A data processing method, device and program product for renal artery FFR evaluation

Abstract

This application relates to the field of intelligent healthcare, specifically to a data processing method, device, and program product for renal artery FFR assessment. The method includes acquiring continuous temporal frame data from renal artery angiography; selecting frames from the temporal frame data of the renal artery angiography to obtain multiple time series, the multiple time series including keyframes and N valid frames, where N is a natural number greater than 1; inputting the multiple time series into a segmentation model to perform renal artery segmentation to obtain renal artery vessel masks for keyframes and valid frames; registering the keyframes and keyframe masks with the N valid frames and valid frame masks sequentially based on L scales to obtain L*N registered valid frame data; and constructing 2.5D data based on the registered valid frame data and keyframe data to obtain 2.5D data for renal artery FFR assessment. This application has significant clinical value.

Description

Technical Field

[0001] This application relates to the field of intelligent healthcare, specifically to a data processing method, device, program product, and computer-readable storage medium for renal artery FFR assessment. Background Technology

[0002] Atherosclerotic renal artery stenosis (ARAS) is a major cause of secondary hypertension and chronic renal insufficiency. Current treatments primarily include pharmacological intervention and percutaneous endovascular renal artery intervention (such as stent placement). Existing research mainly relies on imaging to determine the degree of stenosis, neglecting functional assessment of the stenotic lesion. In recent years, renal fractional flow reserve (rFFR) measured by renal artery pressure guidewires has been considered to more accurately reflect the impact of stenosis on renal hemodynamics and has shown promising prospects in practical applications. Analogous to the FFR technique used in coronary intervention, rFFR can serve as an important decision-making basis for renal artery interventional treatment. The applicant's team's research shows that rFFR-guided renal artery stent placement is significantly more effective than pharmacological treatment at one-year follow-up. Therefore, accurate, efficient, and convenient measurement of rFFR is crucial. Currently, the relevant methods for FFR measurement can be broadly categorized as follows:

[0003] 1) Traditional FFR measurement using pressure guidewires involves injecting vasodilators and using additional guidewires, potentially increasing patient risk and cost. 2) Angiographic image-based FFR estimation methods, similar to coronary QFR (Quantitative Flow Ratio), utilize dual-view images for three-dimensional coronary artery reconstruction and combine this with a physical model for blood flow estimation to predict FFR values. While effective in coronary functional assessment, this approach suffers from complex three-dimensional modeling, high image quality requirements, high computational resource consumption, and poor clinical real-time performance. This is particularly problematic in renal artery assessment, where direct application of this approach is limited due to differences in renal artery anatomy, limited imaging angles, and varying hemodynamic characteristics. 3) CT-FFR is a non-invasive method combining coronary CTA images with computational fluid dynamics models to estimate the degree of functional blood flow restriction caused by vascular stenosis. However, for renal artery assessment, CT image quality is limited by renal function and the risks associated with contrast agent use, making it unsuitable as an intraoperative tool. The rFFR detection process is relatively complex, relying on invasive catheter intervention and drug-induced maximum hyperemia, which increases operational risks and patient burden, limiting its widespread clinical application. Summary of the Invention

[0004] To address the aforementioned issues, there is currently a lack of a deep learning-based method for directly predicting renal artery FFR (rFFR) in renal artery angiography images. This method should be able to achieve non-invasive, rapid, and accurate prediction of functional significance of stenosis without requiring 3D modeling or hydrodynamic model calculations, by learning the mapping relationship between a large number of images and the actual rFFR. Therefore, this invention provides a data processing method for renal artery FFR assessment, specifically including:

[0005] Acquire continuous time-series frame data from renal artery angiography;

[0006] Frame selection is performed on the time-series frame data of the renal artery angiography to obtain multiple time series, which include keyframes and N valid frames, where N is a natural number greater than 1;

[0007] The multiple time series are input into the segmentation model to perform renal artery segmentation to obtain the renal artery mask of key frames and effective frames;

[0008] The keyframes and their masks are registered sequentially with N valid frames and their masks based on L scales to obtain L*N registered valid frame data.

[0009] Based on the registered valid frame data and keyframe data, 2.5D data is constructed to obtain 2.5D data for renal artery FFR assessment.

[0010] Optionally, the registration process is as follows:

[0011] Get keyframes, valid frames, keyframe masks, and valid frame masks;

[0012] Based on the first scale, the deformation field of the first scale is generated by measuring the similarity between the key frame and its mask, and between the effective frame and its mask.

[0013] The initial second-scale deformation field is obtained by upsampling the deformation field at the first scale.

[0014] The second-scale deformation field is generated by performing similarity measurement on keyframes and their masks, and effective frames and their masks, based on the initial second-scale deformation field.

[0015] The initial third-scale deformation field is obtained by upsampling the deformation field at the second scale.

[0016] Based on the initial third-scale deformation field, the similarity between keyframes and their masks, and between valid frames and their masks, is measured to generate the third-scale deformation field, thus obtaining the registered valid frames.

[0017] Optionally, different scales of registration correspond to effective frame masks and keyframe masks of different scales; at the first scale, the mask is reduced by a first factor to obtain the mask of the first keyframe and the mask of the first effective frame, and the similarity measurement of the keyframe and the mask of the first keyframe, the effective frame and the mask of the first effective frame is performed based on the first scale to generate a deformation field at the first scale; at the initial second scale deformation field, the mask is reduced by a second factor to obtain the mask of the second keyframe and the mask of the second effective frame, and the similarity measurement of the keyframe and the mask of the second keyframe, the effective frame and the mask of the second effective frame is performed based on the initial second scale deformation field to generate a deformation field at the second scale; at the initial third scale deformation field, the mask remains unchanged;

[0018] Optionally, the registration process further includes elastic deformation, and the registered effective frames are obtained by performing elastic deformation calculation on the effective frames based on the deformation field of the third scale.

[0019] Optionally, the registration process further includes interpolation calculation, interpolating the registered effective frame and the effective frame mask to obtain the second registered effective frame, and constructing 2.5D data based on the second registered effective frame data and key frame data to obtain 2.5D data for renal artery FFR assessment.

[0020] Optionally, the frame selection process is as follows:

[0021] Acquire time-series frame data from renal artery angiography;

[0022] Keyframes are obtained by selecting keyframes from each time frame segment.

[0023] Each time frame segment is sampled based on a time interval to obtain a valid frame;

[0024] The keyframes are combined with the valid frames to obtain multiple time series;

[0025] Optionally, the key frame can be used as the first frame, and the key frame and the effective frame can be combined to obtain multiple time series;

[0026] Optionally, the timing frame is a frame within the time period from the start of renal artery filling to the disappearance of filling;

[0027] Optionally, the keyframe is an angiographic image taken when the renal artery is first fully filled;

[0028] Optionally, the sampling includes sampling during the time period from the start of renal artery filling to complete filling, and sampling during the time period from complete filling to the disappearance of filling, to obtain a first sampling frame and a second sampling frame, respectively. The first sampling frame and the second sampling frame together constitute a valid frame.

[0029] Optionally, the ratio of the number of frames in the first sampling frame to the number of frames in the second sampling frame is 1:1;

[0030] Optionally, the sampling is replaced by: sampling the continuous time sequence from the first complete filling of the renal artery to the disappearance of filling; selecting key frames for each time frame to obtain key frames and the continuous time sequence after the key frames; and sampling the continuous time sequence after the key frames based on time intervals to obtain valid frames.

[0031] Optionally, the frame selection further includes preprocessing, performing data preprocessing on the key frames and valid frames to obtain processed key frames and valid frames, and combining the processed key frames and valid frames to obtain multiple time series;

[0032] Optionally, the preprocessing includes one or more of the following: spatial normalization, grayscale normalization, and data augmentation;

[0033] Optionally, the keyframe and valid frame are sequentially subjected to spatial normalization, grayscale normalization, and data augmentation to obtain the processed keyframe and valid frame.

[0034] Optionally, the training process of the segmentation model is as follows:

[0035] Obtain a continuous time-series frame dataset of renal artery angiography;

[0036] The continuous time-series frame dataset of the renal artery angiography is delineated to obtain a delineated continuous time-series frame dataset.

[0037] The continuous time-series frame dataset after delineation is input into the segmentation model to be trained for training until the loss function remains unchanged, thus obtaining the segmentation model;

[0038] Optionally, the training process of the segmentation model is replaced by: obtaining a multi-time series set, performing blood vessel delineation on the multi-time series set to obtain a delineated multi-time series set; inputting the delineated multi-time series set into the segmentation model to be trained for training until the loss function remains unchanged, thereby obtaining the segmentation model;

[0039] Optionally, the training process of the segmentation model further includes FFR location labeling. After delineating the vessels in the continuous time-series frame dataset of the renal artery angiography, the FFR measurement locations are labeled to obtain a labeled continuous time-series frame dataset. The labeled continuous time-series frame dataset is then input into the segmentation model to be trained for training to obtain a second segmentation model. Alternatively, after delineating the vessels in the multi-time-series set, the FFR measurement locations are labeled to obtain a labeled multi-time-series set. The labeled multi-time-series set is then input into the segmentation model to be trained for training to obtain a second segmentation model. The multi-time-series set is then input into the second segmentation model for renal artery segmentation to obtain renal artery vessel masks for keyframes and effective frames.

[0040] Optionally, the segmentation model may be one or more of the following: FCN, U-Net, PSPNet, DeepLab;

[0041] Optionally, the segmentation model includes a decoding layer and an encoding layer. The decoding layer and the encoding layer each include K 2.5D convolutional modules, where K is a natural number greater than 1. In the encoding layer, the K 2.5D convolutional modules are connected in series, and each 2.5D convolutional module includes a 2.5D convolutional layer and a downsampling layer. In the decoding layer, the K 2.5D convolutional modules are connected in series, and each 2.5D convolutional module includes a 2.5D convolutional layer and an upsampling layer. In the encoding layer, each 2.5D convolutional layer is connected to the 2.5D convolutional layer of the decoding layer of the same dimension through a skip connection. Multiple time series are sequentially passed through the encoding layer and the decoding layer to perform renal artery segmentation to obtain renal artery masks for keyframes and effective frames.

[0042] The purpose of this invention is to provide a method for constructing an assessment model for renal artery FFR, comprising:

[0043] Obtain a continuous time-series frame dataset of renal arteriography and the corresponding FFR value for each frame;

[0044] The continuous temporal frame dataset is processed using the data processing method described above for renal artery FFR assessment to obtain renal artery vessel masks and 2.5D data for key frames and effective frames; the centerline is calculated based on the renal artery vessel masks of the key frames and effective frames to generate a centerline distance map;

[0045] The renal artery mask, 2.5D data, and centerline distance map are input into the FFR prediction model to be trained to obtain the FFR prediction model. The FFR prediction model includes a first branch and a second branch in parallel. The renal artery mask, 2.5D data, and centerline distance of the key frame and the effective frame are composed of three-channel data and input into the first branch and the second branch. The first branch performs global feature extraction on the three-channel data to obtain global features, and the second branch performs local feature extraction on the three-channel data to obtain local features. The local features and global features are fused and then used to make a prediction to obtain the prediction result. The prediction result is compared with the FFR value corresponding to each frame, and the FFR prediction model is iteratively trained to obtain the FFR prediction model.

[0046] The local feature extraction involves sequentially and intermittently cropping the three-channel data along the centerline of the blood vessel to obtain cropped samples of the three-channel blood vessel sequence, and then performing local feature extraction on the cropped samples to obtain local features.

[0047] Optionally, the process of generating the centerline distance map is as follows: the renal artery vessel mask of the key frame and the effective frame is skeletonized to obtain the centerline of the renal artery vessel mask of the key frame and the effective frame, the distance from each pixel in the mask to the centerline is calculated, and the centerline distance map is generated based on the distance.

[0048] Optionally, the generation of the centerline distance map further includes connected component analysis, performing connected component analysis on the renal artery vascular mask of the keyframe and the effective frame, removing isolated pixels smaller than a preset threshold to obtain the analyzed renal artery vascular mask of the keyframe and the effective frame, and performing skeletonization calculation on the analyzed renal artery vascular mask of the keyframe and the effective frame to obtain the centerline of the renal artery vascular mask of the keyframe and the effective frame.

[0049] The purpose of this invention is to provide a method for assessing renal artery FFR, comprising:

[0050] Acquire continuous time-series frame data from renal artery angiography;

[0051] The continuous temporal frame dataset is processed using the data processing method described above for renal artery FFR assessment to obtain renal artery vessel masks and 2.5D data for key frames and effective frames; the centerline is calculated based on the renal artery vessel masks of the key frames and effective frames to generate a centerline distance map;

[0052] The renal artery vascular mask, 2.5D data, and centerline distance map are input into the FFR prediction model obtained by the above-mentioned renal artery FFR assessment model construction method to obtain the predicted FFR value.

[0053] Optionally, the method can be replaced by: acquiring continuous time-series frame data of renal artery angiography and marking preset positions to obtain marked continuous time-series frame data;

[0054] The marked continuous temporal frame data is processed by the above-mentioned data processing method for renal artery FFR assessment to obtain renal artery vascular masks and 2.5D data of key frames and effective frames;

[0055] Based on the keyframes and valid frames, the renal artery vascular mask is used to calculate the centerline and generate a centerline distance map.

[0056] The renal artery vascular mask, 2.5D data, and centerline distance map are input into the FFR prediction model obtained by the above-mentioned renal artery FFR evaluation model construction method to predict the FFR value at the preset location.

[0057] The purpose of this invention is to provide a computer program product comprising a computer program or instructions, wherein the computer program or instructions are executed by a processor to implement the above-described data processing method for renal artery FFR assessment, or to implement the above-described renal artery FFR assessment model construction method, or to implement the above-described renal artery FFR assessment.

[0058] The purpose of this invention is to provide a computer device comprising a memory, a processor, and a computer program or instructions stored in the memory, wherein the computer program or instructions are executed by the processor to implement the above-described data processing method for renal artery FFR assessment, or to implement the above-described renal artery FFR assessment model construction method, or to implement the above-described renal artery FFR assessment.

[0059] The purpose of this invention is to provide a computer-readable storage medium having a computer program or instructions stored thereon, wherein the computer program or instructions are executed by a processor to implement the above-described data processing method for renal artery FFR assessment, or to implement the above-described renal artery FFR assessment model construction method, or to implement the above-described renal artery FFR assessment.

[0060] Advantages of this invention:

[0061] 1. To address the series of technical challenges in the current renal artery FFR assessment process, such as high invasiveness, complex operation, lack of non-invasive alternatives, high technical threshold of 3D modeling, and insufficient diagnostic efficiency and real-time performance, this invention generates 2.5D spatiotemporal volume data by performing frame selection, segmentation, and registration on the temporal frame data of renal artery angiography. This data contains both temporal and spatial information and is used for renal artery FFR assessment, thereby improving the predictive accuracy and reliability of the FFR model.

[0062] 2. For continuous time-series frame data of renal artery angiography, this invention selects key frames and sampled valid frames for screening, which retains important frame data while reducing the number of frames with the same information, thereby improving the efficiency and accuracy of model processing.

[0063] 3. For FFR prediction in renal arteriography, this invention registers the segmented N effective frames with keyframes and the N effective frame masks with keyframe masks to remove respiratory or motion deformations of the patient during the scanning process, thereby improving the accuracy and reliability of FFR prediction.

[0064] 4. This invention not only predicts renal artery FFR at the training site, but also assists in decision-making and recommendations for treatment plans based on different surgical environments and requirements for rFFR at different locations, thereby improving diagnostic efficiency and assisting clinical decisions on whether to perform renal artery stent implantation.

[0065] 5. Due to fundamental pathophysiological differences between the kidneys and the heart, the renal artery possesses a dual arterial blood supply system consisting of afferent and efferent arterioles. Furthermore, the drugs that induce maximal congestion in the renal artery are entirely different from those used in the renal artery. Therefore, techniques and methods previously employed in the coronary artery field cannot be directly applied to the renal artery, which is a crucial issue that needs to be addressed. This invention discovers that high-quality, uniquely paired data specific to the renal artery (renal arteriography + renal artery FFR measured by a pressure guidewire) is key to solving this problem. Using renal arteriography data and renal artery FFR measured by a pressure guidewire as a label, a non-invasive method for predicting and assessing renal artery FFR has been constructed. Attached Figure Description

[0066] To more clearly illustrate the technical solutions in the embodiments of the present invention, 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 the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0067] Figure 1 This is a schematic diagram of a data processing method for renal artery FFR assessment provided in an embodiment of the present invention;

[0068] Figure 2 This is a schematic diagram of a data processing system for renal artery FFR assessment provided in an embodiment of the present invention;

[0069] Figure 3 This is a schematic diagram of a data processing device for renal artery FFR assessment provided in an embodiment of the present invention;

[0070] Figure 4 This is a network structure diagram of the segmentation model provided in an embodiment of the present invention;

[0071] Figure 5 The image registration process provided in this embodiment of the invention;

[0072] Figure 6 The flowchart of the rFFR prediction model structure provided in the embodiments of the present invention. Detailed Implementation

[0073] To enable those skilled in the art to better understand the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings.

[0074] In some of the processes described in the specification, claims, and accompanying drawings of this invention, multiple operations appearing in a specific order are included. However, it should be clearly understood that these operations may not be executed in the order they appear herein, or may be executed in parallel. The operation numbers, such as S101, S102, etc., are merely used to distinguish different operations and do not represent any execution order. Furthermore, these processes may include more or fewer operations, and these operations may be executed sequentially or in parallel. It should be noted that the descriptions such as "first," "second," etc., in this document are used to distinguish different messages, devices, modules, etc., and do not represent a sequential order, nor do they limit "first" and "second" to different types.

[0075] Figure 1 A schematic diagram of a data processing method for renal artery FFR assessment provided in this embodiment of the invention specifically includes:

[0076] S1: Acquire continuous time-series frame data from renal artery angiography;

[0077] In one embodiment, the continuous time frames of the renal artery angiography are obtained by one or more of the following techniques: DSA, CT, MRI;

[0078] Preferably, the continuous time frames of the renal arteriography are a single-plane renal arteriography sequence output by DSA.

[0079] S2: Frame selection is performed on the time-series frame data of the renal artery angiography to obtain multiple time series, the multiple time series including keyframes and N valid frames, where N is a natural number greater than 1;

[0080] In one embodiment, the frame selection process is as follows:

[0081] Acquire time-series frame data from renal artery angiography;

[0082] Keyframes are obtained by selecting keyframes from each time frame segment.

[0083] Each time frame segment is sampled based on a time interval to obtain a valid frame;

[0084] The keyframes are combined with the valid frames to obtain multiple time series.

[0085] In one embodiment, the keyframe is used as the first frame, and the keyframe and valid frames are combined to obtain multiple time series. The keyframe and N valid frames each correspond to a time point, and each frame represents the renal artery angiography status at the current time point. The keyframe and valid frames are combined to obtain multiple time series, which represent the changes in the renal artery angiography status on a time axis.

[0086] In one embodiment, the keyframe is used as the first frame, and keyframes and valid frames are stacked to obtain initial 2.5D data. The initial 2.5D data is then input into a segmentation model for segmentation. The 2.5D convolutional block in the segmentation model includes multiple channels, each operating in parallel. Renal artery vascular mask segmentation is performed on any frame of the initial 2.5D data to obtain an initial 2.5D segmentation mask. The initial 2.5D data and the initial 2.5D mask are registered at different scales to obtain 2.5D data for renal artery FFR assessment. Each channel processes different frames.

[0087] In one embodiment, the keyframe is used as the first frame, and the keyframe and the effective frame are combined to obtain multiple time series.

[0088] In one embodiment, the timing frame is a frame within the time period from when the contrast agent begins to fill the renal artery to when the filling disappears.

[0089] In one embodiment, the keyframe is an angiographic image taken when the renal artery is first fully filled.

[0090] In one embodiment, the sampling includes sampling during the period from the start of renal artery filling to full filling and sampling during the period from full filling to the disappearance of filling, to obtain a first sampling frame and a second sampling frame, respectively. The first sampling frame and the second sampling frame together constitute a valid frame.

[0091] In one embodiment, the ratio of the number of the first sampling frame to the number of the second sampling frame is 1:1.

[0092] In one embodiment, the sampling is replaced by: sampling a continuous time series from the first complete filling of the renal artery to the disappearance of filling; selecting keyframes for each time frame to obtain keyframes and continuous time series after the keyframes; and sampling the continuous time series after the keyframes based on time intervals to obtain valid frames.

[0093] In one embodiment, the frame selection further includes preprocessing, performing data preprocessing on the keyframes and valid frames to obtain processed keyframes and valid frames, and combining the processed keyframes and valid frames to obtain multiple time series.

[0094] Optionally, the preprocessing includes one or more of the following: spatial normalization, grayscale normalization, and data augmentation;

[0095] Optionally, the keyframe and valid frame are sequentially subjected to spatial normalization, grayscale normalization, and data augmentation to obtain the processed keyframe and valid frame.

[0096] In one embodiment, the image preprocessing process ensures the standardization and diversity of images, making the model more robust. The frame selection process can better learn the characteristics of the renal artery vessels in the angiography image, which can be used for subsequent network model training and inference.

[0097] In one specific embodiment, image preprocessing:

[0098] Image preprocessing of the contrast images (time-series frames) is performed according to the following steps:

[0099] (1) Key frame selection: The algorithm automatically extracts one frame of the renal artery when the contrast agent first fully fills the renal artery as the key frame to reduce redundancy (the key frame of each angiography sequence is unique and is one angiography image when the renal artery is first fully filled).

[0100] (2) Spatial normalization: Resample the pixel spacing to a specified reasonable value (e.g., 0.2mm / px).

[0101] (3) Gray-level normalization: Gray-level normalization is performed using Z-score standardization (μ=0, σ=1).

[0102] (4) Data augmentation during training, including but not limited to:

[0103] ① Random Affine: Performs random affine transformations on the image, such as ±15° rotation, ±5% scaling, and ±10 px translation.

[0104] ② Random Flip: Randomly flip the image horizontally or vertically.

[0105] ③ Elastic deformation: Apply random elastic deformation to the image, such as controlling the spacing between points at 32 px and σ=6 px.

[0106] ④ Noise perturbation: Apply random noise perturbation to the image, such as Gamma 0.7–1.3, random Gaussian noise sigma 0–2, or salt and pepper noise 0–0.5%.

[0107] (During training, data augmentation is performed with a 50% probability, and one of the above data augmentation methods is randomly selected with an equal probability for each method).

[0108] In one specific embodiment, the sampling frames are selected from the period from the start of renal artery filling to the disappearance of filling. The algorithm automatically extracts the sampling frames. The time from the start of filling to complete filling and the time from complete filling to the disappearance of filling are recorded for each training set of data. The number of valid sampling frames is N. N / 2 frames are sampled during the time from the start of filling to complete filling, and N / 2 frames are sampled during the time from complete filling to the disappearance of filling. (For example, if the number of valid sampling frames is set to 10, and the time from the start of filling to complete filling in a certain angiography sequence is 1 second, 10 / 2 = 5 frames are sampled during this time period, with valid frames sampled at 1 / 5 = 0.2s intervals. Similarly, the time from complete filling to the disappearance of filling is 0.2s, and 10 / 2 = 5 frames are sampled during this time period, with valid frames sampled at 0.2 / 5 = 0.04s intervals, resulting in a total of 10 sampling frames.)

[0109] Training sample construction: The keyframes and sampled frames of a single angiography sequence are combined sequentially to form an S-frame training sample (the keyframe is the first frame, the sampled frames are after the keyframes, and the training sample is a multi-time series DSA), and finally the training dataset for subsequent input is obtained, where S = 1 + N.

[0110] S3: Input the multi-time series into the segmentation model to perform renal artery segmentation to obtain the renal artery mask of key frames and effective frames;

[0111] In one embodiment, the training process of the segmentation model is as follows:

[0112] Obtain a continuous time-series frame dataset of renal artery angiography;

[0113] The continuous time-series frame dataset of the renal artery angiography is delineated to obtain a delineated continuous time-series frame dataset.

[0114] The continuous time-series frame dataset after delineation is input into the segmentation model to be trained until the loss function remains unchanged, thus obtaining the segmentation model.

[0115] In one embodiment, the training process of the segmentation model is replaced by: obtaining a multi-time series set, performing blood vessel delineation on the multi-time series set to obtain a delineated multi-time series set; inputting the delineated multi-time series set into the segmentation model to be trained for training until the loss function remains unchanged, thereby obtaining the segmentation model.

[0116] In one embodiment, the training process of the segmentation model further includes FFR location labeling. After delineating the vessels in the continuous temporal frame dataset of the renal artery angiography, FFR measurement locations are labeled to obtain a labeled continuous temporal frame dataset. This labeled continuous temporal frame dataset is then input into the segmentation model to be trained to obtain a second segmentation model. The process involves: acquiring continuous temporal frame data of the renal artery angiography; performing FFR location labeling on the continuous temporal frame data of the renal artery angiography to obtain labeled continuous temporal frame data; selecting frames from the labeled continuous temporal frame data to obtain labeled multi-time series, which includes keyframes and N valid frames, where N is a natural number greater than 1; inputting the labeled multi-time series into the second segmentation model for renal artery segmentation to obtain renal artery vessel label masks for keyframes and valid frames; and registering the labeled keyframes and their labels with the N labeled valid frames and their labels sequentially at L scales to obtain L*N labeled registered valid frame data.

[0117] Based on the effective frame data and keyframe data after the label registration, 2.5D data is constructed to obtain labeled 2.5D data for renal artery FFR assessment;

[0118] Alternatively, after delineating blood vessels in the multi-time series set, mark the FFR measurement positions to obtain a labeled multi-time series set. Input the labeled multi-time series set into the segmentation model to be trained to obtain a second segmentation model. Input the multi-time series into the second segmentation model to perform renal artery segmentation to obtain renal artery mask for key frames and effective frames.

[0119] In one embodiment, the segmentation model or the second segmentation model adopts one or more of the following: FCN, U-Net, PSPNet, DeepLab, nnUNet, medNext.

[0120] In one embodiment, the segmentation model or the second segmentation model includes a decoding layer and an encoding layer. The decoding layer and the encoding layer each include K 2.5D convolutional modules, where K is a natural number greater than 1. In the encoding layer, the K 2.5D convolutional modules are connected in series, and each 2.5D convolutional module includes a 2.5D convolutional layer and a downsampling layer. In the decoding layer, the K 2.5D convolutional modules are connected in series, and each 2.5D convolutional module includes a 2.5D convolutional layer and an upsampling layer. In the encoding layer, each 2.5D convolutional layer is connected to the 2.5D convolutional layer of the decoding layer of the same dimension through a skip connection. Multiple time series are sequentially passed through the encoding layer and the decoding layer to perform renal artery segmentation to obtain renal artery masks for keyframes and effective frames.

[0121] In one embodiment, when the segmentation model or the second segmentation model is a model constructed from a 2.5D convolutional module, multiple time series are in the 2.5D convolutional layer, and the channels of the 2.5D convolutional layer perform feature processing on each frame (key frame and valid frame).

[0122] In one specific embodiment, the training data for the model (segmentation model / second segmentation model / FFR prediction model) is acquired by collecting renal artery angiography images from multiple centers and the corresponding rFFR values ​​measured by pressure guidewires, as follows:

[0123] ①Acquisition source: Single-plane renal artery angiography sequence output from a multi-center DSA machine (≥15 fps, 512×512 pixels, 8-bit grayscale).

[0124] ② Synchronous recording: Invasive FFR value (gold standard) measured by pressure guidewire, frame synchronization timestamp.

[0125] (2) Data cleaning: The collected data is cleaned to ensure data quality and usability, and sequences with insufficient exposure, severe motion artifacts, and duct obstruction are automatically removed.

[0126] (3) Data annotation: Construct a standardized dataset, perform annotation and label matching, including outlining blood vessels in multiple frames of images and selecting points for the corresponding rFFR value measurement locations.

[0127] In one specific embodiment, a 2D UNet architecture CNN network is used to segment the renal artery. The network structure includes, but is not limited to, commonly used segmentation models such as FCN, U-Net, PSPNet, DeepLab, nnUNet, and medNext. Since the renal artery is relatively clear when filled with contrast agent, a model with a small number of parameters can be selected to directly segment and train the entire image, ensuring the integrity of the blood vessels. At the same time, since the blood vessels are thin, the downsampling scale is minimized to ensure that the blood vessel information is not lost as much as possible.

[0128] In one specific embodiment, a 2.5D network based on the U-Net architecture, such as Figure 4 As shown:

[0129] ① Backbone: Uses 2D-UNet with separable 2.5D convolution and 2.5D downsampling and upsampling.

[0130] ② Encoder: Uses separable 2.5D convolution to extract spatial texture features and capture temporal dimension information, and then uses 2.5D downsampling to extract iconic features and reduce feature size.

[0131] ③ Decoder: Uses separable 2.5D convolution to recover spatial texture features and recovers feature size through 2.5D upsampling.

[0132] ④ Output: Sigmoid activation, single-channel probability map, threshold 0.5 to obtain a binary mask.

[0133] Loss function:

[0134] Main loss: Dice Loss + Binary Cross-Entropy (weight 3:1).

[0135] Boundary loss: Add an extra 0.1×Boundary Focal Loss to improve accuracy at narrow edges.

[0136] The training dataset is divided into batches (each batch contains B training samples), and the batches are input into the constructed segmentation model (2D-UNet / 2.5D-UNet) to finally obtain the segmented results.

[0137] Input and output of the segmentation model:

[0138] Input: B × H × W × S;

[0139] Output: B × H × W × S;

[0140] Where B is the number of training samples of the angiography images in the batch, H and W are the number of pixels in the height and width of the angiography images, S is 1 + N, the sum of the number of keyframes and sampled frames (effective frames), and N is the number of samples.

[0141] In one specific embodiment, the renal artery vessel mask of the valid frames or keyframes obtained by the segmentation model or the second segmentation model is post-processed, including:

[0142] ① Connectivity analysis: Remove isolated pixels <20 px.

[0143] ②Skeletonization: The segmented results are processed by a skeletonization algorithm to obtain a 1-pixel centerline, which is then used as input to the rFFR network (the skeletonization algorithm can directly calculate the centerline mask from the segmented mask).

[0144] S4: The keyframe and the keyframe mask are registered with N valid frames and the valid frame mask in sequence based on L scales to obtain L*N registered valid frame data.

[0145] In one embodiment, the registration process is as follows:

[0146] Get keyframes, valid frames, keyframe masks, and valid frame masks;

[0147] Based on the first scale, the deformation field of the first scale is generated by measuring the similarity between the key frame and its mask, and between the effective frame and its mask.

[0148] The initial second-scale deformation field is obtained by upsampling the deformation field at the first scale.

[0149] The second-scale deformation field is generated by performing similarity measurement on keyframes and their masks, and effective frames and their masks, based on the initial second-scale deformation field.

[0150] The initial third-scale deformation field is obtained by upsampling the deformation field at the second scale.

[0151] Based on the initial third-scale deformation field, a similarity measurement is performed on the keyframe and its mask, and the effective frame and its mask to generate the third-scale deformation field, thus obtaining the registered effective frame.

[0152] In one embodiment, different scales correspond to effective frame masks and keyframe masks of different scales under different scale registration. At the first scale, the mask is reduced by a first factor to obtain the mask of the first keyframe and the mask of the first effective frame. Based on the first scale, the keyframe and the mask of the first keyframe, and the effective frame and the mask of the first effective frame are similarly measured to generate a deformation field at the first scale. At the initial second scale deformation field, the mask is reduced by a second factor to obtain the mask of the second keyframe and the mask of the second effective frame. Based on the initial second scale deformation field, the keyframe and the mask of the second keyframe, and the effective frame and the mask of the second effective frame are similarly measured to generate a deformation field at the second scale. At the initial third scale deformation field, the mask remains unchanged.

[0153] In one embodiment, the registration process further includes elastic deformation, whereby elastic deformation calculation is performed on the effective frame based on the deformation field of the third scale to obtain the registered effective frame.

[0154] In one embodiment, the registration process further includes interpolation calculation, interpolating the registered effective frame and the effective frame mask to obtain a second registered effective frame, and constructing 2.5D data for renal artery FFR assessment based on the second registered effective frame data and key frame data.

[0155] In a specific implementation, due to the patient's breathing or motion deformation during the scanning process, the necessary image registration step is indispensable. In the algorithm flow of this scheme, non-rigid elastic registration is used to register key frames and other sampling frames to ensure that DSA images scanned at different times can be well matched to the same coordinate space, so that they have effective spatial features.

[0156] In one specific embodiment, DSA multi-frame image registration:

[0157] (1) Non-rigid body registration of key frame Reference and effective frame Floatings is performed one by one based on Free-Form Deformation (FFD);

[0158] (2) For images of the same modality, use CC / MI / NGF as image similarity measures to optimize registration.

[0159] (3) Use the pyramid model to perform multi-level registration from small to large to improve the control of registration over deformation.

[0160] (4) Use WarpField to elastically deform the Floating frame to generate the registered image T(Fs), and construct 2.5D data as input for the rFFR prediction model.

[0161] Specifically:

[0162] (1) Registration input:

[0163] ① Keyframe DSA single-frame image;

[0164] ② Keyframe DSA single-frame vessel segmentation mask results;

[0165] ③ The other N DSA multi-frame images to be registered;

[0166] ④ Results of the other N DSA multi-frame vessel segmentation masks to be registered;

[0167] (2) Registration output:

[0168] ①N deformation fields between N DSA multi-frame data (including images and masks) to be registered and keyframe data (including images and masks);

[0169] ② Obtain N registered DSA multi-frame images and masks through the deformation field;

[0170] Due to the characteristics of DSA scanning, images from different frames within the same sequence exist within the same unified patient coordinate system. Therefore, in the registration of images of the same sequence and modality, only the deformations caused by the patient's respiratory movements or slight body displacement need to be registered. Thus, this registration assumes that the keyframes and effective frames already exist in the same coordinate space.

[0171] (3) Registration process, such as Figure 5 As shown:

[0172] ① Input keyframe R and valid frames Fs and their corresponding segmentation masks. For each valid frame data, perform registration sequentially. The process is as follows:

[0173] ② Construct a multi-scale pyramid model and perform registration sequentially from small scale to large scale. The process includes:

[0174] 1) Minimum scale: Reduce the input image and mask by a factor of 4;

[0175] 2) Perform image registration at a small scale. The registration process is based on optimization of similarity measures such as MI / CC / NGF to generate a small-scale WarpField deformation field Ws1.

[0176] 3) Upsample the generated deformation field Ws1 by a factor of 2 to generate the mesoscale deformation field Ws2_init;

[0177] 4) Reduce the input image and mask by a factor of 2, and use the mesoscale deformation field Ws2_init as the initial registration deformation, and input it into the registration algorithm for registration.

[0178] 5) Registration is completed at the mesoscale, generating the deformation field Ws2;

[0179] 6) Upsample the generated deformation field Ws2 by a factor of 2 to generate the original scale deformation field Ws3_init;

[0180] 7) Use the original input image and mask, and the original scale deformation field Ws3_init as the initial registration deformation, and input them into the registration algorithm for registration;

[0181] 8) Registration is completed at the original scale, generating the deformation field Ws3, which is then used as the final output T of the registration algorithm.

[0182] ③ The registered output Ts of multiple valid frames Fs is used to interpolate all the original corresponding valid frame images and masks to generate the final registered data Ts(Fs);

[0183] ④ Combine multiple registered valid frames Ts (Fs) with key frames R, for a total of S frames, and construct 2.5D data in sequence as input to the rFFR calculation model.

[0184] S5: Based on the registered valid frame data and keyframe data, 2.5D data is constructed to obtain 2.5D data for renal artery FFR assessment.

[0185] In one specific embodiment, since DSA data consists of multiple frames of time-series images, representing images scanned at different times during the flow of contrast agent through the renal artery, this invention, in order to better predict renal artery FFR and obtain more useful feature information for the model, uses inter-frame elastic registration → spatiotemporal feature extraction to treat the registered frame sequence (512×512×N) in DSA angiography as 2.5D spatiotemporal volume data. This enables renal artery FFR calculation without guidewires, congestion, or three-dimensional reconstruction under single-plane angiography conditions.

[0186] In one specific embodiment, taking a keyframe as the center, n valid frames are taken from the same position above, below (or in front of, behind) it, and stacked onto the keyframe to form 2.5D data, where n is less than or equal to N. The 2.5D data contains two-dimensional information that changes over time.

[0187] This invention provides a method for constructing an assessment model for renal artery FFR, comprising:

[0188] Obtain a continuous time-series frame dataset of renal arteriography and the corresponding FFR value for each frame;

[0189] The continuous temporal frame dataset is processed using the data processing method described above for renal artery FFR assessment to obtain renal artery vessel masks and 2.5D data for key frames and effective frames; the centerline is calculated based on the renal artery vessel masks of the key frames and effective frames to generate a centerline distance map;

[0190] The renal artery mask, 2.5D data, and centerline distance map are input into the FFR prediction model to be trained to obtain the FFR prediction model. The FFR prediction model includes a first branch and a second branch in parallel. The renal artery mask, 2.5D data, and centerline distance of the key frame and the effective frame are composed of three-channel data and input into the first branch and the second branch. The first branch performs global feature extraction on the three-channel data to obtain global features, and the second branch performs local feature extraction on the three-channel data to obtain local features. The local features and global features are fused and then used to make a prediction to obtain the prediction result. The prediction result is compared with the FFR value corresponding to each frame, and the FFR prediction model is iteratively trained to obtain the FFR prediction model.

[0191] In one embodiment, the local feature extraction involves sequentially and intermittently cropping the three-channel data along the centerline of the blood vessel to obtain a cropped sample of the three-channel blood vessel sequence, and then performing local feature extraction on the cropped sample to obtain local features.

[0192] In one embodiment, the process of generating the centerline distance map is as follows: skeletonization calculation is performed on the renal artery vascular mask of the keyframe and the effective frame to obtain the centerline of the renal artery vascular mask of the keyframe and the effective frame; the distance from each pixel in the mask to the centerline is calculated; and a centerline distance map is generated based on the distance.

[0193] In one embodiment, the generation of the centerline distance map further includes connected component analysis. Connected component analysis is performed on the renal artery vessel masks of the keyframes and valid frames to remove isolated pixels smaller than a preset threshold, obtaining the analyzed renal artery vessel masks of the keyframes and valid frames. Skeletonization calculation is then performed on the analyzed renal artery vessel masks of the keyframes and valid frames to obtain the centerline of the renal artery vessel masks of the keyframes and valid frames. The preset threshold is 20 pixels.

[0194] In another embodiment, the model is constructed as follows: acquiring a continuous time-series frame dataset of renal arteriography and the corresponding FFR value for each frame;

[0195] A labeled continuous temporal frame dataset is obtained by performing FFR position labeling on the continuous temporal frame dataset of renal arteriography;

[0196] The labeled continuous temporal frame dataset is processed using the data processing method described above for renal artery FFR assessment to obtain renal artery vessel labeling masks and labeled 2.5D data of labeled keyframes and labeled valid frames; the centerline is calculated based on the renal artery vessel labeling masks of the labeled keyframes and labeled valid frames to generate a centerline distance map;

[0197] The renal artery vascular marker mask, 2.5D marker data, and centerline distance map are input into the FFR prediction model to be trained to obtain the second FFR prediction model; the second FFR prediction model and the FFR prediction model have the same model architecture.

[0198] This invention provides a method for assessing renal artery free flow factor (FFR), comprising:

[0199] Acquire continuous time-series frame data from renal artery angiography;

[0200] The continuous temporal frame dataset is processed using the data processing method described above for renal artery FFR assessment to obtain renal artery vessel masks and 2.5D data for key frames and effective frames; the centerline is calculated based on the renal artery vessel masks of the key frames and effective frames to generate a centerline distance map;

[0201] The renal artery vascular mask, 2.5D data, and centerline distance map are input into the FFR prediction model obtained by the above-mentioned renal artery FFR assessment model construction method to obtain the predicted FFR value.

[0202] In one embodiment, the method for evaluating renal artery FFR is replaced by: acquiring continuous time-series frame data of renal artery angiography and marking preset positions to obtain marked continuous time-series frame data; the preset positions are any positions where the pressure guidewire detects renal artery FFR.

[0203] The labeled continuous temporal frame data is processed by the above-described data processing method for renal artery FFR assessment to obtain renal artery vascular labeling masks and labeled 2.5D data of labeled key frames and labeled valid frames;

[0204] Based on the renal artery vessel marker mask of the marked keyframe and marked valid frame, the centerline is calculated to generate a centerline distance map;

[0205] The renal artery vascular marker mask, 2.5D marker data, and centerline distance map are input into the second FFR prediction model obtained by the above-mentioned renal artery FFR evaluation model construction method to predict the FFR value at the preset location.

[0206] In one specific embodiment, a convolutional neural network (CNN) or a Transformer-based network or an FFR prediction model is constructed to extract features of blood vessels along the centerline of the vessels and automatically learn structural information such as vessel morphology, stenosis, and bifurcation location in the angiographic images. Optionally, the FFR prediction model also includes an attention mechanism when performing global feature extraction, which is connected between the feature extraction layer and the global pooling layer to improve the model's understanding of the topology of blood vessels. The output layer adopts a regression prediction form to directly give the rFFR prediction value of the corresponding centerline point in the image.

[0207] In one specific embodiment, rFFR is calculated as follows:

[0208] Construct an end-to-end rFFR prediction model (regression network rFFR-Net), such as Figure 6 As shown.

[0209] (1) Input:

[0210] ① The effective frames Ts (Fs) of the registration output are combined with the key frame R, for a total of S frames, to construct a 2.5D image in sequence. The image data dimension is (512×512×S).

[0211] ② The mask of the effective frame Ts(Fs) of the registration output is combined with the mask of the key frame R, for a total of N frames of masks, and a 2.5D mask is constructed in sequence, with the mask dimension being (512×512×S).

[0212] ③ Based on the 2.5Dmask constructed above, the center line distance map is extracted (512×512×S, pixel value = Euclidean distance to the center line).

[0213] ④ Three-channel concat → Network input tensor (512×512×S×3);

[0214] (2) Network structure:

[0215] ① Backbone networks: EfficientNet, ResNet, etc.

[0216] ② Global-local dual path:

[0217] 1) Global path: GAP (Global Average Pooling) → 256-dimensional vector, encoding the overall geometry.

[0218] 2) Local path: 32×32 patches (8 px interval) were cropped under the guidance of the center line, and local patch features were extracted using VisionTransformer (ViT-Tiny). The sequence length was 128.

[0219] ③ Fusion: Concat(global, local_avg) → 512-dimensional → 3-layer MLP → 1-dimensional output (predicting FFR).

[0220] (3) Loss function: Smooth L1 Loss + 0.01×Rank Loss (to ensure ranking consistency).

[0221] (4) Optimizer: AdamW, initial lr=1e-4, cosine annealing.

[0222] (5) Training strategy:

[0223] ①5-fold cross-validation.

[0224] ② Early stop patience = 15 epochs, monitoring validation set MAE.

[0225] In one specific embodiment, the model is evaluated as follows:

[0226] Supervised learning was performed using the training set, and hyperparameter tuning was used to maximize the Dice coefficient of the renal artery and minimize the error between the predicted rFFR and the true rFFR.

[0227] The model performance was evaluated in the validation set, including segmentation Dice coefficient (DSC), mean squared error (MSE), and classification (positive / negative) performance (AUC), with the following specific objectives:

[0228] (1) Segmentation: DSC;

[0229] (2) Regression: MAE, RMSE, Pearsonr.

[0230] (3) Classification: FFR ≤0.8 is considered positive, and AUC, sensitivity and specificity are calculated.

[0231] In one embodiment, the method further includes an auxiliary decision-making system that integrates the renal artery FFR assessment model into an image workstation or clinical operating system; the renal artery angiography images are automatically uploaded to the renal artery FFR assessment model for processing and rFFR prediction, and the rFFR prediction is used to assist in determining whether stent implantation treatment should be performed.

[0232] In one specific embodiment, the decision support system is constructed as follows:

[0233] Integrate the trained model into an image workstation or clinical operating system;

[0234] Renal artery angiography images are automatically uploaded to the trained model for processing and rFFR prediction;

[0235] Provide real-time rFFR assessment results during interventional procedures to help determine whether stent implantation is necessary.

[0236] In one embodiment, the trained DL model is deployed as a software tool to support doctors in quickly uploading angiographic images. The system outputs a predicted rFFR, labels and segments the renal artery region, and assists clinicians in determining whether to perform interventional procedures. Specific modules are as follows:

[0237] 1. Image receiving service: Monitors PACS and automatically retrieves sequences;

[0238] 2. Inference Engine: GPU Docker container, loading dual models of segmentation and FFR;

[0239] 3. Results storage: PostgreSQL (patient ID, predicted value, timestamp, thumbnail);

[0240] 4. Frontend: Single-page application, supports drag-and-drop upload of offline sequences.

[0241] The present invention also discloses a computer program product or system, including a computer program that, when executed by a processor, implements the above-described method steps.

[0242] Figure 2 A schematic diagram of a data processing system for renal artery FFR assessment provided in this embodiment of the invention specifically includes:

[0243] Acquisition Unit: Acquires continuous time-series frame data from renal arteriography;

[0244] Selection unit: Frame selection is performed on the time-series frame data of the renal artery angiography to obtain multiple time series, the multiple time series including keyframes and N valid frames, where N is a natural number greater than 1;

[0245] Segmentation Unit: Inputs the multi-time series into the segmentation model to perform renal artery segmentation to obtain renal artery masks for keyframes and effective frames;

[0246] Registration unit: The keyframe and the keyframe mask are sequentially registered with N valid frames and the valid frame masks based on L scales to obtain L*N registered valid frame data;

[0247] Construction Unit: Based on the registered valid frame data and key frame data, 2.5D data is constructed to obtain 2.5D data for renal artery FFR assessment.

[0248] Figure 3 An embodiment of the present invention provides a schematic diagram of a computer device, specifically including:

[0249] The system includes a memory and a processor; the memory is used to store program instructions; the processor is used to invoke the program instructions, which, when executed, perform the data processing method for renal artery FFR assessment described above, or perform the renal artery FFR assessment model construction method described above, or perform the renal artery FFR assessment described above.

[0250] The present invention also discloses a computer-readable storage medium storing a computer program. When the computer program is executed by a processor, it performs the above-described data processing method for renal artery FFR assessment, or performs the above-described renal artery FFR assessment model construction method, or performs the above-described renal artery FFR assessment.

[0251] The verification results of this verification embodiment show that assigning inherent weights to indications can improve the performance of this method compared to the default settings. Those skilled in the art will understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here. In the several embodiments provided in this application, it should be understood that the disclosed systems, devices, and methods can be implemented in other ways. For example, the device embodiments described above are merely illustrative; for example, the division of units is merely a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the displayed or discussed mutual coupling or direct coupling or communication connection may be through some interfaces, indirect coupling or communication connection of devices or units, and may be electrical, mechanical, or other forms. The units described as separate components may or may not be physically separated; 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 can be selected according to actual needs to achieve the purpose of this embodiment. Furthermore, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated units described above can be implemented in hardware or as software functional units. Those skilled in the art will understand that all or part of the steps in the various methods of the above embodiments can be implemented by a program instructing related hardware. This program can be stored in a computer-readable storage medium, which may include: read-only memory (ROM), random access memory (RAM), a magnetic disk, or an optical disk, etc.

[0252] Those skilled in the art will understand that all or part of the steps in the methods of the above embodiments can be implemented by a program instructing related hardware. The program can be stored in a computer-readable storage medium, such as a read-only memory, a disk, or an optical disk.

[0253] The computer device provided by the present invention has been described in detail above. For those skilled in the art, there will be changes in the specific implementation and application scope based on the ideas of the embodiments of the present invention. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A data processing method for assessing renal artery FFR, characterized in that, include: Acquire continuous time-series frame data from renal artery angiography; Frame selection is performed on the time-series frame data of the renal artery angiography to obtain multiple time series, which include keyframes and N valid frames, where N is a natural number greater than 1; The multiple time series are input into the segmentation model to perform renal artery segmentation to obtain the renal artery mask of key frames and effective frames; The keyframes and their masks are registered sequentially with N valid frames and their masks based on L scales to obtain L. N registered valid frame data; Based on the registered valid frame data and keyframe data, 2.5D data is constructed to obtain 2.5D data for renal artery FFR assessment; The registration process is as follows: Get keyframes, valid frames, keyframe masks, and valid frame masks; Based on the first scale, the deformation field of the first scale is generated by measuring the similarity between the key frame and its mask, and between the effective frame and its mask. The initial second-scale deformation field is obtained by upsampling the deformation field at the first scale. The second-scale deformation field is generated by performing similarity measurement on keyframes and their masks, and effective frames and their masks, based on the initial second-scale deformation field. The initial third-scale deformation field is obtained by upsampling the deformation field at the second scale. Based on the initial third-scale deformation field, a similarity measurement is performed on the keyframe and its mask, and the effective frame and its mask to generate the third-scale deformation field, thus obtaining the registered effective frame.

2. The data processing method for renal artery FFR assessment according to claim 1, characterized in that, Effective frame masks and keyframe masks at different scales under different scale registration; At the first scale, the mask is reduced by a first factor to obtain the mask of the first keyframe and the mask of the first valid frame. Based on the first scale, the keyframe and the mask of the first keyframe, and the valid frame and the mask of the first valid frame are similarly measured to generate the deformation field of the first scale. At the initial second scale deformation field, the mask is reduced by a second factor to obtain the mask of the second keyframe and the mask of the second valid frame. Based on the initial second scale deformation field, the keyframe and the mask of the second keyframe, and the valid frame and the mask of the second valid frame are similarly measured to generate the deformation field of the second scale. At the initial third scale deformation field, the mask remains unchanged.

3. The data processing method for renal artery FFR assessment according to claim 1, characterized in that, The registration process also includes elastic deformation, which calculates the elastic deformation of the effective frame based on the deformation field of the third scale to obtain the registered effective frame.

4. The data processing method for renal artery FFR assessment according to claim 1, characterized in that, The registration process also includes interpolation calculation, interpolating the registered effective frame and the effective frame mask to obtain the second registered effective frame, and constructing 2.5D data based on the second registered effective frame data and key frame data to obtain 2.5D data for renal artery FFR assessment.

5. The data processing method for renal artery FFR assessment according to claim 1, characterized in that, The frame selection process is as follows: Acquire time-series frame data from renal artery angiography; Keyframes are obtained by selecting keyframes from each time frame segment. Each time frame segment is sampled based on a time interval to obtain a valid frame; The keyframes are combined with the valid frames to obtain multiple time series.

6. The data processing method for renal artery FFR assessment according to claim 1, characterized in that, The keyframe is used as the first frame, and the keyframe and the effective frame are combined to obtain multiple time series.

7. The data processing method for renal artery FFR assessment according to claim 1, characterized in that, The time frame refers to the frame within the time period from the start of renal artery filling to the disappearance of filling.

8. The data processing method for renal artery FFR assessment according to claim 1, characterized in that, The keyframe is an angiographic image taken when the renal artery is fully filled for the first time.

9. The data processing method for renal artery FFR assessment according to claim 5, characterized in that, The sampling includes sampling during the time period from the start of renal artery filling to complete filling, and sampling during the time period from complete filling to the disappearance of filling, to obtain a first sampling frame and a second sampling frame, which together constitute a valid frame.

10. The data processing method for renal artery FFR assessment according to claim 9, characterized in that, The ratio of the number of frames in the first sampling frame to the number of frames in the second sampling frame is 1:

1.

11. The data processing method for renal artery FFR assessment according to claim 9, characterized in that, The sampling is replaced by sampling of a continuous time series from the first complete filling of the renal artery to the disappearance of filling; For each time frame segment, key frames are selected to obtain key frames and continuous time sequences following the key frames. The continuous time sequences following the key frames are sampled based on time intervals to obtain valid frames.

12. The data processing method for renal artery FFR assessment according to claim 5, characterized in that, The frame selection also includes preprocessing, which involves preprocessing the key frames and valid frames to obtain processed key frames and valid frames, and then combining the processed key frames and valid frames to obtain multiple time series.

13. The data processing method for renal artery FFR assessment according to claim 12, characterized in that, The preprocessing includes one or more of the following: spatial normalization, grayscale normalization, and data augmentation.

14. The data processing method for renal artery FFR assessment according to claim 1, characterized in that, The training process of the segmentation model is as follows: Obtain a continuous time-series frame dataset of renal artery angiography; The continuous time-series frame dataset of the renal artery angiography is delineated to obtain a delineated continuous time-series frame dataset. The continuous time-series frame dataset after delineation is input into the segmentation model to be trained until the loss function remains unchanged, thus obtaining the segmentation model.

15. The data processing method for renal artery FFR assessment according to claim 14, characterized in that, The training process of the segmentation model is replaced by: obtaining a multi-time series set, performing blood vessel delineation on the multi-time series set to obtain a delineated multi-time series set; inputting the delineated multi-time series set into the segmentation model to be trained for training until the loss function remains unchanged, thus obtaining the segmentation model.

16. The data processing method for renal artery FFR assessment according to claim 14, characterized in that, The training process of the segmentation model also includes FFR position labeling. After delineating the blood vessels in the continuous temporal frame dataset of the renal artery angiography, the FFR measurement positions are labeled to obtain a labeled continuous temporal frame dataset. The labeled continuous temporal frame dataset is then input into the segmentation model to be trained to obtain the second segmentation model.

17. The data processing method for renal artery FFR assessment according to claim 15, characterized in that, The training process of the segmentation model also includes FFR position marking. After delineating the blood vessels in the multi-time series set, the FFR measurement positions are marked to obtain a marked multi-time series set. The marked multi-time series set is input into the segmentation model to be trained for training to obtain a second segmentation model. The multi-time series is input into the second segmentation model to perform renal artery segmentation to obtain renal artery mask for key frames and effective frames.

18. The data processing method for renal artery FFR assessment according to claim 1, characterized in that, The segmentation model adopts one or more of the following: FCN, U-Net, PSPNet, DeepLab.

19. The data processing method for renal artery FFR assessment according to claim 1, characterized in that, The segmentation model includes a decoding layer and an encoding layer. Each decoding layer and encoding layer includes K 2.5D convolutional modules, where K is a natural number greater than 1. In the encoding layer, the K 2.5D convolutional modules are connected in series, and each 2.5D convolutional module includes a 2.5D convolutional layer and a downsampling layer. In the decoding layer, the K 2.5D convolutional modules are connected in series, and each 2.5D convolutional module includes a 2.5D convolutional layer and an upsampling layer. In the encoding layer, each 2.5D convolutional layer is connected to the 2.5D convolutional layer of the decoding layer of the same dimension through a skip connection. Multiple time series are sequentially passed through the encoding layer and the decoding layer to perform renal artery segmentation to obtain renal artery masks for keyframes and effective frames.

20. A method for constructing an assessment model for renal artery FFR, characterized in that, include: Obtain a continuous time-series frame dataset of renal arteriography and the corresponding FFR value for each frame; The continuous temporal frame dataset is processed by the data processing method for renal artery FFR assessment as described in any one of claims 1-19 to obtain renal artery vessel masks and 2.5D data for key frames and effective frames; the centerline is calculated based on the renal artery vessel masks of the key frames and effective frames to generate a centerline distance map; The renal artery mask, 2.5D data, and centerline distance map are input into the FFR prediction model to be trained to obtain the FFR prediction model. The FFR prediction model includes a first branch and a second branch in parallel. The renal artery mask, 2.5D data, and centerline distance of the key frame and the effective frame are composed of three-channel data and input into the first branch and the second branch. The first branch performs global feature extraction on the three-channel data to obtain global features, and the second branch performs local feature extraction on the three-channel data to obtain local features. The local features and global features are fused and then used to make a prediction to obtain the prediction result. The prediction result is compared with the FFR value corresponding to each frame, and the FFR prediction model is iteratively trained to obtain the FFR prediction model.

21. The method for constructing an assessment model for renal artery FFR according to claim 20, characterized in that, The local feature extraction involves sequentially and intermittently cropping the three-channel data along the centerline of the blood vessel to obtain a cropped sample of the three-channel blood vessel sequence, and then performing local feature extraction on the cropped sample to obtain local features.

22. The method for constructing an assessment model for renal artery FFR according to claim 20, characterized in that, The process of generating the centerline distance map is as follows: skeletonization calculation is performed on the renal artery vascular mask of the keyframe and the effective frame to obtain the centerline of the renal artery vascular mask of the keyframe and the effective frame; the distance from each pixel in the mask to the centerline is calculated; and a centerline distance map is generated based on the distance.

23. The method for constructing an assessment model for renal artery FFR according to claim 22, characterized in that, The generation of the centerline distance map also includes connected component analysis. Connected component analysis is performed on the renal artery vascular mask of the key frame and the effective frame to remove isolated pixels smaller than a preset threshold to obtain the renal artery vascular mask of the analyzed key frame and the effective frame. Skeletonization calculation is performed on the renal artery vascular mask of the analyzed key frame and the effective frame to obtain the centerline of the renal artery vascular mask of the key frame and the effective frame.

24. The method for constructing an assessment model for renal artery FFR according to claim 20, characterized in that, The method also includes an auxiliary decision-making system that integrates the renal artery FFR assessment model into an image workstation or clinical operating system; the renal artery angiography images are automatically uploaded to the renal artery FFR assessment model for processing and rFFR prediction, and the rFFR prediction is used to assist in determining whether to perform stent implantation treatment.

25. A method for assessing renal artery FFR, characterized in that, include: Acquire continuous time-series frame data from renal artery angiography; The continuous temporal frame dataset is processed by the data processing method for renal artery FFR assessment as described in any one of claims 1-19 to obtain renal artery vessel masks and 2.5D data for key frames and effective frames; the centerline is calculated based on the renal artery vessel masks of the key frames and effective frames to generate a centerline distance map; The renal artery vascular mask, 2.5D data, and centerline distance map are input into the FFR prediction model obtained by the renal artery FFR assessment model construction method according to any one of claims 20-24 to obtain the predicted FFR value.

26. The method for assessing renal artery FFR according to claim 25, characterized in that, The method is replaced by: acquiring continuous time-series frame data of renal artery angiography and marking preset positions to obtain marked continuous time-series frame data; The marked continuous temporal frame data is processed by the data processing method for renal artery FFR assessment according to any one of claims 1-19 to obtain renal artery vascular masks and 2.5D data of key frames and effective frames; Based on the keyframes and valid frames, the renal artery vascular mask is used to calculate the centerline and generate a centerline distance map. The renal artery vascular mask, 2.5D data, and centerline distance map are input into the FFR prediction model obtained by the renal artery FFR assessment model construction method according to any one of claims 21-24 to predict the FFR value at the preset location.

27. A computer program product comprising a computer program or instructions, characterized in that, The computer program or instructions are executed by a processor to implement the data processing method for renal artery FFR assessment according to any one of claims 1-19, or to implement the renal artery FFR assessment model construction method according to any one of claims 20-24, or to implement the renal artery FFR assessment method according to any one of claims 25-26.

28. A computer device comprising a memory, a processor, and a computer program or instructions stored in the memory, characterized in that, The computer program or instructions are executed by a processor to implement the data processing method for renal artery FFR assessment according to any one of claims 1-19, or to implement the renal artery FFR assessment model construction method according to any one of claims 20-24, or to implement the renal artery FFR assessment method according to any one of claims 25-26.

29. A computer-readable storage medium having a computer program or instructions stored thereon, characterized in that, The computer program or instructions are executed by a processor to implement the data processing method for renal artery FFR assessment according to any one of claims 1-19, or to implement the renal artery FFR assessment model construction method according to any one of claims 20-24, or to implement the renal artery FFR assessment method according to any one of claims 25-26.

Images