A Method and System for Assessing Postoperative Abdominal Organ Ischemia Risk Based on Image Analysis

Abstract

This invention discloses a method and system for assessing postoperative abdominal organ ischemia risk based on image analysis, belonging to the field of image analysis technology. The method includes acquiring enhanced abdominal CT images of the target patient and outputting them after standardization processing; employing an improved 3D U-Net++ architecture with multi-task collaborative learning to jointly segment the target patient's organs and target regions, integrating geometric priors and topological inference mechanisms; its key technical points are: using multi-task segmentation to focus RPPR calculation on the real ischemic area, avoiding average dilution of the signal across all organs, and topological correction to ensure that vascular features reflect the real anatomy; furthermore, through the joint analysis of low-perfusion area prediction masks and vascular VTIF, it reveals two ischemic subtypes: structural occlusion and functional hypoperfusion, promoting the individualization of clinical intervention strategies and defining the necessary vascular intervention or conservative treatment, making the overall solution both innovative and clinically applicable.

Description

Technical Field

[0001] This invention relates to the field of image analysis technology, specifically to a method and system for assessing the risk of postoperative abdominal organ ischemia based on image analysis. Background Technology

[0002] Image analysis refers to the process of using computer technology to process, identify, quantify, and interpret images, including medical images, remote sensing images, and industrial inspection images, in order to extract meaningful information or assist in decision-making. In the medical field, image analysis specifically refers to the automatic or semi-automatic analysis of medical images such as X-rays, CT scans, MRI scans, and ultrasound images for disease detection, diagnosis, treatment planning, and efficacy evaluation.

[0003] Current medical imaging analysis techniques have significant limitations in assessing postoperative peritoneal ischemia: First, traditional segmentation methods focus only on anatomical structures. For example, using U-Net to delineate the entire kidney fails to distinguish or effectively differentiate normal renal parenchyma from areas of localized hypoperfusion, resulting in RPPR calculations being diluted across the entire organ and masking early ischemic lesions. Second, vascular reconstruction relies on fixed-scale Frangi filtering or simple skeletonization, which is highly susceptible to interruption by metal artifacts after aortic stenting. For instance, the distal superior mesenteric artery is often invisible due to ray sclerosis, leading to misjudgment of branch continuity and subsequent complications. The current risk models are mostly based on a single perfusion indicator, such as judging ischemia by only the decrease in CT value. They cannot identify occlusion of the feeding artery in patients with occlusion but still good enhancement. These patients often have atypical CT manifestations due to chronic compensation and are often missed until acute kidney injury occurs. The existing process for assessing the risk of postoperative abdominal organ ischemia is disconnected from the functional and structural analysis. It lacks modeling of the perfusion-pathway coupling relationship and makes it difficult to distinguish between functional hypoperfusion that can be treated conservatively and structural occlusion that requires emergency intervention. This can easily lead to problems such as over-operation or delayed intervention. Summary of the Invention

[0004] To achieve the above objectives, the present invention provides the following technical solution:

[0005] A method for assessing postoperative abdominal organ ischemia risk based on image analysis, comprising:

[0006] Acquire enhanced abdominal CT images of the target patient, and output them after standardization processing;

[0007] An improved 3D U-Net++ architecture using multi-task collaborative learning is employed to jointly segment organs and target regions of the target patient. It integrates geometric priors and topological reasoning mechanisms to achieve topological correction and obtains the blood supply artery to each organ, thereby deriving the corresponding branch continuity index.

[0008] The dynamic perfusion feature extraction process is completed, and the input features are comprehensively acquired. An ischemia risk scoring model is constructed using an integrated XGBoost classifier, and a preliminary abdominal organ ischemia risk score is output. The input features include at least the completion results of the dynamic perfusion feature extraction process and the branch continuity index.

[0009] Using the branch continuity index as a basis, a vascular flow correction factor is introduced to perform a correction process on the preliminary abdominal organ ischemia risk score to generate the final abdominal organ ischemia risk score, and simultaneously trigger a visualization risk warning strategy to assess the risk status of the target patient.

[0010] Furthermore, the standardization process for abdominal enhanced CT images is as follows:

[0011] Image registration: A non-rigid registration algorithm based on mutual information is used to unify the abdominal enhanced CT images at each time point to the preoperative baseline space. Window width and window level normalization: A liver window is set to ensure consistent grayscale between different scans. Noise suppression and edge enhancement: An improved nonlocal means denoising combined with the Laplacian sharpening operator is applied. In the denoising stage, an improved nonlocal means algorithm is used, and a local variance adaptive factor is introduced into the similarity weight. In the enhancement stage, the Laplacian sharpening operator is superimposed on the denoised image, and the standardized image is output.

[0012] Furthermore, the improved 3D U-Net++ architecture configuration includes a backbone encoder and a dual-task decoder head, and a total loss function is introduced simultaneously. The backbone encoder is built based on 3D ResNet-34, and extracts multi-scale spatial-semantic features through stepwise downsampling. The dual-task decoder head includes branch A and branch B. Branch A is an organ segmentation head, which reconstructs the U-Net++ path through nested dense skip connections and outputs a predicted whole organ mask y^seg containing several types of organ masks. Branch B is a target region detection head, and the target region is a low-perfusion region. It uses parallel lightweight convolutional layers to predict a binary functional abnormality mask y^perf.

[0013] Furthermore, the total loss function is a weighted sum of three terms: organ segmentation loss L_seg, low perfusion detection loss L_perf, and consistency constraint loss L_consist. The corresponding weights used all have values ​​in the range [0, 1]. Among them, organ segmentation loss L_seg is used as the main loss to ensure the accuracy of anatomical structures; low perfusion detection loss L_perf is used as an auxiliary loss to guide the network to focus on pathological areas; and consistency constraint loss is used as a regularization term to introduce physical priors.

[0014] Furthermore, the process of integrating geometric priors and topological reasoning mechanisms is as follows:

[0015] Improved Frangi filter: Direction-adaptive multi-scale enhancement, applying Hessian matrix analysis to the standardized abdominal enhanced CT image to dynamically select the optimal scale σ. ∗ (x);

[0016] GNN-driven vascular topology correction: The input is the initial vascular centerline point cloud output by a lightweight 3D ResUNet, represented as a set of three-dimensional spatial coordinate point clouds: {p1, p2, ..., p...} N}, where p i1 Let be the physical coordinates of the i1th centerline point; transform the point cloud into a graph structure G=(V, E) for GNN inference, if d(v i1 v j If the value is less than the set value, then add edge e. i1j Where V represents the set of nodes, and each node v i1 ∈V corresponds to a centerline point p i1 E represents the set of edges, where each edge e i1j ∈E represents node v i1 With v j There is a potential connection between them; e i1j Indicates the connection node v i1 With v j Edges are established only if the two edges satisfy the proximity condition; a message-passing graph neural network is used to process each edge e. i1j Perform binary classification to determine whether it represents a real blood vessel connection, and predict the edge connectivity confidence level c. i1j ∈[0,1]; Edge connectivity confidence c based on GNN output i1j For automatic repair to be performed in conjunction with geometric constraints, two repair conditions must be met simultaneously:

[0017] Condition A is c i1j <0.4 indicates low confidence level;

[0018] Condition B is d(v) i1 v j ) < 1.6 × set value, indicating a short-distance clearance; d(v i1 v j ) represents node v i1 With v j The Euclidean distance between them.

[0019] Furthermore, the corresponding branch continuity index Cr is derived. i The basis is: Cr i=L_de(i) / L_ex(i); where L_de(i) represents the effective length of the corresponding artery actually detected in the repaired vessel centerline, and L_ex(i) represents the expected full length of the corresponding artery based on the standard anatomical template; the process of completing the dynamic perfusion feature extraction is as follows:

[0020] Constructing organ-specific functional sub-region masks: Perform mask-limiting actions on each organ i, extract the predicted whole organ mask y^seg, and obtain y^seg(i)∈{0,1}. Simultaneously extract the predicted binary functional abnormality mask y^perf, and obtain y^perf∈{0,1}. Generate organ-limited low-perfusion binary mask y^seg(i); Calculate the two-level relative perfusion retention rate: Define two types of RPPR and obtain the required RPPR_local after filtering. (i) (t); where the two defined RPPRs are the global RPPR used for clinical overview, namely RPPR_global. (i) (t) and the local RPPR used for risk modeling, namely RPPR_local (i) (t); where the selection criteria are: the required RPPR_local under the condition that the number of voxels N_perf(i) in the low-perfusion region of organ i is ≥ the minimum effective voxel number threshold N_min. (i) (t) takes the value of local RPPR; otherwise, the required RPPR_local (i) (t) takes the value of global RPPR.

[0021] Furthermore, the process of running the ischemia risk scoring model is as follows:

[0022] For each target patient, the input features are a multi-dimensional feature vector, including the required RPPR_local for each organ i. (i) (t) and the continuity index Cr of each branch i The input features are fed into a pre-trained ischemia risk scoring model to generate an ischemia risk probability Pr, which is then linearly mapped to a preliminary abdominal organ ischemia risk score VIRScore_0. The ischemia risk scoring model uses XGBoost as the base learner.

[0023] Furthermore, the vascular flow domain correction factor VTIF is introduced, defined as: the branch continuity index Cr corresponding to different organs i. i Its corresponding weight w iMultiply the results to obtain the cumulative value for each target organ; finally, multiply the cumulative value by the reciprocal of the total number of target organs to generate the required vascular flow correction factor VTIF; perform correction processing on the preliminary abdominal organ ischemia risk score VIRScore_0 to obtain the required final abdominal organ ischemia risk score VIRScore based on: VIRScore = VIRScore_0 + Δ × (1 - VTIF); where Δ represents the maximum correction magnitude.

[0024] Furthermore, the process of synchronously triggering the visual risk warning strategy is as follows:

[0025] The final abdominal organ ischemia risk score (VIRScore) is compared with a preset risk threshold range. If the final VIRScore is greater than the upper limit of the risk threshold range, it is considered high risk; if the final VIRScore is less than the lower limit of the risk threshold range, it is considered low risk. If the final VIRScore is within the risk threshold range, a secondary judgment mechanism is initiated. The vascular flow correction factor (VTIF) is extracted and compared with a preset retrospective indicator. If the VTIF does not exceed the retrospective indicator, it is considered occult high risk; otherwise, it is considered moderate risk. A color-coded ischemia risk heatmap is generated according to different risk levels, defined as: high risk in red, occult high risk in light red, moderate risk in yellow, and low risk in green. High-risk and occult high-risk vascular segments are automatically extracted and labeled.

[0026] A postoperative abdominal organ ischemia risk assessment system based on image analysis, comprising:

[0027] Image preprocessing module: Acquires enhanced abdominal CT images of the target patient, performs standardization processing, and outputs the images.

[0028] Multi-task segmentation module: The improved 3D U-Net++ architecture of multi-task collaborative learning is used to jointly segment the organs and target regions of the target patient. It integrates geometric prior and topological reasoning mechanisms to achieve topological correction and obtains the blood supply artery to each organ to obtain the corresponding branch continuity index.

[0029] Risk modeling module: Completes dynamic perfusion feature extraction, comprehensively acquires input features, constructs an ischemia risk scoring model using an integrated XGBoost classifier, and outputs a preliminary abdominal organ ischemia risk score; among which, the input features include at least: the completion result of the dynamic perfusion feature extraction and the branch continuity index;

[0030] Dynamic correction module: It calls the branch continuity index as the basis, introduces the vascular flow correction factor, performs correction processing on the preliminary abdominal organ ischemia risk score to generate the final abdominal organ ischemia risk score, and simultaneously triggers a visual risk warning strategy to assess the risk status of the target patient.

[0031] The present invention provides a method and system for assessing the risk of postoperative abdominal organ ischemia based on image analysis, which has the following beneficial effects: (1) This scheme effectively overcomes the limitations of postoperative CT image quality by modeling the vascular centerline as a graph structure and introducing GNN for intelligent repair with topological perception. The repair strategy takes into account both anatomical rationality and algorithm robustness, significantly improves the integrity of the vascular tree, and provides a reliable structural basis for subsequent hemodynamic analysis and risk scoring.

[0032] (2) In this scheme, multi-task segmentation is used to focus RPPR calculation on the real ischemic area, avoiding the average dilution of the signal across the whole organ. Topological correction ensures that the vascular features reflect the real anatomy. In addition, the combined analysis of low perfusion area prediction mask and vascular VTIF reveals two ischemic subtypes: structural occlusion and functional hypoperfusion, promotes the individualization of clinical intervention strategies, and defines the situations requiring vascular intervention or conservative treatment, making the overall scheme both innovative and clinically applicable.

[0033] (3) This scheme uses the branch continuity index to dynamically correct the preliminary abdominal organ ischemia risk score, realizing a single source for dual use. It can be used for the original feature input and can also extract the vascular flow correction factor. This not only allows related data belonging to the same data stream to be deeply mined without increasing the computational overhead, but also improves the accuracy of risk stratification. Furthermore, it reveals the value of structural integrity as an independent prognostic factor, promoting clinical shift from only looking at perfusion to a two-dimensional assessment of perfusion + pathway, achieving a dual breakthrough in clinical cognition and original effect.

[0034] (4) This scheme introduces the branch continuity index to construct the vascular flow correction factor, which dynamically corrects the preliminary abdominal organ ischemia risk score. In some extreme cases, it can effectively identify structural ischemia caused by stent coverage. Even if the CT manifestation is atypical, it can intervene in advance to avoid serious consequences such as acute kidney injury, significantly improve the sensitivity of early diagnosis, and avoid missing high-risk patients. Attached Figure Description

[0035] Figure 1 This is a flowchart illustrating the postoperative abdominal organ ischemia risk assessment method based on image analysis in this invention. Detailed Implementation

[0036] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0037] Example 1:

[0038] Please see Figure 1 This embodiment provides a method for assessing the risk of postoperative ischemic lesions in abdominal organs based on image analysis. This method is applicable to patients undergoing cardiovascular surgery, especially after type A aortic dissection repair. It aims to achieve early, accurate, and dynamic assessment of the perfusion status of abdominal organs, such as the liver, spleen, kidney, and mesentery, through deep learning-driven multimodal image analysis technology, thereby predicting the risk of postoperative ischemic complications. One of its core features is that it can selectively solve the two major technical challenges of functional-anatomical separation and vascular rupture in specific scenarios through multi-task segmentation and topology-aware vascular reconstruction. On the one hand, it can improve the spatial specificity of organ perfusion assessment, and on the other hand, it can enhance the anatomical rationality of vascular structure inference. Specific details of the scheme can be found in the following description.

[0039] The main steps of this evaluation method are as follows:

[0040] S1. Acquire enhanced abdominal CT images of the target patient, and output them after standardization processing;

[0041] In this embodiment, the target patient is in the range of preoperative and postoperative status of cardiovascular surgery, that is, the preoperative baseline and the immediate postoperative state are obtained, such as abdominal enhanced CT images within 0 to 72 hours; the specific time points are selected as multiple time points such as within 24 hours before surgery and 6, 24, 48 and 72 hours after surgery; the abdominal enhanced CT images include the entire abdominal cavity, with a slice thickness ≤3mm, and use standard reconstruction algorithms;

[0042] Specifically, this step emphasizes the precision of the time window and the consistency of the scanning protocol to ensure that perfusion changes can be reliably captured. Therefore, preoperative baseline images must be completed within 24 hours before surgery to reflect the true physiological state of the target patient and exclude interference from underlying diseases such as chronic liver disease and renal atrophy on perfusion assessment. The selected postoperative follow-up time points cover the key pathological window of abdominal ischemia after aortic repair: early stage (<24 hours) is often caused by stent obstruction or embolism leading to acute ischemia, while late stage (>48 hours) may be due to secondary mesenteric vein thrombosis or reperfusion injury. All scans of abdominal enhanced CT images use the portal venous phase, usually 70±5 seconds after contrast agent injection, as the organ parenchyma enhancement is most uniform during this phase, which is beneficial for comparing CT values ​​across time points. The image parameter requirements are not only slice thickness ≤3mm and standard reconstruction algorithm, but more specifically: slice thickness ≤3 mm, slice interval ≤1.5 mm, to avoid partial volume effect obscuring small blood vessels or thin-walled intestinal segments; tube voltage 100 to 120. kV, automatic tube current modulation, balancing radiation dose and image noise; the standard reconstruction algorithm in this embodiment uses a standard soft tissue convolution kernel, such as Siemens B30f or GE Standard, and disables bone algorithms or excessively high iterative reconstruction intensity settings to prevent texture distortion from affecting subsequent segmentation; the above specific settings are because if the arterial phase or delayed phase is used, the difference in output volume and cycle time at different time points will cause the enhancement phase shift, making the CT values ​​incomparable; while excessive slice thickness will blur the details of the intestinal wall or renal cortex, reducing the detection rate of low perfusion.

[0043] The process of standardizing abdominal enhanced CT images is as follows:

[0044] S1.1 Image Registration: A non-rigid registration algorithm based on mutual information was used to unify the enhanced abdominal CT images at each time point to the preoperative baseline space to eliminate displacement caused by body position and respiratory movements. Specifically, the preoperative enhanced abdominal CT image was used as a fixed reference image, and the images at each postoperative time point were used as floating images for registration. An affine transformation + B-spline non-rigid deformation field combination model was used to first correct the overall displacement / rotation / scaling, and then compensate for local deformation. The similarity measure used was Normalized Mutual Information (NMI), which is robust to nonlinear grayscale changes in multimodal or multi-temporal images. Subsequent registration accuracy verification involved selecting 5 anatomical landmarks in the right lobe of the liver, such as the portal vein bifurcation and the entrance of the right hepatic vein. The average Euclidean distance after registration was ≤1.2 mm. S1.2 Window Width and Level Normalization: A liver window was set, for example: window width 150 HU, window level 50. HU ensures grayscale consistency across different scans. The reason for this setting is that the liver is the largest solid organ in the abdominal cavity, and its CT value range is central, making it suitable as a grayscale reference. This window width can clearly display the liver parenchyma, spleen, renal cortex, and intestinal wall enhancement simultaneously, accommodating the visualization of multiple organs. Therefore, the above specific value was chosen in this embodiment. S1.3, Noise Suppression and Edge Enhancement: An improved nonlocal means denoising combined with the Laplacian sharpening operator is applied to improve the contrast between small blood vessels and parenchymal boundaries. In the denoising stage, an improved nonlocal means algorithm is used, with a search window set to 21×21×7 voxels and a neighborhood block size of 5×5×3. A local variance adaptive factor is introduced for similarity weights to avoid excessive smoothing in low-contrast areas, such as edematous intestinal walls. In the enhancement stage, the Laplacian sharpening operator is superimposed on the denoised image to enhance organ boundaries and small blood vessel contours. The final output image can achieve a signal-to-noise ratio improvement of over 35% and an edge gradient amplitude improvement of over 20%, significantly improving the sensitivity of the corresponding model to weak ischemic signs in subsequent S2.

[0045] In summary, the specific S1 scheme proposed above all serve a core objective: to construct a set of spatiotemporally aligned, grayscale consistent, and structurally clear multi-phase CT image sequences to ensure that the output image results not only retain the true physiological perfusion information but also eliminate technical interference factors, thus laying a solid data foundation for subsequent analysis or adjustment.

[0046] S2. An improved 3D U-Net++ architecture using multi-task collaborative learning is employed to jointly segment the organs and target regions of the target patient. Geometric priors and topological inference mechanisms are integrated to achieve topological correction. Based on this, the blood supply arteries to each organ are obtained, and the corresponding branch continuity index Cr is derived. i ;

[0047] The improved 3D U-Net++ architecture includes a backbone encoder and a dual-task decoder head, with a total loss function introduced simultaneously. The backbone encoder is built on 3D ResNet-34, extracting multi-scale spatial-semantic features through stepwise downsampling. The dual-task decoder head consists of branch A and branch B. Branch A is the organ segmentation head, which reconstructs the U-Net++ path through nested dense skip connections, outputting several organ masks, as follows:

[0048] y^seg=[y^liver, y^spleen, y^kidney_L, y^kidney_R, y^meentery];

[0049] y^seg represents the prediction of whole organ mask, y^liver represents the liver mask, y^spleen represents the spleen mask, y^kidney_L represents the left kidney mask, y^kidney_R represents the right kidney mask, and y^mesentery represents the mesentery mask.

[0050] Branch B is the target region detection head. In this embodiment, the target region is specifically the low-perfusion region: parallel lightweight convolutional layer, predicting binary functional abnormality mask y^perf. This region is defined as a region where the CT value is lower than the mean of the same organ by 2 standard deviations and the boundary is blurred. The rationale for the above network architecture design is as follows: traditional single-task models only learn anatomical boundaries and cannot distinguish between normal kidneys and low-perfusion kidneys. The multi-task design allows the network to learn structure and function at the same time, improving the specificity of subsequent perfusion quantification.

[0051] The specific explanation of the total loss function introduced subsequently is as follows:

[0052] The total loss function is a weighted sum of three terms: organ segmentation loss L_seg, low perfusion detection loss L_perf, and consistency constraint loss L_consist. The corresponding weights used all range from [0, 1]. The specific values ​​are set according to actual needs or optimized on the validation set through grid search. This will not be elaborated on here.

[0053] Among them, organ segmentation loss L_seg is used as the main loss to ensure the accuracy of anatomical structure; low perfusion detection loss L_perf is used as an auxiliary loss to guide the network to focus on pathological areas; and consistency constraint loss is used as a regularization term to introduce physical priors and improve generalization.

[0054] (1) Regarding organ segmentation loss L_seg: L_seg=(1-Dice(y_seg,y^seg))+λ1×FocalLoss(y_seg,y^seg); where y_seg represents the ground truth labels of several types of organs manually labeled, y^seg is the predicted overall organ mask, i.e. organ probability map, Dice() represents the measurement of segmentation overlap, which is sensitive to small organs, such as the mesentery, FocalLoss() represents the solution to the problem of extreme imbalance between foreground and background, λ1 is the empirical weight, which is used to balance Dice's preference for the overall structure and Focal Loss' attention to edge details, and its value range is also in [0,1].

[0055] (2) Regarding the low perfusion detection loss L_perf: L_perf = BinaryFocalLoss(y_perf, y^perf); where y_perf is a pseudo-label generated from the standardized image, defined as:

[0056] y_perf(p)=1, if I(p)<μ i -2σ i and ∣∇I(p)∣ <τ g ;

[0057] y_ perf(p)=0, if otherwise;

[0058] In the formula, the pseudo-label value of the low-perfusion region at spatial location p is 1, indicating that the voxel at spatial location p is identified as a suspected ischemic / low-perfusion region; a value of 0 indicates normal tissue; spatial location p is a three-dimensional coordinate in the image; I(p) represents the gray value of the standardized abdominal enhanced CT image I at spatial location p, i.e., the HU value; μ i σ represents the arithmetic mean of the HU values ​​of all voxels in the preoperative baseline enhanced abdominal CT image of organ i; i μ represents the standard deviation of the HU values ​​of all voxels in the preoperative baseline enhanced abdominal CT image of organ i; i -2σ i This falls under the dynamic threshold, defined as the preoperative mean minus two standard deviations. According to statistical principles, approximately 95% of normal voxel HU values ​​fall within [μ...]. i -2σ i μ i +2σ i Within the interval, therefore, if postoperative I(p) < μ i -2σ iThis indicates that the enhancement at this point is significantly lower than the normal range, suggesting reduced perfusion; ∇I(p) represents the gradient vector of the abdominal enhanced CT image I at spatial location p after standardization, representing the rate of change of the HU value in three-dimensional space. In this embodiment, the calculation method selected is usually Sobel or Scharr operator convolution. A large gradient results in a clear boundary, while a small gradient results in a blurred boundary; τ g This represents the gradient magnitude threshold, which is set empirically based on the actual situation, usually set to 5 HU / mm. It should be noted that this pseudo-label generation strategy uses a two-condition AND logic to ensure high specificity, that is, it captures under-enhancement and boundary blurring respectively by using the two corresponding density conditions and texture conditions before and after AND. Overall, it suppresses a large number of negative samples by using BinaryFocalLoss, that is, normal tissue dominates the training.

[0059] (3) Regarding the consistency constraint loss L_consist: L_consist=‖y^perf-M(y^seg)⊙G(I)‖2; y^perf represents the prediction mask for low perfusion areas, output by Sigmoid; M(y^seg) represents binarizing and cropping the organ probability map to ensure that y^perf is only effective within the organ; G(I)=‖∇I‖ represents the gradient magnitude map of the enhanced abdominal CT image I after standardization, with low-value areas corresponding to blurred boundaries; ⊙ represents element-wise multiplication, forcing y^perf to align with the underlying texture of the image. This loss prevents the model from falsely reporting low perfusion outside the organ or at clear boundaries.

[0060] The process of integrating geometric priors and topological reasoning mechanisms is as follows:

[0061] S2.1 Improved Frangi Filtering: Direction-Adaptive Multi-Scale Enhancement. Hessian matrix analysis is applied to the standardized abdominal enhanced CT image, but the optimal scale σ is dynamically selected. ∗ (x), based on:

[0062] ;

[0063] In the formula, λ2(x, σ) represents the second largest eigenvalue of the Hessian matrix at scale σ, with a larger absolute value indicating a stronger tubular structure; x represents another spatial location; argmax represents the parameter that maximizes the expression; σ specifically represents the Gaussian scale parameter, controlling the standard deviation of the Gaussian function; [σ_min, σ_min] [_max] represents the candidate scale search interval, a preset closed interval, and the specific endpoints can be set according to actual needs; θ(x) represents the estimated local vessel orientation angle at position x, determined by the eigenvector of the Hessian matrix or smoothed through time series information after preoperative CT registration, simply put, it is the direction the currently observed vessel is heading; θ0 represents the prior expected vessel direction, derived from the standard aortic anatomy template; for example, the superior mesenteric artery, after originating from the celiac trunk, usually runs downwards at approximately 30°; the right renal artery runs horizontally to the right; this template is based on the statistical average of a large amount of normal human CTA data, encoding typical vessel orientation knowledge to guide scale selection and avoid misjudging tortuosity artifacts as vessels; κ represents the direction tolerance half-width parameter, set to κ=15°, controlling the sensitivity of the exponential term to direction deviation; then, for exp(-((θ(x)-θ0)) 2 ) / 2κ 2 Overall, it represents the directional consistency weight factor, which is a Gaussian penalty function. When θ(x) = θ0, the weight is 1 and there is no penalty. As the directional deviation increases, the weight decreases exponentially. Its function is to suppress those structures that have strong tubular responses but unreasonable orientations, such as the strip-shaped pseudo-blood vessels caused by metal artifacts that may exist.

[0064] It should be noted that traditional Frangi filtering uses a fixed set of σ for each location and takes the maximum response, without considering the anatomical rationality of the vessel's orientation. This can easily lead to false positives or weakened responses in areas with bends, bifurcation, or artifact interference. This improved method introduces a priori directional constraints, ensuring that even if |λ2| is slightly low at locations conforming to the anatomical orientation, the vessel may still be selected. Conversely, at locations with abnormal orientation, even if |λ2| is high, such as stent artifacts, the exponential term will suppress the selection. The final selected σ... ∗ (x) Simultaneously satisfying the two conditions of structural resemblance to blood vessels and reasonable orientation, so as to achieve a balance between the two, thereby significantly improving the specificity and continuity of vascular enhancement. This method prioritizes the selection of scales that both conform to tubular intensity and match anatomical orientation, which also improves the signal-to-noise ratio to a certain extent.

[0065] S2.2, GNN-driven vascular topology correction:

[0066] S2.2.1 Input Data: The input is the initial blood vessel centerline point cloud output by the lightweight 3D ResUNet, represented as a set of three-dimensional spatial coordinate points: {p1, p2, ..., p...}N}, where p i1 The physical coordinates of the i1th centerline point are in mm; this point cloud may contain gaps, i.e. missing segments on the actual continuous vascular path; N is the total number of centerline points, which is also the maximum value of i1.

[0067] S2.2.2 Graph Construction: The point cloud is transformed into a graph structure G=(V, E) for GNN inference. If d(v i1 v j If the value is less than the set value, then add edge e. i1j At this point, with node v i1 Let e ​​be the coordinates of the centerline. i1j To connect neighboring points; the above setting is typically set to 5mm in this embodiment; Note: G represents the constructed undirected graph used to represent the topology of the blood vessel centerline, V represents the set of nodes, each node v i1 ∈V corresponds to a centerline point p i1 E represents the set of edges, where each edge e i1j ∈E represents node v i1 With v j There is a potential connection between them; e i1j Indicates the connection node v i1 With v j An edge is established only if the two edges satisfy the proximity condition;

[0068] S2.2.3, GNN Inference: A message-passing graph neural network is used for each edge e. i1j Perform binary classification to determine whether it represents a real blood vessel connection, and predict the edge connectivity confidence level c. i1j ∈[0,1]; where the GNN architecture uses a two-layer GAT, and each node v i1 The initial features are their corresponding attributes, and the node representation is updated through neighbor aggregation. The edge features are represented by the concatenation of the two endpoints and output c via MLP. i1j ;

[0069] S2.2.4, Fracture Repair Strategy: Based on the edge connectivity confidence level c output by GNN i1j For automatic repair to be performed in accordance with geometric constraints, two repair conditions must be met simultaneously; where condition A is c. i1j A value <0.4 indicates low confidence, meaning the GNN considers there to be no direct connection, i.e., a break exists; condition B is d(v i1 v j ) < 1.6 × set value, representing a short-distance gap, used to ensure that only small areas of loss are repaired, avoiding incorrect connections across blood vessels; where, combined with the previously stated set value of 5mm, then 1.6 × 5 = 8mm, which is the maximum permissible repair gap length clinically; d(vi1 v j ) then represents node v i1 With v j The Euclidean distance between them.

[0070] Based on the above explanation, the root cause of the problem it addresses is that the aortic endovascular stent graft contains metallic markers, which produce radiographic artifacts on CT scans, masking the enhancement signals of the distal mesenteric or renal arteries. Traditional skeletonization algorithms fail at this point. Therefore, using a Generative Neural Network (GNN) ensures consistent local orientation and, through tangent direction constraints, ensures the interpolation path conforms to the natural curvature of the vessel. Furthermore, the GNN learns the fractal patterns of thousands of normal blood vessels during training, such as branch angles and curvature distribution, allowing for the inference of reasonable missing segments. Moreover, it avoids over-repair through c i1j The dual thresholds of distance prevent incorrect connection of different blood vessels. In summary, this scheme effectively overcomes the limitations of postoperative CT image quality by modeling the vascular centerline as a graph structure and introducing GNN for topology-aware intelligent repair. The repair strategy takes into account both anatomical rationality and algorithm robustness, significantly improves the integrity of the vascular tree, and provides a reliable structural basis for subsequent hemodynamic analysis and risk scoring.

[0071] On the one hand, multi-task segmentation allows RPPR calculation to focus on the real ischemic area, avoiding signal dilution across all organs; topology correction ensures that vascular features reflect the real anatomy, resulting in a VIRScore AUC of over 0.9 and a sensitivity of around 90%. On the other hand, through the joint analysis of the hypoperfusion region prediction mask y^perf and vascular VTIF, it reveals two ischemic subtypes: structural occlusion and functional hypoperfusion, promoting the individualization of clinical intervention strategies and facilitating a clear distinction between the former requiring vascular intervention and the latter being an option for conservative treatment. Overall, the technology achieves both innovation and clinical applicability.

[0072] The corresponding branch continuity index Cr is obtained. i The basis is: Cr i =L_de(i) / L_ex(i); where L_de(i) represents the effective length of the artery actually detected in the repaired vessel centerline, that is, starting from the aortic opening and tracing along the centerline to the end of the most distal visible branch; isolated short segments are excluded to prevent noise interference; the standard for a short segment is <3mm; L_ex(i) represents the expected total length of the artery based on a standard anatomical template, such as the Michels classification or population average CTA data; this artery is also the corresponding artery of organ i.

[0073] S3. Complete the dynamic perfusion feature extraction process and comprehensively acquire the input features. Use an integrated XGBoost classifier to construct an ischemia risk scoring model and output a preliminary abdominal organ ischemia risk score VIRScore_0. The input features must include at least the results of the dynamic perfusion feature extraction process and the branch continuity index Cr from S2. i ;

[0074] The process of completing the dynamic perfusion feature extraction is based on the following:

[0075] S3.1 Constructing functional sub-region masks within organs: Perform mask constraint actions for each organ i, i.e., extract the predicted whole organ mask y^seg from S2, obtaining y^seg(i)∈{0,1}, and simultaneously extract the low-perfusion region prediction mask, i.e., the binary functional abnormality mask y^perf, obtaining y^perf∈{0,1}, generating the organ-constrained low-perfusion binary mask y^seg(i):

[0076] y^seg(i)=Ⅱ(y^perf>Qt)⊙y^seg(i);

[0077] In the formula, Ⅱ() is an exponential function, Qt represents an empirical threshold, which is obtained by ROC curve optimization and has a value range of [0, 1]. In this embodiment, the value is selected as 0.55. ⊙ represents element-wise multiplication to ensure that only suspicious areas inside organs are retained; to avoid misjudging adjacent organs or background noise as low perfusion areas, and to ensure spatial semantic consistency.

[0078] S3.2 Calculation of two-level relative perfusion retention rate: To take into account both the overall perfusion trend and the local ischemic focus, two types of RPPR are defined and screened to obtain the required RPPR_local. (i) (t); where the two defined RPPRs are the global RPPR used for clinical overview, namely RPPR_global. (i) (t), based on:

[0079] ;

[0080] In the formula, N_seg(i) represents the total number of elements in organ i, that is, the cardinality of the set Ω_seg(i): N_seg(i) = |Ω_seg(i)|, and the set Ω_seg(i) represents the set of all elements in organ i; t (p) represents the HU value at spatial location or voxel p in the CT image at postoperative time point t, reflecting the tissue density and contrast enhancement at that point; I pre(p) represents the HU value at voxel p at the same anatomical location in the preoperative baseline CT image, which serves as an individualized perfusion reference. The aforementioned CT images are all standardized enhanced abdominal CT images.

[0081] Local RPPR used for risk modeling, i.e., RPPR_local (i) (t), based on:

[0082] ;

[0083] In the formula, i specifically refers to the target organ index. In this embodiment, the value can be selected from five categories: {liver, spleen, left kidney, right kidney, root of small intestine mesentery}. t represents the postoperative time point, such as t=6 h, 24 h, 48 h, etc., indicating the postoperative time of the CT scan. The set Ω_perf(i) is the set of voxels in the low-perfusion region within organ i. N_perf(i) represents the number of voxels in the low-perfusion region within organ i. N_min represents the minimum effective voxel count threshold. In this embodiment, the value is usually 50. Regions with values ​​less than this may be caused by noise or false positives in the model and are insufficient to represent true pathological changes.

[0084] It should be noted that, while ensuring robustness, highly specific local information should be prioritized for risk modeling to select the required RPPR_local. (i) When N_perf(i) ≥ N_min, if the number of voxels in the detected low-perfusion area is ≥ 50, the area is considered anatomically significant, and its RPPR value is reliable. In this case, local RPPR is used because it can better reflect the true ischemic focus and avoids dilution by normal tissue. Conversely, if the number of voxels is ≥ 50, it indicates that the low-perfusion area is too small, which may be due to isolated low-HU voxels caused by image noise, the model's oversensitivity to blurred boundaries, or non-pathological uneven enhancement. If local RPPR is forcibly used in this case, the statistical instability will be caused by the small sample size. Therefore, it is safe to revert to global RPPR to ensure feature robustness. The above scheme demonstrates clinical value. When dealing with high-risk patients, local RPPR is significantly lower than global RPPR, triggering an early warning. When dealing with low-risk patients, there is no significant low-perfusion area or the area is too small. Global RPPR is used to avoid false alarms, thereby achieving a balance between sensitivity and specificity, which is superior to single-index schemes.

[0085] The process of running the ischemia risk scoring model is as follows:

[0086] For each target patient, the input feature is a multi-dimensional feature vector, including the completed result of the dynamic perfusion feature extraction action, that is, the required RPPR_local for each organ i. (i) (t) and the continuity index Cr of each branch iThe input features are fed into a pre-trained ischemia risk scoring model to generate an ischemia risk probability Pr, which is then linearly mapped to a clinically friendly integer score, i.e., the preliminary abdominal organ ischemia risk score VIRScore_0. The ischemia risk scoring model uses XGBoost as the base learner and employs an early stopping strategy with 5-fold cross-validation to prevent overfitting. The target variable y is a binary label; where y=1 indicates that the target patient has clinically confirmed abdominal organ ischemia within 72 hours post-surgery, including conditions such as intestinal necrosis, acute renal infarction, and lactate >4 mmol / L confirmed by imaging; y=0 indicates no ischemic event. XGBoost automatically calculates the information gain of each input feature to explain model decisions and address the weights w in the subsequently generated VTIF. i The ischemic risk probability Pr is mapped to the preliminary abdominal organ ischemic risk score VIRScore_0 based on the following formula: VIRScore_0 = round(100 × Pr).

[0087] Specifically, the ischemia risk scoring model proposed in this scheme is not a simple weighted average, but rather uses machine learning to automatically learn the nonlinear contribution of the perfusion-structure combination of each organ to the ischemic outcome. Its core is: using RPPR to reflect functional status, using the branch continuity index to reflect the integrity of anatomical pathways, capturing complex interactions through XGBoost, and finally outputting a preliminary actionable risk score, supporting subsequent dynamic correction. This design realizes a complete process from imaging representation to physiological inference to clinical decision-making, serving as a key link in achieving early and accurate early warning in this scheme.

[0088] By adopting the above technical solutions, the limitations of single perfusion assessment can be overcome, and organ function and vascular anatomy information can be integrated simultaneously. On the one hand, low-perfusion areas can be accurately located through multi-task segmentation, so that RPPR reflects the true ischemic focus. On the other hand, the integrity of the vascular tree can be ensured through GNN topology correction, and the branch continuity index can accurately quantify the status of blood supply pathways. The combination of the two can effectively distinguish between functional hypoperfusion and structural occlusion, providing a basis for subsequent development of differentiated clinical strategies, avoiding over-operation or delayed treatment, and improving the precision of intervention and the efficiency of resource utilization.

[0089] S4, call the branch continuity index Cr output by S2. i As a foundation, a vascular flow correction factor is introduced to perform correction processing on the preliminary abdominal organ ischemia risk score VIRScore_0 to generate the final abdominal organ ischemia risk score VIRScore, and simultaneously trigger a visualization risk warning strategy to assess the risk status of the target patient.

[0090] Among them, the vascular flow domain correction factor VTIF is introduced and defined as:

[0091] The branch continuity index Cr corresponding to different organs i i Its corresponding weight w i The products are multiplied and summed to obtain the cumulative value for each target organ. Finally, the cumulative value is multiplied by the reciprocal of the total number of target organs to generate the required vascular flow correction factor VTIF; where the weight w i The importance of the corresponding RPPR feature is normalized in XGBoost to ensure that high-contribution organs have a greater impact on VTIF. This will not be elaborated on further here.

[0092] The preliminary abdominal organ ischemia risk score VIRScore_0 was corrected to obtain the final abdominal organ ischemia risk score VIRScore based on the following formula: VIRScore = VIRScore_0 + Δ × (1 - VTIF); where Δ represents the maximum correction range, which can be determined by cross-validation. In this embodiment, Δ = 15; (1 - VTIF) reflects the degree of vascular tree rupture, and the larger the value, the more incomplete the blood supply pathway. It should be noted that even if the organ CT value decreases, i.e., low RPPR, if its supply artery structure is intact, the corresponding VTIF ≈ 1, then the ischemia may be transient hypoperfusion, and the risk is low. Conversely, if the RPPR decreases slightly but the supply artery is severely interrupted, the corresponding VTIF ≈ 0.5 indicates structural blood flow obstruction, the true risk is underestimated, and the score needs to be adjusted upward.

[0093] Specifically, this scheme achieves secondary utilization of the generation results through S2, using the branch continuity index Cr. i Dynamic correction of the preliminary abdominal organ ischemia risk score VIRScore_0 achieves dual-use from a single source. It can be used for its own original feature input and can also extract vascular flow domain correction factors. This not only allows for in-depth mining of related data belonging to the same data stream without additional acquisition or computational overhead, but also improves the accuracy of risk stratification. More unexpectedly, it reveals the value of structural integrity as an independent prognostic factor, promoting clinical shift from perfusion-only assessment to a two-dimensional assessment of perfusion and pathway, achieving a dual breakthrough in technical effectiveness and clinical understanding. For the original ischemia risk scoring model, although XGBoost can learn the association between RPPR and clinical outcome, it is difficult to explicitly model the high-dimensional topological constraint of vascular structural integrity. VTIF provides interpretable physical priors, achieving limitation compensation.

[0094] The synchronously triggered visual risk warning strategy is explained as follows:

[0095] The final abdominal organ ischemia risk score (VIRScore) is compared with a preset risk threshold range. If the final VIRScore is greater than the upper limit of the risk threshold range, it is considered high risk; if the final VIRScore is less than the lower limit of the risk threshold range, it is considered low risk. In the remaining cases, where the final VIRScore is within the risk threshold range, a secondary assessment mechanism is initiated. The vascular flow correction factor (VTIF) is extracted and compared with a preset retrospective indicator. If the VTIF does not exceed the threshold value... If the retrospective indicators are not met, it is determined to be a hidden high risk; otherwise, it is determined to be a medium risk. A color-coded ischemic risk heat map is generated according to different risk levels, defined as: high risk is red, hidden high risk is light red, medium risk is yellow, and low risk is green. High-risk and hidden high-risk vascular segments are automatically extracted and labeled, and a structured report conforming to the HL7FHIR standard can be selectively output and pushed to the existing electronic medical record system to trigger a clinical alarm. In this embodiment, the risk threshold interval is set to [50, 65], so the upper value of the risk threshold interval is 65 and the lower value of the risk threshold interval is 50.

[0096] It should be noted that retrospective analysis in actual clinical practice revealed that even with a VIRS score of <65, patients with VTIF <0.7 still had a 40% risk of elevated lactate levels within 72 hours; the 0.7 mentioned here is the retrospective indicator set in this embodiment. This leads to an early warning strategy targeting hidden high risks, which can, to some extent, advance the early intervention window and reduce the rate of unplanned exploratory laparotomy by approximately 20%. The following specific example illustrates the technical solution implemented in S4 mentioned above:

[0097] Case: A 72-year-old female, 24 hours post-type A renal dissection, with no abdominal pain; right renal local RPPR=0.78, not reaching the typical ischemic threshold; the origin of the right renal artery was covered by a stent, and the corresponding branch continuity index Cr... i =0.42; XGBoost output: VIRScore_0=61, calculated VTIF=0.63, therefore VIRScore=61+15×(1−0.63)≈66.6 (retain 1 decimal place), judged as high risk, emergency CTA confirmed right renal artery occlusion, and renal function stabilized after balloon dilation; without VTIF correction, this patient would have been missed, and acute kidney injury may have occurred after 48 hours.

[0098] Compared to traditional methods that rely on the overall enhancement of organs to determine ischemia, which can easily lead to missed cases where perfusion is adequate but the blood supply pathway has been interrupted, this approach introduces a branch continuity index to construct a vascular flow correction factor. This dynamically corrects the initial abdominal organ ischemia risk score, and in some extreme cases, it can effectively identify structural ischemia caused by stent coverage. Even if the CT findings are atypical, early intervention can be provided to avoid serious consequences such as acute kidney injury, significantly improving the sensitivity of early diagnosis and avoiding the missed diagnosis of high-risk patients.

[0099] Example 2:

[0100] Based on Example 1, this example also provides a postoperative abdominal organ ischemia risk assessment system based on image analysis. This system includes: an image preprocessing module: acquiring enhanced abdominal CT images of the target patient and outputting them after standardization processing; a multi-task segmentation module: using an improved 3D U-Net++ architecture with multi-task collaborative learning to jointly segment the target patient's organs and target regions, integrating geometric priors and topological inference mechanisms to achieve topological correction, and acquiring the blood supply arteries to each organ to derive the corresponding branch continuity index; a risk modeling module: completing dynamic perfusion feature extraction, comprehensively acquiring input features, constructing an ischemia risk scoring model using an integrated XGBoost classifier, and outputting a preliminary abdominal organ ischemia risk score; wherein, the input features include at least: the completion result of the dynamic perfusion feature extraction and the branch continuity index; and a dynamic correction module: using the branch continuity index as a basis, introducing a vascular flow domain correction factor, performing correction processing on the preliminary abdominal organ ischemia risk score to generate the final abdominal organ ischemia risk score, and simultaneously triggering a visual risk warning strategy to assess the risk status of the target patient.

[0101] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented in software, the above embodiments can be implemented, in whole or in part, as a computer program product. Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution.

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

[0103] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.

Claims

1. A method for assessing the risk of postoperative abdominal organ ischemia based on image analysis, characterized in that, The method includes: Acquire enhanced abdominal CT images of the target patient, and output them after standardization processing; An improved 3D U-Net++ architecture employing multi-task collaborative learning is used to jointly segment organs and target regions of a target patient. It integrates geometric priors and topological inference mechanisms to achieve topological correction and obtains the blood supply artery to each organ, deriving the corresponding branch continuity index. The improved 3D U-Net++ architecture configuration includes a backbone encoder and a dual-task decoder head, with a total loss function introduced simultaneously. The backbone encoder is built based on 3D ResNet-34, progressively downsampling to extract multi-scale spatial-semantic features. The dual-task decoder head includes branches A and B. Branch A is the organ segmentation head, reconstructing the U-Net++ path through nested dense skip connections, outputting a predicted whole organ mask y^seg containing several organ masks. Branch B is the target region detection head, targeting a low-perfusion region, using parallel lightweight convolutional layers to predict a binary functional abnormality mask y^perf. The total loss function is a weighted sum of three terms: organ segmentation loss L_seg, low-perfusion detection loss L_ The weighted sum of the three terms, perf, consistency constraint loss L_consist, and the corresponding weights all have values ​​in the range of [0, 1]. Among them, organ segmentation loss L_seg is used as the main loss to ensure the accuracy of anatomical structure; low perfusion detection loss L_perf is used as an auxiliary loss to guide the network to focus on pathological areas; and consistency constraint loss is used as a regularization term to introduce physical priors. The process of integrating geometric priors and topological reasoning mechanisms is as follows: Improved Frangi filter: Direction-adaptive multi-scale enhancement, applying Hessian matrix analysis to the standardized abdominal enhanced CT image to dynamically select the optimal scale σ. ∗ (x); GNN-driven vascular topology correction: The input is the initial vascular centerline point cloud output by a lightweight 3D ResUNet, represented as a set of three-dimensional spatial coordinate point clouds: {p1, p2, ..., p...} N }, where p i1 Let be the physical coordinates of the i1th centerline point; transform the point cloud into a graph structure G=(V, E) for GNN inference, if d(v i1 v j If the value is less than the set value, then add edge e. i1j Where V represents the set of nodes, and each node v i1 ∈V corresponds to a centerline point p i1 E represents the set of edges, where each edge e i1j ∈E represents node v i1 With v j There is a potential connection between them; e i1j Indicates the connection node v i1 With v j Edges are established only if the two edges satisfy the proximity condition; a message-passing graph neural network is used to process each edge e. i1j Perform binary classification to determine whether it represents a real blood vessel connection, and predict the edge connectivity confidence level c. i1j ∈[0,1]; Edge connectivity confidence c based on GNN output i1j For automatic repair to be performed in conjunction with geometric constraints, two repair conditions must be met simultaneously: Condition A is c i1j <0.4 indicates low confidence level; Condition B is d(v) i1 v j ) < 1.6 × set value, indicating a short-distance clearance; d(v i1 v j ) represents node v i1 With v j The Euclidean distance between them; The dynamic perfusion feature extraction process is completed, and the input features are comprehensively acquired. An ischemia risk scoring model is constructed using an integrated XGBoost classifier, and a preliminary abdominal organ ischemia risk score is output. The input features include at least the completion results of the dynamic perfusion feature extraction process and the branch continuity index. Using the branch continuity index as a basis, a vascular flow correction factor is introduced to correct the preliminary abdominal organ ischemia risk score, generating the final abdominal organ ischemia risk score. Simultaneously, a visual risk warning strategy is triggered to assess the risk status of target patients; the corresponding branch continuity index Cr is then derived. i The basis is: Cr i =L_de(i) / L_ex(i); where L_de(i) represents the effective length of the corresponding artery actually detected in the repaired vascular centerline, and L_ex(i) represents the expected total length of the corresponding artery based on the standard anatomical template; the vascular flow domain correction factor VTIF is introduced, defined as: the branch continuity index Cr corresponding to different organs i i Its corresponding weight w i The products are multiplied and summed to obtain the cumulative value for each target organ. Finally, the cumulative value is multiplied by the reciprocal of the total number of target organs to generate the required vascular flow correction factor (VTIF).

2. The method for assessing postoperative abdominal organ ischemia risk based on image analysis according to claim 1, characterized in that, The process of standardizing abdominal enhanced CT images is as follows: Image registration: A non-rigid registration algorithm based on mutual information was used to unify the abdominal enhanced CT images at each time point to the preoperative baseline space. Each time point included preoperative and postoperative time points. Window width and window level normalization: A liver window was set to ensure consistent grayscale between different scans. Noise suppression and edge enhancement: An improved nonlocal means denoising combined with the Laplacian sharpening operator was applied. In the denoising stage, an improved nonlocal means algorithm was used, and a local variance adaptive factor was introduced into the similarity weight. In the enhancement stage, the Laplacian sharpening operator was superimposed on the denoised image, and the standardized image was output.

3. The method for assessing postoperative abdominal organ ischemia risk based on image analysis according to claim 1, characterized in that, The process of completing the dynamic perfusion feature extraction is as follows: Constructing organ-specific functional sub-region masks: Perform mask-limiting actions on each organ i, extract the predicted whole organ mask y^seg, and obtain y^seg(i)∈{0,1}. Simultaneously extract the predicted binary functional abnormality mask y^perf, and obtain y^perf∈{0,1}. Generate organ-limited low-perfusion binary mask y^seg(i); Calculate the two-level relative perfusion retention rate: Define two types of RPPR and obtain the required RPPR_local after filtering. (i) (t); where the two defined RPPRs are the global RPPR used for clinical overview, namely RPPR_global. (i) (t) and the local RPPR used for risk modeling, namely RPPR_local (i) (t); where the selection criteria are: the required RPPR_local under the condition that the number of voxels N_perf(i) in the low-perfusion region of organ i is ≥ the minimum effective voxel number threshold N_min. (i) (t) takes the value of local RPPR; otherwise, the required RPPR_local (i) (t) takes the value of global RPPR.

4. The method for assessing postoperative abdominal organ ischemia risk based on image analysis according to claim 3, characterized in that, The process of running the ischemia risk scoring model is as follows: For each target patient, the input features are a multi-dimensional feature vector, including the required RPPR_local for each organ i. (i) (t) and the continuity index Cr of each branch i The input features are fed into a pre-trained ischemia risk scoring model to generate an ischemia risk probability Pr, which is then linearly mapped to a preliminary abdominal organ ischemia risk score VIRScore_0. The ischemia risk scoring model uses XGBoost as the base learner.

5. The method for assessing postoperative abdominal organ ischemia risk based on image analysis according to claim 1, characterized in that, The preliminary abdominal organ ischemia risk score VIRScore_0 is corrected to obtain the final abdominal organ ischemia risk score VIRScore based on the formula: VIRScore = VIRScore_0 + Δ × (1 - VTIF); where Δ represents the maximum correction range.

6. The method for assessing postoperative abdominal organ ischemia risk based on image analysis according to claim 5, characterized in that, The process of synchronously triggering a visual risk warning strategy is as follows: The final abdominal organ ischemia risk score (VIRScore) is compared with a preset risk threshold range. If the final VIRScore is greater than the upper limit of the risk threshold range, it is considered high risk; if the final VIRScore is less than the lower limit of the risk threshold range, it is considered low risk. If the final VIRScore is within the risk threshold range, a secondary judgment mechanism is initiated. The vascular flow correction factor (VTIF) is extracted and compared with a preset retrospective indicator. If the VTIF does not exceed the retrospective indicator, it is considered occult high risk; otherwise, it is considered moderate risk. A color-coded ischemia risk heatmap is generated according to different risk levels, defined as: high risk in red, occult high risk in light red, moderate risk in yellow, and low risk in green. High-risk and occult high-risk vascular segments are automatically extracted and labeled.

7. A postoperative abdominal organ ischemia risk assessment system based on image analysis, characterized in that, The system includes: Image preprocessing module: Acquires enhanced abdominal CT images of the target patient, performs standardization processing, and outputs the images. Multi-task segmentation module: An improved 3D U-Net++ architecture using multi-task collaborative learning is employed to jointly segment organs and target regions of the target patient. It integrates geometric priors and topological inference mechanisms to achieve topological correction and obtains the blood supply arteries to each organ, deriving the corresponding branch continuity index. The improved 3D U-Net++ architecture configuration includes a backbone encoder and a dual-task decoder head, with a total loss function introduced simultaneously. The backbone encoder is built based on 3D ResNet-34, progressively downsampling to extract multi-scale spatial-semantic features. The dual-task decoder head includes branches A and B. Branch A is the organ segmentation head, reconstructing the U-Net++ path through nested dense skip connections, outputting a predicted whole organ mask y^seg containing several organ masks. Branch B is the target region detection head, targeting a low-perfusion region, using parallel lightweight convolutional layers to predict a binary functional abnormality mask y^perf. The total loss function is a weighted sum of three terms: organ segmentation loss L_seg, low-perfusion detection loss L_ The weighted sum of the three terms, perf, consistency constraint loss L_consist, and the corresponding weights all have values ​​in the range of [0, 1]. Among them, organ segmentation loss L_seg is used as the main loss to ensure the accuracy of anatomical structure; low perfusion detection loss L_perf is used as an auxiliary loss to guide the network to focus on pathological areas; and consistency constraint loss is used as a regularization term to introduce physical priors. The process of integrating geometric priors and topological reasoning mechanisms is as follows: Improved Frangi filter: Direction-adaptive multi-scale enhancement, applying Hessian matrix analysis to the standardized abdominal enhanced CT image to dynamically select the optimal scale σ. ∗ (x); GNN-driven vascular topology correction: The input is the initial vascular centerline point cloud output by a lightweight 3D ResUNet, represented as a set of three-dimensional spatial coordinate point clouds: {p1, p2, ..., p...} N }, where p i1 Let be the physical coordinates of the i1th centerline point; transform the point cloud into a graph structure G=(V, E) for GNN inference, if d(v i1 v j If the value is less than the set value, then add edge e. i1j Where V represents the set of nodes, and each node v i1 ∈V corresponds to a centerline point p i1 E represents the set of edges, where each edge e i1j ∈E represents node v i1 With v j There is a potential connection between them; e i1j Indicates the connection node v i1 With v j Edges are established only if the two edges satisfy the proximity condition; a message-passing graph neural network is used to process each edge e. i1j Perform binary classification to determine whether it represents a real blood vessel connection, and predict the edge connectivity confidence level c. i1j ∈[0,1]; Edge connectivity confidence c based on GNN output i1j For automatic repair to be performed in conjunction with geometric constraints, two repair conditions must be met simultaneously: Condition A is c i1j <0.4 indicates low confidence level; Condition B is d(v) i1 v j ) < 1.6 × set value, indicating a short-distance clearance; d(v i1 v j ) represents node v i1 With v j The Euclidean distance between them; Risk modeling module: Completes dynamic perfusion feature extraction, comprehensively acquires input features, constructs an ischemia risk scoring model using an integrated XGBoost classifier, and outputs a preliminary abdominal organ ischemia risk score; among which, the input features include at least: the completion result of the dynamic perfusion feature extraction and the branch continuity index; Dynamic correction module: Using the branch continuity index as a base, it introduces a vascular flow correction factor to correct the initial abdominal organ ischemia risk score, generating the final abdominal organ ischemia risk score. Simultaneously, it triggers a visual risk warning strategy to assess the risk status of the target patient; and derives the corresponding branch continuity index Cr. i The basis is: Cr i =L_de(i) / L_ex(i); where L_de(i) represents the effective length of the corresponding artery actually detected in the repaired vascular centerline, and L_ex(i) represents the expected total length of the corresponding artery based on the standard anatomical template; the vascular flow domain correction factor VTIF is introduced, defined as: the branch continuity index Cr corresponding to different organs i i Its corresponding weight w i The products are multiplied and summed to obtain the cumulative value for each target organ. Finally, the cumulative value is multiplied by the reciprocal of the total number of target organs to generate the required vascular flow correction factor (VTIF).

Images