation model to gradually generate the target PET image.

7. A CT-to-PET image generation system based on multi-scale conditional diffusion according to claim 6, characterized in that, The feature extraction network has several feature extraction branches with different receptive fields or different spatial resolutions, which are used to generate several CT feature maps with different spatial scales.

8. A CT-to-PET image generation system based on multi-scale conditional diffusion according to claim 6, characterized in that, The diffusion generation model adopts the Brownian bridge diffusion generation model. The Brownian bridge diffusion generation model is equipped with an encoder and a decoder. It takes the current diffusion state, time step information and condition information as input, and gradually removes noise so that the samples converge from the random state to the target image distribution along the Brownian bridge path. The Brownian bridge diffusion generation model is also equipped with a condition constraint channel for incorporating the information of the input image into the reverse diffusion process.

Images