Self-supervised angiogram segmentation method based on x-ray attenuation modeling and frequency domain decoupling

Abstract

This disclosure relates to a self-supervised angiography image segmentation method based on X-ray attenuation modeling and frequency domain decoupling, addressing the shortcomings of existing vessel segmentation models in terms of accuracy, robustness, and versatility in real clinical scenarios. The method first renders the radial brightness gradient of simulated vessels and their realistic anatomical morphology, giving the generated vessels realistic and interpretable visual features. Then, it accurately extracts low-frequency information from the real angiography background and high-frequency features from the vessels. Next, it performs cross-domain information fusion based on the product relationship of X-ray attenuation rates, ensuring the physical rigor of the synthesized image. Finally, by simulating dynamic attenuation changes under different contrast agent concentrations and applying consistency regularization constraints, it achieves decoupling of the vessel's geometric features from its contrast features. Experiments show that vessel-aware synthetic learning can significantly reduce segmentation errors and capture more vascular details.

Description

Technical Field

[0001] This disclosure relates to medical artificial intelligence, particularly to the field of cardiovascular angiography image segmentation, specifically a self-supervised angiography image segmentation method based on X-ray attenuation modeling and frequency domain decoupling. Background Technology

[0002] Coronary artery disease is one of the leading causes of death worldwide. In clinical treatment, percutaneous coronary intervention (PCI) is a crucial means of restoring blood flow and alleviating symptoms. This procedure is highly dependent on real-time X-ray coronary angiography (XCA). To ensure the safety and effectiveness of the procedure, physicians need to identify complex vascular anatomy, plan the guidewire path, and assess the extent of lesions in real time based on the angiographic images. Therefore, accurate and automated vascular segmentation technology is of great significance in assisting clinical decision-making. However, existing segmentation methods for X-ray angiography images all have certain technical limitations.

[0003] Supervised learning methods relying on manual annotation typically employ deep neural networks such as U-Net, using large amounts of image data with masks manually drawn by radiologists for training. While they perform well on specific datasets, they face significant annotation bottlenecks: on the one hand, the vascular structures in XCA images are extremely complex and contain numerous fine branches, making manual annotation extremely time-consuming; on the other hand, the equipment parameters and contrast agent concentrations vary among different clinical institutions, limiting the model's generalization ability and making it difficult to solve the problem of scarce labeled data through simple transfer learning.

[0004] Domain adaptation and cross-modal segmentation methods attempt to transfer knowledge from a labeled source domain (such as 3D CT images) to an unlabeled XCA target domain. However, the physical imaging principles of CT and XCA are quite different (3D voxels versus 2D projection). This cross-modal texture difference and geometric distortion makes it easy for the model to lose fine peripheral vascular information during the transfer process.

[0005] Traditional and unsupervised methods based on handcrafted features rely on pre-defined mathematical models to identify tubular structures, but they are extremely sensitive to noise and have poor robustness to overlapping blood vessels and motion artifacts. Unsupervised learning (such as clustering-based or generative adversarial networks-based methods) do not require labels, but they often struggle to capture the long-range topological connectivity of vascular trees.

[0006] Self-supervised learning methods based on fractal synthesis are currently a promising direction, generating synthetic data with vascular features through algorithms. However, existing self-supervised synthesis methods suffer from three core drawbacks. First, the geometric simulation is unrealistic. Existing vascular synthesis methods are mostly based on simplified fractal geometry, resulting in vessels that are mostly connected by straight line segments, lacking the physiological curvature and random curvature of real coronary arteries. Second, physical modeling is lacking. Existing methods typically use simple addition or linear weighting when fusing synthetic vessels with the real background, which violates the Beer-Lambert law in X-ray imaging. In actual imaging, brightness is due to the exponential decay of X-rays after penetrating tissue and contrast agents; ignoring this physical principle leads to a significant "texture gap" between the synthetic image and the real image. Third, there is the problem of feature entanglement. During training, the model often memorizes specific artifacts in the synthetic image (overfitting to the synthetic domain) and cannot extract robust geometric features under different contrast agent concentrations and background interference.

[0007] In summary, existing vessel segmentation models still face significant challenges in real-world clinical coronary artery segmentation, exhibiting deficiencies in accuracy, robustness, and versatility. Summary of the Invention

[0008] To address the aforementioned problems in the existing technology, the purpose of this disclosure is to provide a self-supervised angiography image segmentation method based on X-ray attenuation modeling and frequency domain decoupling. Under the premise of completely eliminating manual annotation, this method integrates physical imaging models with advanced frequency domain decoupling technology to construct a self-supervised learning framework that can both simulate real vascular texture and adapt to changing clinical environments.

[0009] Firstly, this disclosure proposes a self-supervised angiography image segmentation method based on X-ray attenuation modeling and frequency domain decoupling, including the following steps: segmenting the real angiography image using Fourier transform. The background signal and vascular signal are decoupled in the frequency domain to generate a background image. Synthesize blood vessel contours with geometric curvature, utilizing X-ray radiation brightness attenuation rate. Generate blood vessel texture and compare it with the background image. Synthetic angiography images ; to display real angiography images The synthesized angiography image will be used as the target image. Using the student network as the source images, features of individual pixels in the target images were obtained. Features of a single pixel in the source image Based on features and characteristics Predict the prediction mask for the output real image respectively. Predictive mask for synthetic images Calculate the adversarial learning loss of the student network Predicting mask and predictive mask An input discriminator network determines whether the input mask is a predicted mask of the real image. Or a predictive mask for synthetic images And based on the prediction mask and predictive mask Calculate the discriminator loss Adjusting the concentration of the contrast agent to obtain a contrast image Perturbed images with different contrast agent concentrations To disturb the image Input the teacher network to obtain the features of a single pixel in the corresponding perturbation image. Based on features and characteristics Constructing a concentration perturbation consistency regularization constraint Projecting the 3D blood vessel object onto a 2D plane yields the actual mask for the synthesized image. Calculate the Dice loss of the predicted mask and the synthesized image. and cross-entropy loss ; Calculate the total loss of the teacher-student segmentation network. , , and The weight parameters are used. During training, the teacher-student segmentation network is trained based on the total loss of the teacher-student segmentation network, and the discriminator network is trained based on the discriminator loss. The teacher-student segmentation network and the discriminator network are trained alternately and iteratively. After training, the acquired angiography image is input into the student network to obtain the segmented angiography image.

[0010] Secondly, this disclosure relates to a computer-readable storage medium storing a computer program that can be loaded by a processor and execute any of the methods of this disclosure.

[0011] Thirdly, this disclosure relates to a self-supervised angiography image segmentation system based on X-ray attenuation modeling and frequency domain decoupling, comprising: a background and vessel signal separation module configured to separate the real angiography image using Fourier transform. The background signal and vascular signal are decoupled in the frequency domain to generate a background image. The synthetic angiography module is configured to use synthesized vascular contours with geometric curvature, based on X-ray radiation attenuation rate. Generate blood vessel texture and compare it with the background image. Synthetic angiography images The student network was configured to display real angiographic images. The synthesized angiography image will be used as the target image. Using the student network as the source images, features of individual pixels in the target images were obtained. Features of a single pixel in the source image The teacher network is configured to regulate the concentration of the contrast agent to obtain an image consistent with the contrast agent concentration. Perturbed images with different contrast agent concentrations To disturb the image Input the teacher network to obtain the features of a single pixel in the corresponding perturbation image. The discriminator network is configured to predict the mask. and predictive mask An input discriminator network determines whether the input mask is a predicted mask of the real image. Or a predictive mask for synthetic images The training inference module is configured to train the teacher-student segmentation network based on the total loss of the teacher-student segmentation network, and train the discriminator network based on the discriminator loss, with the teacher-student segmentation network and discriminator network being trained alternately and iteratively. After training, the acquired angiography images are input into the student network to obtain segmented angiography images. The loss calculation module is configured to calculate based on features... and characteristics Predict the prediction mask for the output real image respectively. Predictive mask for synthetic images Calculate the adversarial learning loss of the student network Projecting the 3D blood vessel object onto a 2D plane yields the actual mask for the synthesized image. Calculate the prediction mask of the image and the synthesized image. Dice loss and cross-entropy loss Based on features and characteristics Constructing a concentration perturbation consistency regularization constraint ; Calculate the total loss of the teacher-student segmentation network. , , and Weight parameters; based on the prediction mask and predictive mask Calculate the discriminator loss .

[0012] Compared with the prior art, the present invention has the following advantages: 1) This invention presents a novel self-supervised student-teacher adversarial learning segmentation architecture for synthetic imaging image data, thereby achieving high-quality and low-cost image feature output and segmentation mask prediction.

[0013] 2) This invention designs a new method for synthesizing pseudo-vascular angiography images. By combining the volume rendering texture of the three-dimensional cylindrical structure of blood vessels with the topologically guided curved geometric structure, a realistic vascular angiography pattern is synthesized.

[0014] 3) This invention designs a new deconstruction algorithm for blood vessel information and background information. It uses the principle of perspective imaging to model X-ray attenuation to reduce overfitting and introduces double elliptical frequency constraints to achieve more accurate blood vessel-background decoupling than the previous rectangular constraints.

[0015] 4) This invention provides a mechanism to improve the segmentation network against contrast agent concentration perturbations. By simulating the change in contrast agent concentration in blood vessels and forcing the consistency of segmentation features of images corresponding to different contrast agent concentrations in the same blood vessel, the robustness of the segmentation model to contrast agent concentration is ensured. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0017] Figure 1 This is a flowchart of an algorithm for extracting background information from a real contrast image in one embodiment.

[0018] Figure 2 This is a flowchart of an algorithm for synthesizing perspective texture and geometric curvature characteristics in angiography images, as described in one embodiment.

[0019] Figure 3 This is a schematic diagram illustrating the learning process to resist contrast agent concentration disturbances in one implementation method.

[0020] Figure 4 This is a diagram of a self-supervised student-teacher adversarial learning segmentation architecture for synthetic imaging image data in one implementation. Detailed Implementation

[0021] Traditional self-supervised methods primarily rely on simple fractal geometry synthesis, focusing only on the topological tree structure of blood vessels while neglecting crucial physical characteristics in X-ray angiography: namely, the non-uniform tubular texture within blood vessels caused by contrast agent concentration distribution, and the complex geometric curvature induced by cardiac motion and physiological anatomy. This neglect of physical realism leads to significant differences in texture features between synthetic data and real images, making deep neural networks prone to overfitting artifacts in the synthetic domain and failing to generalize to complex clinical environments. Furthermore, existing hybridization mechanisms often employ simple additive models, violating the Beer-Lambert law (exponential decay property) that governs X-ray penetration of tissue.

[0022] To correct these model biases, this invention focuses on "physically consistent synthesis," synthesizing blood vessels from a background image obtained through frequency domain coupling. During the vessel synthesis stage, geometric curvature modeling and a multiplicative attenuation model are introduced to simulate the geometric curvature and physical attenuation patterns in real X-ray coronary angiography (XCA), forcing the model to learn truly discriminative vascular anatomical features. Based on this, the constructed vessel-aware synthesis learning framework eliminates the dependence on expensive manually labeled data and solves the "domain gap" problem between synthesized images and real images in self-supervised learning. This enables high-precision and robust coronary artery segmentation without any manual labeling, providing reliable real-time navigation feedback for interventional surgery.

[0023] The following provides a clear and complete description of how the technical solution of this case is implemented. Obviously, the described implementation methods are only a part of the implementation methods of this case, and not all of them. Based on the implementation methods in this case, all other implementation methods obtained by those skilled in the art without inventive effort are within the scope of protection of this application.

[0024] (a) Decoupling the background image from real data First, the actual angiography image is obtained through Fourier transform. The background signal and vascular signal in the angiography image are decoupled in the frequency domain to obtain the signal from the angiography image. The background information map is decoupled from the background information map, and then the background information map is enhanced by perturbing the detail information to generate the background map. Then, based on the theory of volumetric rendering (the attenuation ratio of light radiation after passing through a fixed medium is also fixed), the background image is... The 3D blood vessel object is rendered onto the background image by multiplying it with the vascular shadow, thus obtaining a synthetic angiographic image. Simultaneously, by adjusting the concentration of contrast agent in the 3D blood vessels (and thus the transparency of the 3D blood vessel object), a result consistent with the contrast image can be obtained. Perturbed images with different contrast agent concentrations .

[0025] In one implementation, the actual angiography image is transformed using Fourier transform. The background signal and vascular signal are decoupled in the frequency domain to generate a background image. The steps include: First, the FFT algorithm is used to transform the two-dimensional spatial domain image. Convert to complex spectrum By rotating the angiography image at any angle, and using the trajectories formed by the highest / lowest frequency points of the vascular pattern in the spectrogram, the frequency range of the vascular object in the spectrogram is defined. This range has inner and outer edges on the spectrogram, which are two ellipses centered at the center of the spectrogram. This can be represented by a single masking image. To describe this spectrogram region, in the occlusion map A value of 1 indicates that the frequency is within this range (the frequency range required by this solution), while a value of 0 indicates that it is not within this range. (1) in, This represents the frequency along both the horizontal and vertical axes, with each frequency... All correspond to the spectrum diagram A point on the top, and These represent the lower and upper limits of the frequency corresponding to the vascular pattern, respectively. and These represent the width and height of the image, respectively.

[0026] Complex spectrum Decomposed into amplitude spectrum Phase spectrum diagram For the phase spectrum diagram Multiply the value of each pixel by the imaginary unit i, then calculate the natural exponent, and finally compare the result with the amplitude spectrum. Multiplying the corresponding pixel values ​​yields the complex spectral value at that location:

[0027] By setting the phase position to zero and keeping the amplitude unchanged, a noise signal with the same frequency as the signal corresponding to the blood vessel pattern is obtained. This noise signal is then used to replace the signal corresponding to the blood vessel pattern, resulting in a complex spectrum diagram with a synthetic background without the blood vessel pattern. :

[0028] The complex spectrum of the synthesized background is obtained by inverse Fourier transform. Convert to composite background image .

[0029] The above steps can decompose the image information of real coronary angiography images into anatomical structures. For example... Figure 1 For example, using the FFT algorithm to transform a two-dimensional spatial domain image Convert to complex spectrum .

[0030] The principle of the FFT algorithm is Fourier transform:

[0031] Where i represents the imaginary unit of the complex number, and n This represents the two-dimensional normalized coordinate vector corresponding to each pixel in the spectrogram. This means transpose the column vector n into a row vector. This represents the frequency along the horizontal and vertical axes corresponding to each pixel in the spectrum. The previous double integration work is too cumbersome to write out, so the double integration of the FFT algorithm will be denoted by the symbol "". Thus, the Fourier transform of the FFT algorithm can be simplified as:

[0032] The specific implementation of the FFT algorithm is as follows: firstly, through linear transformation... Original image coordinates Mapping to a new coordinate range Then, the double integral is approximated using a double accumulation algorithm:

[0033] Where w represents the two-dimensional frequency (complex spectrum). (coordinates on the x-axis), n Representing a two-dimensional spatial domain image The coordinates on the [space].

[0034] Spectrum The middle part corresponds to the angiography image. The low-frequency signal corresponds to the edge, while the high-frequency signal corresponds to the edge. Traditional FDA methods use rectangular bounding boxes to define the frequency range of the target object. FDA needs to process photographic images, which have vertical and horizontal distinctions (usually determined by the positions of the sky and ground in the image). This implementation differs from FDA in that, since angiography images do not have this directional distinction, they can be rotated at any angle during data augmentation. Using the orthogonality property of the rotation matrix, it can be deduced that after rotating the spatial domain image, the spectrogram will also rotate by the same angle, that is:

[0035] in, It represents rotation at any angle.

[0036] After the angiogram image is rotated at any angle, the highest / lowest frequency points of the vascular pattern in the spectrogram (the points of the vascular object farthest / nearest from the center) will also rotate synchronously, thus forming two circular trajectories left on the spectrogram. If the length and width of the image are inconsistent, the trajectories left will be elliptical. Therefore, in this problem, the double ellipse constraint can most accurately define the frequency range of the vascular object in the spectrogram, as shown in Equation (1).

[0037] Next, in order to synthesize a background image based on the angiogram, it is best to replace the signal corresponding to the blood vessel pattern in the angiogram with a noise signal of the same frequency, thereby removing the target blood vessel from the angiogram. Therefore, a noise map with a signal similar to the source angiogram, but with patterns as unrelated as possible, is needed. Complex spectrogram According to Euler's formula, it can be decomposed into an amplitude spectrum. Phase spectrum diagram And the decomposition method is as follows:

[0038]

[0039] in, Representing complex numbers The corresponding modulus, arg, is used to obtain the principal argument value of the complex number.

[0040] The background information decoupled from the real image is The decoupling information does not contain information resonating with blood vessels. However, real-world backgrounds often contain a small amount of information resonating with blood vessels. To make the synthesized background more realistic, supplementary information resonating with blood vessels needs to be appropriately added to the background information. This enhances the synthesized background. To construct this supplementary information for background enhancement, we first need to understand the phase and amplitude in the spectrum. Spectral information includes both phase and amplitude information. Phase represents the positional structure of the target pattern, while amplitude represents the brightness contrast of the current pattern. If we set the phase to zero while keeping the amplitude unchanged, the target pattern structure will be completely destroyed, but the new pattern signal can be made very close to the source angiographic image. The information corresponding to this new pattern is the amplitude information of the original image. Extracting the portion that resonates with vascular information. As supplementary information.

[0041] Through inverse Fourier transform The pattern corresponding to the supplementary information can be parsed. See Figure 3 Supplementary image Spectrum visualization:

[0042] like Figure 1 As shown, supplementary information To enhance the background information decoupled from the real image This process yields the enhanced background image information. In other words, the background enhancement operation replaces the signal corresponding to the blood vessel pattern in the angiography image with a noise signal of the same frequency, thus obtaining a complex spectrum of the artificially synthesized background without the blood vessel pattern. :

[0043] Then, the complex spectrum of the artificially synthesized background can be obtained using the Inv-FFT algorithm. Two-dimensional spatial domain background image converted to artificially synthesized background The Inv-FFT algorithm is based on the inverse Fourier transform (...). ):

[0044] The specific implementation of the inverse Fourier transform double integral described above is also a double accumulation algorithm:

[0045] The above steps convert the image signal from the spatial domain to the frequency domain using Fourier transform. Then, a double-elliptic frequency constraint mask is used to extract the low-frequency components of the real background (representing tissue background texture) and the high-frequency components of the synthetic blood vessels (representing blood vessel edges and details). By recombining information in the frequency domain and applying rotation invariance constraints, this operation effectively decouples background interference from the target signal, prevents the deep model from overfitting to the synthetic artifacts during training, and significantly enhances the robustness of the model in complex clinical contexts.

[0046] In one embodiment, the above steps are encapsulated as an anti-overfit mixer (SORMix).

[0047] (ii) Synthetic blood vessels In existing self-supervised coronary vessel segmentation methods, the fractal vessel patterns lack curvature. The fractal vessels in this scheme can be simulated using the L-system fractal algorithm. To better simulate the real coronary vessel structure, this paper replaces the line segments in the fractal vessels with curves, such as... Figure 2 Illustration. A point on the angiogram before adding the bending rule. The conditions within a blood vessel are:

[0048] in, This indicates the center point of the blood vessel. It is a 2D unit vector representing the direction of the blood vessel axis. Indicates the direction perpendicular to the blood vessel axis. This indicates the width of the blood vessel. This indicates the length of the blood vessel.

[0049] To describe the tortuosity of blood vessels, a tortuosity rule function `fun` was designed. `fun` describes the change in the deviation of the blood vessel from its central axis as it travels from one end to the other. After adding the curve rule... After that, points in space The conditions within this blood vessel are:

[0050] Where DIST represents points The distance to the midline of the blood vessel, the curve rule fun is a domain of Continuous functions, This indicates the center point of the blood vessel. It is a 2D unit vector representing the direction of the blood vessel axis. Indicates the direction perpendicular to the blood vessel axis. This indicates the width of the blood vessel. This indicates the length of the blood vessel.

[0051] The specific expression of the curve rule fun is as follows:

[0052] Where k is the curve normalization parameter, k∈[0,1] represents a point on the curve, k=0 represents the starting point of the curve, and k=1 represents the ending point of the curve. , ,... , It is a set of parameters that control the shape of the curve, and I represents the number of these parameters. Specifically, the process of drawing each curve segment is roughly as follows: a point moves from one position to another, and the trajectory left by the point forms the curve segment. k∈[0,1] represents the distance traveled by the point. When k equals 0, it represents the starting position of the curve segment. When k equals 1, it represents the ending position of the curve segment. A value of k between 0 and 1 corresponds to the middle position on the curve segment.

[0053] Assuming the cross-section of the blood vessel is circular ( Figure 2 The distance a light ray travels in a blood vessel can be derived from the standard equation of a circle. for:

[0054] Get the travel distance for each pixel. This is equivalent to obtaining a three-dimensional tubular model of the blood vessel. To obtain the blood vessel pattern (i.e., the grayscale values ​​of each point on the image), it is necessary to calculate the distance the X-ray travels within the blood vessel corresponding to each pixel. Furthermore, it is necessary to clarify the correspondence between this distance and the radiation brightness of X-rays after they pass through the human body—this is actually the process of rendering a three-dimensional tubular model into a two-dimensional blood vessel texture image. Specifically, this rendering process is based on the Beer-Lambert Law for generating blood vessel textures. The grayscale value of each pixel in the blood vessel pattern corresponds to the radiation brightness of X-rays after they pass through the human body, while the attenuation of light radiation brightness occurs after the light passes through the object. The light absorption rate depends on the material of the object. Light brightness The distance light travels in an object Here, dt can represent the distance light travels through various tissues in the human body. Regarding the change in brightness of light before and after passing through blood vessels, dt is the term mentioned earlier. The specific relationship is as follows: Solve the above ordinary differential equation:

[0055] The derived equation This indicates that an initial radiance is The proportion by which the radiance of light decreases after passing through an object depends on the light absorption rate of the object's material. The main reason for the decrease in radiation brightness during penetration is that some of the light energy is converted into heat energy of the object.

[0056] An X-ray passes through two light absorptivity zones in succession. and For objects with radiant brightness, the attenuation rate is:

[0057] Where m represents the distance the X-ray travels after passing through the first object, and T represents the total distance the X-ray travels after passing through both objects; that is, the thickness of the first object is m, and the thickness of the second object is Tm. Assume this X-ray first passes through the object with the highest light absorption rate... The object, and then For objects with radiant brightness, the attenuation rate is:

[0058] By substituting the integral, we can deduce that these two equations are equal. Therefore, the final radiance of an X-ray passing through multiple objects is independent of the order in which the objects are passed. In actual contrast imaging, X-rays first pass through human tissue, then blood vessels, and finally the remaining human tissue. In the calculation, the light is first allowed to pass through the non-tissue areas of the body to obtain the background image, and then through the blood vessels to obtain the final radiance. This approach yields results consistent with the actual X-ray imaging process.

[0059] The concentration of contrast agent in the synthetic blood vessel determines the light absorption rate of the synthetic blood vessel object. Therefore, the light absorption rate of the blood vessel object can be used to describe the concentration of contrast agent in the synthetic blood vessel. Figure 3 Understood. When passing through a contrast agent concentration of... After the blood vessels, the attenuation rate of X-ray radiation brightness The calculation method is as follows:

[0060] in, It is the distance that light travels in blood vessels. Figure 2 The blood vessel texture synthesis mechanism calculates the attenuation rate of X-Ray radiation in the direction corresponding to each pixel after passing through the blood vessel, based on the formula mentioned above. X-Ray attenuation rate The corresponding image is essentially a description file of a 3D blood vessel object, given the contrast agent concentration. In this case, the geometry of the 3D blood vessels can be reconstructed from this image.

[0061] A method for simulating contrast agent concentration perturbations. X-ray attenuation rate. Perturbation of the corresponding images is a prerequisite for S42 consistency regularization learning. The main purpose of the perturbation is to enable the neural network to learn that "contrast agent concentration does not affect vascular geometry." In this scheme, the contrast agent concentration of the synthesized blood vessels is expressed as the light absorption rate of the vascular object. The light absorption rate of 3D blood vessel objects is increased from... Adjust to Subsequently, the attenuation rate of X-ray radiation as it passes through blood vessels. for:

[0062]

[0063]

[0064] (iii) Fusing the background with the synthetic blood vessels to form an angiographic image The analysis of the synthesized vascular texture above shows that the calculated radiation intensity, obtained by having X-rays pass through non-vascular tissue and then through blood vessels, is the same as the actual value. During this process, the X-ray radiation intensity changes from the initial intensity... Transformed into the final decayed brightness :

[0065] In this process, the initial brightness is The brightness of X-rays after passing through non-vascular tissues of the human body is :

[0066] The brightness of this X-ray after passing through non-vascular tissue in the human body It's actually a composite background image. The upper coordinate is The pixel value corresponding to each pixel. Then, this ray passes through the blood vessel tissue, and the X-ray radiation intensity attenuates to the final value. Following the physical attenuation process of X-rays sequentially passing through non-vascular tissue and blood vessels, the spatial domain background image is... With vascular attenuation rate Perform multiplicative mixing to generate a synthetic contrast image resistant to overfitting:

[0067]

[0068] The final attenuation of the X-ray radiance It is actually a composite image. The upper coordinate is The pixel value corresponding to the pixel point. Figure 1 The process of mixing the real background image with the synthetic blood vessel image involves calculating the final radiation brightness of the X-Ray in the direction corresponding to each pixel after passing through the blood vessel, based on the formula mentioned above.

[0069] The exponential decay characteristic based on X-ray imaging is used to fuse simulated vascular signals with the real background through a multiplicative decay model. This process ensures that the synthetic data is highly consistent with the real angiographic images in terms of anatomical morphology and physical texture, thus providing high-quality pseudo-label data for subsequent self-supervised training.

[0070] (iv) Model Training and Application A student-teacher architecture is constructed for synthetic angiography image data. The teacher network integrates learning weights from historical moments to provide stable pseudo-label guidance for the student network, which is trained using augmented data generated by VASL. To make the distribution of masks generated by the student network on real images indistinguishable from that of synthetic image masks, and to achieve knowledge transfer from the synthetic data domain to the real data domain, a discriminator network is introduced to determine whether the input mask is a predicted mask of the real image or a predicted mask of the synthetic image. Simultaneously, the student network employs adversarial learning.

[0071] The input to this architecture is the target image. (Actual angiography image). See also Figure 4 First, the target image is removed using Fourier transform. Background information is decoupled from the vascular structure in the image, and then supplementary information is used to enhance the background information to generate a background image. Then, based on volumetric rendering theory, the 3D blood vessel object is rendered onto the background image using a cumulative multiplication method, thereby obtaining an artificially synthesized angiographic image. Simultaneously, by adjusting the concentration of contrast agent in the 3D blood vessels (and thus the transparency of the 3D blood vessel object), a result consistent with the contrast image can be obtained. Perturbed images with different contrast agent concentrations Furthermore, by projecting the 3D blood vessel object onto a 2D plane, the actual mask for the synthesized image can be obtained. The framework employs a student-teacher dual neural network model, where both the student and teacher models use the same U-Net backbone network. (Unlabeled target image) and the source image of the synthesis Both methods use student networks for analysis, thereby obtaining the features of individual pixels in the target image. Features of a single pixel in the source image Perturbation image The semantic features of individual pixels in the perturbation image are obtained through the teacher network. By applying fully connected layers and softmax to the semantic features of each pixel, a blood vessel-background score is obtained for that pixel. Pixels with a score higher than 0.5 are predicted as blood vessels, otherwise they are predicted as background. By scoring and predicting the features of the target image and the source image, prediction masks for the real image can be obtained respectively. Predictive mask for synthetic images .

[0072] The overall framework integrates Dice loss, cross-entropy loss, consistency regularization loss, and adversarial loss, forming an optimized closed loop for coronary angiography image segmentation tasks. Through this self-supervised feedback mechanism, the model can automatically extract vascular features from unlabeled angiography video sequences without any manual annotation, ultimately achieving accurate vessel segmentation. This approach is applicable to coronary artery segmentation.

[0073] Specifically, the 3D blood vessel object is projected onto a 2D plane to obtain the actual mask for the synthesized image. Segmentation loss of synthetic images The difference between the predicted mask and the actual mask of the synthesized image was measured. The segmentation loss of this synthesized image can learn the perspective projection rules of tubular tree-like objects from artificially synthesized labeled data. This segmentation loss... Specifically, it is a weighted sum of Dice loss and cross-entropy loss:

[0074] in, These are the weight parameters.

[0075] For blood vessel patterns Perturbation is a prerequisite for consistency regularization learning; neural networks learn from mixed imaging images. semantic features extracted Compared with concentration perturbation images Features extracted Features are constrained by concentration perturbation consistency regularization loss function. and characteristics Both are kept as consistent as possible to encourage the neural network to learn the physical prior knowledge that changes in contrast agent concentration do not alter the vascular geometry. In terms of implementation, the MSE consistency function and the cosine consistency function were tested. Experiments showed that the loss constraint based on the cosine consistency function performed better. Therefore, the cosine form was ultimately chosen to implement the concentration perturbation consistency regularization constraint.

[0076] The concentration perturbation consistency regularization loss measures the difference between the semantic features of the source image and the perturbed image. This consistency constraint teaches the neural network that "contrast agent concentration does not affect vascular geometry," thus improving the framework's robustness. By simulating vascular imaging effects under different attenuation coefficients, pairs of the same blood vessel samples with different contrast features are generated. Then, by introducing the concentration perturbation consistency regularization loss function, the model is forced to extract pure geometric topological features in the feature space that are independent of brightness / contrast. This strategy ensures that the segmentation network maintains stable vascular extraction capabilities even when faced with extremely low contrast agent concentrations or rapid metabolism. This contrast-invariant learning strategy overcomes the impact of contrast agent concentration fluctuations on segmentation accuracy in clinical settings.

[0077] The discriminator network's ability to analyze and distinguish the differences in blood vessel contours between masks in real images and masks in synthetic images is acquired through training the discriminator loss. The discriminator loss takes the form of binary cross-entropy.

[0078] Adversarial loss is used to ensure that the predicted masks of real images generated by the student network are as indistinguishable from the predicted masks of synthetic images as possible. Therefore, adversarial learning enables the student network to transfer knowledge learned from synthetic images to real images. The specific form of this adversarial loss is information entropy:

[0079] Calculate the total loss of the teacher-student segmentation network. , and These are the weight parameters.

[0080] During training, the teacher-student segmentation network is used. Optimization is performed, while the discriminator network uses... The teacher-student segmentation network and the discriminator network are optimized alternately. The optimization goal of the teacher-student segmentation network is to generate high-quality segmentation maps regardless of whether the data is real or synthetic, while the optimization goal of the discriminator is to identify the differences between synthetic and real data.

[0081] During training, the teacher-student segmentation network is trained based on the total loss of the teacher-student segmentation network, the discriminator network is trained based on the discriminator loss, and the teacher-student segmentation network is trained alternately (using...). (for optimization) and discriminator network (using (Optimize).

[0082] After training, the acquired angiography images are input into the student network to obtain segmented angiography images.

[0083] (V) Experiment Extensive experimental validation was conducted on two public datasets and one internal dataset, and quantitative comparative analysis was performed with several existing mainstream methods.

[0084] This experiment evaluated the model on three X-ray coronary angiography (XCA) datasets: the public datasets XCAD and XACV, and the internal dataset DSA_Tongji. As shown in Table 1, XCAD and XACV provide expert-annotated vascular ground truths, while the annotations for DSA_Tongji were manually completed by experts with medical knowledge. XCAD and DSA_Tongji are static image datasets, while XACV is a video sequence dataset. Testing on the XACV dataset involved directly using a model pre-trained on the XCAD dataset for cross-domain evaluation without any fine-tuning or domain adaptation, to verify the strong generalization ability of the invention.

[0085] Table 1: Feature Statistics of XCAD, DSA_Tongji and XACV Datasets

[0086] The evaluation metrics used are five commonly used metrics in the field of medical image segmentation: Jaccard index (JC, i.e., intersection-overlap ratio), F1 score (Dice coefficient), sensitivity (Sn), specificity (Sp), and accuracy (Accuracy). Among them, JC measures overlap, F1 balances precision and recall, Sn assesses vascular integrity, and Sp measures background detection capability.

[0087] In this embodiment of the invention, the topology of the synthetic blood vessels is generated using the L-system fractal iterative algorithm. The total loss function during the student-teacher network training process is defined as...

[0088] After grid search and empirical tuning, the weight hyperparameters were set as follows: , , 0.25. During the background information extraction stage, the frequency threshold of the double-elliptical frequency constraint mask is set to: lower bound... Upper Realm .

[0089] The experimental results are shown below: The self-supervised method proposed in this invention is quantitatively compared with traditional mathematical filtering methods (Frangi), unsupervised methods (DeNVeR), and the latest cutting-edge self-supervised methods (DARL, FreeCOS, etc.).

[0090] As shown in Table 2, the method of this invention achieved the best performance on the three major datasets: JC, F1, Sn, and Acc. In particular, it achieved a significant improvement in sensitivity (Sn), which is attributed to the realistic simulation of the morphology of small distal branches by Vascular Sensing Synthesis Learning (VASL) and the effective resistance of the anti-synthesis overfit mixer (SORMix) to artifacts in complex clinical backgrounds.

[0091] Table 2: Comparison of quantitative evaluations of the present invention and existing methods on three major datasets.

[0092] Since FreeCOS was the best-performing method without manual annotation before, it was used as the baseline method. Table 3 shows a comparison of the significance improvement of the index between the method of this invention and the baseline method based on the saliency pairwise T-detection algorithm. The results show that the index improvement caused by the present invention is sufficiently significant and eliminates the interference of random perturbations of the index on the conclusions.

[0093] Table 3: Paired-samples T-test: Significance analysis of the difference between the present invention and the baseline method

[0094] To verify the independent contributions of each innovative module in this invention (Vascular Sensing Synthesis Learning (VASL), Anti-overfitting Synthesis Mixer SORMix, Domain Adaptive ADA, and Contrast Invariant Learning (CIL)), ablation experiments were designed as shown in Table 4. The ablation results show that removing VASL leads to a significant performance drop, demonstrating the criticality of introducing 3D curvature and physical attenuation texture synthesis; removing SORMix (degenerating into additive attenuation mixing) leads to a significant decrease in JC and F1, demonstrating the necessity of the multiplicative physical attenuation mixing mechanism; and the application of contrast invariant learning (CIL) and a specific cosine loss function independently improves the network's robustness.

[0095] Table 4: Ablation Experiment Results of Model Components (Based on XCAD Dataset)

[0096] The above experiments fully demonstrate that the self-supervised angiography image segmentation method based on X-ray attenuation modeling and frequency domain decoupling provided by this invention can effectively avoid dependence on manual annotation, and effectively bridge the domain gap between the synthetic domain and the real clinical environment through high-precision synthesis and reconstruction at the geometric structure and optical attenuation levels. The model converges faster, and its ability to detect small and terminal blood vessels is significantly stronger than existing methods, meeting the application requirements of high precision, real-time operation, and low deployment cost in clinical interventional surgery.

[0097] This invention can be a system, method, and / or computer program product.

[0098] For example, a self-supervised angiography image segmentation method based on X-ray attenuation modeling and frequency domain decoupling includes the following steps: segmenting the real angiography image using Fourier transform. The background signal and vascular signal are decoupled in the frequency domain to generate a background image. Synthesize blood vessel contours with geometric curvature, utilizing X-ray radiation brightness attenuation rate. Generate blood vessel texture and compare it with the background image. Synthetic angiography images ; to display real angiography images The synthesized angiography image will be used as the target image. Using the student network as the source images, features of individual pixels in the target images were obtained. Features of a single pixel in the source image Based on features and characteristics Predict the prediction mask for the output real image respectively. Predictive mask for synthetic images Calculate the adversarial learning loss of the student network Predicting mask and predictive mask An input discriminator network determines whether the input mask is a predicted mask of the real image. Or a predictive mask for synthetic images And based on the prediction mask and predictive mask Calculate the discriminator loss Adjusting the concentration of the contrast agent to obtain a contrast image Perturbed images with different contrast agent concentrations To disturb the image Input the teacher network to obtain the features of a single pixel in the corresponding perturbation image. Based on features and characteristics Constructing a concentration perturbation consistency regularization constraint Projecting the 3D blood vessel object onto a 2D plane yields the actual mask for the synthesized image. Calculate the Dice loss of the predicted mask and the synthesized image. and cross-entropy loss ; Calculate the total loss of the teacher-student segmentation network. , , and The weight parameters are used. During training, the teacher-student segmentation network is trained based on the total loss of the teacher-student segmentation network, and the discriminator network is trained based on the discriminator loss. The teacher-student segmentation network and the discriminator network are trained alternately and iteratively. After training, the acquired angiography image is input into the student network to obtain the segmented angiography image.

[0099] For example, a self-supervised angiography image segmentation system based on X-ray attenuation modeling and frequency domain decoupling includes: a background and vessel signal separation module configured to separate the real angiography image using Fourier transform. The background signal and vascular signal are decoupled in the frequency domain to generate a background image. The synthetic angiography module is configured to use synthesized vascular contours with geometric curvature, based on X-ray radiation attenuation rate. Generate blood vessel texture and compare it with the background image. Synthetic angiography images The student network was configured to display real angiographic images. The synthesized angiography image will be used as the target image. Using the student network as the source images, features of individual pixels in the target images were obtained. Features of a single pixel in the source image The teacher network is configured to regulate the concentration of the contrast agent to obtain an image consistent with the contrast agent concentration. Perturbed images with different contrast agent concentrations To disturb the image Input the teacher network to obtain the features of a single pixel in the corresponding perturbation image. The discriminator network is configured to predict the mask. and predictive mask An input discriminator network determines whether the input mask is a predicted mask of the real image. Or a predictive mask for synthetic images The training inference module is configured to train the teacher-student segmentation network based on the total loss of the teacher-student segmentation network, and train the discriminator network based on the discriminator loss, with the teacher-student segmentation network and discriminator network being trained alternately and iteratively. After training, the acquired angiography images are input into the student network to obtain segmented angiography images. The loss calculation module is configured to calculate based on features... and characteristics Predict the prediction mask for the output real image respectively. Predictive mask for synthetic images Calculate the adversarial learning loss of the student network Projecting the 3D blood vessel object onto a 2D plane yields the actual mask for the synthesized image. Calculate the prediction mask of the image and the synthesized image. Dice loss and cross-entropy loss Based on features and characteristics Constructing a concentration perturbation consistency regularization constraint ; Calculate the total loss of the teacher-student segmentation network. , , and Weight parameters; based on the prediction mask and predictive mask Calculate the discriminator loss .

[0100] Computer program products may include computer-readable storage media having computer-readable program instructions loaded thereon for causing a processor to implement various aspects of the present invention. The storage medium may be implemented by any type of non-volatile storage device, or a combination thereof. The non-volatile memory may be read-only memory (ROM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), magnetic random access memory (FRAM), flash memory, magnetic surface memory, optical disc, or compact disc read-only memory (CD-ROM); the magnetic surface memory may be a disk storage device or a magnetic tape storage device. The storage media described in the embodiments of the present invention are intended to include, but are not limited to, these and any other suitable types of memory.

[0101] In the several embodiments provided by this invention, it should be understood that the disclosed systems and methods can be implemented in other ways. The device embodiments described above are merely illustrative. For example, the division of units is only a logical functional division, and in actual implementation, there may be other division methods, such as: multiple units or components can be combined, or integrated into another system, or some features can be ignored or not executed. In addition, the coupling, direct coupling, or communication connection between the various components shown or discussed can be through some interfaces, and the indirect coupling or communication connection between devices or units can be electrical, mechanical, or other forms.

[0102] The units described above as separate components may or may not be physically separate. The components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the units may be selected to achieve the purpose of this embodiment according to actual needs.

[0103] In addition, in the various embodiments of the present invention, each functional unit can be integrated into one processing unit, or each unit can be a separate unit, or two or more units can be integrated into one unit; the integrated unit can be implemented in hardware or in the form of hardware plus software functional units.

[0104] Those skilled in the art will understand that all or part of the steps of the above method embodiments can be implemented by hardware related to program instructions. The aforementioned program can be stored in a computer-readable storage medium. When the program is executed, it performs the steps of the above method embodiments. The aforementioned storage medium includes various media capable of storing program code, such as mobile storage devices, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0105] Alternatively, if the integrated units of this invention are implemented as software functional modules and sold or used as independent products, they can also be stored in a computer-readable storage medium. Based on this understanding, the technical solutions of the embodiments of this invention, or the parts that contribute to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the methods described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as mobile storage devices, ROM, RAM, magnetic disks, or optical disks.

[0106] The methods disclosed in the several method embodiments provided by this invention can be arbitrarily combined without conflict to obtain new method embodiments.

[0107] The features disclosed in the several product embodiments provided by this invention can be arbitrarily combined without conflict to obtain new product embodiments.

[0108] The features disclosed in the several method or device embodiments provided by the present invention can be arbitrarily combined without conflict to obtain new method or device embodiments.

[0109] The above description, in conjunction with specific preferred embodiments, provides a further detailed explanation of the present invention. It should not be construed that the specific implementation of the present invention is limited to these descriptions. For those skilled in the art, various equivalent substitutions or obvious modifications can be made without departing from the concept of the present invention, and all such modifications, achieving the same performance or application, should be considered within the scope of protection of the present invention.

[0110] The various embodiments of the present invention have been described above. These descriptions are exemplary and not exhaustive, nor are they limited to the disclosed embodiments. Many modifications and variations will be apparent to those skilled in the art without departing from the scope and spirit of the described embodiments. The terminology used herein is chosen to best explain the principles, practical application, or technical improvements to the embodiments in the market, or to enable others skilled in the art to understand the embodiments disclosed herein. The scope of the invention is defined by the appended claims.

Claims

1. A self-supervised angiography image segmentation method based on X-ray attenuation modeling and frequency domain decoupling, characterized in that, Includes the following steps: Using Fourier transform to obtain the real angiography image The background signal and vascular signal are decoupled in the frequency domain to generate a background image. ; Synthesize blood vessel contours with geometric curvature using X-ray radiation brightness attenuation rate. Generate blood vessel texture and compare it with the background image. Synthetic angiography images ; Real angiography images The synthesized angiography image will be used as the target image. Using the student network as the source images, features of individual pixels in the target images were obtained. Features of a single pixel in the source image Based on features and characteristics Predict the prediction mask for the output real image respectively. Predictive mask for synthetic images Calculate the adversarial learning loss of the student network ; Predicting mask and predictive mask An input discriminator network determines whether the input mask is a predicted mask of the real image. Or a predictive mask for synthetic images And based on the prediction mask and predictive mask Calculate the discriminator loss ; By adjusting the concentration of the contrast agent, a contrast image can be obtained. Perturbed images with different contrast agent concentrations To disturb the image Input the teacher network to obtain the features of a single pixel in the corresponding perturbation image. Based on features and characteristics Constructing a concentration perturbation consistency regularization constraint ; The actual mask for the synthesized image is obtained by projecting the 3D blood vessel object onto a 2D plane. Calculate the Dice loss of the predicted mask and the synthesized image. and cross-entropy loss ; Calculate the total loss of the teacher-student segmentation network. , , and These are weight parameters; During training, the teacher-student segmentation network is trained based on the total loss of the teacher-student segmentation network, and the discriminator network is trained based on the discriminator loss. The teacher-student segmentation network and the discriminator network are trained alternately and iteratively. After training, the acquired angiography images are input into the student network to obtain segmented angiography images.

2. The method according to claim 1, characterized in that, Using Fourier transform to obtain the real angiography image The background signal and vascular signal are decoupled in the frequency domain to generate a background image. The steps include: Using the FFT algorithm to transform the two-dimensional spatial domain image Convert to complex spectrum By rotating the angiography image at any angle, and using the trajectories formed by the highest / lowest frequency points of the vascular pattern in the spectrogram, the frequency range of the vascular object in the spectrogram can be defined: in, It is a spectrum diagram coordinates on and These represent the lower and upper limits of the frequency corresponding to the vascular pattern, respectively. and These represent the width and height of the image, respectively. Complex spectrum Decomposed into amplitude spectrum Phase spectrum diagram , By setting the phase position to zero while keeping the amplitude constant, a noise signal with the same frequency as the signal corresponding to the blood vessel pattern is obtained. This noise signal is then used to replace the signal corresponding to the blood vessel pattern, resulting in a complex spectrum diagram with a synthetic background without the blood vessel pattern. : The complex spectrum of the synthesized background is obtained by inverse Fourier transform. Convert to composite background image .

3. The method according to claim 1, characterized in that, The steps for synthesizing blood vessel contours with geometric curvature include: A point on the imaging image The conditions in a segment of a fractal blood vessel are: Where DIST represents points The distance to the midline of the blood vessel, the curve rule fun is a domain of Continuous functions, This indicates the center point of the blood vessel. It is a 2D unit vector representing the direction of the blood vessel axis. Indicates the direction perpendicular to the blood vessel axis. This indicates the width of the blood vessel. This indicates the length of the blood vessel; For the simulated blood vessels, simulated physiological geometric distortions are randomly introduced to generate the final blood vessel contour; The specific expression of the curve rule fun is as follows: Where k is the curve normalization parameter, k∈[0,1], k=0 indicates the starting point of the curve, and k=1 indicates the ending point of the curve. , ,... , It is a set of parameters that control the shape of the curve, and I represents the number of these parameters.

4. The method according to claim 1, characterized in that, Using X-Ray radiance attenuation rate The steps to generate blood vessel texture include: Calculate the distance light travels in blood vessels : , This indicates the width of the blood vessel, and DIST represents the point. Distance to the central axis of the blood vessel; Calculate the X-ray concentration passing through the contrast agent. After the blood vessels, the attenuation rate of X-ray radiation brightness : 。 5. The method according to claim 1, characterized in that, Angiography image Background image X-Ray radiance attenuation rate Multiplicative mixtures.

6. The method according to claim 1, characterized in that, The student network and the teacher network use the same U-Net backbone network.

7. The method according to claim 1, characterized in that, By adjusting the attenuation coefficient of X-ray, pairs of the same blood vessel samples with different contrast characteristics are generated, and the features of the sample pairs are calculated. and Based on features and characteristics Constructing a concentration perturbation consistency regularization constraint : E represents the expectation, through The model is forced to extract features that are independent of contrast within the feature space.

8. A computer-readable storage medium, characterized in that: The computer program is stored that can be loaded by a processor and executed according to any one of claims 1 to 7.

9. A self-supervised angiography image segmentation system based on X-ray attenuation modeling and frequency domain decoupling, characterized in that, include: Background and vascular signal separation module: configured to separate the real angiographic image from the vascular signal using Fourier transform. The background signal and vascular signal are decoupled in the frequency domain to generate a background image. ; The synthetic angiography module is configured to use synthesized vascular contours with geometric curvature, based on X-ray radiation attenuation rate. Generate blood vessel texture and compare it with the background image. Synthetic angiography images ; The student network was configured to display real angiographic images. The synthesized angiography image will be used as the target image. Using the student network as the source images, features of individual pixels in the target images were obtained. Features of a single pixel in the source image ; The teacher network is configured to regulate the concentration of the contrast agent to obtain an image consistent with the contrast agent concentration. Perturbed images with different contrast agent concentrations To disturb the image Input the teacher network to obtain the features of a single pixel in the corresponding perturbation image. ; The discriminator network is configured to predict the mask. and predictive mask An input discriminator network determines whether the input mask is a predicted mask of the real image. Or a predictive mask for synthetic images ; The training inference module is configured to train the teacher-student segmentation network based on the total loss of the teacher-student segmentation network and the discriminator network based on the discriminator loss, and to train the teacher-student segmentation network and the discriminator network alternately and iteratively. After training, the acquired angiography image is input into the student network to obtain the segmented angiography image. The loss calculation module is configured to be feature-based. and characteristics Predict the prediction mask for the output real image respectively. Predictive mask for synthetic images Calculate the adversarial learning loss of the student network Projecting the 3D blood vessel object onto a 2D plane yields the actual mask for the synthesized image. Calculate the prediction mask of the image and the synthesized image. Dice loss and cross-entropy loss ; Based on features and characteristics Constructing a concentration perturbation consistency regularization constraint ; Calculate the total loss of the teacher-student segmentation network. , , and Weight parameters; based on the prediction mask and predictive mask Calculate the discriminator loss .

Images