A sequential scanning DECT cardiac imaging method based on AIGC condition generation

Abstract

The present invention discloses a sequential scanning DECT cardiac imaging method based on AIGC conditional generation, including sequential CT scanning of projection data of different phases and high and low energy levels of the heart, using the projection data to perform FDK reconstruction to obtain image domain data of high and low energy levels, selecting high and low energy level organ images in a relatively static state in the image domain to train a conditional generation network model Energy-GAN, using low-energy image layers of moving organs as input conditions of Energy-GAN to generate high-energy image layers in the same motion state, and finally performing image domain material decomposition on the generated high-energy image and the corresponding low-energy image, and using a pre-trained diffusion model to denoise the material decomposition image. The present invention realizes sequential scanning DECT cardiac imaging, effectively reducing the material decomposition artifacts caused by cardiac motion and the noise generated by image domain material decomposition without performing flexible 3D image registration.

Description

Technical Field

[0001] The present invention relates to the field of computer image processing, and in particular to a sequential scanning DECT cardiac imaging method based on AIGC condition generation. Background Art

[0002] In recent years, with the continuous development of medical imaging technology, CT (computed tomography) technology has become an important tool in diagnosis and treatment. CT technology uses X-rays to scan the body at different angles to obtain cross-sectional images of the body, thereby helping doctors diagnose diseases and guide surgery. Traditional CT imaging technology is X-ray mixed energy imaging, which leads to two major inherent defects: X-ray hardening effect and average absorption effect [1]. The former will produce beam hardening artifacts, which are more serious for objects with larger atomic weights, such as bones and iodine contrast agents; the latter will cause different substances to have similar CT values, which is the phenomenon of foreign body imagining.

[0003] To overcome these shortcomings of traditional single-energy computed tomography (SECT), dual-energy computed tomography (DECT) was developed. DECT uses two X-ray beams of different energies to scan an object, which can obtain the energy characteristics of different tissues and materials. Compared with SECT, DECT can provide more accurate tissue identification capabilities, reduce the generation of metal artifacts, and improve the imaging quality of low-contrast structures. However, DECT requires the acquisition of two sets of spatially and temporally consistent projection data of different energies, which places higher demands on the data acquisition method.

[0004] DECT data acquisition methods include rapid KVP switching scanning, single-source multi-slice scanning, and dual-source scanning. These methods require specialized hardware, are costly, and technically challenging. In contrast, single-source sequential scanning can be implemented on any conventional SECT scanner, is less expensive, and requires less technical difficulty. It is primarily used for kidney stone and gout examinations. However, this method is more susceptible to the motion of the scanned object during scanning, such as changes in body position, organ motion caused by respiration, and heartbeat. In particular, heartbeat can reach several centimeters within 1 second and is highly nonlinear and localized. Since sequential scans are acquired from different locations of the same anatomical structure, they may produce large global motion artifacts, reducing the quantitative accuracy of DECT imaging.

[0005] For cardiac data collected by sequential scanning DECT, it is necessary to match the motion state before material decomposition, which is generally achieved through image registration algorithms. According to different similarity measurement standards, the traditional registration algorithms currently developed mainly include three categories. The first category is the registration method based on the grayscale features of the image. This type of method matches according to the grayscale distribution characteristics of the image, including Demons, NCC, MSE, MI, etc. The second category is the registration method based on image features, including surf, ORB, HOG, LBP, etc. The third category is image registration based on the transform domain, such as Fourier transform. The above algorithms measure image similarity from different perspectives. Although they can play a good effect in specific registration tasks, they all have their own limitations. For example, the Demons algorithm has the constraints of constant brightness and small motion, and does not work well when the motion amplitude is large; the NCC and MSE algorithms are based on the consistency of the grayscale features of the image, and do not work well when the grayscale values ​​of the images are very different and for multimodal registration; MI, as a global metric, only measures the statistical correlation of the grayscale values ​​between two complete images for local image alignment; HOG, surf, LBP, ORB and other image feature-based registration methods only align important points in the image and tend to ignore global information; methods based on transform domains usually require certain assumptions on the image transformation model, such as affine transformation, polynomial transformation, etc. If the transformation model does not conform to the actual situation, it may affect the accuracy and stability of the registration.

[0006] Currently, for single-modality CT image registration, similarity metrics based on CT intensity values, such as NCC, MSE, and MI, are generally used. This approach has been proven to be feasible for registering CT images scanned at the same energy level. However, in DECT imaging, due to the different energy levels of the two scans, there is a significant difference in intensity values ​​between high-energy and low-energy CT images. This leads to significant CT value drift in the deformed images generated by single-modality CT image registration algorithms based on intensity features, amplifying DECT imaging errors.

[0007] Therefore, traditional registration-based motion state matching methods are difficult to directly use to resolve motion artifacts in sequential scanning DECT cardiac imaging. With the development of AIGC technology, it has become possible to directly generate the required images through conditional generation. Therefore, the present invention combines AIGC conditional generation technology with DECT technology to develop a sequential scanning DECT cardiac imaging method based on AIGC conditional generation, which solves the motion artifact problem in sequential scanning DECT cardiac imaging in a new way. Summary of the Invention

[0008] In order to solve the problem of motion artifacts in sequential scanning DECT cardiac imaging, the present invention discloses a sequential scanning DECT cardiac imaging method based on AIGC conditional generation. The sequential scanning DECT cardiac image domain data with motion state matching is directly generated by conditional generation, avoiding the motion state matching of images through complex 3D flexible alignment under sequential scanning DECT cardiac imaging conditions.

[0009] In order to achieve the above object, the present invention provides the following technical solutions:

[0010] A sequential scanning DECT cardiac imaging method based on AIGC condition generation includes the following steps:

[0011] Step 1: Acquire cardiac high-energy and low-energy CT data obtained by sequential scanning, and perform FDK reconstruction on the high-energy and low-energy CT data respectively;

[0012] Step 2: Separate static image layers and dynamic image layers from the high- and low-energy CT image domain data reconstructed by FDK;

[0013] Step 3: Use the static low-energy image layer as the generation condition and the static high-energy image layer as the generation target to train the Energy-GAN conditional generation model;

[0014] Step 4: Use the dynamic low-energy image layer as the generation condition and input it into Energy-GAN to generate the corresponding high-energy layer;

[0015] Step 5: Perform image domain material decomposition to calculate iodine and water base images based on the dynamic low-energy image layer and the generated high-energy image layer;

[0016] Step 6: Using the simulated iodine-based and water-based images of the heart, gradually add Gaussian noise to train the diffusion denoising model;

[0017] In step 7, the water-based and iodine-based images in step 5 are input into the pre-trained denoising model for denoising.

[0018] Furthermore, in step 1, the cone-beam 3D projection algorithm in the ASTRA Toolbox is used to perform FDK analytical reconstruction.

[0019] Furthermore, in step 2, the image similarity measurement function provided by the scikit-image library is used to separate the static image layer and the dynamic image layer, specifically including the following process: selecting SSIM as the image similarity measure, setting an appropriate threshold as the standard for distinguishing dynamic and static layers, taking the corresponding 2D layers in the 3D high-energy level cardiac image and the 3D low-energy level cardiac image as input, comparing the SSIM values ​​of image pairs from different energy levels and the same image layer, and if the SSIM values ​​are lower than the specified threshold, it is considered that the image pair has inconsistent motion states and belongs to the dynamic layer; otherwise, it belongs to the static layer. Finally, four image sets M(low, t1, static), M(low, t1, dynamic), M(high, t2, static), and M(high, t2, dynamic) are obtained, representing the low-energy static layer, the low-energy dynamic layer, the high-energy static layer, and the high-energy dynamic layer, respectively, where the motion states of M(low, t1, static) and M(high, t2, static) match, and the motion states of M(low, t1, dynamic) and M(high, t2, dynamic) do not match.

[0020] Furthermore, the step 3 specifically includes the following process:

[0021] ① The image layers corresponding to the separated static low-energy image M(low,t1,static) and the static high-energy image M(high,t2,static) are spliced ​​along the channel direction as positive samples;

[0022] ② Input the separated static low-energy image M(low,t1,static) into the generator G, output G(M(low,t1,static)), and concatenate M(low,t1,static) and G(M(low,t1,static)) channel-wise as negative samples;

[0023] ③ Alternately train the generator G and the discriminator D until the loss functions of both the generator and the discriminator converge;

[0024] Furthermore, the step 4 specifically includes the following process:

[0025] The dynamic low-energy image layer is input into the trained Energy-GAN to obtain the corresponding high-energy image in the same motion state generated by the generator based on the dynamic low-energy image.

[0026] Furthermore, the step 5 specifically includes the following process:

[0027] The dynamic low-energy image layer and the corresponding generated high-energy image layer are solved by the image domain material decomposition algorithm based on matrix inversion. According to the image-based decomposition theory, the linear attenuation coefficient of the CT image is approximately the weighted sum of the two basis material images, as shown in Formula 1:

[0028]

[0029] X H,i,j represents the CT image reconstructed at different energy levels; d1 / 2 represents the base material image or decomposition image; i, j are the image pixel indices; the component matrix is ​​represented by X kH / L (k=1 or 2) element composition, Xk H / L is the linear attenuation coefficient or mass attenuation coefficient of the base material measured at high and low energies. At this time, the matrix is ​​inverted, and the process is shown in Formula 2:

[0030]

[0031] In this method, d1 and d2 represent water-based image and iodine-based image, respectively.

[0032] X H , X L They represent the dynamic low-energy image layers M(low,t1,dynamic)[slicer] and G(M(low,t1,dynamic)[slicer]) respectively, where slicer is the layer index.

[0033] Furthermore, in step 6, the simulated heart data in the XCAT phantom is used to generate a material-based image gold standard T, and Gaussian noise is gradually added to the gold standard for training the diffusion denoising model.

[0034] Furthermore, in step 7, the water-based image water and the iodine-based image iodine generated in step 5 are input into the trained diffusion denoising model to reduce noise.

[0035] Compared with the existing technology, the present invention has the following advantages and beneficial effects:

[0036] 1. The present invention applies AIGC technology to sequential scanning DECT cardiac imaging, directly generating motion-state-matched cardiac dual-energy scanning image domain data, avoiding the use of complex 3D flexible registration technology to perform motion state correction on the initial high- and low-energy image domains. The trained model can perform real-time online dual-energy imaging, effectively reducing imaging time compared to 3D flexible registration technology.

[0037] 2. The present invention inputs the separated static low-energy image M(low, t1, static) into the generator G as a generation condition, splices the separated static high- and low-energy image pairs M(low, t1, static) and M(high, t2, static) in the channel direction and inputs them into the discriminator D as positive samples, and splices M(low, t1, static) and G(M(low, t1, static)) in the channel direction and inputs them into the discriminator as negative samples. This effectively changes the CT intensity distribution of M(low, t1, static) while keeping the motion state of M(low, t1, static) unchanged, so that the CT intensity distribution of G(M(low, t1, static)) is consistent with the intensity distribution of M(high, t2, static).

[0038] 3. The diffusion model-based image denoising method implemented in this invention has many advantages. First, it does not require a large amount of paired training data, making it more flexible in practical applications. Second, it can approximate the posterior distribution of the clean image through the reverse diffusion process, thus avoiding the problem of oversmoothing. Finally, the degree of denoising can be controlled by adjusting the number of reverse diffusion steps to meet different needs.

[0039] 4. The sequential scanning DECT cardiac imaging method based on AIGC condition generation proposed in the present invention effectively expands the clinical application scope of ordinary CT and reduces the cost of DECT cardiac imaging. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] Figure 1 The overall flow chart of the sequential scanning DECT cardiac imaging method based on AIGC condition generation provided by the present invention;

[0041] Figure 2 Schematic diagram of Energy-GAN, a network for generating high-energy CT images based on AIGC conditions;

[0042] Figure 3 Schematic diagram of the diffusion denoising model used in the present invention;

[0043] Figure 4 The high-energy image M(high,t1,dynamic) generated by Energy-GAN in this invention and the high-energy image M'(high,t1,dynamic) after deformation of M(high,t2,dynamic) using three different registration algorithms SyNMI, Demons, and VoxelMorph, where Ground Truth is the gold standard and slicer is the layer index. Experimental results show that the proposed scheme is superior to the registration-based method;

[0044] Figure 5 The iodine-based images obtained by the sequential scanning DECT cardiac imaging method based on AIGC conditions in the present invention and the iodine-based material images obtained by three registration-based methods: SyNMI, Demons, and VoxelMorph. GroundTruth is used as the gold standard. Experimental results show that the scheme proposed in this invention is superior to the registration-based methods.

[0045] Figure 6 The water-based images obtained by the sequential scanning DECT cardiac imaging method generated based on AIGC conditions in the present invention and the water-based material images obtained by three registration-based methods SyNMI, Demons, and VoxelMorph, among which GroundTruth is the gold standard. The experimental results show that the scheme proposed in the present invention is superior to the registration-based method. DETAILED DESCRIPTION

[0046] The present invention will be further described below with reference to the accompanying drawings and specific embodiments. It should be understood that the following specific embodiments are intended only to illustrate the present invention and are not intended to limit the scope of the present invention. It should be noted that the terms "front," "rear," "left," "right," "up," and "down" used in the following description refer to directions in the accompanying drawings, and the terms "inward" and "outward" refer to directions toward or away from the geometric center of a particular component, respectively.

[0047] This embodiment provides a sequential scanning DECT cardiac imaging method based on AIGC conditional generation. First, sequential scanning DECT myocardial perfusion projection domain data of a human body obtained from clinical scans is used to perform FDK parsing and reconstruction using the cone-beam 3D projection algorithm in the ASTRA Toolbox to obtain image domain data with mismatched motion states. Next, an image similarity measurement function provided by the scikit-image library is used to separate static image layers and dynamic image layers. Next, an image similarity measurement function provided by the scikit-image library is used to separate static image layers and dynamic image layers. Then, low-energy and high-energy static image pairs M(low, t1, static) and M(high, t2, static) are used to train a conditional generative model Energy-GAN. The low-energy dynamic image layer M(low, t1, dynamic) is input into the generator G of the trained Energy-GAN as a generation condition to generate a corresponding high-energy layer M(high, t1, dynamic) with matching motion states.

[0048] Then, M(low,t1,dynamic) and M(high,t1,dynamic) were used to perform material decomposition in the image domain, and the decomposed material-based image was input into the pre-trained diffusion denoising model for denoising; finally, the sequential scanning DECT cardiac imaging method generated based on AIGC conditions was compared with the results obtained using the traditional registration-based algorithm, and the performance of the network model was evaluated.

[0049] Specifically, the overall process of this embodiment is as shown in the attached Figure 1-4 As shown, the following steps are included:

[0050] Step 1: Acquire cardiac high- and low-energy CT data obtained by sequential scanning, and perform FDK analytical reconstruction using the cone-beam 3D projection algorithm in ASTRA Toolbox.

[0051] Step 2: Use the image similarity measurement function provided by the scikit-image library to separate the static image layer and the dynamic image layer. The specific process includes the following: selecting SSIM as the image similarity measure, setting an appropriate threshold as the standard for distinguishing dynamic and static layers, taking the corresponding 2D layers in the 3D high-energy level cardiac image and the 3D low-energy level cardiac image as input, and comparing the SSIM values ​​of image pairs from different energy levels and the same image layer. If the SSIM value is lower than the specified threshold, the image pair is considered to have inconsistent motion states and belongs to the dynamic layer. Otherwise, it belongs to the static layer. Finally, four image sets M(low, t1, static), M(low, t1, dynamic), M(high, t2, static), and M(high, t2, dynamic) are obtained, representing the low-energy static layer, the low-energy dynamic layer, the high-energy static layer, and the high-energy dynamic layer, respectively. Among them, the motion states of M(low, t1, static) and M(high, t2, static) match, and the motion states of M(low, t1, dynamic) and M(high, t2, dynamic) do not match. In this implementation case, the threshold is set to 0.9.

[0052] Step 3: Splice the image layers corresponding to the separated static low-energy image M(low, t1, static) and the static high-energy image M(high, t2, static) along the channel direction as positive samples. Input the separated static low-energy image M(low, t1, static) into the generator G, output G(M(low, t1, static)), and splice M(low, t1, static) and G(M(low, t1, static)) along the channel direction as negative samples. Alternately train the generator G and the discriminator D until the loss functions of both the generator and the discriminator converge.

[0053] In step 4, the dynamic low-energy image layer is input into the trained Energy-GAN to obtain the corresponding high-energy image in the same motion state generated by the generator based on the dynamic low-energy image. In step 5, the dynamic low-energy image layer and the corresponding generated high-energy image layer are solved using the image domain material decomposition algorithm based on matrix inversion. According to image-based decomposition theory, the linear attenuation coefficient of the CT image is approximately the weighted sum of the two basis material images, as shown in Formula 1:

[0054]

[0055] X H,i,j represents the CT image reconstructed at different energy levels; d1 / 2 represents the base material image or decomposition image; i and j are the image x-axis and y-axis pixel indices respectively; the component matrix is ​​composed of X kH / L (k=1 or 2) element composition, Xk H / L is the linear attenuation coefficient or mass attenuation coefficient of the base material measured at high and low energies. At this time, the matrix is ​​inverted, and the process is shown in Formula 2:

[0056]

[0057] In this method, d1 and d2 represent water-based image and iodine-based image, respectively.

[0058] X H , X L They represent the dynamic low-energy image layers M(low,t1,dynamic)[slicer] and G(M(low,t1,dynamic)[slicer]) respectively, where slicer is the layer index.

[0059] In step 6, the simulated heart data in the XCAT phantom is used to generate a material-based image gold standard T, and Gaussian noise is gradually added to the gold standard for training the diffusion denoising model.

[0060] In step 7, the water-based image and the iodine-based image generated in step 5 are input into the trained diffusion denoising model to reduce noise.

[0061] In order to verify the effectiveness of the sequential scanning DECT cardiac imaging method based on AIGC condition generation disclosed in the present invention, a set of human CT data sets obtained by clinical scans was used to verify the significant improvement of the method disclosed in the present invention over the traditional algorithm. Figure 5 、 6A comparison chart of the material decomposition effects obtained by using the traditional algorithm and the conditional generation algorithm of the present invention is shown. Compared with the two, the method provided by the present invention can significantly reduce the motion artifact problem caused by cardiac motion in sequential scanning DECT imaging while ensuring the quantitative accuracy of material decomposition. The imaging effect is significantly improved, which has profound significance for conducting clinical medical research.

[0062] The technical means disclosed in the solution of the present invention are not limited to the technical means disclosed in the above-mentioned embodiment, but also include technical solutions composed of any combination of the above technical features.

Claims

1. A sequential scanning DECT cardiac imaging method based on AIGC condition generation, characterized in that: The following steps are involved: Step 1: Acquire cardiac high-energy and low-energy CT data obtained by sequential scanning, and perform FDK reconstruction on the high-energy and low-energy CT data respectively; Step 2: Separate static image layers and dynamic image layers from the high- and low-energy CT image domain data reconstructed by FDK; Step 3: Use the static low-energy image layer as the generation condition and the static high-energy image layer as the generation target to train the Energy-GAN conditional generation model; Step 4: Use the dynamic low-energy image layer as the generation condition and input it into Energy-GAN to generate the corresponding high-energy layer; Step 5: Perform image domain material decomposition to calculate iodine and water base images based on the dynamic low-energy image layer and the generated high-energy image layer; Step 6: Using the simulated iodine-based and water-based images of the heart, gradually add Gaussian noise to train the diffusion denoising model; In step 7, the water-based and iodine-based images in step 5 are input into the pre-trained denoising model for denoising.

2. The sequential scanning DECT cardiac imaging method based on AIGC condition generation according to claim 1, characterized in that: In step 1, the cone-beam 3D projection algorithm in the ASTRA Toolbox is used to perform FDK reconstruction.

3. The sequential scanning DECT cardiac imaging method based on AIGC condition generation according to claim 1, characterized in that: In step 2, the static image layer and the dynamic image layer are separated using the image similarity measurement function provided by the scikit-image library, which specifically includes the following process: SSIM was selected as the image similarity measure, and a suitable threshold was set based on actual operations as the standard for distinguishing dynamic and static layers. The corresponding 2D layers in the 3D high-energy level cardiac image and the 3D low-energy level cardiac image were used as input, and the SSIM values ​​of image pairs from the same image layer at different energy levels were compared. If the SSIM values ​​were lower than the specified threshold, the image pair was considered to have inconsistent motion states and belonged to a dynamic layer. Otherwise, it was considered to belong to a static layer. Finally, four image sets M(low, t1, static), M(low, t1, dynamic), M(high, t2, static), and M(high, t2, dynamic) were obtained, representing the low-energy static layer, the low-energy dynamic layer, the high-energy static layer, and the high-energy dynamic layer, respectively. The motion states of M(low, t1, static) and M(high, t2, static) matched, while the motion states of M(low, t1, dynamic) and M(high, t2, dynamic) did not match.

4. The sequential scanning DECT cardiac imaging method based on AIGC condition generation according to claim 1, characterized in that: The step 3 specifically includes the following process: ① The image layers corresponding to the separated static low-energy image M(low,t1,static) and the static high-energy image M(high,t2,static) are spliced ​​along the channel direction as positive samples; ② Input the separated static low-energy image M(low,t1,static) into the generator G, output G(M(low,t1,static)), and concatenate M(low,t1,static) and G(M(low,t1,static)) channel-wise as negative samples; ③ Alternately train the generator G and the discriminator D until the loss functions of both the generator and the discriminator converge.

5. The sequential scanning DECT cardiac imaging method based on AIGC condition generation according to claim 1, characterized in that: The step 4 specifically includes the following process: The dynamic low-energy image layer is input into the trained Energy-GAN to obtain the corresponding high-energy image in the same motion state generated by the generator based on the dynamic low-energy image.

6. The sequential scanning DECT cardiac imaging method based on AIGC condition generation according to claim 1, characterized in that: The step 5 specifically includes the following process: The dynamic low-energy image layer and the corresponding generated high-energy image layer are solved by the image domain material decomposition algorithm based on matrix inversion. According to the image-based decomposition theory, the linear attenuation coefficient of the CT image is approximately the weighted sum of the two basis material images, as shown in Formula 1: X H,i,j represents the CT image reconstructed at different energy levels; d1 / 2 represents the base material image or decomposition image; i and j are the image x-axis and y-axis pixel indices respectively; the component matrix is ​​composed of X kH / L (k=1 or 2) element composition, Xk H / L is the linear attenuation coefficient or mass attenuation coefficient of the base material measured at high and low energies; at this time, the matrix is ​​inverted, and the process is shown in (2): Where d1 and d2 represent water-based image and iodine-based image respectively. X H， X L They represent the dynamic low-energy image layers M(low,t1,dynamic)[sl icer] and G(M(low,t1,dynamic)[slicer]) respectively, where slicer is the layer index.

7. The sequential scanning DECT cardiac imaging method based on AIGC condition generation according to claim 1, characterized in that: The step 6 specifically includes the following process: The simulated heart data in the XCAT phantom are used to generate the material-based image gold standard T, and Gaussian noise is gradually added to the gold standard for training the diffusion denoising model.

8. The sequential scanning DECT cardiac imaging method based on AIGC condition generation according to claim 1, characterized in that: The step 7 specifically includes the following process: The water-based image water and the iodine-based image iodine generated in step 5 are input into the trained diffusion denoising model to reduce noise.

Images