A system and method for generating an angiogram image based on noise enhancement difference

Abstract

This invention relates to a system and method for generating angiographic images based on noise enhancement differences. The system employs a data acquisition device to obtain a series of slices composed of multiple contrast-free angiographic images. A decoupled structure-texture network is used to fuse and encode the noise samples of each slice with the conditional embedding of that slice. The encoded results are then dual-headedly decoded to obtain structural differences and noise components. Subsequently, the structural differences and noise components are substituted into a noise-assisted structural difference degradation operator to obtain a preliminary estimate of the angiographic image. Finally, the preliminary estimate of the angiographic image for each slice is subjected to global perception adjustment to obtain the angiographic image for each slice. This achieves cross-modal generation from contrast-free angiographic images to angiographic images, improving the global structural consistency and visual naturalness of the generated results, and significantly enhancing the quality of vascular imaging and structural fidelity.

Description

Technical Field

[0001] This invention relates to the field of computer modeling and systems technology, and more specifically, to a system and method for generating angiographic images based on noise enhancement differences. Background Technology

[0002] Computed tomography angiography (CT angiography) is a non-invasive, effective, and high-resolution vascular imaging technique widely used for the diagnosis of various vascular diseases. However, this technique relies on intravenous injection of contrast agents such as iodinated contrast agents to enhance vascular visualization. This can pose additional risks to patients with iodine allergies, impaired kidney function, or multiple myeloma, and the procedure is relatively expensive and complex. In contrast, non-contrast CT angiography (non-contrast CT angiography) eliminates the need for contrast agents, offering advantages such as ease of operation and high safety. However, due to limited contrast in vascular areas, non-contrast CT angiography images cannot clearly display details such as small vessels, stenosis, and complex branching structures, and are typically only used for preliminary screening or overall assessment.

[0003] However, it is worth noting that contrast-free angiography images still contain a certain amount of potential vascular structural information. If generative modeling techniques can be used to directly reconstruct vascular images with the contrast characteristics of computed tomography angiography from contrast-free angiography images—that is, angiography images—it would not only significantly reduce dependence on contrast agents, but also provide a safer and lower-cost alternative examination option for high-risk groups such as those with contrast agent allergies, renal insufficiency, and multiple myeloma. Furthermore, it would open up new avenues for the development of medical image synthesis and intelligent diagnosis.

[0004] To address this, existing technologies have proposed a diffusion model for constructing angiography images, namely the Diff angiography image model, which applies an adversarial diffusion model to the conversion task of transforming contrast-free angiography images into angiography images. This diffusion model effectively improves the generation accuracy and structural robustness of vascular regions by introducing unsupervised contrast enhancement masks, contrast enhancement attention mechanisms, and a mask-guided PatchNCE loss function.

[0005] However, since the above diffusion model relies entirely on random noise during the forward degradation process, it is difficult to explicitly model the structural differences between contrast-free angiography images and angiography images, and the effective constraints on the angiography region are insufficient, thereby reducing the vascular imaging quality and structural fidelity of the generated images. Summary of the Invention

[0006] The technical problem to be solved by this invention is how to overcome the technical defects of existing diffusion models for constructing angiography images, which have insufficient vascular imaging quality and structural fidelity in the process of converting contrast-free angiography images into angiography images. In order to overcome this technical defect, this invention provides an angiography image generation system and method based on noise enhancement difference, specifically including an angiography image generation system based on noise enhancement difference and an angiography image generation method based on noise enhancement difference.

[0007] This invention provides a system for generating angiographic images based on noise enhancement differences, comprising: The acquisition device is used to acquire contrast-free angiography images of the subject and obtain a series of slices consisting of multiple contrast-free angiography images. A decoupled structure-texture network, electrically connected to the collector, is used to fuse and encode the noise samples of each slice in the slice series with the conditional embedding of that slice, and to perform dual-head decoding on the encoding result to obtain structural differences and noise components. Then, the structural differences and noise components are substituted into a noise-assisted structural difference degradation operator to obtain a preliminary estimate of the angiography image. Finally, the preliminary estimate of the angiography image of each slice is subjected to global perception adjustment to obtain the angiography image of each slice. Here, the conditional embedding of a slice refers to the result of a convolution operation on the intensity change fusion of the slice. The intensity change fusion of the slice is the sum of two differences: the first difference is the difference between the slice and the previous slice in the slice series, and the second difference is the difference between the next slice and the previous slice.

[0008] The angiography image generation system based on noise enhancement difference disclosed in this invention addresses the aforementioned technical deficiencies by setting up a collector and a decoupled structure-texture network. The collector acquires a series of slices composed of multiple contrast-free angiography images. The decoupled structure-texture network fuses and encodes the noise samples of each slice in the slice series with the conditional embedding of that slice. The encoding result is then decoded by two heads to obtain structural differences and noise components. Subsequently, the structural differences and noise components are substituted into a noise-assisted structural difference degradation operator to obtain a preliminary estimate of the angiography image. Finally, the preliminary estimate of the angiography image of each slice is subjected to global perception adjustment to obtain the angiography image of each slice. This achieves cross-modal generation from contrast-free angiography images (i.e., contrast-free angiography images) to angiography images (i.e., angiography images). Considering that the main difference between contrast-free angiography and traditional angiography stems from the visual changes in the vascular region after contrast agent injection, while the overall anatomical structure remains largely consistent, this application uses vascular visual difference as the core of modeling. A decoupled structure-texture network executing a noise-assisted structural difference degradation operator (NASD) is designed. The noise-assisted NASD operator guides the forward degradation process primarily with structural difference information between contrast-free and traditional angiography images, while simultaneously introducing noise samples to assist in modeling the overall texture and high-frequency details of the image. This enhances the ability to recover fine-grained texture information while highlighting changes in vascular structure. Furthermore, before image encoding, the decoupled structure-texture network introduces conditional embedding of slices. This conditional embedding possesses cross-slice structural perception properties, mitigating the spatial discontinuity of blood vessels during generation. Additionally, the decoupled structure-texture network employs a dual-head decoding approach to predict structural differences and noise components separately, enabling the model to simultaneously ensure effective constraints on vascular structure and the recovery of texture details. In addition, to further optimize the overall visual perception of the generated results, the decoupled structure-texture network also adopted a global perception adjustment method. Without interfering with the backbone network prediction, it improved the global structural consistency and visual naturalness of the generated results through perception loss, thereby significantly improving the quality of vascular imaging and structural fidelity.

[0009] In one possible implementation, the decoupled structure-texture network includes: An encoding device, electrically connected to the collector, is used to calculate the conditional embedding and noise samples of the slice, fuse the noise samples with the conditional embedding, and then perform feature encoding on the fusion result to obtain the encoding result; An intermediate module, electrically connected to the encoding device, is used to perform a convolution operation on the encoding result to achieve a fixed resolution and number of channels and obtain an intermediate mapping result; A dual-head decoding module includes decoder A and decoder B, both of which are electrically connected to the intermediate module. Decoder A is used to convert the intermediate mapping result into the noise component through convolution and upsampling, and decoder B is used to convert the intermediate mapping result into the structural difference through convolution and upsampling. A noise-assisted module, which is electrically connected to both decoder A and decoder B, is used to substitute the structural differences and noise components into a noise-assisted structural difference degradation operator to obtain a preliminary estimate of the angiography image. A global perception adjustment module, electrically connected to the noise-assisted module, is used to perform global perception adjustment on the preliminary estimate of the angiography image to obtain the angiography image.

[0010] The decoupled structure-texture network, possessing the aforementioned structure and functions, combined with cross-slice structure-aware conditional embedding computed by the encoding device, a global perception adjustment module, and a dual-decoder design, achieves independent prediction of structural difference information and Gaussian noise components. By strengthening vascular structure constraints and texture detail representation through dual decoders, the overall structural accuracy, detail clarity, and visual realism of the generated angiography images are significantly improved.

[0011] In one possible implementation, the encoding device includes: A cross-slice conditional embedding module, electrically connected to the collector, is used to calculate the difference between the target slice and the previous slice, and the difference between the next slice and the target slice. The two differences are summed to obtain the intensity change fusion. Then, the intensity change fusion is subjected to a convolution operation to obtain the conditional embedding of the slice. The noise module, electrically connected to the collector, is used to substitute the noise components of the target slice and the previous slice into the noise estimation formula to obtain the noise sample of the slice. The fusion unit is electrically connected to both the cross-slice conditional embedding module and the noise module, and is used to fuse the noise samples of the target slice with the conditional embedding to obtain the fusion result. The encoder, which is electrically connected to both the fusion unit and the intermediate module, is used to perform feature encoding on the fusion result through multiple convolutions and downsampling to transform the fusion result into the encoded result.

[0012] The difference between a slice and the previous slice, and the difference between the next slice and the current slice, characterizes the intensity variation along the direction perpendicular to the slice, thus explicitly revealing the extension and disappearance patterns of dynamic structures such as blood vessels between adjacent slices. Then, a noise module and a fusion unit combine these difference information with the features of the current slice to enhance the directionality and discriminativeness of the conditional information. The cross-slice conditional embedding module can more effectively highlight the imaging regions that vary with the slice, accurately capture the spatial continuity of anatomical structures such as blood vessels, and provide more precise conditional signals. This allows the model to better integrate static structural information and cues of dynamic changes in blood vessels during denoising, thereby better capturing information changes along the z-axis within the two-dimensional generative framework.

[0013] In one possible implementation, the noise estimation formula is as follows: , In the formula, For noise samples, For time steps, For slicing, The noise intensity coefficient, This represents the noise component of the slice.

[0014] Introducing this form of noise to help model the overall texture and high-frequency details of an image can enhance the ability to recover fine-grained texture information while highlighting changes in vascular structure.

[0015] In one possible implementation, the expression for the noise-assisted structural difference degradation operator is as follows: , In the formula, Preliminary estimation of angiography images, For control coefficients, Due to structural differences.

[0016] The noise operators described above can not only stimulate the high-frequency recovery capability of the model, but also avoid destroying the overall structural difference degradation process, thereby significantly improving the texture clarity and detail fidelity of the generated image.

[0017] In one possible implementation, the global awareness adjustment module is configured to perform the following steps: A1: The angiography image is initially estimated and then mapped using an embedding layer to obtain embedding information; A2: Perform multiple convolutional layer mapping processes on the embedded information to obtain the convolutional processing result; A3: The embedded information and the convolution processing result are transformed into queries and keys through linear layer operations; A4: The query and key are transformed into a spatial confidence graph and values ​​through attention layer operations; A5: Substitute the spatial confidence map and the preliminary estimate of the angiography image into the adaptive fusion calculation formula to obtain the angiography image.

[0018] The global perception adjustment module, operating according to the above steps, ensures the stability of the backbone network's training process, unaffected by additional gradient perturbations. Simultaneously, it allows the global perception adjustment module to focus on perceptual consistency optimization without interfering with the backbone network's prediction task. Thus, while maintaining stable training, the model significantly enhances the structural fidelity and visual perception quality of the generated images.

[0019] In one possible implementation, the adaptive fusion calculation formula is as follows: , In the formula, The angiography image is shown. For spatial confidence plots.

[0020] Another technical solution of the present invention is to provide a method for generating angiographic images based on noise enhancement differences, the method comprising the following steps: S1: The parameters of the decoupled structure-texture network are optimized by solving for the optimal value of the joint loss function of image estimation and perception. S2: Acquire contrast-free angiography images of the subject using a data acquisition device to obtain a series of slices consisting of multiple contrast-free angiography images; S3: The noise samples of each slice in the slice series are fused and encoded with the conditional embedding of the slice through the decoupled structure-texture network with optimized parameters, and the encoding result is decoded by two heads to obtain the structural differences and noise components; S4: Substitute the structural differences and noise components into the noise-assisted structural difference degradation operator through the parameter-optimized decoupled structure-texture network to obtain a preliminary estimate of the angiography image; S5: The angiography image of each slice is initially estimated by the decoupled structure-texture network with optimized parameters and then subjected to global perception adjustment to obtain the angiography image of each slice.

[0021] The method disclosed in this invention first optimizes the parameters of a decoupled structure-texture network using a joint loss function of image estimation and perception. Then, a collector acquires a series of slices composed of multiple contrast-free angiography images. The decoupled structure-texture network fuses and encodes the noise samples of each slice with the conditional embedding of that slice. The encoding result is then dual-headedly decoded to obtain structural differences and noise components. These structural differences and noise components are then substituted into a noise-assisted structural difference degradation operator to obtain a preliminary estimate of the angiography image. Finally, the preliminary estimate of the angiography image for each slice is subjected to global perception adjustment to obtain the angiography image for each slice, achieving cross-modal generation from contrast-free angiography images to angiography images. Simultaneously, using vascular contrast differences as the core modeling element, a noise-assisted structural difference degradation operator guides the forward degradation process primarily with structural difference information between contrast-free angiography images and angiography images. Noise samples are introduced to assist in modeling the overall texture and high-frequency details of the image, thereby highlighting vascular structural changes while enhancing the recovery capability of fine-grained texture information. Meanwhile, the decoupled structure-texture network introduces slice-based conditional embedding before image encoding. This conditional embedding possesses cross-slice structural perception, mitigating the spatial discontinuity of blood vessels during generation. Furthermore, the decoupled structure-texture network employs a dual-head decoding approach to predict structural differences and noise components separately, enabling the model to simultaneously ensure effective constraint of vascular structure and restoration of texture details. In addition, to further optimize the overall visual perception of the generated results, a global perceptual adjustment method is adopted. Without interfering with the backbone network's prediction, perceptual loss is used to improve the global structural consistency and visual naturalness of the generated results, thereby significantly improving the quality of vascular visualization and structural fidelity.

[0022] In one possible implementation, the joint loss function for image estimation and perception is expressed as follows: , In the formula, For the true structural differences, The true noise components of the slice. These are real angiography images. For a deep feature-based perceptual similarity measure, The value of the joint loss function for image estimation and perception. and All are weighting coefficients.

[0023] This multi-branch loss function design can explicitly separate the optimization directions of different tasks, avoid mutual interference between gradients, and improve the overall training stability and convergence efficiency. Attached Figure Description

[0024] Figure 1 This is a schematic diagram of the angiography image generation system based on noise enhancement differences disclosed in the embodiments of this application; Figure 2 This is a schematic diagram of the decoupled structure-texture network operation process disclosed in the embodiments of this application; Figure 3 This is a schematic diagram of the operation flow of the cross-slice conditional embedding module disclosed in the embodiments of this application; Figure 4 This is a schematic diagram of the operation flow of the global perception and adjustment module disclosed in the embodiments of this application; Figure 5 This is a schematic diagram of the method flow disclosed in the embodiments of this application. Detailed Implementation

[0025] First, those skilled in the art should understand that these embodiments are merely used to explain the technical principles of the embodiments of this application and are not intended to limit the scope of protection of the embodiments of this application. Those skilled in the art can make adjustments as needed to adapt to specific application scenarios.

[0026] In the description of the embodiments of this application, it should be noted that, unless otherwise explicitly specified and limited, the terms "electrical connection" and "electrical connection relationship" should be interpreted broadly, referring to a connection method with an electrical relationship. For example, it can be a circuit connection achieved through conductive wires, an electrical connection achieved through a radio signal channel, or a combination of both. Furthermore, "electrical connection" and "electrical connection relationship" can be based on a mechanical connection (such as conductive wires being placed within a connecting key); it can be a direct connection or an indirect connection through an intermediate medium. Those skilled in the art can understand the specific meaning of the above terms in the embodiments of this application based on the specific circumstances.

[0027] In the embodiments of this application, unless otherwise explicitly specified and limited, "above," "below," "in front of," or "behind" the second feature can mean that the first and second features are in direct contact, or that the first and second features are in indirect contact through an intermediate medium. Furthermore, "above," "on top of," and "over" the second feature can mean that the first feature is directly above or diagonally above the second feature, or simply indicates that the first feature is at a higher horizontal level than the second feature. "Below," "below," and "under" the second feature can mean that the first feature is directly below or diagonally below the second feature, or simply indicates that the first feature is at a lower horizontal level than the second feature. "Before," "in front of," and "in front of" the second feature can mean that the first feature is directly in front of or diagonally in front of the second feature, or simply indicates that the first feature precedes the second feature in sequence. "After," "behind," and "behind" the second feature can mean that the first feature is directly behind or diagonally behind the second feature, or simply indicates that the first feature is after the second feature in sequence.

[0028] The technical solution of this application will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0029] See Figures 1-5 This embodiment discloses a system for generating angiographic images based on noise enhancement differences. Figure 1 This is a schematic diagram of the overall structure of the angiography image generation system. Figure 1 As shown, the angiography image generation system includes a collector and a decoupled structure-texture network, wherein the decoupled structure-texture network is electrically connected to the collector. The collector is used to acquire contrast-free angiography images of the subject, such as a patient, to obtain a series of slices composed of multiple contrast-free angiography images.

[0030] See Figure 1 and Figure 2 In this angiography image generation system, a decoupled structure-texture network is used to fuse and encode the noise samples of each slice in a slice series with the conditional embedding of that slice. The encoding result is then dual-headed decoded to obtain structural differences and noise components. Subsequently, the structural differences and noise components are substituted into a noise-assisted structural difference degradation operator to obtain a preliminary estimate of the angiography image. Finally, the preliminary estimate of the angiography image of each slice is subjected to global perceptual conditioning to obtain the angiography image of each slice, thus obtaining the angiography image series corresponding to the slice series. Here, the conditional embedding of a slice refers to the result of a convolution operation on the intensity change fusion of the slice. The intensity change fusion of the slice is the sum of two differences: the first difference is the difference between the current slice and the previous slice in the slice series, and the second difference is the difference between the next slice and the current slice.

[0031] See Figure 1 and Figure 2In this embodiment, the decoupled structure-texture network includes an encoding device, an intermediate module, a dual-head decoding module, a noise-assisted module, and a global perception adjustment module. The encoding device is electrically connected to the collector, the intermediate module is electrically connected to the encoding device, the dual-head decoding module includes decoder A and decoder B, both of which are electrically connected to the intermediate module, the noise-assisted module is electrically connected to both decoder A and decoder B, and the global perception adjustment module is electrically connected to the noise-assisted module.

[0032] See Figure 1 and Figure 2 In decoupled structure-texture networks, the encoding device computes the conditional embeddings of slices and noise samples, fuses the noise samples with the conditional embeddings, and then performs feature encoding on the fused result to obtain the encoded result. For example... Figure 1 As shown, in this embodiment, the encoding device includes a cross-slice conditional embedding module, a noise module, a fusion unit, and an encoder. The cross-slice conditional embedding module is electrically connected to the collector, the noise module is electrically connected to the collector, the fusion unit is electrically connected to both the cross-slice conditional embedding module and the noise module, and the encoder is electrically connected to both the fusion unit and the intermediate module.

[0033] See Figure 2 and Figure 3 In the encoding device, the cross-slice conditional embedding module is used to calculate the difference between the target slice and the previous slice, and the difference between the next slice and the target slice. The two differences are summed to obtain intensity variation fusion, and then the intensity variation fusion is convolved to obtain the conditional embedding of the slice. To improve the expressive power of the decoupled structure-texture network model in terms of vascular continuity and imaging consistency, this embodiment designs a cross-slice conditional embedding module (CSSAE) in the conditional embedding stage. The operation flow of this module is as follows: Figure 3 As shown, this module is configured to fully utilize the spatial correlation and intensity variation information between adjacent slices. Specifically, as... Figure 2 As shown, let Let each slice represent the previous, current, and next slices, respectively, thus constructing a three-slice input combination as the model's conditional signal, thereby introducing spatial context information from adjacent layers. To further enhance the discriminative and directional nature of the conditional information, we introduce cross-slice differencing operations to capture the changing trends of local structures. The differencing operation is defined as follows: , in, and These two differences characterize the intensity variation along the z-axis (i.e., the direction perpendicular to the slice surface), thus explicitly revealing the extension and disappearance patterns of dynamic structures such as blood vessels between adjacent slices. Then, this difference information is combined with the features of the current slice to enhance the directionality and discriminativeness of the conditional information. The cross-slice conditional embedding module can more effectively highlight the imaging regions that vary with the slice, accurately capture the spatial continuity of anatomical structures such as blood vessels, and provide more precise conditional signals. This allows the model to better integrate static structural information and cues of dynamic changes in blood vessels during denoising, thereby better capturing information changes along the z-axis within the two-dimensional generative framework.

[0034] See Figure 2 In encoding devices, a noise module is used to substitute the noise components of two adjacent slices into the noise estimation formula to obtain noise samples of the slices. For example... Figure 2 As shown. Specifically, in this embodiment, the noise estimation calculation formula is as follows: , In the formula, For noise samples, For time steps, For slicing, The noise intensity coefficient, This represents the noise component of the slice.

[0035] See Figure 2 In the encoding device, the fusion unit is used to fuse the noise samples of the target slice with the conditional embedding to obtain the fusion result; the encoder is used to perform feature encoding on the fusion result through multiple convolutions and downsampling to transform the fusion result into an encoded result.

[0036] See Figure 1 and Figure 2 In the decoupled structure-texture network, the intermediate module performs convolution operations on the encoded result to achieve a fixed resolution and number of channels, obtaining an intermediate mapping result. The dual-head decoding module includes decoder A and decoder B. Decoder A is used to convert the intermediate mapping result into noise components through convolution and upsampling, while decoder B is used to convert the intermediate mapping result into structural differences through convolution and upsampling.

[0037] Existing degradation methods based on structural differences rely solely on linear scaling of these differences to generate intermediate states during the degradation process, without incorporating any Gaussian noise perturbations. This results in a lack of effective handling of high-frequency image components during the forward pass, preventing the model from obtaining supervisory signals for high-frequency information such as texture and edges during training. In the reverse generation stage, the model's optimization primarily depends on low-frequency, structurally consistent components, with insufficient constraints on high-frequency details. This leads the model to prioritize global pixel consistency while neglecting local detail restoration, often resulting in overly smoothed images and blurred blood vessel edges.

[0038] To solve this problem, see Figure 2 In this embodiment, a noise-assisted module is specifically set in the decoupled structure-texture network. This noise-assisted module is used to substitute structural differences and noise components into a noise-assisted structural difference degradation operator to obtain a preliminary estimate of the angiography image. Figure 2 As shown, the Noise-Augmented Structural Difference Degradation Operator (NASD) is an operator that further introduces random noise terms into the linear structural difference degradation framework, enabling the degradation process to simultaneously possess structural difference dominance and high-frequency perturbation capability. Its form is as follows: , In the formula, Preliminary estimation of angiography images, For control coefficients, Due to structural differences.

[0039] To analyze the impact of the noise term on the frequency components, a Fourier transform is performed on both sides of the noise-assisted structural difference degradation operator, yielding the following: , remember , , , Then the noise-assisted structural difference degradation operator in the frequency domain can be expressed as: , Since the power spectral density of random Gaussian noise is constant in the frequency domain, that is: , The power spectral density of medical images is inversely proportional to spatial frequency. , in, and Let be the scaling correction factor. Then the signal-to-noise ratio (SNR) of the degraded image in the frequency domain can be expressed as: , Therefore, it can be seen that with spatial frequency The increase in signal-to-noise ratio It will gradually decrease, meaning that high-frequency regions correspond to low signal-to-noise ratios. As noise is gradually injected, the overall signal-to-noise ratio of the image gradually decreases. When When the signal-to-noise ratio drops below a certain threshold, it indicates that the frequency information has been completely submerged. Since high-frequency information corresponds to a lower signal-to-noise ratio, it is more likely to drop to this threshold first; therefore, during degradation, the high-frequency information of the image is the first to be destroyed. On the other hand, With noise intensity factor Inversely proportional, therefore by adjusting The size of the value can control the final signal-to-noise ratio above a preset threshold, thereby ensuring that high-frequency information is degraded to provide texture supervision while relatively preserving low-frequency structural information.

[0040] In the noise-assisted structural difference degradation operator, by controlling The scope of noise can be selectively adjusted to primarily affect the high-frequency components of the image, thereby perturbing details and texture information without disrupting the overall structure. This design makes the noise injection process more targeted, providing explicit supervision signals for high-frequency information to the model, promoting the recovery and reconstruction of details and textures, while maintaining the consistency and stability of the low-frequency structure. In the design of the noise term, this embodiment gradually increases the noise intensity over time. In the early stages of degradation, the noise amplitude is small, thus avoiding excessive interference with the low-frequency structure and ensuring that the structural difference degradation process dominates the degradation of overall and low-frequency information. As time progresses... As noise increases, its intensity gradually rises until, in the later stages of degradation, the perturbation of high-frequency details reaches a predetermined intensity, providing sufficient high-frequency restoration supervision for the model. This progressive noise enhancement mechanism has a similar effect to the noise addition mechanism of the vanilla diffusion model, effectively balancing high-frequency texture learning and low-frequency structure preservation. Through this design, the auxiliary noise not only stimulates the model's high-frequency restoration ability but also avoids disrupting the overall structural difference degradation process, thereby significantly improving the texture clarity and detail fidelity of the generated image.

[0041] See Figure 2 and Figure 4To further improve the stability of network training and the structural fidelity of generated angiography images, and to prevent the model from tending to learn a smooth, averaged mapping to minimize global loss during training, thus leading to the loss of detail and texture, this embodiment designs a Global Perception Refinement Module (GPRM), whose operation flow is as follows: Figure 4 As shown, this is done to optimize the perceptual quality of the generated image. The module receives a preliminary estimate from the angiography image. and current time step embedded information and output a fine estimate. This is used to introduce perceptual loss supervision. Specifically, the global perception adjustment module is used to perform global perception adjustment on the preliminary estimate of the angiography image to obtain the angiography image. In this embodiment, the global perception adjustment module is configured to perform the following steps: A1: Perform embedding layer mapping processing on the preliminary estimate of the angiography image to obtain embedding information; A2: Perform multiple convolutional layer mapping processing on the embedding information to obtain convolution processing results; A3: Convert the embedding information and convolution processing results into queries and keys through linear layer operations; A4: Convert the queries and keys into spatial confidence maps and values ​​through attention layer operations; A5: Substitute the spatial confidence map and the preliminary estimate of the angiography image into the adaptive fusion calculation formula to obtain the angiography image. The adaptive fusion calculation formula is as follows: , In the formula, For angiography images, For spatial confidence plots.

[0042] It is worth noting that, acting on The loss gradient is not backpropagated to the backbone network, but is only used to update the parameters of the global perception regulation module itself. On one hand, this is based on the experience of existing research, where predicting noise often exhibits better stability and convergence compared to directly predicting angiographic images. On the other hand, explicitly predicting structural differences in the parallel decoding branch helps introduce stronger structural constraints to the model, thereby guiding the network to focus more on modeling vascular regions. This design ensures that the training process of the backbone network remains stable and is not affected by additional gradient perturbations, while allowing the global perception regulation module to focus on perceptual consistency optimization without interfering with the prediction task of the backbone network. Thus, the model significantly enhances the structural fidelity and visual perception quality of the generated images while maintaining stable training.

[0043] See Figure 5 The following further discloses the usage method of the angiography image generation system based on noise enhancement difference in this embodiment, such as... Figure 5 As shown, the method includes the following steps: S1: The parameters of the decoupled structure-texture network are optimized by solving for the optimal value of the joint loss function of image estimation and perception.

[0044] During the parameter optimization phase, this embodiment applies independent constraints to the three sub-tasks of noise prediction, structural difference recovery, and target angiography image generation to achieve collaborative optimization among multiple objectives. Specifically, the overall training objective is decomposed into three types of non-interfering loss functions: noise prediction error... Structural differences and reconstruction errors and image estimation and perception joint loss Its definition is as follows: , In the formula, For the true structural differences, The true noise components of the slice. These are real angiography images. It is a perceptual similarity measure based on deep features, used to measure the visual consistency between generated images and real angiography images. The value of the joint loss function for image estimation and perception. and These are all weight coefficients used to balance the optimization objectives between pixel-level reconstruction accuracy and perceptual quality. This multi-branch loss function design can explicitly separate the optimization directions of different tasks, avoid mutual interference between gradients, and improve the overall training stability and convergence efficiency. In terms of optimization strategy, we configure an independent optimizer for each loss function, used to update the parameters of the noise prediction branch, structural difference recovery branch, and target image optimization branch, respectively. This design can effectively decouple the gradient propagation paths between different tasks, avoid gradient interference problems that may be caused by a single global loss function, and enable each branch network to converge independently and stably on its specific task objective.

[0045] S2: Acquire contrast-free angiography images of the subject using a data acquisition device to obtain a series of slices consisting of multiple contrast-free angiography images.

[0046] S3: The noise samples of each slice in the slice series are fused and encoded with the conditional embedding of the slice through the decoupled structure-texture network with optimized parameters, and the encoding results are decoded by two heads to obtain structural differences and noise components.

[0047] S4: By substituting the structural differences and noise components into the noise-assisted structural difference degradation operator through the parameter-optimized decoupled structure-texture network, a preliminary estimate of the angiography image is obtained.

[0048] S5: The angiography image of each slice is initially estimated by the decoupled structure-texture network with optimized parameters and then subjected to global perception adjustment to obtain the angiography image of each slice.

[0049] To verify the effectiveness of the proposed decoupled structure-texture network in the task of generating angiography images from contrast-free angiography images, this embodiment systematically compares it with several mainstream image generation and modality transfer models under the same dataset and experimental settings. These include traditional adversarial generative frameworks such as Pix2Pix, CycleGAN, angiography-GAN, UNIT, and MUNIT, as well as diffusion-based generative models that have gradually become mainstream in recent years, including DDPM, SynDiff, Fast-DDPM, and CoreDiff. The experiments evaluate the model performance from multiple levels, including pixel, structure, and perception, using evaluation metrics such as peak signal-to-noise ratio, structural similarity index, root mean square error, feature similarity index, visual information fidelity, and gradient magnitude similarity deviation. Table 1 summarizes the quantitative evaluation results of all compared models.

[0050] Table 1: Experimental Results Overall, the decoupled structure-texture network achieved state-of-the-art or near-optimal performance across multiple key metrics, demonstrating significant comprehensive advantages. In terms of reconstruction accuracy, the decoupled structure-texture network achieved a peak signal-to-noise ratio of 28.92, surpassing all compared methods, and achieved the lowest root mean square error (RMSE) of 9.75, indicating optimal pixel-level reconstruction accuracy. Regarding structural similarity, the decoupled structure-texture network achieved a structural similarity index of 0.9372, the highest among all models, demonstrating better recovery capabilities for vascular structures and detailed textures. Correspondingly, the feature similarity index score was 0.9394, further highlighting the decoupled structure-texture network's structural fidelity advantage in edge information and high-frequency textures. In terms of perceptual quality, the decoupled structure-texture network achieved a visual information fidelity score of 0.5224, also the highest value, indicating that the model retained more information and visual saliency in the generated angiography images. In the gradient magnitude similarity deviation index, which reflects local structural distortion, the decoupled structure-texture network once again obtained the lowest value of 0.0588, demonstrating its superiority in high-frequency structure recovery.

[0051] Compared to traditional GAN-based methods such as Pix2Pix, CycleGAN, and angiography image-GAN, decoupled structure-texture networks significantly improve performance across multiple metrics, including peak signal-to-noise ratio, structural similarity index, visual information fidelity, and root mean square error. This indicates that relying solely on discriminator supervision is insufficient to capture modal differences between contrast-free and contrast-enhanced angiography images, while the structural difference modeling strategy of decoupled structure-texture networks is more effective in angiographic regions. Compared to UNIT and MUNIT, decoupled structure-texture networks exhibit lower root mean square error and higher structural similarity index, demonstrating that explicitly modeling the degradation path of modal differences effectively improves the reliability and structural consistency of cross-modal mapping.

[0052] Compared to diffusion models such as DDPM, SynDiff, Fast-DDPM, and CoreDiff, the decoupled structure-texture network maintains a leading position in both structural and perceptual metrics. For example, CoreDiff's structural similarity index is 0.9345, while the decoupled structure-texture network improves to 0.9372. In gradient magnitude similarity bias, the decoupled structure-texture network achieves the lowest global distortion of 0.0588, outperforming Fast-DDPM, SynDiff, and CoreDiff, indicating that the noise-assisted structural difference degradation process can effectively improve the recovery of high-frequency information. Furthermore, the decoupled structure-texture network also achieves the highest score in visual information fidelity, further demonstrating its significant advantages in the visualization of microvessels and the discernibility of key anatomical structures.

[0053] Overall, the decoupled structure-texture network exhibits superior performance across multiple dimensions, including peak signal-to-noise ratio, root mean square error, structural similarity index, feature similarity index, visual information fidelity, and gradient magnitude similarity deviation, demonstrating its significant comprehensive advantages in cross-modal generation tasks. The model excels in structural consistency, texture clarity, and vascular continuity, accurately recovering subtle anatomical structures in angiography images while maintaining good visual quality and numerical stability. Experimental results fully demonstrate the superiority of the decoupled structure-texture network in cross-modal generation tasks from contrast-free angiography images to angiography images, validating the effectiveness and application potential of noise-assisted structural difference degradation processes in the field of medical image generation.

[0054] In the description of this application, the references to terms such as "an embodiment," "some embodiments," "in this embodiment," "specific example," or "some examples," etc., refer to specific features, structures, materials, or characteristics described in connection with that embodiment or example, which are included in at least one embodiment or example of this application. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in a suitable manner in any one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0055] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A system for generating angiographic images based on noise enhancement differences, characterized in that, include: The acquisition device is used to acquire contrast-free angiography images of the subject and obtain a series of slices consisting of multiple contrast-free angiography images. A decoupled structure-texture network, electrically connected to the collector, is used to fuse and encode the noise samples of each slice in the slice series with the conditional embedding of that slice, and to perform dual-head decoding on the encoding result to obtain structural differences and noise components. Then, the structural differences and noise components are substituted into a noise-assisted structural difference degradation operator to obtain a preliminary estimate of the angiography image. Finally, the preliminary estimate of the angiography image of each slice is subjected to global perception adjustment to obtain the angiography image of each slice. Here, the conditional embedding of a slice refers to the result of a convolution operation on the intensity change fusion of the slice. The intensity change fusion of the slice is the sum of two differences: the first difference is the difference between the slice and the previous slice in the slice series, and the second difference is the difference between the next slice and the previous slice.

2. The angiography image generation system based on noise enhancement difference according to claim 1, characterized in that, The decoupled structure-texture network includes: An encoding device, electrically connected to the collector, is used to calculate the conditional embedding and noise samples of the slice, fuse the noise samples with the conditional embedding, and then perform feature encoding on the fusion result to obtain the encoding result; An intermediate module, electrically connected to the encoding device, is used to perform a convolution operation on the encoding result to achieve a fixed resolution and number of channels and obtain an intermediate mapping result; A dual-head decoding module includes decoder A and decoder B, both of which are electrically connected to the intermediate module. Decoder A is used to convert the intermediate mapping result into the noise component through convolution and upsampling, and decoder B is used to convert the intermediate mapping result into the structural difference through convolution and upsampling. A noise-assisted module, which is electrically connected to both decoder A and decoder B, is used to substitute the structural differences and noise components into a noise-assisted structural difference degradation operator to obtain a preliminary estimate of the angiography image. A global perception adjustment module, electrically connected to the noise-assisted module, is used to perform global perception adjustment on the preliminary estimate of the angiography image to obtain the angiography image.

3. The angiography image generation system based on noise enhancement difference according to claim 2, characterized in that, The encoding device includes: A cross-slice conditional embedding module, electrically connected to the collector, is used to calculate the difference between the target slice and the previous slice, and the difference between the next slice and the target slice. The two differences are summed to obtain the intensity change fusion. Then, the intensity change fusion is subjected to a convolution operation to obtain the conditional embedding of the slice. The noise module, electrically connected to the collector, is used to substitute the noise components of the target slice and the previous slice into the noise estimation formula to obtain the noise sample of the slice. The fusion unit is electrically connected to both the cross-slice conditional embedding module and the noise module, and is used to fuse the noise samples of the target slice with the conditional embedding to obtain the fusion result. The encoder, which is electrically connected to both the fusion unit and the intermediate module, is used to perform feature encoding on the fusion result through multiple convolutions and downsampling to transform the fusion result into the encoded result.

4. The angiography image generation system based on noise enhancement difference according to claim 3, characterized in that, The noise estimation formula is as follows: , In the formula, For noise samples, For time steps, For slicing, The noise intensity coefficient, This represents the noise component of the slice.

5. The angiography image generation system based on noise enhancement difference according to any one of claims 2-4, characterized in that, The expression for the noise-assisted structural difference degradation operator is as follows: ， In the formula, Preliminary estimation of angiography images, For control coefficients, Due to structural differences.

6. The angiography image generation system based on noise enhancement difference according to claim 5, characterized in that, The global awareness adjustment module is configured to perform the following steps: A1: The angiography image is initially estimated and then mapped using an embedding layer to obtain embedding information; A2: Perform multiple convolutional layer mapping processes on the embedded information to obtain the convolutional processing result; A3: The embedded information and the convolution processing result are transformed into queries and keys through linear layer operations; A4: The query and key are transformed into a spatial confidence graph and values ​​through attention layer operations; A5: Substitute the spatial confidence map and the preliminary estimate of the angiography image into the adaptive fusion calculation formula to obtain the angiography image.

7. The angiography image generation system based on noise enhancement difference according to claim 6, characterized in that, The adaptive fusion calculation formula is as follows: ， In the formula, The angiography image is shown. For spatial confidence plots.

8. A method for generating angiographic images based on noise enhancement differences, characterized in that, The angiography image generation system based on noise enhancement difference according to any one of claims 1-7 includes the following steps: S1: The parameters of the decoupled structure-texture network are optimized by solving for the optimal value of the joint loss function of image estimation and perception. S2: Acquire contrast-free angiography images of the subject using a data acquisition device to obtain a series of slices consisting of multiple contrast-free angiography images; S3: The noise samples of each slice in the slice series are fused and encoded with the conditional embedding of the slice through the decoupled structure-texture network with optimized parameters, and the encoding result is decoded by two heads to obtain the structural differences and noise components; S4: Substitute the structural differences and noise components into the noise-assisted structural difference degradation operator through the parameter-optimized decoupled structure-texture network to obtain a preliminary estimate of the angiography image; S5: The angiography image of each slice is initially estimated by the decoupled structure-texture network with optimized parameters and then subjected to global perception adjustment to obtain the angiography image of each slice.

9. The method for generating angiographic images based on noise enhancement differences according to claim 8, characterized in that, The expression for the joint loss function of image estimation and perception is as follows: ， In the formula, For the true structural differences, The true noise components of the slice. These are real angiography images. For a deep feature-based perceptual similarity measure, The value of the joint loss function for image estimation and perception. and All are weighting coefficients.

Images