A temporal super-resolution imaging method for angiography images

By constructing temporal prior prediction images and encoding them into latent space representations, and combining residual effectiveness weights and reliability assessments, the limitations of traditional methods in dynamic computed tomography angiography sequences are overcome. This approach achieves small corrections in vascular regions of interest and stability in non-vascular regions, thereby improving the reliability and accuracy of medical image reconstruction.

CN123118359APending Publication Date: 2026-11-17INST OF COMPUTING TECH CHINESE ACAD OF SCI +1
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611416751.0
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-09-14
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

In dynamic computed tomography angiography sequences, existing technologies lack the ability of traditional image interpolation methods to locally adaptively correct errors. Motion compensation or registration-driven methods can easily confuse contrast agent enhancement changes with spatial displacement. Methods based on GANs and diffusion models can produce uncontrollable modifications in medical images, resulting in insufficient reliability and stability of temporal super-resolution in angiography sequences.

Method used

By constructing a temporal prior prediction image and encoding it into a latent space representation, a temporal residual prediction model is used to predict candidate latent space residuals and reliability scores. The residuals are combined with effective weights for amplitude-limited effective processing and then superimposed onto the temporal prior latent space representation for decoding. By combining reliability judgment and backoff mechanism, the depth model is ensured to make small corrections in the blood vessel region of interest and maintain stability in non-blood vessel regions.

Benefits of technology

It improves the reliability and stability of temporal super-resolution of angiography sequences, reduces uncontrollable modifications to background tissue and non-target areas, and enhances the accuracy and controllability of medical image reconstruction.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123118359A_ABST
    Figure CN123118359A_ABST
Patent Text Reader

Abstract

This invention provides a temporal super-resolution imaging method for angiography images, comprising: acquiring an angiography sequence and its acquisition time, and identifying visible observation frames and target time points therefrom; constructing a temporal prior prediction image of the target time point based on the visible observation frames and their acquisition times; encoding the visible observation frames and the temporal prior prediction image into a visible frame latent space representation and a temporal prior latent space representation, respectively; determining candidate latent space residuals and reliability scores using a temporal residual prediction model based on the visible frame latent space representation, target time point information, target time point mask, and temporal prior latent space representation; determining residual effectiveness weights based on the reliability scores, and performing amplitude-limited effectiveness processing on the candidate latent space residuals based on the residual effectiveness weights to obtain effective latent space residuals; superimposing the effective latent space residuals onto the temporal prior latent space representation to obtain a predicted latent space representation, and decoding the predicted latent space representation to obtain a candidate image of the target time point.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of medical image processing, specifically to the field of dynamic medical image sequence reconstruction, and more specifically, to a temporal super-resolution imaging method for angiography images. Background Technology

[0002] Dynamic computed tomography (CT) angiography sequences acquire or reconstruct angiographic images at multiple time points to observe the entry, peak enhancement, and drainage of contrast agents within blood vessels. Unlike ordinary natural video, the main temporal changes in dynamic CT angiography sequences are not necessarily large-scale object motions, but often local grayscale or density changes caused by variations in contrast agent concentration within the blood vessels. For computed tomography perfusion imaging, to reduce radiation dose or shorten the acquisition process, the actual number of dynamic frames acquired may be limited, thus sacrificing temporal resolution. Therefore, in dynamic CT angiography sequences or angiography sequences derived from computed tomography perfusion imaging, how to reconstruct images of the target time point from a limited number of observation frames is a crucial issue in temporal super-resolution of dynamic medical imaging sequences.

[0003] From the perspective of more general temporal image reconstruction and super-resolution techniques, video frame interpolation (VFI) typically refers to synthesizing one or more intermediate frames between existing adjacent frames to improve the temporal resolution of the sequence. In recent years, video frame interpolation methods have evolved from traditional motion compensation methods to various deep learning schemes based on optical flow estimation, adaptive convolutional kernels, deep feature alignment, Generative Adversarial Networks (GANs), Transformer models, and diffusion models. Depending on the form of the target time point, related methods can also be divided into center-time interpolation and arbitrary-time interpolation.

[0004] The first category of existing methods can be summarized as traditional image interpolation methods. These methods include spatial linear interpolation, piecewise cubic Hermite conformal interpolation, spline interpolation, polynomial interpolation, and interpolation based on time-intensity curves. Their common characteristic is that they estimate the voxel values ​​at the target time point based on the voxel values ​​at known time points according to a pre-defined mathematical curve, without explicitly estimating the motion field of blood vessels or tissues, nor synthesizing new local structures through depth models. The advantages of these methods are their simplicity, high stability, strong interpretability, and low computational complexity. Especially in dynamic computed tomography angiography scenarios with short time intervals, minimal spatial structural motion, and where the main changes originate from contrast agent enhancement, traditional image interpolation can often produce stable temporal prior prediction results. However, these methods typically only predict pre-defined fixed change patterns and are difficult to adaptively correct for different change patterns in local vascular branches, contrast agent arrival delays, abnormal perfusion in lesion areas, or different change patterns in the region of interest. Therefore, when the enhancement curves of different vascular regions exhibit non-uniform changes, relying solely on traditional image interpolation methods may not accurately recover the vascular enhancement details at the target time point.

[0005] The second category of existing methods can be summarized as motion-deformation driven temporal super-resolution methods. These methods estimate the optical flow, deformation field, or registration transformation between adjacent frames, deform the reference frame to the target time point, and then fuse multiple deformed reference frames to obtain the image at the target time point. In general video frame interpolation, optical flow or deformation methods have the advantage of explicitly describing pixel-level displacement, possessing a certain degree of interpretability, and facilitating the reconstruction of intermediate images at arbitrary time points. However, the performance of these methods is highly dependent on the accuracy of motion estimation or registration results. When there is occlusion, large motion, brightness changes, or nonlinear deformation, motion estimation errors directly affect the quality of the reconstructed image. For dynamic computed tomography angiography sequences, motion-deformation driven methods have additional limitations. In dynamic computed tomography angiography sequences, the vascular morphology between adjacent frames is usually relatively stable; the key changes come more from changes in intravascular contrast agent concentration than from large displacements of the vascular structure itself. If the algorithm misinterprets the grayscale changes caused by contrast agent enhancement as spatial motion or tissue deformation, it may introduce unnecessary deformation corrections, leading to vessel edge mismatch, local artifacts in vessel branches, or distortion of the vessel enhancement curve. Therefore, while this type of method is suitable for sequences with significant spatial motion, it is not necessarily suitable for temporal super-resolution tasks in dynamic computed tomography angiography sequences that rely on contrast agent enhancement changes.

[0006] The third category of existing methods can be summarized as temporal super-resolution methods based on Generative Adversarial Networks (GANs). These methods typically involve a generator synthesizing images of the target time point based on adjacent frames, multi-frame conditions, or temporal conditions, while a discriminator constrains the reconstruction results to more closely approximate real data in terms of image distribution, local texture, or temporal continuity. Compared to traditional interpolation and simple motion compensation methods, GAN-based methods have stronger texture completion and detail synthesis capabilities, enabling the acquisition of visually natural intermediate time point images even with insufficient input information or complex local variations. However, the adversarial training objective of GANs primarily constrains overall distribution and perceptual realism, and does not inherently guarantee pixel-level, voxel-level, or vascular enhancement curve consistency at the target time point. For dynamic computed tomography angiography sequences, visually natural vascular texture does not equate to diagnostic reliability; when training data is limited, the discriminator focuses more on global appearance, or the model lacks medical region constraints, the model may generate pseudo-vascular enhancements in locations with insufficient evidence, or excessively smooth real small branches, stenosis, occlusion, and local enhancement delays. Therefore, temporal super-resolution methods based solely on GANs still require temporal prior preservation, blood vessel region of interest constraints, and reliability control mechanisms.

[0007] The fourth category of existing methods can be summarized as temporal super-resolution methods based on diffusion models. These methods typically learn the data distribution of the image or latent space representation through progressive noise addition and denoising, and then gradually recover the target time point image by combining adjacent frames, temporal conditions, masking conditions, or latent space conditions during the inference phase. Diffusion models have strong distribution modeling and conditional reconstruction capabilities, are generally more stable in training processes than GANs, and are also easier to guide reconstruction results through conditional inputs. However, the multi-step sampling process and high degrees of freedom of diffusion models also bring the problem of difficulty in guaranteeing the reliability of medical image reconstruction. Without reliable temporal prior preservation, vascular region constraints, residual amplitude constraints, and backoff mechanisms in case of insufficient reliability, the model may still generate non-existent vascular enhancements in areas with insufficient evidence, or erase real vascular details, thereby reducing the reliability of medical image reconstruction. In the fields of medical dynamic computed tomography or computed tomography perfusion imaging, some studies have also attempted to use deep convolutional neural networks to recover temporally downsampled dynamic image sequences. For example, a dynamic computed tomography perfusion sequence with a high temporal sampling rate can be downsampled in time, and then a residual deep convolutional neural network can be used to recover the high temporal sampling rate sequence. Furthermore, perfusion parameters such as cerebral blood flow, cerebral blood volume, mean transit time, and time to peak flow can be calculated. This type of method demonstrates that deep learning can be used to recover low temporal resolution dynamic medical image sequences, but its focus is usually on directly recovering missing images or perfusion parameters.

[0008] The main drawback of traditional temporal intensity interpolation methods is that they typically assume the intensity changes of each pixel, voxel, or local region can be approximated by a pre-defined mathematical curve. This method does not explicitly model the vascular tree structure, nor does it distinguish between regions of interest (ROIs) and non-vascular regions. In dynamic computed tomography (CT) angiography sequences, the arrival time of contrast agent at different vascular branches may vary, and the enhancement changes in local lesion areas may differ from those in normal vascular areas. Simply relying on linear, spline, or other deterministic curve interpolation makes it difficult to specifically correct for these local differences.

[0009] The main drawback of motion-compensated or registration-driven methods lies in their core assumption that the target time-point image can be obtained by spatially shifting, deforming, or registering it with a reference frame. This assumption applies to video or medical sequences with significant structural motion, but the main changes in dynamic computed tomography angiography sequences are often changes in contrast agent concentration within the vessels, rather than significant movement of the vessel structure itself. When the algorithm misinterprets changes in contrast agent enhancement as spatial motion, it may produce erroneous deformations, leading to mismatches at vessel edges, branches, or localized enhancement areas.

[0010] The main drawback of GAN-based and diffusion-based methods is that while the model's expressive power is enhanced, it also increases the risk of unconstrained reconstruction and over-modification, leading to decreased reliability. For medical imaging, the reconstruction results should not only aim for visual naturalness or a reduction in average error, but also avoid producing false vascular enhancement, erroneous enhancement, or erasure of real vascular details in areas with insufficient evidence.

[0011] Furthermore, existing methods often fail to fully utilize the characteristics of dynamic computed tomography angiography sequences. Key changes in these sequences are typically concentrated in vessels and their surrounding enhanced regions, while non-vascular background areas should largely remain stable. If the depth model is modified without constraints across the entire image, it may introduce irrelevant changes to bony structures, background tissues, noisy textures, or non-target regions, thereby reducing the reliability of medical image reconstruction.

[0012] In summary, among existing technologies, traditional temporal intensity interpolation methods are stable but lack local adaptive correction capabilities; motion compensation or registration-driven methods are suitable for scenarios with significant structural motion, but may confuse contrast agent enhancement changes with spatial displacement in dynamic computed tomography angiography sequences; temporal super-resolution methods based on GANs and diffusion models have strong expressive capabilities, but are prone to producing uncontrollable modifications in medical images.

[0013] It should be noted that the background information presented here is only for illustrating relevant information about the present invention to aid in understanding the technical solution of the present invention, and does not imply that the relevant information is necessarily prior art. The relevant information was submitted and disclosed together with the present invention, and should not be considered prior art unless there is evidence that the relevant information was disclosed before the filing date of the present invention. Summary of the Invention

[0014] Therefore, the purpose of this invention is to overcome the shortcomings of the prior art and provide a temporal super-resolution imaging method for angiography images.

[0015] The objective of this invention is achieved through the following technical solution:

[0016] According to a first aspect of the present invention, a temporal super-resolution imaging method for angiography images is provided, comprising: acquiring an angiography sequence and its acquisition time, and identifying visible observation frames and target time points therefrom; constructing a temporal prior prediction image of the target time point based on the visible observation frames and their acquisition times; encoding the visible observation frames and the temporal prior prediction image into a visible frame latent space representation and a temporal prior latent space representation, respectively; determining candidate latent space residuals and corresponding reliability scores using a temporal residual prediction model based on the visible frame latent space representation, target time point information, target time point mask, and temporal prior latent space representation; determining residual activation weights based on the reliability scores, and performing amplitude-limited activation processing on the candidate latent space residuals based on the residual activation weights to obtain activated latent space residuals; superimposing the activated latent space residuals onto the temporal prior latent space representation to obtain a predicted latent space representation, and decoding the predicted latent space representation to obtain a candidate image of the target time point. This scheme can achieve at least the following beneficial technical effects: by first constructing a temporal prior prediction image and encoding it into a temporal prior latent space representation, and then using a temporal residual prediction model to predict candidate latent space residuals and reliability scores in the latent space, and after the residual effective weights are processed with amplitude-limited effectiveness, the images are superimposed and decoded, so that the depth model changes from freely reconstructing the entire image to performing controlled local corrections based on a stable temporal prior. This balances the stability of traditional interpolation methods with the local expressive power of the depth model, reduces the rewriting of background tissue, bony structures and non-target regions by the model, and improves the reliability of temporal super-resolution of angiography sequences.

[0017] Optionally, the method further includes: performing a reliability judgment on the candidate image at the target time point; when the candidate image meets a preset reliability condition, using it as the final output predicted image at the target time point; when the candidate image does not meet the preset reliability condition, using the temporal prior prediction image or the image after decoding the temporal prior latent space representation as the final output predicted image at the target time point. This scheme can achieve at least the following beneficial technical effects: by performing a reliability judgment on the candidate image at the target time point, outputting the candidate result when the candidate correction meets the reliability condition, and backing down to the temporal prior prediction image or the image after decoding the temporal prior latent space representation when the condition is not met, thus avoiding the unreliable candidate latent space residual from destroying the relatively stable temporal prior prediction result, ensuring that the final output predicted image at the target time point is not inferior to the temporal prior prediction result, and improving the stability and controllability of temporal super-resolution of medical images.

[0018] Optionally, the temporal residual prediction model is trained as follows: the true target time point image at the target time point in the angiography sequence sample is encoded as a target latent space representation; a target latent space residual is constructed based on the difference between the target latent space representation and the temporal prior latent space representation of the target time point in the angiography sequence sample; the training of the temporal residual prediction model is guided by a total loss, wherein the total loss includes: a main loss, which is used to constrain the difference between the candidate latent space residual and the target latent space residual; a vascular region of interest reconstruction loss, which is used to constrain the reconstruction error between the candidate image at the target time point and the true target time point image within the vascular region of interest; and a non-vascular region temporal prior preservation loss, which is used to constrain the consistency between the candidate image at the target time point and the temporal prior predicted image within the non-vascular region. This scheme can achieve at least the following beneficial technical effects: by using the target latent space residual as the supervision target and training it in combination with the main loss, the vascular region of interest reconstruction loss and the non-vascular region temporal prior preservation loss, the temporal residual prediction model learns small transitional corrections around the temporal prior latent space representation, focusing the model's correction capability on the diagnostic enhancement changes of the vascular region of interest, while maintaining the stability of the temporal prior prediction image in the non-vascular region, and reducing the model's irrelevant modifications to background tissue, bony structures and noise textures.

[0019] Optionally, the total loss also includes: residual magnitude difference constraint loss, used to constrain the magnitude difference between candidate latent space residuals and target latent space residuals; effective weight magnitude constraint loss, used to constrain the magnitude of the effective weights of the residuals; and effective residual magnitude constraint loss, used to constrain the magnitude of the effective latent space residuals. This scheme can achieve at least the following beneficial technical effects: by constraining the magnitude difference between candidate latent space residuals and target latent space residuals through residual magnitude difference constraint loss, and by constraining the magnitudes of effective weights and effective latent space residuals through effective weight magnitude constraint loss and effective residual magnitude constraint loss, respectively, a model modification permission control mechanism is established from multiple levels of residual prediction, weight generation, and effective magnitude, reducing over-correction caused by candidate residual direction errors or magnitude anomalies, and reducing the possibility of relative time series prior degradation.

[0020] Optionally, the total loss may also include one or more of the following losses: effective weight spatial smoothing loss, used to constrain the smoothness of the spatial distribution of the effective residual weights; residual direction consistency loss, used to constrain the directional consistency between the candidate latent space residuals and the target latent space residuals; and degradation loss, used to compare the errors of applying correction to the effective latent space residuals and not applying correction, and to penalize when the error of applying correction is greater. This scheme can achieve at least the following beneficial technical effects: by constraining the smoothness of the spatial distribution of the effective residual weights through the effective weight spatial smoothing loss, it reduces discontinuous jumps in the effective residual weights; by constraining the directional consistency between the candidate latent space residuals and the target latent space residuals through the residual direction consistency loss, it reduces the relative temporal prior degradation caused by candidate residual direction errors; and by penalizing when the error of applying correction is greater than the error of not applying correction through the degradation loss, it suppresses the effectiveness of candidate corrections that are not better than the temporal prior prediction results, so that the model only increases the effective residual weights when it can bring benefits or does not cause significant degradation.

[0021] Optionally, the method further includes: during the verification phase, comparing the error differences between the candidate images at the target time point and the temporal prior predicted images relative to the real target time point images in pairs to determine the relative temporal prior degradation evaluation result; and based on the relative temporal prior degradation evaluation result, performing model hyperparameter selection, reliability threshold adjustment, or backoff decision to improve the quality of the final output target time point predicted image, wherein the reliability threshold is used to perform reliability judgment on the candidate images at the target time point. This scheme can achieve at least the following beneficial technical effects: by comparing the error differences between the candidate images at the target time point and the temporal prior predicted images relative to the real target time point images in pairs during the verification phase to determine the relative temporal prior degradation evaluation result, and performing model hyperparameter selection, reliability threshold adjustment, or backoff decision accordingly, the degradation risk of the model can be identified and quantified before deployment, the reliability judgment threshold and backoff strategy can be optimized, the reliability judgment and backoff decision in the inference phase can be made more accurate, and the quality of the final output target time point predicted image can be further improved.

[0022] Optionally, the effective latent space residuals can be obtained as follows:

[0023]

[0024]

[0025] in, Indicates the effective latent space residual. Indicates the residual effective weight, ⊙ indicates element-wise multiplication. Represents the candidate latent space residual. Indicates the maximum effective coefficient. This represents the Sigmoid function. This represents the reliability score. The scheme achieves at least the following beneficial technical effects: by mapping the reliability score using the Sigmoid function and multiplying it by the maximum effectiveness coefficient to obtain the residual effectiveness weight, and then applying it element-wise to the candidate latent space residuals to obtain the effective latent space residuals, it achieves position-wise adaptive control of the residual effectiveness weight. This allows the model to allow residuals to be effective with a limited magnitude in local regions with high reliability, while automatically suppressing residual effectiveness in regions with insufficient reliability. Simultaneously, the maximum effectiveness coefficient hard-limits the maximum effectiveness ratio of candidate latent space residuals at the architectural level, preventing candidate residuals from excessively modifying the temporal prior latent space representation.

[0026] According to a second aspect of the present invention, a computer device / apparatus / system is provided, including a memory, a processor, and a computer program / instructions stored in the memory, wherein the processor executes the computer program / instructions to implement the steps of the method described in the first aspect.

[0027] According to a third aspect of the present invention, a computer program product is provided, comprising a computer program / instructions that, when executed by a processor, implement the steps of the method described in the first aspect. Attached Figure Description

[0028] The embodiments of the present invention will be further described below with reference to the accompanying drawings, wherein:

[0029] Figure 1 This is a schematic flowchart of an angiography method according to an embodiment of the present invention;

[0030] Figure 2 This is a schematic diagram illustrating the temporal prior construction and latent space residuals according to an embodiment of the present invention;

[0031] Figure 3 This is a schematic diagram of the reliability assessment residual fusion structure according to an embodiment of the present invention;

[0032] Figure 4 This is a schematic diagram of the ROI constraint and relative temporal prior degradation evaluation process according to an embodiment of the present invention;

[0033] Figure 5 This is a schematic diagram of the training input and diffusion residual prediction process according to an embodiment of the present invention;

[0034] Figure 6 This is a schematic diagram of the modules of a system according to an embodiment of the present invention. Detailed Implementation

[0035] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0036] As mentioned in the background section, in existing technologies, traditional temporal intensity interpolation methods are stable but lack local adaptive correction capabilities; motion compensation or registration-driven methods are suitable for scenarios with significant structural motion, but may confuse contrast agent enhancement changes with spatial displacement in dynamic computed tomography angiography sequences; GAN-based and diffusion model-based temporal super-resolution methods have strong expressive capabilities, but are prone to uncontrollable modifications in medical images. To address this, this invention first constructs a temporal prior prediction image for the target time point based on visible dynamic computed tomography angiography frames and their acquisition times; then, this temporal prior prediction image is encoded into a temporal prior latent space representation; next, a temporal residual prediction model predicts candidate latent space residuals and reliability scores; subsequently, the residual effectiveness weights are calculated based on the reliability scores, and the candidate latent space residuals are subjected to amplitude-limited effectiveness processing, with the effective latent space residuals superimposed on the temporal prior latent space representation, and the candidate image for the target time point is obtained after decoding. Furthermore, when the reliability of the candidate image at the target time point is insufficient, the candidate residual is suppressed, causing the output to fall back to the temporal prior predicted image, or the decoded image represented by the temporal prior latent space. In this way, the present invention transforms the depth model from an unconstrained direct reconstructor into a local residual corrector constrained by reliability assessment. This scheme, while inheriting the stability of traditional temporal intensity interpolation methods, can perform small, constrained corrections to high-confidence vascular enhancement regions, and reduces uncontrollable modifications in temporal super-resolution medical images through relative temporal prior degradation assessment and a fallback mechanism for insufficient reliability.

[0037] The angiography method of the present invention will be described below with reference to specific embodiments. It should be understood that the following embodiments are used to explain the technical solution of the present invention and should not be construed as limiting the scope of protection. Without departing from the core idea of ​​the present invention, the temporal prior prediction construction method, encoder type, temporal residual prediction model structure, reliability assessment form, loss function combination, and backoff decision conditions can all be equivalently replaced or combined.

[0038] I. Overall Technical Solution

[0039] The overall approach of this invention is as follows: For the input angiography sequence (e.g., dynamic computed tomography angiography sequence), a temporal prior prediction image for the target time point is first constructed using visible observation frames and their acquisition times. Then, the visible observation frames are encoded into a visible frame latent space representation, and the temporal prior prediction image is encoded into a temporal prior latent space representation. Next, a temporal residual prediction model is used to predict candidate latent space residuals and their corresponding reliability scores. Then, the residual effectiveness weights are calculated based on the reliability scores, and the candidate latent space residuals are subjected to amplitude-limited effectiveness processing. Finally, the effective latent space residuals are superimposed onto the temporal prior latent space representation and decoded to obtain the candidate image for the target time point. If the reliability score, candidate residual amplitude, region of interest confidence, or relative temporal prior degradation evaluation result does not meet the preset reliability conditions, the effectiveness of the candidate latent space residuals is suppressed, causing the output to fall back to the temporal prior prediction image or the decoded image of the temporal prior latent space representation.

[0040] See Figure 1 The above process can be summarized as follows: Step S1: Obtain the angiography sequence (i.e., the angiography image sequence) and its acquisition time, and identify the visible observation frames and target time points. Step S2: Construct the temporal prior prediction image of the target time point based on the visible observation frames and their acquisition times. Step S3: Encode the visible observation frames and the temporal prior prediction image into visible frame latent space representation and temporal prior latent space representation, respectively. Step S4: Use the temporal residual prediction model to determine the candidate latent space residuals and their corresponding reliability scores based on the visible frame latent space representation, target time point information, target time point mask, and temporal prior latent space representation. Step S5: Determine the residual activation weight based on the reliability score, and perform amplitude-limited activation processing on the candidate latent space residuals based on the residual activation weight to obtain the activated latent space residuals. Step S6: Superimpose the activated latent space residuals onto the temporal prior latent space representation to obtain the predicted latent space representation, and decode the predicted latent space representation to obtain the candidate image of the target time point. Step S7: Perform a reliability judgment on the candidate image at the target time point. When it meets the preset reliability conditions, the candidate image at the target time point (residual correction result) is used as the final output target time point prediction image. When it does not meet the preset reliability conditions, the temporal prior prediction image or the image after decoding the temporal prior latent space representation is used as the final output target time point prediction image.

[0041] II. Input Data and Time Modeling

[0042] According to one embodiment of the present invention, the input is a sequence of dynamic computed tomography angiography images obtained at multiple time points from the same subject, the same scanning site, and the same spatial location. , denoted as:

[0043]

[0044] in, Indicates the first Images at each time point, This indicates the number of images in the sequence. Each image corresponds to a real acquisition time or a normalized acquisition time. , denoted as:

[0045]

[0046] in, Indicates the first The time corresponding to each image. The acquisition time can be the actual acquisition time recorded by the scanning device, or a relative time after normalization.

[0047] During training or inference, the system constructs frame labels according to task requirements. Visible observation frames refer to frames whose image content is used as known input to the model in a temporal super-resolution task, and can serve as time anchors for constructing temporal priors for target time points, used to construct temporal prior prediction images and provide temporal context. Conversely, invisible frames refer to frames whose image content is not used as known observation input for the current task; target frames occluded during training, although real and valid, are invisible to the model and are only used to construct supervised targets. Target time point images (target frames) are images that need to be reconstructed. Valid frames refer to frames that were actually acquired, have valid acquisition time, and whose image quality meets processing requirements. Conversely, invalid frames include padded frames, missing frames, damaged frames, or frames whose quality does not meet requirements, used to exclude padded frames, missing frames, or frames whose quality does not meet requirements, for example, from prior construction, training loss, and evaluation. Through these frame labels, the system can clearly identify which frames can serve as time anchors, which time points require temporal super-resolution, and which frames should be excluded from training or evaluation.

[0048] In one implementation, the system requires the target time point to be located between at least two visible observation frames to ensure that the temporal prior prediction image has clear temporal support. When the number of visible observation frames is insufficient, the target time point is not surrounded by left and right visible observation frames, the acquisition time does not satisfy the increasing relationship, or the input data contains non-finite values, the system can refuse to execute the residual correction process or only output a prompt result, thereby avoiding blindly reconstructing medical images when temporal constraints are not met. The left visible observation frame is the most recent frame whose acquisition time is less than the target time, and the right visible observation frame is the most recent frame whose acquisition time is greater than the target time.

[0049] III. Construction of Temporal Prior Prediction Images

[0050] According to one embodiment of the present invention, constructing a temporal prior prediction image of a target time point based on visible observation frames and their acquisition times includes: determining the left and right visible frames adjacent to the target time point from the angiography sequence, and obtaining the acquisition times corresponding to the left and right visible frames; determining the interpolation weights of the left and right visible frames based on the time interval ratio between the target time point and the acquisition times of the left and right visible frames; and weighting and summing the angiography intensities of corresponding pixels or voxels of the left and right visible frames according to the interpolation weights to generate a temporal prior prediction image of the target time point.

[0051] illustrative, see Figure 2 For the target time point First, find the target time point in the visible observation frames. The left visible frame and the visible frame on the right The corresponding collection times are respectively and And satisfy:

[0052]

[0053] In a preferred embodiment, interpolation weights are calculated based on the relative temporal positions of the target time point and the left and right visible frames:

[0054]

[0055] Subsequently, a temporal prior prediction image for the target time point is constructed in the image space:

[0056]

[0057] in, Indicates the target time point The temporal prior prediction image is used as the basis for subsequent residual correction, rather than being completely replaced by the deep model.

[0058] In other embodiments, the temporal prior prediction image can also be constructed using piecewise cubic Hermite conformal interpolation, spline interpolation, polynomial interpolation, interpolation based on vessel time-intensity curves, temporal interpolation based on vessel centerline, interpolation based on perfusion dynamics models, or a combination of the above interpolation methods. Regardless of the temporal prior prediction construction method used, their common function is to provide stable and interpretable initial prediction results for the target time point and to provide a temporal prior preservation path for subsequent reliability constraint residual correction.

[0059] IV. Construction of Temporal Prior Latent Space Representation and Target Residual

[0060] See you again Figure 2 After obtaining the temporal prior prediction image, an image encoder is used to map the temporal prior prediction image to the latent space, resulting in the temporal prior latent space representation:

[0061]

[0062] in, Indicates an image encoder. This represents the temporal prior latent space representation. The image encoder can be a variational autoencoder, a general autoencoder, a vector quantization autoencoder, a medical image fundamental model encoder, or other learnable image coding networks.

[0063] During the training phase, images of the real target time points will also be used. Encoded as a target latent space representation (or target frame latent space representation):

[0064]

[0065] Then, the target latent space residual is constructed:

[0066]

[0067] in, Represents the target latent space residual. This represents the residual scaling factor. The residual scaling factor is used to adjust the numerical range of the residual training target and can be used to control the magnitude of residual correction during the inference phase.

[0068] Through the above processing, this invention transforms the temporal super-resolution problem of target time point images into a small residual correction problem relative to the temporal prior latent space representation. The model no longer needs to freely reconstruct the complete target image, but instead learns how to perform constrained corrections based on the temporal prior predicted image.

[0069] V. Inputs and Outputs of the Temporal Residual Prediction Model

[0070] In one embodiment, the visible frame latent space representation, the temporal prior latent space representation, the target time point information, the sequence position encoding, the target time point mask, and the diffusion time step information are input into the temporal residual prediction model. The target time point information is a numerical description of the time position to be reconstructed, and includes at least the target acquisition time. This time can be the actual scan time, relative time, or normalized time. Depending on the implementation, it may also include information such as the time interval between the target time and the left and right visible observation frames. Sequence position encoding refers to the dynamic CTA frame sequence. The temporal position or chronological order of each frame is encoded, not the two-dimensional spatial position encoding in the image, nor the diffusion time step. In the embodiment, the actual or relative acquisition time of each frame can be used. Continuous-time encoding is performed. This temporal residual prediction model can employ a diffusion model, a diffusion Transformer model, a video Transformer model, a 3D convolutional network, a temporal convolutional network, a recurrent neural network, a multi-scale attention network, or a combination thereof.

[0071] In a preferred embodiment, it is set This is the target time point mask, which is set to 1 at the target frame position and 0 at the visible observation frame position; The temporal prior latent space representation is represented; the temporal prior latent space representation is extended to the target frame position to form a temporal prior latent space sequence. Indicates the diffusion time step The perturbation residual signal is obtained by adding noise to the target latent space residual signal. The model input can be expressed as:

[0072]

[0073] in, This represents element-wise multiplication. This represents the residual scaling factor.

[0074] Therefore, it can be seen that the observed frame position maintains the true latent space representation; the target time point position adopts the form of temporal prior latent space representation superimposed with perturbation residual signal, so that the model always learns residual correction around the temporal prior prediction results, rather than freely reconstructing the entire target image from random noise or blank states. The model also receives the acquisition time series, diffusion time steps and frame mask as conditions.

[0075] Temporal residual prediction model The output consists of two parts. The first part is the candidate latent space residual, denoted as... , representing the amount of correction the model proposes to the temporal prior latent space representation. The second part is the reliability score, denoted as . The reliability score reflects the reliability of the candidate latent space residuals at different time points and spatial locations. The effectiveness of the candidate latent space residuals at different time points and spatial locations can then be calculated based on the reliability score. The above conditional prediction relationship can be expressed as:

[0076]

[0077] In one implementation, the temporal residual prediction model has a shared backbone network and two output branches. The shared backbone network is used to extract cross-frame temporal features, spatial context features, and target time-point conditional features; the residual prediction branch is used to predict candidate latent spatial residuals based on the cross-frame temporal features, spatial context features, and target time-point conditional features; and the reliability assessment branch is used to predict a reliability score based on the cross-frame temporal features, spatial context features, and target time-point conditional features. The two branches share the temporal context but respectively handle the functions of "suggesting how much correction to make" and "whether the candidate correction is reliable."

[0078] VI. Reliability Assessment and Constrained Residual Fusion

[0079] See Figure 3 The residual effectiveness weight is obtained from the reliability score through nonlinear mapping and the maximum effectiveness coefficient. In one embodiment, the residual effectiveness weight is calculated as follows:

[0080]

[0081] in, Indicates the residual effective weight. Indicates the maximum effective coefficient. express function, This represents the reliability score. The maximum effectiveness coefficient is used to limit the maximum effectiveness ratio of candidate latent space residuals, thereby preventing candidate residuals from excessively modifying the temporal prior latent space representation.

[0082] Calculate the effective latent space residuals based on the residual effective weights:

[0083]

[0084] in, represents the effective latent space residual, and ⊙ represents element-wise multiplication. The effective latent space residual is then superimposed onto the temporal prior latent space representation:

[0085]

[0086] in, The predicted latent space representation for the target time point is given. Finally, the decoder is used to decode the predicted latent space representation into candidate images for the target time point:

[0087]

[0088] in, Indicates decoder, This represents the candidate image for the target time point (the reconstructed image for the target time point).

[0089] The above structure fundamentally distinguishes this invention from unconstrained complete residual prediction. Unconstrained complete residual prediction, by default, can directly superimpose complete candidate residuals onto the temporal prior latent space representation. In contrast, this invention controls the effective range and magnitude of candidate residuals through residual effectiveness weights. When model reliability is insufficient, the residual effectiveness weights can approach zero, making the output resemble the temporal prior prediction image. When the model has a high reliability score in local vascular regions, the residual effectiveness weights can be appropriately increased in those regions, allowing candidate residuals to be effective with a limited magnitude.

[0090] In a preferred embodiment, the reliability evaluation branch is set to near-zero activation during the initial training phase. For example, the weights of the reliability scoring output layer can be initialized to small values, and the bias can be initialized to negative values, resulting in low initial residual activation weights. Assume the initial bias of the reliability score is... Then the expected initial effective weight can be expressed as:

[0091]

[0092] For example, in a preferred embodiment, the maximum effectiveness coefficient can be taken. Set to 0.1 and take a negative bias. The value is approximately -4.595, making the initial effective weight approximately 0.001. Thus, during the initial training phase, the model retains the temporal prior latent space representation by default. Only after learning reliable residual signals gradually during training is the residual allowed to take effect with a small magnitude in regions of higher reliability.

[0093] VII. Constraints on the Region of Interest for Vascular Regions and Temporal Prior Preservation for Non-Vascular Regions

[0094] Key changes in dynamic computed tomography angiography sequences are typically concentrated in the vessels and surrounding enhanced regions. To ensure that model corrections focus on medically significant vascular enhancement changes, this invention can introduce a region of interest constraint for the vessels during the training phase.

[0095] set up This represents the region of interest mask for blood vessels. This represents the candidate image for the target time point obtained through reconstruction (reconstruction result). This represents the real target image at a specific time point (the real target image). Since this embodiment constructs a temporal prior representation in the latent space, the preferred target for preserving non-vascular regions is the temporal prior prediction image obtained by decoding the temporal prior latent space representation. :

[0096]

[0097] in, Indicates decoder, This represents the temporal prior latent space representation.

[0098] Within the region of interest (ROI) of a blood vessel, the reconstruction loss between the reconstructed result and the true target image can be calculated; this is known as the ROI reconstruction loss. :

[0099]

[0100] in, This represents the region of interest mask for blood vessels. This represents smoothed L1 loss or Huber loss. This represents the candidate image for the target time point obtained from the reconstruction. Represents the actual target image at a specific time point. The total number of effective pixels / voxels representing the region of interest for blood vessels.

[0101] In non-vascular regions, the preservation loss between the reconstructed results and the temporal prior predicted images can be calculated:

[0102]

[0103] in, This represents the smoothed L1 loss or Huber loss, with the denominator used for normalization according to the effective mask region. This represents the temporal prior prediction image. Through the above constraints, the model learns the real vascular enhancement changes at the target time point within the vascular region of interest, while preserving the decoded temporal prior prediction image as much as possible in non-vascular regions. This constraint reduces irrelevant modifications made by the model to background tissue, bony structures, noise textures, and non-target regions.

[0104] It should be noted that the region of interest mask for blood vessels can be obtained through manual annotation, threshold segmentation, blood vessel segmentation networks, maximum density projection, blood vessel centerline extraction, temporal intensity curve analysis, or a combination of the above methods. In other embodiments, instead of explicitly constructing a binary mask, a continuous blood vessel weight map can be constructed, giving different blood vessel interest weights at different spatial locations.

[0105] VIII. Relative Temporal Prior Degradation Evaluation and Training Constraints

[0106] To determine whether candidate residual corrections are superior to temporal prior predictions, this invention introduces a relative temporal prior degradation evaluation and non-degradation constraints. This evaluation does not calculate the error between the reconstructed result and the true target image individually, but rather compares the error differences between the candidate residual correction result and the temporal prior prediction result relative to the true target image in pairs. During the training phase, a corresponding degradation loss can also be constructed on the latent space residual target.

[0107] See Figure 4 During the training phase, set This represents the average error of the effective latent space residual relative to the target latent space residual. This represents the average error when no residual correction is applied and the temporal prior latent space representation is directly preserved; the above average error is calculated according to the target time point mask or effective region, and can be written as:

[0108]

[0109]

[0110]

[0111] in, This represents the average error relative to the target latent space residual after applying the effective latent space residual correction. This means taking the average of the absolute values ​​of each element over the target time point mask or valid region. Indicates the effective latent space residual. Represents the target latent space residual. This represents the average error relative to the target latent space residual when the temporal prior latent space representation is directly preserved without applying residual correction. In this case, the residual corresponding to no correction is zero, and the error is the target latent space residual. The absolute average; Δ represents the error difference between the two cases mentioned above, used to determine whether the candidate residual correction has deteriorated relative to the time prior. When A value >0 indicates that the error of the candidate residual correction result has increased relative to the temporal prior prediction result, and relative degradation of the temporal prior may occur. When the value is ≤0, it indicates that the candidate residual correction result is not worse than the time series prior prediction result, or is an improvement over the time series prior prediction result.

[0112] During the training phase, the following degradation loss can be optionally constructed:

[0113]

[0114] in, This represents the function that takes the maximum value.

[0115] Alternatively, a non-negative reliability boundary can be set. The following degradation loss is constructed:

[0116]

[0117] This loss is used to encourage the model to increase the residual activation weight only when the candidate residual can bring benefits or does not cause significant degradation; when the candidate residual may make the result worse than the temporal prior predicted image, the model should reduce the residual activation weight or suppress the activation of the candidate residual. In a preferred embodiment, It can be set to 0 to directly penalize the increase in error relative to the timing prior; or it can be set to a more stringent reliability boundary according to the deployment reliability requirements.

[0118] In the verification or playback evaluation phase with real target images, let the candidate residual correction result be... Decoding the temporal prior prediction image is Then, the relative temporal prior degradation of blood vessels within the region of interest in the image space can be statistically analyzed:

[0119]

[0120] in, This represents the relative temporal prior degradation assessment results within the region of interest in the blood vessel. This represents the mean absolute error calculated within the region of interest of the blood vessel. This represents the candidate correction result, i.e., the candidate image of the target time point obtained after decoding after residual correction; Represents the actual target image at a specific time point. This represents the temporal prior prediction image.

[0121] Based on this pairwise difference, metrics such as relative temporal prior degradation rate, significant degradation rate, improvement rate, mean relative error change, and effective residual magnitude within the blood vessel region of interest can be statistically analyzed. These metrics can be used for model selection, reliability threshold adjustment, and case-level rejection strategies. In the actual deployment phase where real target images are unavailable, the aforementioned supervision errors can be avoided by directly calculating them. Instead, reliability score distribution, residual effectiveness weight distribution, effective residual magnitude, input time interval, blood vessel region confidence, and model quality surrogate metrics can be used for backoff judgment.

[0122] IX. Residual Amplitude and Direction Constraints

[0123] To further avoid abnormal magnitudes or incorrect orientations of candidate latent space residuals, this invention can introduce residual magnitude and orientation constraints. In one embodiment, a smoothed L1 loss is calculated for the predicted residual representation and the target residual representation:

[0124]

[0125] in, This represents the residual obtained by model backpropagation or prediction. This represents the target latent space residual. This constraint can directly constrain the magnitude difference between the predicted residual and the actual residual.

[0126] During the training phase or the validation phase with real target frames, the directional consistency between the predicted residual and the target residual can also be calculated, for example, through cosine similarity constraints.

[0127]

[0128] in, This represents cosine similarity. By using directional constraints, the possibility of relative temporal prior degradation caused by incorrect candidate residual directions can be reduced.

[0129] During training and validation with real target images, the average magnitude of the predicted residual, the average magnitude of the target residual, the quantile magnitude of the predicted residual, the quantile magnitude of the target residual, and the directional consistency between the predicted and target residuals can be recorded. When the magnitude of the predicted residual is significantly greater than the magnitude of the target residual, or the directional consistency is lower than a preset threshold, the residual activation weight can be reduced or the non-deterioration loss weight can be increased. In the actual inference phase, since the real target residual is unavailable, real directional consistency is not used as a necessary criterion. Instead, backoff is performed based on observable surrogate metrics such as reliability score, residual activation weight, residual magnitude, and input quality.

[0130] 10. Training Process

[0131] In one embodiment, the training process of the present invention is as follows: Figure 5 As shown, angiography sequences are sampled from the training set, and visible observation frames and target time point images are selected according to a preset sampling protocol. Subsequently, a temporal prior prediction image for the target time point is constructed based on the visible observation frames and their acquisition times, and this temporal prior prediction image is encoded into a temporal prior latent space representation. If a real target image is available during the training phase, the real target image is also encoded into a target latent space representation, and the difference between the two is used to construct the target latent space residual. Training samples for the training phase are constructed accordingly. Each training sample consists of angiography sequences from the same subject and the same scanning site, along with the acquisition times corresponding to each frame. During training, one or more valid frames are selected as target time point images, and the remaining valid frames are used as visible observation frames. A temporal prior prediction image of the target time point is constructed based on the visible observation frames and their acquisition times, and the visible observation frames and the temporal prior prediction images are encoded as visible frame latent space representations and temporal prior latent space representations, respectively. The input of the temporal residual prediction model includes the visible frame latent space representation, target time point information, target time point mask, and the temporal prior latent space representation. The supervised targets include the real target time point images and their corresponding target latent space residuals.

[0132] Next, the latent space representation of the visible frame, the temporal prior latent space representation, the target time point conditions, the frame mask, and the perturbation residual information are input into the temporal residual prediction model. For the visible observation frame, the input retains its true latent space representation; for the target time point location, the input is in the form of a temporal prior latent space representation superimposed with a noisy residual signal. The model simultaneously receives the acquisition time conditions and the diffusion time step conditions, and outputs candidate latent space residuals and a reliability score. Based on the reliability score, the residual activation weight and the activated latent space residual are calculated and superimposed onto the temporal prior latent space representation to obtain the predicted latent space representation. After decoding, the predicted latent space representation yields the target time point image.

[0133] The training loss can consist of multiple losses, including the main loss for temporal residual prediction, the residual amplitude difference constraint loss, the residual direction consistency loss, the vessel region of interest reconstruction loss, the temporal prior preservation loss for non-vessel regions, the effective weight amplitude constraint loss, the effective weight spatial smoothing loss, and the degradation loss. The overall training objective can be expressed as:

[0134]

[0135] in, This indicates the main loss predicted by the time-series residuals. This represents the residual magnitude difference constraint loss. This indicates the loss of reconstructed blood vessel regions of interest. This indicates the temporal prior preservation loss in non-vascular regions. This indicates the loss due to the effective weight range constraint. This indicates the loss due to the effective residual magnitude constraint. This indicates the effective weight space smoothing loss. This represents the loss of consistency in the direction of the residuals. Indicates degradation loss, , , , , , , and These are the corresponding weighting coefficients.

[0136] It should be noted that the above loss terms do not necessarily all need to be used simultaneously. In different embodiments, one or more of these terms can be selected in combination, depending on the training data size, image quality, model type, and deployment reliability requirements. Illustratively, the following are some illustrative alternative total losses:

[0137]

[0138]

[0139]

[0140]

[0141] In addition, implementers can set the weight coefficients as needed. For example, when the amount of data is small or the temporal prior prediction images are already relatively reliable, the weights of the effective weight constraints and the temporal prior preservation constraints for non-vascular regions can be increased; in scenarios where vascular enhancement changes are complex, the weights of the vascular region of interest reconstruction loss and the residual orientation consistency loss can be increased.

[0142] XI. Reasoning Process

[0143] During the inference phase, the system first receives the angiography sequence to be processed and the acquisition time corresponding to each image. For the target time point to be reconstructed, the system finds the visible observation frames to its left and right and constructs the temporal prior prediction image for the target time point. Then, the temporal prior prediction image is encoded into a temporal prior latent space representation, and the latent space representations of the visible frames, the target time point conditions, the frame mask, and the temporal prior latent space representation are input into the trained temporal residual prediction model.

[0144] The temporal residual prediction model outputs candidate latent space residuals and a reliability score. Based on the reliability score, the residual effectiveness weights and effective latent space residuals are calculated, resulting in a predicted latent space representation. This predicted latent space representation is then decoded by the decoder to form candidate images for the target time point.

[0145] Subsequently, the system performs a reliability assessment. In the verification phase with real target images, supervisory metrics such as relative temporal prior degradation and residual direction consistency can be used. In the actual inference phase, instead of relying on real target images or real target residuals, the system makes assessments based on reliability scores, residual effective weight distribution, effective weight concentration within the blood vessel region of interest, effective latent space residual magnitude, whether the input sequence time interval exceeds the model's applicable range, input image quality, and whether the relative temporal prior degradation surrogate metric exceeds a preset threshold.

[0146] When the candidate result meets the reliability condition, the target time point candidate image after residual correction is output. When the candidate image does not meet the reliability condition, one of two backoff methods can be adopted: first, directly output the image spatial temporal prior prediction image; second, output the image after decoding the temporal prior latent space representation. Through this inference mechanism, the present invention can maintain stable output when reliability is insufficient, avoiding significant relative temporal prior degradation caused by candidate latent space residuals.

[0147] XII. System Implementation Method

[0148] See Figure 6The present invention also provides a temporal super-resolution system for angiography sequences. This system includes a processor, a memory, a medical image input interface, and a result output interface. The memory stores a computer program, which, when executed by the processor, implements the temporal super-resolution method for angiography sequences described in this invention.

[0149] Functionally, the system may include a data input module, a temporal prior prediction image construction module, a latent space coding module, a temporal residual prediction module, a reliability assessment module, a residual fusion module, an image decoding module, a blood vessel region of interest constraint module, a relative temporal prior degradation evaluation module, and a backoff decision module.

[0150] The data input module receives the angiography sequence, acquisition time, visible frame markers, target time point markers, and valid frame markers. The temporal prior prediction image construction module constructs the temporal prior prediction image for the target time point based on the visible observation frames and their acquisition times. The latent space encoding module maps the temporal prior prediction image and visible frames to latent space representations. The temporal residual prediction module predicts candidate latent space residuals based on the temporal prior latent space representation, visible frame latent space representations, and target time point conditions. The reliability evaluation module calculates the reliability score and residual effectiveness weight corresponding to the candidate latent space residuals. The residual fusion module applies the residual effectiveness weights to the candidate latent space residuals and superimposes the effective latent space residuals onto the temporal prior latent space representation. The image decoding module decodes the predicted latent space representation into the target time point image. The relative temporal prior degradation evaluation module determines whether the candidate correction result has an increased error relative to the temporal prior prediction result. The backoff decision module outputs the temporal prior prediction result or a fusion result close to the temporal prior prediction result when the candidate result is unreliable.

[0151] The system can be deployed on a medical imaging workstation, a medical image post-processing server, a cloud-based medical image analysis platform, or a scientific research computing platform. The input image can be a two-dimensional slice sequence, a three-dimensional volumetric data sequence, or a four-dimensional dynamic volumetric data sequence; the output can be an image at a single target time point or multiple images at multiple target time points, thereby achieving temporal super-resolution of angiography sequences.

[0152] XIII. Optional Implementation Methods

[0153] The implementer may also adjust some implementation details of this invention to obtain alternative implementation methods. For example:

[0154] In constructing temporal prior prediction images, in addition to image spatial linear interpolation, piecewise cubic Hermite conformal interpolation, spline interpolation, polynomial interpolation, interpolation based on time-intensity curves, temporal interpolation based on vascular tree structures, interpolation based on local contrast agent kinetic models, or combinations of the above methods can also be used. As long as the method first constructs the temporal prior prediction image at the target time point and then performs reliability constraint residual correction on it, it falls within the scope of this invention.

[0155] In terms of model structure, the temporal residual prediction model is not limited to the diffusion Transformer model; it can also employ convolutional neural networks, 3D U-Net, video Transformer, temporal attention networks, recurrent neural networks, generator structures in generative adversarial networks, diffusion models, or combinations of the above models. The reliability score can be output from the shared backbone of the temporal residual prediction model or from an independent reliability assessment network.

[0156] Regarding the form of reliability assessment, reliability scores and residual effectiveness weights can be applied at the pixel, voxel, block, channel, frame, or multi-scale levels; they can be continuous weights or discrete selection signals; they can be a single weight map or a composite weight combining spatial location, temporal location, and channel features. The maximum effectiveness coefficient can be a fixed value or adaptively adjusted based on the target time interval, input image quality, confidence level of the blood vessel region of interest, or validation set statistical results.

[0157] Regarding the constraint of regions of interest (ROIs) for blood vessels, ROIs can be derived from manual annotation, automatic segmentation, threshold extraction, maximum density projection, vessel centerlines, blood vessel treemap structures, or temporal intensity curve analysis. ROI constraints can be applied to the image space, as well as the latent space, feature space, or blood vessel map structure space.

[0158] Regarding backoff strategies, the system can backoff to the temporal prior prediction image in the image space, or to the temporal prior latent space representation in the latent space. It can also employ a weighted fusion of the temporal prior prediction image and the candidate residual correction results. Backoff triggering conditions may include excessively low reliability scores, excessively low residual effectiveness weights, excessively high residual amplitudes, insufficient consistency in residual directions, excessively high relative temporal prior degradation, input time intervals exceeding the applicable range, or input image quality not meeting requirements.

[0159] In terms of application targets, although the present invention mainly focuses on dynamic computed tomography angiography sequences, it can also be extended to computed tomography perfusion imaging-derived angiography sequences, dynamic contrast-enhanced magnetic resonance angiography sequences, digital subtraction angiography sequences, and other dynamic angiography sequences characterized by temporal enhancement changes.

[0160] In summary, some embodiments of the present invention can achieve at least one or more of the following beneficial effects compared to the prior art:

[0161] (1) This invention improves the reliability of temporal super-resolution in dynamic computed tomography angiography sequences. Traditional depth reconstruction methods tend to freely synthesize image content at target time points in regions with insufficient evidence, while this invention first constructs a temporal prior prediction image and then performs small residual corrections under reliability assessment constraints. When the candidate residual does not meet the reliability conditions, the system can backtrack to the temporal prior prediction result, thereby reducing uncontrollable modifications in medical image reconstruction.

[0162] (2) This invention combines the stability of traditional interpolation methods with the local expressive power of deep models. Traditional temporal intensity interpolation methods can provide stable and interpretable temporal prior results, but they are difficult to handle local vascular enhancement differences. This invention, through latent space residual correction, enables the deep model to mainly learn local changes that are difficult to cover by temporal prior predictions, rather than reconstructing the entire image. This avoids the insufficient reliability of directly reconstructing the complete image while retaining the ability to correct local vascular enhancement details.

[0163] (3) This invention establishes a model modification permission control mechanism through reliability scoring, residual effectiveness weighting, maximum effectiveness limit, and near-zero initialization mechanism. This mechanism prevents the model from having the default permission to make large-scale modifications across the entire map, and instead restricts its modifications to areas that meet reliability conditions. Compared to unconstrained residual prediction, this invention can reduce the possibility of relative temporal prior degradation caused by candidate latent space residual direction errors, abnormal amplitudes, or local miscorrections.

[0164] (4) This invention is more consistent with the medical imaging characteristics of dynamic computed tomography angiography sequences. The key changes in dynamic computed tomography angiography sequences are usually concentrated in the vessels and their surrounding enhancement areas, while non-vascular areas should mostly remain stable. This invention uses the region of interest constraint for vessels and the temporal prior constraint for non-vascular regions to make the model correction focus on the enhancement-related areas of vessels, which helps to reduce irrelevant changes in background tissue, bony structures and noise textures.

[0165] (5) This invention incorporates relative temporal prior degradation assessment into the training, validation, and inference decision-making processes. Compared to evaluation methods that only focus on global average error, peak signal-to-noise ratio, or structural similarity, this invention can clearly determine whether candidate residual corrections degrade relative to temporal prior prediction results, and accordingly perform model selection, reliability threshold adjustment, or temporal prior backoff. This mechanism helps improve the reliability, stability, and interpretability of temporal super-resolution results in medical imaging.

[0166] It should be noted that although the steps are described in a specific order above, it does not mean that the steps must be executed in the above specific order. In fact, some of these steps can be executed concurrently, or even in a different order, as long as the required function can be achieved.

[0167] This invention can be a system, method, electronic device, computing device, computer program product and / or computer-readable medium.

[0168] Computer program products mainly refer to software products that implement various aspects of the present invention through computer programs, or hardware products that carry software that implements various aspects of the present invention.

[0169] Computer-readable storage media can be tangible devices that hold and store instructions for use by an instruction execution device. Computer-readable storage media can include, for example, but not limited to, electrical storage devices, magnetic storage devices, optical storage devices, electromagnetic storage devices, semiconductor storage devices, or any suitable combination thereof. More specific examples (a non-exhaustive list) of computer-readable storage media include: portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), static random access memory (SRAM), portable compact disc read-only memory (CD-ROM), digital multifunction disc (DVD), memory sticks, floppy disks, mechanical encoding devices, such as punch cards or recessed protrusions storing instructions thereon, and any suitable combination thereof.

[0170] 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.

Claims

1. A temporal super-resolution imaging method for angiography images, characterized in that, include: Acquire dynamic angiography sequences and their acquisition times, and identify visible observation frames and target time points from them; Based on the visible observation frames and their acquisition times, construct a temporal prior prediction image for the target time point; The visible observation frame and the temporal prior prediction image are encoded into the visible frame latent space representation and the temporal prior latent space representation, respectively. The temporal residual prediction model is used to determine candidate latent space residuals and corresponding reliability scores based on the latent space representation of visible frames, target time point information, target time point mask, and temporal prior latent space representation. The residual activation weight is determined based on the reliability score. Based on the residual activation weight, the candidate latent space residuals are subjected to amplitude-limited activation processing to obtain the activated latent space residuals. The effective latent space residual is superimposed onto the temporal prior latent space representation to obtain the predicted latent space representation, and the predicted latent space representation is decoded to obtain the candidate image of the target time point.

2. The method according to claim 1, characterized in that, The method further includes: performing a reliability assessment on the candidate images at the target time point. When the preset reliability conditions are met, the candidate image of the target time point will be used as the final output predicted image of the target time point. When the preset reliability conditions are not met, the temporal prior prediction image or the decoded image of the temporal prior latent space representation is used as the final output target time point prediction image.

3. The method according to claim 2, characterized in that, The time-series residual prediction model is trained as follows: Encode the real target time point image at the target time point in the angiography sequence sample into a target latent space representation; The target latent space residual is constructed based on the difference between the target latent space representation and the temporal prior latent space representation of the target time point of the angiography sequence sample; The training of the time-series residual prediction model is guided by the total loss, wherein the total loss includes: The principal loss is used to constrain the difference between the candidate latent space residuals and the target latent space residuals; The region of interest (ROI) reconstruction loss is used to constrain the reconstruction error between candidate images at the target time point and the real target time point images within the ROI region. The non-vascular region temporal prior preservation loss is used to constrain the consistency between the candidate image at the target time point and the temporal prior prediction image within the non-vascular region.

4. The method according to claim 3, characterized in that, The total losses also include: The residual magnitude difference constraint loss is used to constrain the magnitude difference between the candidate latent space residual and the target latent space residual; Effective weight magnitude constraint loss, used to constrain the magnitude of the residual effective weights; Effective residual magnitude constraint loss is used to constrain the magnitude of effective latent space residuals.

5. The method according to claim 3 or 4, characterized in that, The total loss also includes one or more of the following losses: The effective weight spatial smoothing loss is used to constrain the smoothness of the spatial distribution of the effective weights of the residuals. The residual orientation consistency loss is used to constrain the orientation consistency between the candidate latent space residuals and the target latent space residuals. Degradation loss is used to compare the error with and without applying correction to the effective latent space residuals, and a penalty is applied when the error with correction is greater.

6. The method according to any one of claims 1-4, characterized in that, The method further includes: During the verification phase, the error differences between the candidate images and temporal prior predicted images at the target time point are compared in pairs with those of the real target time point images to determine the relative temporal prior degradation evaluation results. Based on the relative temporal prior degradation evaluation results, model hyperparameter selection, reliability threshold adjustment, or backoff decision are performed to improve the quality of the final output target time point prediction image. The reliability threshold is used to perform reliability judgment on the target time point candidate image.

7. The method according to any one of claims 1-4, characterized in that, The effective latent space residuals are obtained as follows: in, Indicates the effective latent space residual. Indicates the residual effective weight, ⊙ indicates element-wise multiplication. Represents the candidate latent space residual. Indicates the maximum effective coefficient. This represents the Sigmoid function. This indicates the reliability score.

8. A computer device / apparatus / system, comprising a memory, a processor, and a computer program / instructions stored in the memory, wherein the processor executes the computer program / instructions to implement the steps of the method according to any one of claims 1-7.

9. A computer program product comprising a computer program / instructions that, when executed by a processor, implement the steps of the method according to any one of claims 1-7.

10. A computer-readable storage medium having a computer program / instructions stored thereon, the computer program / instructions being executed by a processor to implement the steps of the method according to any one of claims 1-7.