Method and system for generating ccta images based on multi-phase ct-mpi images

Abstract

The application provides a method and system for generating CCTA images based on multi-phase CT-MPI images, the method comprising: acquiring multi-phase CT-MPI image and CCTA image data of a clinical patient; performing a preprocessing operation on the acquired multi-phase CT-MPI image data to obtain n-phase preprocessed CT-MPI images; the preprocessing comprises coronary artery region segmentation, phase number screening based on HU value, image registration and normalization processing; based on the n-phase preprocessed CT-MPI images, a conditional diffusion generative adversarial network is used to generate corresponding n-phase CCTA images; and a weighted fusion strategy is adopted to perform pixel-level fusion on the n-phase CCTA images to form a high-quality CCTA image.

Description

Technical Field

[0001] This invention relates to the fields of medical image processing and deep learning, specifically to a method and system for generating high-quality CT angiography (CCTA) images using multi-phase myocardial CT perfusion imaging (CT-MPI or CTP) combined with diffusion generative adversarial networks, and also provides a corresponding computer terminal and computer-readable storage medium. Background Technology

[0002] Currently, cardiovascular disease is the leading cause of death worldwide, with coronary artery disease (CAD) being one of the leading causes of significant cardiovascular morbidity and mortality. Comprehensive imaging examinations of its morphology and function are crucial for accurate diagnosis and assessment of CAD. Coronary computed tomography angiography (CCTA) and computed tomography myocardial perfusion imaging (CT-MPI or CTP) are two major non-invasive examination methods, each with significant advantages in coronary anatomy and function analysis.

[0003] The combined use of CCTA and CT-MPI can improve the specificity for detecting significant hemodynamic stenosis and enhance the specificity and overall accuracy for detecting areas of coronary artery disease or myocardial infarction with restricted blood flow. However, the high total effective radiation dose and contrast agent burden limit the widespread clinical application of combined CCTA and CT-MPI examinations. Currently, a satisfactory non-invasive, one-stop cardiovascular examination method is still lacking in clinical diagnosis.

[0004] CT-MPI contains more dynamic information than CCTA. Therefore, developing a technique for generating CCTA using CT-MPI is a promising solution. Current research has explored methods for generating CCTA using CT-MPI. In one study (DAI X, YU M, PAN J, et al. Image quality and diagnostic accuracy of coronary CT angiography derived from low-dose dynamic CT myocardial perfusion: a feasibility study with comparison to invasive coronary angiography[J]. European Radiology, 2019, 29: 4349-4356.), researchers used the ADMIRE algorithm to place a circular region of interest at the ascending aorta location in the CT-MPI image, evaluated the time decay curve of the CT-MPI, and selected the first-phase image corresponding to the peak phase of the curve for CCTA generation. This study, along with other studies using CT-MPI to generate CCTA, employed the same algorithm: the iterative reconstruction ADMIRE algorithm. However, this algorithm suffers from problems such as long computation time and high computational resource consumption. Furthermore, because CT-MPI is a dynamic, continuous scanning imaging process, each patient's CT-MPI data contains multiple phases of images. These studies only selected single-phase CT-MPI data for CCTA generation, ignoring information contained in other phases of CT-MPI. This may lead to the omission of important vascular dynamic information, reducing the quality and accuracy of CCTA generation. Moreover, iterative reconstruction methods require CT-MPI projection domain data as input. During research, due to commercial privacy and technical limitations, CT-MPI projection domain data is often unavailable. In contrast, image processing algorithms can directly process CT-MPI images reconstructed from the projection domain without accessing the projection domain data, thereby reducing the difficulty and cost of data acquisition and providing greater flexibility and convenience for research.

[0005] In medical image analysis, deep learning-assisted medical diagnosis has become one of its important applications. Deep learning systems can automatically analyze and interpret medical images, which greatly improves the accuracy and efficiency of diagnosis. Currently, research on deep learning-based medical image translation has achieved some preliminary results. For example, studies on magnetic resonance imaging (MRI) enhancement (JAYACHANDRAN PREETHA C, MEREDIG H, BRUGNARA G, et al. Deep-learning-based synthesis of post-contrast T1-weighted MRI for tumourresponse assessment in neuro-oncology: a multicentre, retrospective cohortstudy[J]. The Lancet Digital Health, 2021, 3(12): e784-e794.) and adversarial network-based image translation (Karim Armanious , Chenming Yang , et al. MedGAN: Medical ImageTranslation using GANs) have all demonstrated the potential of deep learning in image translation.

[0006] Despite significant advancements in medical image analysis and enhancement using deep learning, generating CCTA from CT-MPI remains a formidable challenge. The main issues include: matching CT-MPI and CTA images (inconsistent Z-axis, slice count, and overall workflow); difficulty in achieving a one-to-one match between single-phase CT-MPI and CTA images due to inconsistencies in slice count and spacing; and the common practice of processing single-phase CT-MPI data for CCTA generation, often neglecting multi-phase data from the CT-MPI sequence, which may lead to the omission of crucial dynamic information and reduced image quality and accuracy. Summary of the Invention

[0007] To address the aforementioned shortcomings in the prior art, this invention provides a method and system for generating CCTA through multi-phase CT-MPI based on diffusion generative adversarial networks, along with a corresponding computer terminal and computer-readable storage medium.

[0008] According to one aspect of the present invention, a method for generating CCTA images from multi-phase CT-MPI images based on diffusion generative adversarial networks is provided, comprising: Acquire multi-phase CT-MPI and CCTA image data of clinical patients; wherein, the multi-phase CT-MPI images are multiple images obtained by performing multiple CT scans of the heart over a period of time after the patient is injected with contrast agent; the CCTA images are coronary artery images obtained by scanning the heart when the concentration of coronary contrast agent reaches its peak after the patient is injected with contrast agent; the spatial resolution and radiation dose of any single phase CT-MPI image are lower than those of the CCTA image; The acquired multi-phase CT-MPI image data is preprocessed to obtain n preprocessed CT-MPI images; the preprocessing includes coronary artery region segmentation, phase selection based on HU value, image registration and normalization processing; Based on the n-stage preprocessed CT-MPI images, a conditional diffusion generative adversarial network is used to generate corresponding n-stage CCTA images. A weighted fusion strategy is adopted to perform pixel-level fusion of the n-period CCTA images to form a high-quality CCTA image.

[0009] Preferably, the preprocessing includes: The coronary artery region of the multi-phase CT-MPI images is segmented by combining a 3D slicer and an artificial intelligence network; Based on the HU values ​​of the coronary artery region in the multi-phase CT-MPI images, the n-phase CT-MPI images with the highest HU values ​​were selected; Based on the scanning parameter information of the CT-MPI image and the CCTA image, the n-stage CT-MPI image and the CCTA image are scaled and aligned on the Z-axis; Based on the segmented coronary artery region, the n-phase CT-MPI images are cropped to remove the background region.

[0010] Preferably, the segmentation of the coronary artery region in the multi-phase CT-MPI images by combining a 3D slicer and an artificial intelligence network includes: The multi-phase CT-MPI images of the patient were manually annotated using 3D slicer software. Using an artificial intelligence-based algorithm, a model capable of segmenting the coronary artery region in CT-MPI images was trained based on existing labeled CT-MPI image data. The segmentation results obtained from the trained model are checked and corrected.

[0011] Preferably, the n-phase CT-MPI image with the highest HU value among the coronary artery regions of the multi-phase CT-MPI images is selected, including: The average HU value of the coronary artery region is used as the evaluation index from the multi-phase CT-MPI images, and the nth phase with the highest value is selected as the subsequent input.

[0012] Preferably, the training process of the conditional diffusion generative adversarial network includes: The CCTA image is noise-added through multiple forward diffusion processes to obtain a noisy CCTA. b image; The multi-phase CT-MPI images were compared with noisy CCTA images as a condition. b The images are input into the generator model to generate CCTA' images with any number of slices. The CCTA image or the CCTA image is compared with a noisy CCTA image as a condition. b The image input discriminator model is trained adversarially.

[0013] Preferably, the step of adding noise to the CCTA image through multiple forward diffusion processes includes: Perform on the CCTA image The diffusion process involves several consecutive diffusion steps, modeled as a parameterized Markov chain. In the t-th step of the diffusion process, Gaussian noise is added to the image, forming a chain composed of... to Forward noise addition process Its definition is: ; in, For the first The variance of each diffusion step, It is the identity matrix. This indicates a normal distribution.

[0014] Preferably, the training process of the generator model includes: Using a pseudo-3D training method, adjacent Using CT-MPI slices as input, the output is adjacent slices. One CCTA slice, among which ; A feature extraction network was used to analyze CT-MPI images and noisy CCTA images. b The images are subjected to feature extraction, and the features are fused and then upsampled to generate a CCTA' image; the loss function of the generator is: ; in, MSE = .

[0015] Preferably, the training process of the discriminator model includes: Noisy CCTA b The image serves as a condition, with either the generated CCTA image or a real CCTA image as input; Wasserstein distance is used as a measure of divergence to distinguish between generated and real images, guiding the generator to produce realistic blood vessel details. This loss function is specifically designed to... ; This means the discriminator must be sufficiently smooth, and the output is passed through the softplus function to obtain the discriminator's loss function: .

[0016] Preferably, the step of employing a weighted fusion strategy to perform pixel-level fusion of the n-period CCTA images to form a high-quality CCTA image includes: Based on the HU values ​​of the coronary artery region in each phase of CT-MPI images, the weight parameters corresponding to each phase of images are obtained. Based on the weight parameters corresponding to each period's images, the generated CCTA images from each period are fused pixel-by-pixel to obtain the final high-quality CCTA image. The expression for this fusion process is as follows: ; in, This indicates the location of the final fused CCTA image. pixel values, Indicates the first At what point in time is the location pixel values, Indicates the first Weight parameters for each time point.

[0017] According to another aspect of the present invention, a system for generating CCTA images from multi-phase CT-MPI images based on diffusion generative adversarial networks is provided, comprising: The data acquisition module is used to acquire multi-phase CT-MPI and CCTA image data of clinical patients. The multi-phase CT-MPI images are multiple images obtained from repeated CT scans of the heart over a period of time after contrast agent injection. The CCTA images are coronary artery images obtained by scanning the heart when the coronary contrast agent concentration reaches its peak after contrast agent injection. The spatial resolution and radiation dose of any single-phase CT-MPI image are lower than those of the CCTA image. The data preprocessing module is used to preprocess the acquired multi-phase CT-MPI image data to obtain n preprocessed CT-MPI images; the preprocessing includes coronary artery region segmentation, phase selection based on HU value, image registration and normalization processing; The image generation module is used to generate corresponding n-stage CCTA images based on the n-stage preprocessed CT-MPI images using a conditional diffusion generative adversarial network. The weighted fusion module is used to perform pixel-level fusion of the n-period CCTA images to form a high-quality CCTA image by employing a weighted fusion strategy.

[0018] According to a third aspect of the present invention, a computer terminal is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it can be used to perform the method described in any one of the above-described inventions, or to run the system described in any one of the above-described inventions.

[0019] According to a fourth aspect of the present invention, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, can be used to perform the method described in any one of the preceding claims of the present invention, or to run the system described in any one of the preceding claims of the present invention.

[0020] By adopting the above technical solution, the present invention has at least one of the following beneficial effects compared with the prior art: The present invention provides a method and system for generating high-quality CT angiography based on CT perfusion imaging, which can effectively utilize the feature information of multi-phase CT perfusion imaging to generate high-quality CCTA images. This method significantly improves the accuracy of CCTA generation, providing doctors with medical images that are closer to real CT angiography.

[0021] This invention provides a method and system for generating high-quality CT angiography images based on CT perfusion imaging, proposing an end-to-end CT angiography generation system architecture. This architecture leverages the advantages of deep learning to achieve efficient image translation and generation, producing corresponding CT angiography images from CT perfusion imaging, exhibiting both robustness and high quality. This intelligent medical auxiliary diagnostic system can effectively handle information acquisition, fusion, and diagnostic decision-making in different vascular lesion scenarios, promoting the advancement of integrated non-invasive cardiovascular examination. Attached Figure Description

[0022] Other features, objects, and advantages of the present invention will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings: Figure 1This is a schematic diagram illustrating an embodiment of the present invention for generating CT angiography images based on multi-phase CT perfusion imaging using diffusion generative adversarial networks. Figure 2 This is a schematic diagram of the CT perfusion imaging preprocessing workflow provided in an embodiment of the present invention; Figure 3 A schematic diagram of a diffusion-generative adversarial network model provided in an embodiment of the present invention; Figure 4 A schematic diagram of a diffusion-generative adversarial network model provided in an embodiment of the present invention; Figure 5 This is a schematic diagram showing the alignment of CT-MPI and CCTA images on the Z-axis according to an embodiment of the present invention; Figure 6 This is a schematic diagram illustrating forward diffusion noise addition and reverse denoising of an image according to an embodiment of the present invention; Figure 7 This is a schematic diagram of a multi-phase weighted fusion strategy provided in an embodiment of the present invention; Figure 8 Quality assessment of CCTA images generated in embodiments of the present invention; Figure 9 The diagnostic performance of CCTA images generated in this embodiment of the invention on lesions. Detailed Implementation

[0023] The embodiments of the present invention are described in detail below: These embodiments are implemented based on the technical solution of the present invention, and provide detailed implementation methods and specific operation processes. It should be noted that those skilled in the art can make several modifications and improvements without departing from the concept of the present invention, and these all fall within the protection scope of the present invention.

[0024] One embodiment of the present invention provides a method for generating CT angiography images based on multi-phase CT perfusion imaging using diffusion generative adversarial networks. This method utilizes the characteristics of multi-phase CT perfusion imaging, combined with image generation technology and weighted fusion strategy, to generate high-quality CT angiography images, avoiding the problems of high radiation dose and contrast agent burden, effectively shortening the examination process, and assisting doctors in the diagnosis of coronary artery diseases.

[0025] like Figure 1 As shown, this embodiment provides a method for generating CT angiography using multi-phase CT perfusion imaging based on diffusion generative adversarial networks. This method may include the following operations: S100 acquires CT perfusion imaging and CT angiography image data of patients in clinical practice; S200 performs preprocessing operations on the acquired CT perfusion images, including partial segmentation of the coronary artery region of the CT perfusion images, period selection based on the HU value of the coronary artery region of the CT perfusion images, image registration of CT perfusion images and CT angiography, center cropping, normalization and other operations. S300, based on multi-phase preprocessed CT perfusion imaging images, constructs a diffusion-driven adversarial generative framework: forward diffusion is performed on real CT angiography images to obtain noisy conditional inputs, which are then input into the generator along with the multi-phase CT perfusion imaging images to generate corresponding CT angiography images; the generated results and real images are simultaneously input into the discriminator for adversarial training. After iterative optimization, a high-quality CCTA generation model can be obtained.

[0026] The S400 employs a weighted fusion strategy to perform pixel-level fusion of the results generated from multi-phase CT perfusion imaging, resulting in high-quality CT angiography imaging.

[0027] In some preferred embodiments, S200 above performs a preprocessing operation on the acquired CT perfusion images, including: S201, Combining 3D slicer annotation and intelligent network segmentation, the coronary artery region is segmented from the CT perfusion imaging data: S2011 utilizes the Segment Editor module of the 3D Slicer software for blood vessel segmentation. First, a threshold is used to limit the range of CT values, and then the blood vessel contour is manually marked using the paint tool.

[0028] S2012 uses a machine learning-based algorithm to train a model capable of segmenting the coronary artery region in CT perfusion imaging based on existing labeled CT perfusion imaging data. S2013, the results obtained from model segmentation are checked and optimized again.

[0029] In some preferred embodiments, the above-mentioned S200, which involves period selection for CT perfusion imaging, includes: S202 uses the HU value of the coronary artery region as the evaluation index of the quality of CT perfusion imaging from all periods of CT perfusion imaging data, and selects the highest N period as the subsequent input.

[0030] S203, based on the Pixel Spacing and Image Position Patient information in the image header file, scale and align the CT perfusion imaging and CT angiography images on the Z-axis; S204, based on the segmented coronary artery region, the CT perfusion image is cropped to remove the background region; In some preferred embodiments, S300 above, based on multi-stage processed CT perfusion imaging images, generates CT angiography images using a diffusion-based generative adversarial network, including: S301, execute the raw CCTA image data The diffusion process involves a series of diffusion steps, modeled as a parameterized Markov chain. In the t-th step of the diffusion process, Gaussian noise is added to the image, forming a chain composed of... to Forward noise addition process Its definition is as follows: .in, For the first The variance of each diffusion step, It is the identity matrix. This indicates a normal distribution.

[0031] S302, employs a pseudo-3D training method, which combines adjacent... Using CTP slices as input, the output is adjacent slices. One CCTA slice, among which (The values ​​of m and k are related to the Z-axis spatial resolution of CTP and CCTA images, see reference) Figure 5 CCTA images have higher Z-axis resolution and more slice layers; a feature extraction network is used to compare CTP images with noisy CCTA images. b Feature extraction is performed separately, and the features are fused and then upsampled to generate a CCTA image; the loss function of the generator is the sum of the adversarial loss and the mean squared error.

[0032] S3021 uses a pseudo-3D training method to train adjacent... Using CTP slices as input, the output is adjacent slices. One CCTA slice, among which During training, the spatial intersection of CTP and CCTA is selected as the input-output range of the network, and the structural features of CCTA in the intermediate layer are inferred using information from adjacent CTP layers. For overlapping slices generated during multiple inference processes, pixel-wise averaging is used for fusion to improve the consistency of the generated results.

[0033] S3022, using a feature extraction network to process CTP images and noisy conditional CCTA images respectively. b Encoding is performed to extract its low-dimensional feature representation. The two types of features are concatenated or weighted and fused through a feature fusion module, and then the predicted CCTA image CCTA′ is recovered through a decoder based on a U-Net variant structure. The generator's loss function consists of reconstruction loss (MSE) and adversarial loss (GAN Loss). The mean squared error is... ,in: Image size; For real CCTA pixels To generate CCTA' pixels. Adversarial loss. ,in: For discriminator To reconstruct the distribution in reverse for the diffusion model; The distribution is a forward-noisy distribution. The generator's total loss function is... ; S303, the generated image CCTA' and the real image CCTA are input into the discriminator for adversarial training. Wasserstein distance is used as a metric for divergence to distinguish between the generated and real images, thus guiding the generator to produce realistic vascular details. This loss function specifically... This means the discriminator must be sufficiently smooth, and the output is passed through the oftplus function to obtain the discriminator's loss function. ; In some preferred embodiments, the above-mentioned S400 employs a weighted fusion strategy to perform pixel-level fusion of the results generated from multi-phase CT perfusion imaging to form high-quality CT angiography imaging, including: S401, within the segmented coronary artery region, assesses the quality of CT perfusion imaging at each time point based on Hounsfield units (HU values). By analyzing these HU values, corresponding weighting parameters can be generated for each time point. The weighting parameters can be set based on the mean, variance, or other statistical characteristics of the HU values ​​to reflect the reliability and quality of perfusion imaging at each time point.

[0034] S402, such as Figure 7 As shown, based on the calculated weight parameters, pixel-level fusion is performed on the CT perfusion imaging results at each time point according to the weights. Specifically, for each pixel, a weighted average is calculated using the HU value of the corresponding pixel at each time point and the weight parameters to obtain the final pixel value. This process can be represented by the following formula: in, This indicates the final fused CT angiography image at the location pixel values, Indicates the first At what point in time is the location pixel values, Indicates the first Weight parameters for each time point.

[0035] The technical solution provided by the above embodiments of the present invention will be further described in detail below with reference to a preferred embodiment.

[0036] The preferred embodiment of the method for generating CT angiography images based on diffusion generative adversarial networks through multi-phase CT perfusion imaging includes the following steps: Step 1: Acquire CT perfusion imaging and CT angiography image data from clinical patients. The data consists of paired CTP-CCTA scans collected at Shanghai Sixth People's Hospital between 2017 and 2024 using Siemens equipment, and saved in Medical Digital Imaging and Communication (DICOM) format. Each CTP data set comprises 520 to 780 slices, containing 10 to 15 phases of CTP data, with 52 slices per phase. Each CCTA data set comprises 195 to 306 slices. The slice size for both CTP and CCTA is 512 × 512.

[0037] During CTP acquisition, all patients received 50 ml of contrast agent at a rate of 6 ml / s, followed by flushing with 40 ml of normal saline using a dual-barrel dynamic injector. CTP acquisition was performed 5 seconds after contrast agent injection. In all patients, end-systole was set to use shuttle mode for dynamic acquisition with a coverage area of ​​10.5 cm for complete imaging of the entire left ventricle. Scans were initiated every one or two seconds based on the patient's heart rate, acquiring 10 to 15 phases of CTP sequence within a fixed 31 seconds. The CTP acquisition parameters during this process were as follows: collimation = 96 × 0.6 mm, tube voltage = 70 kVp, rotation time = 250 ms, effective current = 250 mAs, reconstructed slice thickness = 0.75 mm, and reconstructed slice interval = 2 mm.

[0038] During CCTA acquisition, all patients received 20 ml of contrast agent at a rate of 4.55 ml / s, followed by flushing with 2040 ml of normal saline using a dual-barreled dynamic injector. Prospective, sequential ECG-triggered acquisition was performed on patients with a heart rate ≥ 70 bpm. CCTA acquisition parameters were as follows: tube voltage = 100 kVp, effective current = 320 mAs, reconstructed slice thickness = 0.75 mm, reconstructed slice interval = 0.5 mm, rotation time = 250 ms.

[0039] Step 2: Perform preprocessing operations on the acquired CT perfusion images, including partial segmentation of the coronary artery region of the CT perfusion images, period selection based on the HU value of the coronary artery region of the CT perfusion images, image registration of CT perfusion images and CT angiography, center cropping, normalization and other operations.

[0040] Step 3: Based on the CT perfusion imaging images processed in Step 2, the input and output networks are designed according to the spatial relationship between CT perfusion imaging and CT angiography, using the image generation technology of deep learning models, so as to generate corresponding CT angiography images.

[0041] Step 4: Using a weighted fusion strategy, based on the HU value of the cut coronary artery region in Step 2, the results generated by multi-phase CT perfusion imaging are fused pixel-level according to weight to form high-quality CT angiography imaging.

[0042] like Figure 2 , Figure 3 As shown, in a preferred embodiment, step 2 further includes the following steps: Step 2.1: Combine 3D slicer annotation and intelligent network segmentation to segment the coronary artery region of the CT perfusion imaging data, specifically including the three coronary arteries: the left anterior descending artery (LAD), the left annular branch (LCx), and the right coronary artery (RCA). Step 2.1.1: Blood vessel segmentation was performed using the Segment Editor module of the 3D Slicer software. Since the CT values ​​of blood vessels are often higher than those of surrounding tissues, the Threshold tool was first used to limit the range of CT values, ensuring coverage of the blood vessel area while minimizing interference from other surrounding tissues. Next, the blood vessel outline was manually marked using the Paint tool. The segmentation results were refined and filled by adjusting the brush size and shape to ensure accuracy. Finally, the segmentation results were checked and adjusted, and the output was saved in .nrrd format.

[0043] Step 2.1.2: Using a machine learning-based algorithm, a Unet network with a symmetric encoder and decoder structure is employed. The encoder extracts image features, which are then passed to the decoder, which outputs the segmentation result, achieving pixel-level image segmentation. Based on existing labeled CT perfusion imaging data, a model capable of segmenting the coronary artery region in CT perfusion imaging is trained using binary cross-entropy loss as the loss function; specifically, .in: It is the probability value at each pixel predicted by the model (obtained through the sigmoid function). It is the ground truth label, with a value of 0 or 1. It represents the total number of pixels.

[0044] Step 2.1.3: The clinician re-examines and corrects the results obtained from the model segmentation; Step 2.2: From all phases of CT perfusion imaging data, the HU value of the coronary artery region is used as an indicator of the quality of CT perfusion imaging. The phase with the highest N value is selected as the subsequent input. According to the CT perfusion imaging coronary artery HU value phase curve, N is generally set to 3 or 4. Step 2.3: Ensuring proper alignment between CTP and CCTA data is crucial during training, requiring appropriate translation and scaling of the CTP data. First, we retrieve the Pixel Spacing and Image Position Patient information from the DICOM files of CTP and CCTA from the same patient. Pixel Spacing represents the size of each pixel in physical space, while Image Position Patient represents the image's position in physical space. Based on Image Position Patient, we calculate the spatial offset between CTP and CCTA. Based on Pixel Spacing, we calculate the scaling factor for both pixels, and then determine the scale of translation and scaling of the CTP data needed to achieve CTP-CCTA alignment.

[0045] Step 2.4: Based on the segmented coronary artery region, the CT perfusion image is cropped to remove the background region; See Figure 3 , Figure 4 , Figure 5 , Figure 6 In a preferred embodiment, step 3 further includes the following steps: Step 3.1: Design the input / output network. Since the number of slices for CTP and CCTA is not consistent in the sample data, and the location range of the slices often differs, we choose the intersection of their spatial locations as the input / output slice range of the network. For example... Figure 5 Since the spacing between adjacent slices in CCTA is 0.5mm and that in CTP is 2.0mm, after removing redundant slices, the ratio of slice count between the two is 4:1. To avoid increasing the complexity of the problem by using a 3D network, we adopted a pseudo-3D training method, such as... Figure 5As shown, in network training, each round simultaneously inputs three adjacent CTP slices and outputs five adjacent CCTA slices. In this training method, the input of each round overlaps with the input of the previous round by two slices, and the output overlaps with the output of the previous round by one slice. This special input-output design has two advantages. First, this design ensures that the CCTA slice output in each round is exactly in the middle of the three adjacent CTP slices, thus ensuring the proper positional relationship between CCTA and CTP, improving the matching accuracy of the generated image. Second, the overlapping CCTA slice is located in the middle of the two preceding and following CTP slices. Since the middle CCTA slice is relatively far from the two adjacent CTP slices and has low similarity, it is the most difficult to generate. Therefore, it is prioritized for generation, and the method of averaging the two generation processes is used to improve the accuracy of generating this slice.

[0046] Step 3.2: Using the single-stage CT perfusion imaging processed in Step 2, and following the input-output design described in Step 3.1, generate CT angiography images using a diffusion-based generative adversarial network model. (See [link to relevant documentation]). Figure 6 .

[0047] Step 3.2.1: Let the real CCTA image be... Define the length as The forward diffusion (noise addition) process is a parameterized Markov chain: The conditional distribution for each step is a Gaussian distribution: and define but about The marginal distribution can be written in closed form (for easier training sampling): here For preset noise scheduling (e.g., linear or cosine scheduling), the typical value range is... , The value is typically 100–1000 (preferably 100–400).

[0048] Step 3.2.2: Inverse (denoising) model with parameters Modeling conditional probability distributions Usually Assuming a Gaussian distribution: in Represents conditional information (e.g., multi-period CTP features and noisy CCTA). (Itself), the generator needs to learn (can be parameterized as neural network output) and optional variance term .

[0049] A commonly used equivalent parameterization for prediction noise Format (commonly used in DDPM): In this invention, the generator receives joint conditions. It outputs a denoised prediction, thus obtaining the result according to the above formula. .

[0050] The generator's loss function consists of reconstruction loss (MSE) and adversarial loss (GAN Loss). The mean squared error is one of the loss functions. ,in: Image size; For real CCTA pixels To generate CCTA' pixels. Adversarial loss. ,in: For discriminator To reconstruct the distribution in reverse for the diffusion model; The distribution is a forward-noisy distribution. The generator's total loss function is... ; Step 3.2.3: Input the generated image CCTA' and the real image CCTA into the discriminator for adversarial training. The discriminator structure uses Wasserstein distance as a measure of divergence to distinguish between the generated image and the real image, thus guiding the generator to generate realistic vascular details. - Lipschitz requires stable training using Wasserstein divergence. The discriminator loss is defined as: Where: the first term is the average output of the discriminator to the generated samples (expected generated score), and the second term is the average output of the discriminator to the real samples (expected real score). The Wasserstein objective is to maximize the difference between the real sample score and the generated sample score.

[0051] In a preferred embodiment, step 4 further includes the following steps: Step 4.1: Within the segmented coronary artery region, the quality of CT perfusion imaging at each time point is evaluated based on Hounsfield units (HU values). HU values ​​reflect the radioactivity of the tissue, and the HU values ​​for CT perfusion imaging within the coronary artery region may vary at different time points. By analyzing these HU values, corresponding weighting parameters can be generated for each time point. The weighting parameters can be set based on the mean, variance, or other statistical characteristics of the HU values ​​to reflect the reliability and quality of perfusion imaging at each time point.

[0052] Step 4.2: As Figure 7 As shown, based on the calculated weight parameters, pixel-level fusion is performed on each of the single-stage CT perfusion imaging results obtained through the above steps. Specifically, for each pixel, a weighted average is calculated using the HU value of the corresponding pixel at each time point and the weight parameters to obtain the final pixel value, thus obtaining the final CT perfusion imaging. This process can be represented by the following formula: in, This indicates the final fused CT angiography image at the location pixel values, Indicates the first At what point in time is the location pixel values, Indicates the first Weight parameters for each time point.

[0053] See Figure 8 , Figure 9 This is a synthetic CCTA image effect generated using the method of the embodiments of the present invention.

[0054] This study included 840 eligible patients (mean age ± standard deviation: 61.91 ± 12 years, 566 males), with 640 patients in the training / validation set, 120 in the internal test set, and 80 in the external test set. The CCTA case training / validation set contained 1920 vessels from 640 patients (mean age 61.84 ± 13 years, 418 males); the test sets consisted of 360 vessels from 120 patients in the internal test set (mean age 62.01 ± 12 years, 90 males) and 240 vessels from 80 patients in the external test set (mean age 62.30 ± 8 years, 58 males). There were no significant differences in patient characteristics between the training set and the two test sets (P>0.05).

[0055] In quantitative evaluations on both internal and external test sets, the proposed model outperformed StyleGAN and RegGAN. Overall, the NMAE of Syn-CCTA images (i.e., synthetic CCTA images generated using the method of this invention) in the internal test set was 0.035, the mean PSNR was 32.251 dB, and the mean SSIM was 0.939; in the external test set, the NMAE of Syn-CCTA images was 0.044, the mean PSNR was 30.198 dB, and the mean SSIM was 0.918. CCTA images from all 120 patients in the internal test dataset and 80 patients in the external test dataset were successfully reconstructed from CT-MPI data, and the image quality met clinical diagnostic requirements. Subjective evaluation results showed that the average subjective scores for the raw axial image quality of the Syn-CCTA group and the real CCTA group were comparable [Internal dataset: 2.285±0.443 vs 2.700±0.559, P=0.051; External dataset: 2.750±0.436 vs 2.600±0.493, P=0.043]. The high-quality score (score=3) rate for the Syn-CCTA group was 75% (90 / 120) in the internal dataset and 60% (48 / 80) in the external dataset; while the high-quality score rate for the real CCTA group was 85% (102 / 120) in the internal dataset and 75% (60 / 80) in the external dataset. See details for the score distribution. Figure 8 A in the middle. Figure 8 Figure B shows a comparison between two sets of original and post-processed images, demonstrating that the Syn-CCTA images and the real CCTA images have comparable reconstruction quality for distal small vessels.

[0056] The diagnostic efficacy of Syn-CCTA images based on CT-MPI for stenosis and high-risk plaque features was highly consistent with that based on real CCTA in both internal and external datasets. When using traditional CCTA diagnostic results as a reference standard (stenosis threshold of 50%), the AUC values ​​for stenosis diagnosis in the Syn-CCTA group ranged from 0.941 to 0.989 in the internal dataset and from 0.880 to 0.972 in the external dataset in vascular level analysis. This study included DSA images from 41 patients (involving 123 vessels). Using digital subtraction angiography (DSA) as a reference, Syn-CCTA performed similarly to CCTA in both patient-level and vascular level analysis at stenosis thresholds of 50% and 75% (all P values ​​> 0.05). In the internal dataset, 10 patients were diagnosed with high-risk plaque features on CCTA images, and all lesions were clearly visible on Syn-CCTA images; no patients in the external dataset showed high-risk plaque features. Figure 9 It shows a detailed image of a patient with high-risk characteristics.

[0057] One embodiment of the present invention provides a system for generating high-quality CT angiography images based on CT perfusion imaging, the system comprising: The data acquisition module is used to acquire patient CT perfusion imaging and CT angiography image data.

[0058] The data preprocessing module is used to perform operations on the CT perfusion imaging data, including partial segmentation of the coronary artery region of CT perfusion imaging, period filtering based on the HU value of the coronary artery region of CT perfusion imaging, image registration of CT perfusion imaging and CT angiography, center cropping, normalization, etc.

[0059] A medical image generation module is provided to provide a medical image generation model. The model is trained using the CT perfusion imaging and CT angiography data. The model generates CT angiography images based on the processed CT perfusion imaging data in a single phase.

[0060] The weighted fusion module is used to perform pixel-level fusion of the generated results of multi-phase CT perfusion imaging based on the HU value of the cut coronary artery region, using a weighted fusion strategy to obtain the final high-quality CT angiography image.

[0061] In some preferred embodiments, the data preprocessing module utilizes the SegmentEditor module of the 3D Slicer software for vessel segmentation. First, a threshold is used to limit the CT value range, and then the vessel contours are manually annotated using the paint tool. Next, a machine learning-based algorithm is used to train a model capable of segmenting the coronary artery region in CT perfusion imaging based on existing annotated CT perfusion imaging data. Finally, the segmentation results are checked and optimized again. From all periods of the CT perfusion imaging data, the HU value of the coronary artery region is used as an evaluation index for the quality of CT perfusion imaging, and the highest N-period is selected as subsequent input. Based on the Pixel Spacing and Image Position Patient information in the image header file, the CT perfusion imaging and CT angiography images are scaled and aligned on the Z-axis. Based on the segmented coronary artery region, the CT perfusion imaging is cropped to remove background areas. In some preferred embodiments, the aforementioned medical image generation module employs a unique input-output design and a pseudo-3D training method. Specifically, during network training, each round simultaneously inputs three adjacent CTP slices and outputs five adjacent CCTA slices. Under this training method, the input of each round overlaps with the input of the previous round by two slices, and the output overlaps with the output of the previous round by one slice, thereby improving the accuracy of the generated slices.

[0062] In some preferred embodiments, the weighted fusion module evaluates the quality of CT perfusion imaging at each time point within the segmented coronary artery region based on Hounsfield units (HU values). By analyzing these HU values, corresponding weight parameters can be generated for each phase of CT perfusion imaging. Based on the calculated weight parameters, pixel-level fusion is performed on the CT perfusion imaging results at each time point according to the weights. Specifically, for each pixel, a weighted average is calculated using the HU value and weight parameters of the corresponding pixel at each time point to obtain the final pixel value. This process can be represented by the following formula: in, This indicates the final fused CT angiography image at the location pixel values, Indicates the first At what point in time is the location pixel values, Indicates the first Weight parameters for each time point.

[0063] It should be noted that the steps in the method provided by the present invention can be implemented using corresponding modules, devices, units, etc. in the system. Those skilled in the art can refer to the technical solution of the method to realize the composition of the system. That is, the embodiments in the method can be understood as preferred examples for building the system, and will not be elaborated here.

[0064] An embodiment of the present invention provides a computer terminal, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it can be used to perform any of the methods in the above embodiments of the present invention, or to run any of the systems in the above embodiments of the present invention.

[0065] Optionally, the memory is used to store programs; the memory may include volatile memory, such as random-access memory (RAM), such as static random-access memory (SRAM), double data rate synchronous dynamic random-access memory (DDR SDRAM), etc.; the memory may also include non-volatile memory, such as flash memory. The memory is used to store computer programs (such as application programs, functional modules, etc. that implement the above methods), computer instructions, etc., and the aforementioned computer programs, computer instructions, etc., can be partitioned and stored in one or more memories. Furthermore, the aforementioned computer programs, computer instructions, data, etc., can be accessed by the processor.

[0066] The aforementioned computer programs, computer instructions, etc., can be stored in partitions within one or more memory locations. Furthermore, the aforementioned computer programs, computer instructions, data, etc., can be accessed by a processor.

[0067] A processor is used to execute computer programs stored in memory to implement the various steps of the methods or various modules of the systems involved in the above embodiments. For details, please refer to the relevant descriptions in the preceding method and system embodiments.

[0068] The processor and memory can be separate structures or integrated structures. When the processor and memory are separate structures, they can be coupled together via a bus.

[0069] An embodiment of the present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, can be used to perform the method of any of the above embodiments of the present invention, or to run the system of any of the above embodiments of the present invention.

[0070] The method and system for generating high-quality CT angiography images based on CT perfusion imaging provided in the above embodiments of the present invention are based on deep learning. Targeting CT perfusion imaging, it utilizes its multi-phase features to sequentially complete preprocessing, image generation model training, and weighted fusion, realizing end-to-end auxiliary algorithm research, and generating high-quality images that are close to real CT angiography, thus promoting the advancement of integrated non-invasive cardiovascular examination.

[0071] Those skilled in the art will understand that, in addition to implementing the system and its various devices provided by this invention in the form of purely computer-readable program code, the same functions can be achieved entirely through logical programming of the method steps, making the system and its various devices of this invention function as logic gates, switches, application-specific integrated circuits, programmable logic controllers, and embedded microcontrollers. Therefore, the system and its various devices provided by this invention can be considered as a hardware component, and the devices included therein for implementing various functions can also be considered as structures within the hardware component; alternatively, the devices for implementing various functions can be considered as both software modules implementing the method and structures within the hardware component.

[0072] The specific embodiments of the present invention have been described above. It should be understood that the present invention is not limited to the specific embodiments described above, and those skilled in the art can make various modifications or variations within the scope of the claims, which do not affect the essence of the present invention.

Claims

1. A method for generating CCTA images from multi-phase CT-MPI images based on diffusion generative adversarial networks, characterized in that, include: Acquire multi-phase CT-MPI and CCTA image data of clinical patients; wherein, the multi-phase CT-MPI images are multiple images obtained by performing multiple CT scans of the heart over a period of time after the patient is injected with contrast agent; the CCTA images are coronary artery images obtained by scanning the heart when the concentration of coronary contrast agent reaches its peak after the patient is injected with contrast agent; the spatial resolution and radiation dose of any single phase CT-MPI image are lower than those of the CCTA image; The acquired multi-phase CT-MPI image data is preprocessed to obtain n preprocessed CT-MPI images; the preprocessing includes coronary artery region segmentation, phase selection based on HU value, image registration and normalization processing; Based on the n-stage preprocessed CT-MPI images, a conditional diffusion generative adversarial network is used to generate corresponding n-stage CCTA images. A weighted fusion strategy is adopted to perform pixel-level fusion of the n-period CCTA images to form a high-quality CCTA image.

2. The method according to claim 1, characterized in that, The preprocessing includes: The coronary artery region of the multi-phase CT-MPI images is segmented by combining a 3D slicer and an artificial intelligence network; Based on the HU values ​​of the coronary artery region in the multi-phase CT-MPI images, the n-phase CT-MPI images with the highest HU values ​​were selected; Based on the scanning parameter information of the CT-MPI image and the CCTA image, the n-stage CT-MPI image and the CCTA image are scaled and aligned on the Z-axis; Based on the segmented coronary artery region, the n-phase CT-MPI images are cropped to remove the background region.

3. The method according to claim 2, characterized in that, The segmentation of the coronary artery region in the multi-phase CT-MPI images by combining a 3D slicer and an artificial intelligence network includes: The multi-phase CT-MPI images of the patient were manually annotated using 3D slicer software. Using an artificial intelligence-based algorithm, a model capable of segmenting the coronary artery region in CT-MPI images was trained based on existing labeled CT-MPI image data. The segmentation results obtained from the trained model are checked and corrected.

4. The method according to claim 2, characterized in that, Based on the HU values ​​of the coronary artery region in the multi-phase CT-MPI images, the n-phase CT-MPI images with the highest HU values ​​were selected, including: The average HU value of the coronary artery region is used as the evaluation index from the multi-phase CT-MPI images, and the nth phase with the highest value is selected as the subsequent input.

5. The method according to claim 1, characterized in that, The training process of the conditional diffusion generative adversarial network includes: The CCTA image is noise-added through multiple forward diffusion processes to obtain a noisy CCTA. b image; The multi-phase CT-MPI images were compared with noisy CCTA images as a condition. b The images are input into the generator model to generate CCTA' images with any number of slices. The CCTA image or the CCTA image is compared with a noisy CCTA image as a condition. b The image input discriminator model is trained adversarially.

6. The method according to claim 5, characterized in that, The process of adding noise to the CCTA image through multiple forward diffusion processes includes: Perform on the CCTA image The diffusion process involves several consecutive diffusion steps, modeled as a parameterized Markov chain. In the t-th step of the diffusion process, Gaussian noise is added to the image, forming a chain composed of... to Forward noise addition process Its definition is: ； in, For the first The variance of each diffusion step, It is the identity matrix. This represents a normal distribution.

7. The method according to claim 6, characterized in that, The training process of the generator model includes: Using a pseudo-3D training method, adjacent Using CT-MPI slices as input, the output is adjacent slices. One CCTA slice, among which ; A feature extraction network was used to analyze CT-MPI images and noisy CCTA images. b The images are subjected to feature extraction, and the features are fused and then upsampled to generate a CCTA' image; the loss function of the generator is: ； in, MSE = 。 8. The method according to claim 7, characterized in that, The training process of the discriminator model includes: Noisy CCTA b The image serves as a condition, with either the generated CCTA image or a real CCTA image as input; Wasserstein distance is used as a measure of divergence to distinguish between generated and real images, guiding the generator to produce realistic blood vessel details. This loss function is specifically designed to... ； This means the discriminator must be sufficiently smooth, and the output is passed through the softplus function to obtain the discriminator's loss function: 。 9. The method according to claim 1, characterized in that, The weighted fusion strategy, which fuses the n-period CCTA images at the pixel level to form a high-quality CCTA image, includes: Based on the HU values ​​of the coronary artery region in each phase of CT-MPI images, the weight parameters corresponding to each phase of images are obtained. Based on the weight parameters corresponding to each period's images, the generated CCTA images from each period are fused pixel-by-pixel to obtain the final high-quality CCTA image. The expression for this fusion process is as follows: ； in, This indicates the location of the final fused CCTA image. pixel values, Indicates the first At what point in time is the location pixel values, Indicates the first Weight parameters for each time point.

10. A system for generating CCTA images from multi-phase CT-MPI images based on diffusion generative adversarial networks, characterized in that, include: The data acquisition module is used to acquire multi-phase CT-MPI and CCTA image data of clinical patients. The multi-phase CT-MPI images are multiple images obtained from repeated CT scans of the heart over a period of time after contrast agent injection. The CCTA images are coronary artery images obtained by scanning the heart when the coronary contrast agent concentration reaches its peak after contrast agent injection. The spatial resolution and radiation dose of any single-phase CT-MPI image are lower than those of the CCTA images. The data preprocessing module is used to preprocess the acquired multi-phase CT-MPI image data to obtain n preprocessed CT-MPI images; the preprocessing includes coronary artery region segmentation, phase selection based on HU value, image registration and normalization processing; The image generation module is used to generate corresponding n-stage CCTA images based on the n-stage preprocessed CT-MPI images using a conditional diffusion generative adversarial network. The weighted fusion module is used to perform pixel-level fusion of the n-period CCTA images to form a high-quality CCTA image by employing a weighted fusion strategy.

Images