A post-plastic surgery infection early warning method and system based on multi-modal data fusion

Abstract

The method of the embodiment of the application proposes a multi-modal data fusion postoperative infection early warning method and system. First, the postoperative wound image, clinical text, wearable vital signs and baseline features are quality evaluated and repaired. The repaired image is double-flow encoded and the coarse mask is obtained by prior segmentation. The superpixel graph is constructed and the topologically consistent fine mask is obtained by the graph neural network to quantify the erythema / effusion area. The text and vital signs are encoded and fused with the image features and the mask through cross-attention and self-attention, and the fused hidden vector and uncertainty are obtained through the quality weight gate. The individualized dynamic threshold is constructed by combining the baseline risk, the logit calibration and the sliding window trend, the shape-preserving prediction output confidence band is used to realize the early and reliable infection early warning.

Description

Technical Field

[0001] This invention belongs to the field of medical model processing technology, and in particular relates to a method and system for early warning of postoperative infection in plastic surgery based on multimodal data fusion. Background Technology

[0002] Postoperative infection after plastic surgery is a serious complication affecting surgical outcomes and patient safety. Timely detection and treatment of infection are crucial for reducing risk. However, traditional postoperative monitoring relies primarily on single-modality information, such as body temperature, blood counts, or visual observation of the wound, making it difficult to capture early signs of infection in a timely and comprehensive manner. With the development of digital healthcare, postoperative patients can generate multimodal data: such as photographs and medical images of the wound site, continuous vital sign monitoring data, electronic medical records, and activity and sleep data recorded by wearable devices. How to integrate and utilize these heterogeneous data to improve the accuracy and timeliness of infection prediction has become an urgent problem to be solved in postoperative care after plastic surgery. Summary of the Invention

[0003] In view of this, embodiments of the present invention provide a method and system for early warning of postoperative infection in plastic surgery based on multimodal data fusion.

[0004] The multimodal data fusion-based postoperative infection early warning method for plastic surgery provided in this embodiment of the invention includes:

[0005] S1. Perform quality assessment and repair on the postoperative data after pretreatment. The postoperative data includes postoperative wound images, clinical text, wearable vital signs, and patient baseline characteristics.

[0006] S2. Perform dual-stream visual coding on the repaired wound image to obtain local and global features. After gating fusion, generate a coarse mask of the wound through prior segmentation.

[0007] S3. The fused features and coarse mask are used to form a superpixel map through mapping. A topologically consistent fine mask is obtained based on graph neural network message passing, and the erythema area and exudation area of ​​the wound are calculated.

[0008] S4. After text encoding and temporal encoding of clinical text and wearable vital signs, cross-attention writing and self-attention fusion are performed with fusion features and coarse mask as tokens, and the modality contribution is gated with quality weights to output fusion latent vector and probability-evidence uncertainty.

[0009] S5. After estimating individual baseline risk based on patient baseline characteristics, log-odds calibration is performed on fusion risk, and an individualized dynamic threshold is constructed based on sliding window statistics and trend terms. The confidence band is then output by combining slope and conformal prediction.

[0010] In addition to the aspects and any possible implementations described above, another implementation is provided.

[0011] S1 specifically includes:

[0012] S11. Obtain the same patient at any given time. Postoperative wound images Clinical texts Wearable vital signs sequence and patient baseline characteristics And complete the timestamp alignment and identifier association;

[0013] S12. Image quality scores are obtained after image quality assessment, text quality assessment, and temporal quality assessment. Text quality score and timing quality score ;

[0014] S13, according to , and Generate mass vector And generate a set of defect labels based on the sub-indicators. ;

[0015] S14. Based on the repair decision, select one of the following methods for repair: conditional diffusion repair, conditional normalization flow repair, or attention-guided infilling. After repair ;

[0016] in, , , , and The weights corresponding to each indicator, For the Sigmoid function; The full-image variance of the sharpness score, i.e., the Laplace response, is expressed as... , For sample variance, Laplace response; Occlusion rate , The visible area of ​​the wound is obtained from the segmentation network. For the image domain; exposure sub-rating ,in, The percentage of overexposed pixels is expressed by the formula: , This represents the intensity value at pixel position x. For indicator functions; The percentage of underexposed pixels is expressed by the formula: , and All are brightness thresholds; color deviation sub-scores , Color difference for white balance reference;

[0017] , , , and These are, respectively, completeness, structural correctness, semantic validity, and spelling correctness. , , and These are their respective weighting factors;

[0018] , The missing rate, For sensor drift, Noise level, , and These are their respective weighting factors.

[0019] In addition to the aspects and any possible implementations described above, a further implementation is provided in which the repair decision is:

[0020] a. Calculate the required metrics for the determination:

[0021] Clarity and blur ;

[0022] Coverage and wound overlap rate ;

[0023] Exposure anomaly ;

[0024] ROI saturation ratio ;

[0025] Structural discernibility missing ;

[0026] Nonlocal similarity ;

[0027] in, For the normalization operator, To cover the mask, Let x be the gradient vector at pixel position x. For normalized cross-correlation, For the target small block adjacent to the occluded area, Candidate small patches that are not occluded in the entire image;

[0028] b. When any of the following conditions are met, a resampling command will be output without entering the repair process:

[0029] b1、 ;

[0030] b2, ;

[0031] b3、 ;

[0032] c. If none of the conditions in b are met, proceed to the repair selection:

[0033] c1. If the set is hit Then select conditional normalization flow repair, set Defined as:

[0034] ;

[0035] c2, if the set is hit Then choose attention-guided filling, collection Defined as:

[0036] ;

[0037] c3. If there is no set hit It also failed to hit the target. Then choose conditional diffusion repair;

[0038] in, , , , , , , , , , and All are calibration thresholds.

[0039] As described above, and in accordance with any possible implementation, a further implementation is provided, wherein the conditional diffusion repair specifically includes:

[0040] Based on mass vector and defect tag set Generate condition vector ;

[0041] Set the diffusion step number T and the dynamic noise schedule, and use dynamic noise scheduling during the forward diffusion process. and calculate and ;

[0042] In the inverse denoising process, a hybrid noise prediction network combining U!-!Net and Visual Transformer is used. Calculate the denoised mean ,right After the area was reconstructed, ;

[0043] The conditional normalization flow repair specifically includes:

[0044] Generate condition vector Then, define conditionally invertible mappings. ;

[0045] base distribution As a priori, and by maximizing the conditional log-likelihood, structure fidelity is preserved, i.e., maximizing... ;

[0046] exist Internal flow through local geometry Maintain the continuity of the creation edge, and through global illumination flow. Unified correction of brightness, contrast and color temperature, and then reconstruction. ;

[0047] The attention-guided filling specifically includes:

[0048] According to the search database Sure Area Mask And retrieve candidate fragments ;

[0049] Calculate candidate weighted correlation ,based on right After weighted fusion and detail enhancement of the regions, the result is .

[0050] In addition to the aspects and any possible implementations described above, a further implementation is provided, wherein S2 specifically includes:

[0051] S21, will Input Local Encoder With global encoder Local features were then obtained through calculation. and global features ;

[0052] S22. Constructing channel gating vectors based on quality vectors and defect label sets. ;

[0053] S23, will and The spatial dimensions are stitched together, and then convolution is used to generate pixel-by-pixel spatial gating:

[0054] ;

[0055] S24. Based on the joint control of channel gating and spatial gating to determine the proportion of local / global features, the fused features are obtained. ;

[0056] S25, Based on fusion features Build a pixel-level logit with prior validation data and obtain the wound probability:

[0057] ;

[0058] S26. Using a threshold Binarization yields a coarse mask. ;

[0059] in, and To generate parameters for channel gating and spatial gating. This is the channel gating vector; , and In pixels The fusion of feature vectors, color priors, and edge priors at the location. for The corresponding channel weight vector, and They are respectively and The corresponding scalar weights, For bias terms, This is the probability threshold.

[0060] In addition to the aspects and any possible implementations described above, a further implementation is provided, wherein S3 specifically includes:

[0061] S31, to Superpixel segmentation is performed to obtain a superpixel set. Then, fused features are extracted from each superpixel. Color Priority Edge Prior and the mean probability of coarse mask and form node features. ;

[0062] S32, Construction Diagram The edge weights of a graph are defined as follows:

[0063] ;

[0064] S33. The node update representation based on graph neural network message passing is as follows:

[0065] ;

[0066] Calculate the superpixel wound probability And projected onto pixels to obtain a probability map. ;

[0067] S34. A topologically consistent fine mask is obtained after topologically consistent energy minimization based on the energy function. :

[0068] ;

[0069] S35, in fine mask Internally, the probability of erythema is determined based on color and texture features. And calculate the area of ​​erythema. ;

[0070] S36. Calculate the bleed probability based on brightness, saturation, specular highlights, and texture flatness. And calculate the seepage area. ;

[0071] in, This is the weight matrix. , and For their respective weighting coefficients, The average gradient intensity on the common edge of two adjacent superpixels. and The weight vector and bias of the superpixel classification head. Let p be the embedding representation of node p at layer l. Let p be the set of adjacent nodes. Let be the attention weights from neighbor q to p. For residual gate coefficients, Let be the energy function. Let p be the binary label of the superpixel. For the adjacency weight in the regularization term, and To adjust the weights for smoothing regularization and topological regularization, Let be an Euler characteristic. For the topological reference from the coarse mask, Logit for erythema discrimination, The threshold for the probability of erythema. Logit for leakage detection This is the threshold for the probability of leakage.

[0072] In addition to the aspects and any possible implementations described above, an implementation is further provided, wherein S4 specifically includes:

[0073] S41. Clinical text and wearable vital sign sequences They are coded as follows and The missing rate is calculated by the missing test estimator. ;

[0074] S42, Fusion Features Pooling into image tokens And generate mask tokens from coarse masks. With attention bias matrix ;

[0075] S43. Use the image token as the context token for the query pair. Perform a biased cross-attention write:

[0076] ;

[0077] S44. Calculate modal gating based on the quality vector and the missing rate:

[0078] ;

[0079] After weighting, we get ;

[0080] S45, Fusion through Self-Attention And pooling output fused latent vectors ;

[0081] S46. Output non-negative evidence from the evidence classification head. and And calculate:

[0082] ;

[0083] Where p is the probability of infection and U is the probability-evidence uncertainty.

[0084] In addition to the aspects and any possible implementations described above, an implementation is further provided, wherein S5 specifically includes:

[0085] S51, Based on patient baseline characteristics Estimate individual baseline risk and logarithmic odds ;

[0086] S52. Convert the infection probability p to the fusion logarithmic odds. ;

[0087] S53, Incorporating Logarithmic Probabilities Anchored to After temperature scaling and bias correction, the calibrated probability is obtained. This can be expressed as a formula: ;

[0088] S54, Based on post-calibration probability Constructing time series And in the sliding window Internal calculation of EWMA mean With variance The trend slope is estimated by robust regression. ;

[0089] S55. Constructing individualized dynamic thresholds:

[0090] ;

[0091] S56. By using inductive conformal prediction to link the confidence band with the slope formed by the residuals and quantiles, an alarm will be triggered under the following conditions: ;

[0092] in, and These are all baseline model parameters. For baseline feature mapping, For quantile levels, , and All are log-odds calibration parameters. , and All are weighting coefficients. For the look-ahead lower bound probability, For the purpose of The prediction threshold.

[0093] In addition to the aspects described above and any possible implementations, a further implementation is provided in which the residuals and quantiles are expressed by the formula:

[0094] ;

[0095] The slope-linked confidence band can be expressed by the formula:

[0096] ;

[0097] in, For non-consistent residuals, For the observation results of the i-th sample, Let be the predicted probability of the i-th sample. Therefore The residual quantiles at the confidence level, For the purpose of The upper and lower bounds of the probability confidence band. Let be the smoothed probability at time t. The slope of the trend within the sliding window. To predict the time interval, To effectively divide the bandwidth, The uncertainty inflation coefficient is... This is probabilistic-evidence uncertainty.

[0098] On the other hand, embodiments of the present invention provide a multimodal data fusion-based postoperative infection early warning system for plastic surgery, the system comprising:

[0099] The repair module is used to perform quality assessment and repair on the pre-processed postoperative data, which includes postoperative wound images, clinical text, wearable vital signs, and patient baseline characteristics.

[0100] The segmentation module is used to perform dual-stream visual encoding on the repaired wound image to obtain local and global features. After gating fusion, a coarse mask of the wound is generated through prior segmentation.

[0101] The mapping module is used to form a superpixel map by combining fused features and a coarse mask, obtain a topologically consistent fine mask based on graph neural network message passing, and calculate the erythema area and exudation area of ​​the wound.

[0102] The fusion module is used to perform text encoding and temporal encoding on clinical text and wearable vital signs, and then perform cross-attention writing and self-attention fusion with fusion features and coarse mask as tokens. It also uses quality weights to gating modal contributions and outputs fusion latent vectors and probabilistic-evidence uncertainty.

[0103] The early warning module is used to estimate individual baseline risk based on patient baseline characteristics, perform log-odds calibration on fusion risk, construct individualized dynamic thresholds based on sliding window statistics and trend terms, and output confidence bands by combining slope and conformal prediction.

[0104] One of the above technical solutions has the following beneficial effects:

[0105] (1) The segmentation framework of “mask-aware superpixel-graph neural network-topology consistent regularization” is adopted to suppress cross-boundary leakage and hole artifacts at the structural level, output a stable and coherent fine wound mask, and can simultaneously provide quantifiable indicators such as erythema area and exudation area; it is still robust under low resolution and uneven lighting.

[0106] (2) Using image terms as queries and combining masking and temporal bias cross-attention mechanism, key information from clinical text and wearable vital signs is directed into the visual representation of the wound neighborhood, significantly reducing cross-modal semantic mismatch; the accompanying attention heatmap and regional evidence visualization improve the interpretability and auditability of the model.

[0107] (3) Introduce a gating strategy for data quality and missing data, and provide both probability and uncertainty through evidence-based output; when a certain modality is of poor quality or missing, the system automatically reduces the weight and increases the uncertainty, thereby reducing the risk of false alarms and false negatives and enhancing the generalization ability in real complex scenarios.

[0108] (4) Based on the patient’s baseline characteristics, the individual prior risk is estimated first, and then the multimodal fusion probability is log-odds calibrated. The individual dynamic threshold is constructed by combining the upper quantile of the sliding window, the trend slope, recent fluctuations and relative baseline deviation, thereby alleviating the threshold mismatch caused by population differences and time drift, and ensuring the consistency of decision-making for different patients and different stages.

[0109] (5) Use shape-preserving prediction to generate a forward confidence band that is linked to the trend, and use the dual criteria of "currently crossing the line" or "predicting the lower bound to cross the line" to trigger early warning. This provides both theoretical coverage and advance warning, avoids false alarms caused by occasional spikes, facilitates connection with the clinical workflow closed loop, and supports deployment on in-hospital servers, edge devices, and wearable terminals. Attached Figure Description

[0110] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the embodiments will be briefly introduced below. Obviously, the 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.

[0111] Figure 1 This is a flowchart illustrating a multimodal data fusion-based method for early warning of postoperative infection in plastic surgery, provided in an embodiment of the present invention.

[0112] Figure 2 This is a schematic block diagram of a multimodal data fusion postoperative infection early warning system for plastic surgery provided in an embodiment of the present invention;

[0113] Figure 3 This is a schematic diagram of the hardware structure of a multimodal data fusion postoperative infection early warning system for plastic surgery provided in an embodiment of the present invention. Detailed Implementation

[0114] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below in conjunction with specific embodiments and corresponding drawings. Obviously, the described embodiments are only a part of the embodiments of this invention, and not all of them. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention.

[0115] Please refer to Figure 1 Its essence Figure 1 This is a flowchart illustrating a multimodal data fusion-based postoperative infection early warning method for plastic surgery provided in an embodiment of the present invention. As shown in the figure, the method includes the following steps:

[0116] S1. Perform quality assessment and repair on the postoperative data after pretreatment. The postoperative data includes postoperative wound images, clinical text, wearable vital signs, and patient baseline characteristics.

[0117] S2. Perform dual-stream visual coding on the repaired wound image to obtain local and global features. After gating fusion, generate a coarse mask of the wound through prior segmentation.

[0118] S3. The fused features and coarse mask are used to form a superpixel map through mapping. A topologically consistent fine mask is obtained based on graph neural network message passing, and the erythema area and exudation area of ​​the wound are calculated.

[0119] S4. After text encoding and temporal encoding of clinical text and wearable vital signs, cross-attention writing and self-attention fusion are performed with fusion features and coarse mask as tokens, and the modality contribution is gated with quality weights to output fusion latent vector and probability-evidence uncertainty.

[0120] S5. After estimating individual baseline risk based on patient baseline characteristics, log-odds calibration is performed on fusion risk, and an individualized dynamic threshold is constructed based on sliding window statistics and trend terms. The confidence band is then output by combining slope and conformal prediction.

[0121] S1 is a preliminary step for establishing intelligent repair. It first performs temporal pairing and labeling association between postoperative wound photos, clinical text, wearable vital signs, and patient baseline characteristics. Then, it provides quantifiable quality scores for each modality and identifies defect types (such as blur, overexposure / underexposure, color cast, occlusion, missing / structural errors in text, missing / drifting / noise in vital signs). Based on the defect type, it adaptively selects and performs repairs: conditional diffusion repair for noise and blur, conditional normalization flow repair for brightness / contrast / color temperature unification, and attention-guided infilling for small missing and occluded areas. It outputs the repaired wound image and quality weight vector, and generates a set of defect labels as priors for subsequent modules. On the image side, it focuses on maintaining the continuity of incision and tissue boundaries, unifying illumination and color, and reducing overexposure / underexposure and color cast; on the text and temporal side, it performs integrity and validity checks, missing segment imputation, and outlier suppression.

[0122] S1 specifically includes:

[0123] S11. Obtain the same patient at any given time. Postoperative wound images Clinical texts Wearable vital signs sequence and patient baseline characteristics And complete the timestamp alignment and identifier association;

[0124] S12. Image quality scores are obtained after image quality assessment, text quality assessment, and temporal quality assessment. Text quality score and timing quality score ;

[0125] S13, according to , and Generate mass vector And generate a set of defect labels based on the sub-indicators. ;

[0126] S14. Based on the repair decision, select one of the following methods for repair: conditional diffusion repair, conditional normalization flow repair, or attention-guided infilling. After repair ;

[0127] in, , , , and The weights corresponding to each indicator, For the Sigmoid function; The full-image variance of the sharpness score, i.e., the Laplace response, is expressed as... , For sample variance, Laplace response; Occlusion rate , The visible area of ​​the wound is obtained from the segmentation network. For the image domain; exposure sub-rating ,in, The percentage of overexposed pixels is expressed by the formula: , Let x be the intensity value at pixel position x. For indicator functions; The percentage of underexposed pixels is expressed by the formula: , and All are brightness thresholds; color deviation sub-scores , Color difference for white balance reference;

[0128] , , , and These are, respectively, completeness, structural correctness, semantic validity, and spelling correctness. , , and These are their respective weighting factors;

[0129] , The missing rate, For sensor drift, Noise level, , and These are their respective weighting factors.

[0130] Specifically, the repair decision is as follows:

[0131] a. Calculate the required metrics for the determination:

[0132] Clarity and blur ;

[0133] Coverage and wound overlap rate ;

[0134] Exposure anomaly ;

[0135] ROI saturation ratio ;

[0136] Structural discernibility missing ;

[0137] Nonlocal similarity ;

[0138] in, For the normalization operator, To cover the mask, Let x be the gradient vector at pixel position x. For normalized cross-correlation, For the target small block adjacent to the occluded area, Candidate small patches that are not occluded in the entire image;

[0139] b. When any of the following conditions are met, a resampling command will be output without entering the repair process:

[0140] b1、 ;

[0141] b2, ;

[0142] b3、 ;

[0143] c. If none of the conditions in b are met, proceed to the repair selection:

[0144] c1. If the set is hit Then select conditional normalization flow repair, set Defined as:

[0145] ;

[0146] c2, if the set is hit Then choose attention-guided filling, collection Defined as:

[0147] ;

[0148] c3. If there is no set hit It also failed to hit the target. Then choose conditional diffusion repair;

[0149] in, , , , , , , , , , and All are calibration thresholds.

[0150] Conditional diffusion repair specifically includes:

[0151] Based on mass vector and defect tag set Generate condition vector ;

[0152] Set the diffusion step number T and the dynamic noise schedule, and use dynamic noise scheduling during the forward diffusion process. and calculate and ;

[0153] In the inverse denoising process, a hybrid noise prediction network combining U!-!Net and Visual Transformer is used. Calculate the denoised mean ,right After the area was reconstructed, ;

[0154] The conditional normalization flow repair specifically includes:

[0155] Generate condition vector Then, define conditionally invertible mappings. ;

[0156] base distribution As a priori, and by maximizing the conditional log-likelihood, structure fidelity is preserved, i.e., maximizing... ;

[0157] exist Internal flow through local geometry Maintain the continuity of the creation edge, and through global illumination flow. Unified correction of brightness, contrast and color temperature, and then reconstruction. ;

[0158] The attention-guided filling specifically includes:

[0159] According to the search database Sure Area Mask And retrieve candidate fragments ;

[0160] Calculate candidate weighted correlation ,based on right After weighted fusion and detail enhancement of the regions, the result is .

[0161] After S1 repair, the wound image has clearer and more coherent boundaries, more balanced exposure, and more consistent colors; the text is more complete and the structure is more standardized; the rate of missing vital signs decreases, drift and noise are suppressed, reducing the interference of "garbage input" on subsequent judgments and making subsequent fine masking more stable. Defect labels are used as prior input across the attention module, making it easier to write key text points and abnormal vital signs into the wound neighborhood. When the quality of the image, text, or vital signs is poor, their contribution to the final risk automatically decreases.

[0162] S2 specifically includes:

[0163] S21, will Input Local Encoder With global encoder Local features were then obtained through calculation. and global features ;

[0164] S22. Constructing channel gating vectors based on quality vectors and defect label sets. ;

[0165] S23, will and The spatial dimensions are stitched together, and then convolution is used to generate pixel-by-pixel spatial gating:

[0166] ;

[0167] S24. Based on the joint control of channel gating and spatial gating to determine the proportion of local / global features, the fused features are obtained. ;

[0168] S25, Based on fusion features Build a pixel-level logit with prior validation data and obtain the wound probability:

[0169] ;

[0170] S26. Using a threshold Binarization yields a coarse mask. ;

[0171] in, and To generate parameters for channel gating and spatial gating. This is the channel gating vector; , and In pixels The fusion of feature vectors, color priors, and edge priors at the location. for The corresponding channel weight vector, and They are respectively and The corresponding scalar weights, For bias terms, This is the probability threshold.

[0172] S2 constructs a visual encoding and prior segmentation module, enabling the fusion process to adaptively perceive quality and defects (such as blurring, overexposure, and occlusion). It dynamically adjusts the ratio of "local flow and global flow" in the channel dimension and retains only the regional response related to the cut in the spatial dimension, thus suppressing background noise and irrelevant tissues from the source.

[0173] S2 also assigns weights to each semantic channel based on image clarity, exposure, color cast, and defect type. For scenes with poor quality or defects, it automatically increases the global stream ratio to stabilize semantics, while amplifying local details when the quality is good. A lightweight pixel-level discriminator is built on the fused features to output a wound probability map; a coarse mask is then obtained based on a probability threshold, which can be adaptively adjusted according to image quality to avoid false positives or false negatives.

[0174] Spatial gating focuses attention near the incision, significantly reducing the impact of background, instrument highlights, and dressing edges on the segmentation starting point. Channel gating incorporates quality and defect information into feature fusion; it provides reliable priors even in blurred, overexposed / underexposed, color-shifted, or partially occluded scenes, minimizing crashes. Coarse mask boundaries are closer to the true boundaries, superpixel generation is more edge-aligned, and subsequent message passing and topology-consistent optimization in the graph neural network converge more easily. The prior segmentation head is lightweight and has low latency, adaptable to in-hospital servers and edge terminals; thresholds and gating strategies can be quickly calibrated by department or equipment, facilitating large-scale deployment.

[0175] Specifically, S3 includes:

[0176] S31, to Superpixel segmentation is performed to obtain a superpixel set. Then, fused features are extracted from each superpixel. Color Priority Edge Prior and the mean probability of coarse mask and form node features. ;

[0177] S32, Construction Diagram The edge weights of a graph are defined as follows:

[0178] ;

[0179] S33. The node update representation based on graph neural network message passing is as follows:

[0180] ;

[0181] Calculate the superpixel wound probability And projected onto pixels to obtain a probability map. ;

[0182] S34. A topologically consistent fine mask is obtained after topologically consistent energy minimization based on the energy function. :

[0183] ;

[0184] S35, in fine mask Internally, the probability of erythema is determined based on color and texture features. And calculate the area of ​​erythema. ;

[0185] S36. Calculate the bleed probability based on brightness, saturation, specular highlights, and texture flatness. And calculate the seepage area. ;

[0186] in, This is the weight matrix. , and For their respective weighting coefficients, The average gradient intensity on the common edge of two adjacent superpixels. and The weight vector and bias of the superpixel classification head. Let p be the embedding representation of node p at layer l. Let p be the set of adjacent nodes. Let be the attention weights from neighbor q to p. For residual gate coefficients, Let be the energy function. Let p be the binary label of the superpixel. For the adjacency weight in the regularization term, and To adjust the weights for smoothing regularization and topological regularization, Let be an Euler characteristic. For the topological reference from the coarse mask, Logit for erythema discrimination, The threshold for the probability of erythema. Logit for leakage detection This is the threshold for the probability of leakage.

[0187] S3 segments the repaired wound image into a group of superpixels based on appearance similarity and builds an adjacency graph accordingly. Using the superpixel graph as a carrier, it refines the coarse mask obtained in the previous stage into a topologically consistent fine mask using message passing and attention in a graph neural network. Within the wound area defined by the fine mask, it distinguishes between erythema and exudation based on color and texture, and outputs quantifiable indicators such as erythema area and exudation area, providing interpretable evidence for subsequent multimodal fusion and early warning decision-making.

[0188] Mask-aware superpixels prioritize aligning block boundaries with the actual incision / tissue boundaries; graph neural networks reduce cross-domain propagation at boundaries, significantly minimizing "leakage" from background or neighboring tissue into the wound. Incorporating topological consistency constraints into energy optimization effectively avoids breaks, voids, and flying spots, resulting in a more refined mask consistent with surgical anatomy. Edge weights and attention weights, constructed by integrating boundary gradients, color, and texture differences, maintain contour stability even in complex scenarios such as strong reflections, dressing edges, and blood clots, without drastic fluctuations due to local noise. Pixels are first clustered into superpixels before graph construction, reducing node size from pixel-level to the order of thousands, making the computational cost of the graph neural network linearly controllable; energy minimization converges more easily with good priors, reducing repeated patching. Erythema and exudate areas are estimated within the effective region defined by the refined mask, and can be converted to physical units using pixel scales, forming traceable and comparable quantitative evidence for doctors. Fine-grained masks and area metrics serve directly as significant priors and sources of evidence for multimodal fusion, improving the localization accuracy of cross-modal attention and reducing the reliance of subsequent risk assessments on low-quality visual information.

[0189] Specifically, S4 includes:

[0190] S41. Clinical text and wearable vital sign sequences They are coded as follows and The missing rate is calculated by the missing test estimator. ;

[0191] S42, Fusion Features Pooling into image tokens And generate mask tokens from coarse masks. With attention bias matrix ;

[0192] S43. Use the image token as the context token for the query pair. Perform a biased cross-attention write:

[0193] ;

[0194] S44. Calculate modal gating based on the quality vector and the missing rate:

[0195] ;

[0196] After weighting, we get ;

[0197] S45, Fusion through Self-Attention And pooling output fused latent vectors ;

[0198] S46. Output non-negative evidence from the evidence classification head. and And calculate:

[0199] ;

[0200] Where p is the probability of infection and U is the probability-evidence uncertainty.

[0201] S4 encodes clinical text and wearable vital signs into sequence representations, which, together with image terms and coarse masks, form a context. Through cross-attention-oriented writing "using image terms as queries," it accurately infuses incision-related information into the visual representation of the wound neighborhood. It also dynamically adjusts the contributions of the three modalities based on quality signals such as image sharpness / exposure, text integrity, missing vital signs, and noise, selectively discarding at the term level when necessary to suppress the negative impact of low-quality modalities. After self-attention fusion, it outputs fused latent vectors, infection risk probability, and evidential uncertainty, providing reliable input and confidence for subsequent baseline anchoring calibration and individualized thresholding.

[0202] Masking and temporal priors serve as attention biases, prioritizing key textual points and abnormal physical signs near the incision site, significantly reducing semantic mismatches with irrelevant areas such as background, instrument reflections, and dressings. Quality gating and missing information estimation de-weight or mask low-quality information before fusion; the overall judgment remains stable even when a modality is blurred, overexposed, missing, or experiences sudden noise. The system can output attention heatmaps, visualizations of gating weights and masking biases, illustrating "which text fragments, which time periods of physical signs, and which regions of the image" drive the decision; evidential uncertainty provides model confidence and suggestions for re-acquisition / review. The linkage between quality and uncertainty enables the system to adapt to different equipment, lighting, shooting habits, and disease types, reducing performance drift in new departments / patient populations. Dividing images into a finite number of terms and performing global self-attention after cross-attention ensures controllable computational scale, facilitating real-time operation on in-hospital servers or edge terminals. The output probability and uncertainty are directly used as inputs to baseline anchoring, dynamic threshold and conformal confidence band, enabling early warnings to be triggered earlier and more stably under the dual criteria of "current overshoot / look-ahead lower bound overshoot".

[0203] Specifically, S5 includes:

[0204] S51, Based on patient baseline characteristics Estimate individual baseline risk and logarithmic odds ;

[0205] S52. Convert the infection probability p to the fusion logarithmic odds. ;

[0206] S53, Incorporating Logarithmic Probabilities Anchored to After temperature scaling and bias correction, the calibrated probability is obtained. This can be expressed as a formula: ;

[0207] S54, Based on post-calibration probability Constructing time series And in the sliding window Internal calculation of EWMA mean With variance The trend slope is estimated by robust regression. ;

[0208] S55. Constructing individualized dynamic thresholds:

[0209] ;

[0210] S56. By using inductive conformal prediction to link the confidence band with the slope formed by the residuals and quantiles, an alarm will be triggered under the following conditions: ;

[0211] in, and These are all baseline model parameters. For baseline feature mapping, For quantile levels, , and All are log-odds calibration parameters. , and All are weighting coefficients. For the look-ahead lower bound probability, For the purpose of The prediction threshold.

[0212] The residuals and quantiles are expressed by the following formulas:

[0213] ;

[0214] The slope-linked confidence band can be expressed by the formula:

[0215] ;

[0216] in, For non-consistent residuals, For the observation results of the i-th sample, Let be the predicted probability of the i-th sample. Therefore The residual quantiles at the confidence level, For the purpose of The upper and lower bounds of the probability confidence band. Let be the smoothed probability at time t. The slope of the trend within the sliding window. To predict the time interval, To effectively divide the bandwidth, The uncertainty inflation coefficient is... This is probabilistic-evidence uncertainty.

[0217] S5 first estimates the individual prior risk based on the patient's baseline characteristics, providing a reference anchor for subsequent judgments. The risk obtained from multimodal fusion is transformed and calibrated to the individual prior scale, eliminating mismatches caused by population differences and equipment / department biases. The calibrated risk is formed into a time series, simultaneously estimating the recent mean, volatility, and trend slope within a sliding window. An individualized dynamic threshold is constructed by combining "upper quantile (reflecting recent high levels) + trend slope + volatility + relative baseline deviation," and the threshold is adaptively tightened or loosened according to uncertainty and data quality. Conformal prediction is used to generate a forward-looking confidence band, and an early warning is triggered by the dual criteria of "current crossing the line or the predicted lower bound about to cross the line."

[0218] This step, through prior anchoring and log-odds calibration, ensures that risk values ​​across different patients, departments, and devices are comparable, reducing cross-domain drift. Trend slope and volatility terms are introduced into the threshold, triggering earlier when risk is continuously rising and automatically increasing safety margins to suppress false spikes when short-term volatility is large. Shape-preserving prediction provides upper and lower confidence bands for future risk, with distribution-independent coverage guarantees; it is linked to dynamic thresholds to achieve a balance between lead time and robustness. When model evidence is insufficient or data quality is poor, the threshold and confidence band are automatically widened, and the system tends to make "cautious judgments," reducing the risk of false positives and false negatives. The threshold consists of four parts: upper quantile, slope, volatility, and baseline deviation, with the source of each item displayed; the confidence band and triggering conditions are clearly defined, facilitating physician understanding and verification. Based on sliding window and exponential weighting, online statistics and simple linear extrapolation have low computational costs and are suitable for the real-time requirements of continuous uploads from in-hospital servers, edge terminals, and wearable devices.

[0219] Through the above steps, the present invention achieves the following technical effects:

[0220] (1) The segmentation framework of “mask-aware superpixel-graph neural network-topology consistent regularization” is adopted to suppress cross-boundary leakage and hole artifacts at the structural level, output a stable and coherent fine wound mask, and can simultaneously provide quantifiable indicators such as erythema area and exudation area; it is still robust under low resolution and uneven lighting.

[0221] (2) Using image terms as queries and combining masking and temporal bias cross-attention mechanism, key information from clinical text and wearable vital signs is directed into the visual representation of the wound neighborhood, significantly reducing cross-modal semantic mismatch; the accompanying attention heatmap and regional evidence visualization improve the interpretability and auditability of the model.

[0222] (3) Introduce a gating strategy for data quality and missing data, and provide both probability and uncertainty through evidence-based output; when a certain modality is of poor quality or missing, the system automatically reduces the weight and increases the uncertainty, thereby reducing the risk of false alarms and false negatives and enhancing the generalization ability in real complex scenarios.

[0223] (4) Based on the patient’s baseline characteristics, the individual prior risk is estimated first, and then the multimodal fusion probability is log-odds calibrated. The individual dynamic threshold is constructed by combining the upper quantile of the sliding window, the trend slope, recent fluctuations and relative baseline deviation, thereby alleviating the threshold mismatch caused by population differences and time drift, and ensuring the consistency of decision-making for different patients and different stages.

[0224] (5) Use shape-preserving prediction to generate a forward confidence band that is linked to the trend, and use the dual criteria of "currently crossing the line" or "predicting the lower bound to cross the line" to trigger early warning. This provides both theoretical coverage and advance warning, avoids false alarms caused by occasional spikes, facilitates connection with the clinical workflow closed loop, and supports deployment on in-hospital servers, edge devices, and wearable terminals.

[0225] The present invention further provides an apparatus embodiment for implementing the steps and methods in the above method embodiments.

[0226] Please refer to Figure 2 It is provided by the embodiments of the present invention. Figure 2 A multimodal data fusion-based postoperative infection early warning system for plastic surgery is provided in this embodiment of the invention. The system includes:

[0227] Repair module 210 is used to perform quality assessment and repair on the pre-processed postoperative data, which includes postoperative wound images, clinical text, wearable vital signs and patient baseline characteristics.

[0228] The segmentation module 220 is used to perform dual-stream visual encoding on the repaired wound image to obtain local and global features. After gating fusion, a coarse mask of the wound is generated through prior segmentation.

[0229] The mapping module 230 is used to form a superpixel map by mapping the fused features and the coarse mask, obtain a topologically consistent fine mask based on graph neural network message passing, and calculate the erythema area and exudation area of ​​the wound.

[0230] The fusion module 240 is used to perform text encoding and temporal encoding on clinical text and wearable vital signs, and then perform cross-attention writing and self-attention fusion with fusion features and coarse mask as tokens. It also uses quality weights to gating modal contributions and outputs fusion latent vectors and probabilistic-evidence uncertainty.

[0231] The early warning module 250 is used to estimate individual baseline risk through patient baseline characteristics, perform log-odds calibration on fusion risk, construct individualized dynamic thresholds based on sliding window statistics and trend terms, and output confidence bands by combining slope and conformal prediction.

[0232] Because each unit module in this embodiment can execute Figure 1 For the methods shown, parts not described in detail in this embodiment, please refer to the [examples / descriptions]. Figure 1 Related explanations.

[0233] Please refer to Figure 3 This is a schematic diagram of the hardware structure of a multimodal data fusion-based postoperative infection early warning system for plastic surgery provided in an embodiment of the present invention. The data prediction device includes at least one processor and a memory. The at least one processor is coupled to the memory and is used to read and execute instructions in the memory to execute the multimodal data fusion-based postoperative infection early warning method for plastic surgery provided in this embodiment of the present invention.

[0234] This invention provides a computer-readable medium. The computer-readable medium stores program code that, when executed on a computer, causes the computer to perform the multimodal data fusion-based postoperative infection early warning method for plastic surgery provided in this invention.

[0235] At the hardware level, the central server may include a processor, and optionally also an internal bus, network interface, and memory. The memory may include main memory, such as high-speed random-access memory (RAM), or non-volatile memory, such as at least one disk drive. Of course, the central server may also include other hardware required for other business operations.

[0236] The processor, network interface, and memory can be interconnected via an internal bus, which can be an ISA (Industry Standard Architecture) bus, a PCI (Peripheral Component Interconnect) bus, or an EISA (Extended Industry Standard Architecture) bus, etc. This bus can be categorized into address bus, data bus, control bus, etc.

[0237] Memory is used to store programs. Specifically, programs may include program code, which includes computer operation instructions. Memory may include main memory and non-volatile memory, and provides instructions and data to the processor.

[0238] The steps of the method disclosed in the embodiments of this invention can be directly manifested as being executed by a hardware decoding processor, or executed by a combination of hardware and software modules in the decoding processor. The software modules can reside in random access memory, flash memory, read-only memory, programmable read-only memory, electrically erasable programmable memory, registers, or other mature storage media in the art. This storage medium is located in memory, and the processor reads information from the memory and, in conjunction with its hardware, completes the steps of the above method.

[0239] The systems, devices, modules, or units described in the above embodiments can be implemented by computer chips or entities, or by products with certain functions. A typical implementation device is a computer. Specifically, a computer can be, for example, a personal computer, laptop computer, cellular phone, camera phone, smartphone, personal digital assistant, media player, navigation device, email device, game console, tablet computer, wearable device, or any combination of these devices.

[0240] For ease of description, the above apparatus is described by dividing it into various units or modules according to their functions. Of course, in implementing this invention, the functions of each unit or module can be implemented in one or more software and / or hardware.

[0241] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0242] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0243] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0244] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0245] In a typical configuration, a computing device includes one or more processors (CPU), input / output interfaces, network interfaces, and memory.

[0246] Memory may include non-persistent storage in computer-readable media, such as random access memory (RAM) and / or non-volatile memory, such as read-only memory (ROM) or flash RAM. Memory is an example of computer-readable media.

[0247] Computer-readable media includes both permanent and non-permanent, removable and non-removable media that can store information using any method or technology. Information can be computer-readable instructions, data structures, modules of programs, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic magnetic disk storage or other magnetic storage devices, or any other non-transferable medium that can be used to store information accessible by a computing device. As defined herein, computer-readable media does not include transient computer-readable media, such as modulated data signals and carrier waves.

[0248] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0249] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0250] This invention can be described in the general context of computer-executable instructions, such as program modules, that are executed by a computer. Generally, program modules include routines, programs, objects, components, data structures, etc., that perform a specific task or implement a specific abstract data type. This invention can also be practiced in distributed computing environments where tasks are performed by remote processing devices connected via a communication network. In distributed computing environments, program modules can reside in local and remote computer storage media, including storage devices.

[0251] The various embodiments in this invention are described in a progressive manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the system embodiments are basically similar to the method embodiments, so the description is relatively simple; relevant parts can be referred to the descriptions in the method embodiments.

[0252] The above description is merely an embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principle of the present invention should be included within the scope of the claims of the present invention.

Claims

1. A method for early warning of postoperative infection in plastic surgery based on multimodal data fusion, characterized in that, The method includes: S1. Perform quality assessment and repair on the postoperative data after pretreatment. The postoperative data includes postoperative wound images, clinical text, wearable vital signs, and patient baseline characteristics. S2. Perform dual-stream visual coding on the repaired wound image to obtain local and global features. After gating fusion, generate a coarse mask of the wound through prior segmentation. S3. The fused features and coarse mask are used to form a superpixel map through mapping. A topologically consistent fine mask is obtained based on graph neural network message passing, and the erythema area and exudation area of ​​the wound are calculated. S4. After text encoding and temporal encoding of clinical text and wearable vital signs, cross-attention writing and self-attention fusion are performed with fusion features and coarse mask as tokens, and the modality contribution is gated with quality weights to output fusion latent vector and probability-evidence uncertainty. S5. After estimating individual baseline risk based on patient baseline characteristics, log-odds calibration is performed on fusion risk, and an individualized dynamic threshold is constructed based on sliding window statistics and trend terms. The confidence band is then output by combining slope and conformal prediction. S4 specifically includes: S41. Clinical text and wearable vital sign sequences They are coded as follows and The missing rate is calculated by the missing test estimator. ; S42, Fusion Features Pooling into image tokens And generate mask tokens from coarse masks. With attention bias matrix ; S43. Use the image token as the context token for the query pair. Perform a biased cross-attention write: ； S44. Calculate modal gating based on the quality vector and the missing rate: ； After weighting, we get ; S45, Fusion through Self-Attention And pooling output fused latent vectors ; S46. Output non-negative evidence from the evidence classification head. and And calculate: ； Where p is the probability of infection and U is the probability-evidence uncertainty; S5 specifically includes: S51, Based on patient baseline characteristics Estimate individual baseline risk and logarithmic odds ; S52. Convert the infection probability p to the fusion logarithmic odds. ; S53, Incorporating Logarithmic Probabilities Anchored to After temperature scaling and bias correction, the calibrated probability is obtained. This can be expressed as a formula: ; S54, Based on post-calibration probability Constructing time series And in the sliding window Internal calculation of EWMA mean With variance The trend slope is estimated by robust regression. ; S55. Constructing individualized dynamic thresholds: ； S56. By using inductive conformal prediction to link the confidence band with the slope formed by the residuals and quantiles, an alarm will be triggered under the following conditions: ; in, and These are all baseline model parameters. For baseline feature mapping, For quantile levels, , and All are log-odds calibration parameters. , and All are weighting coefficients. For the look-ahead lower bound probability, For the purpose of The prediction threshold; The residuals and quantiles are expressed by the following formulas: ； The slope-linked confidence band can be expressed by the formula: ； in, For non-consistent residuals, For the observation results of the i-th sample, Let be the predicted probability of the i-th sample. Therefore The residual quantiles at the confidence level, For the purpose of The upper and lower bounds of the probability confidence band. Let be the smoothed probability at time t. The slope of the trend within the sliding window. To predict the time interval, To effectively divide the bandwidth, The uncertainty inflation coefficient is... This is probabilistic-evidence uncertainty.

2. The method for early warning of postoperative infection in plastic surgery based on multimodal data fusion according to claim 1, characterized in that, S1 specifically includes: S11. Obtain the same patient at any given time. Postoperative wound images Clinical texts Wearable vital signs sequence and patient baseline characteristics And complete the timestamp alignment and identifier association; S12. Image quality scores are obtained after image quality assessment, text quality assessment, and temporal quality assessment. Text quality score and timing quality score ; S13, according to , and Generate mass vector And generate a set of defect labels based on the sub-indicators. ; S14. Based on the repair decision, select one of the following methods for repair: conditional diffusion repair, conditional normalization flow repair, or attention-guided infilling. After repair ; in, , , , and The weights corresponding to each indicator, For the Sigmoid function; The full-image variance of the sharpness score, i.e., the Laplace response, is expressed as... , For sample variance, Laplace response; Occlusion rate , The visible area of ​​the wound is obtained from the segmentation network. For the image domain; exposure sub-rating ,in, The percentage of overexposed pixels is expressed by the formula: , Let x be the intensity value at pixel position x. For indicator functions; The percentage of underexposed pixels is expressed by the formula: , and All are brightness thresholds; color deviation sub-scores , Color difference for white balance reference; , , , and These are, respectively, completeness, structural correctness, semantic validity, and spelling correctness. , , and These are their respective weighting factors; , The missing rate, For sensor drift, Noise level, , and These are their respective weighting factors.

3. The method for early warning of postoperative infection in plastic surgery based on multimodal data fusion according to claim 2, characterized in that, The repair decision is as follows: a. Calculate the required metrics for the determination: Clarity and blur ; Coverage and wound overlap rate ; Exposure anomaly ; ROI saturation ratio ; Structural discernibility missing ; Nonlocal similarity ; in, For the normalization operator, To cover the mask, Let x be the gradient vector at pixel position x. For normalized cross-correlation, For the target small block adjacent to the occluded area, Candidate small patches that are not occluded in the entire image; b. When any of the following conditions are met, a resampling command will be output without entering the repair process: b1、 ； b2、 ； b3、 ； c. If none of the conditions in b are met, proceed to the repair selection: c1. If the set is hit Then select conditional normalization flow repair, set Defined as: ； c2, if the set is hit Then choose attention-guided filling, collection Defined as: ； c3. If there is no set hit It also failed to hit the target. Then choose conditional diffusion repair; in, , , , , , , , , , and All are calibration thresholds.

4. The method for early warning of postoperative infection in plastic surgery based on multimodal data fusion according to claim 3, characterized in that, The conditional diffusion repair specifically includes: Based on mass vector and defect tag set Generate condition vector ; Set the diffusion step number T and the dynamic noise schedule, and use dynamic noise scheduling during the forward diffusion process. and calculate and ; In the inverse denoising process, a hybrid noise prediction network combining U!-!Net and Visual Transformer is used. Calculate the denoised mean ,right After the area was reconstructed, ; The conditional normalization flow repair specifically includes: Generate condition vector Then, define conditionally invertible mappings. ; base distribution As a priori, and by maximizing the conditional log-likelihood, structure fidelity is preserved, i.e., maximizing... ; exist Internal flow through local geometry Maintain the continuity of the creation edge, and through global illumination flow. Unified correction of brightness, contrast and color temperature, and then reconstruction. ; The attention-guided filling specifically includes: According to the search database Sure Area Mask And retrieve candidate fragments ; Calculate candidate weighted correlation ,based on right After weighted fusion and detail enhancement of the regions, the result is .

5. The method for early warning of postoperative infection in plastic surgery based on multimodal data fusion according to claim 4, characterized in that, S2 specifically includes: S21, will Input Local Encoder With global encoder Local features were then obtained through calculation. and global features ; S22. Constructing channel gating vectors based on quality vectors and defect label sets. ; S23, will and The spatial dimensions are stitched together, and then convolution is used to generate pixel-by-pixel spatial gating: ； S24. Based on the joint control of channel gating and spatial gating to determine the proportion of local / global features, the fused features are obtained. ; S25, Based on fusion features Build a pixel-level logit with prior validation data and obtain the wound probability: ； S26. Using a threshold Binarization yields a coarse mask. ; in, and To generate parameters for channel gating and spatial gating. This is the channel gating vector; , and In pixels The fusion of feature vectors, color priors, and edge priors at the location. for The corresponding channel weight vector, and They are respectively and The corresponding scalar weights, For bias terms, This is the probability threshold.

6. The method for early warning of postoperative infection in plastic surgery based on multimodal data fusion according to claim 5, characterized in that, S3 specifically includes: S31, to Superpixel segmentation is performed to obtain a superpixel set. Then, fused features are extracted from each superpixel. Color Priority Edge Prior and the mean probability of coarse mask and form node features. ; S32, Construction Diagram The edge weights of a graph are defined as follows: ； S33. The node update representation based on graph neural network message passing is as follows: ； Calculate the superpixel wound probability And projected onto pixels to obtain a probability map. ; S34. A topologically consistent fine mask is obtained after topologically consistent energy minimization based on the energy function. : ； S35, in fine mask Internally, the probability of erythema is determined based on color and texture features. And calculate the area of ​​erythema. ; S36. Calculate the bleed probability based on brightness, saturation, specular highlights, and texture flatness. And calculate the seepage area. ; in, This is the weight matrix. , and For their respective weighting coefficients, The average gradient intensity on the common edge of two adjacent superpixels. and The weight vector and bias of the superpixel classification head. Let p be the embedding representation of node p at layer l. Let p be the set of adjacent nodes. Let be the attention weights from neighbor q to p. For residual gate coefficients, Let be the energy function. Let p be the binary label of the superpixel. For the adjacency weight in the regularization term, and To adjust the weights for smoothing regularization and topological regularization, Let be an Euler characteristic. For the topological reference from the coarse mask, Logit for erythema discrimination, The threshold for the probability of erythema. Logit for leakage detection This is the threshold for the probability of leakage.

7. A postoperative infection early warning system for plastic surgery using multimodal data fusion according to any one of claims 1-6, characterized in that, The system includes: The repair module is used to perform quality assessment and repair on the pre-processed postoperative data, which includes postoperative wound images, clinical text, wearable vital signs, and patient baseline characteristics. The segmentation module is used to perform dual-stream visual encoding on the repaired wound image to obtain local and global features. After gating fusion, a coarse mask of the wound is generated through prior segmentation. The mapping module is used to form a superpixel map by combining fused features and a coarse mask, obtain a topologically consistent fine mask based on graph neural network message passing, and calculate the erythema area and exudation area of ​​the wound. The fusion module is used to perform text encoding and temporal encoding on clinical text and wearable vital signs, and then perform cross-attention writing and self-attention fusion with fusion features and coarse mask as tokens. It also uses quality weights to gating modal contributions and outputs fusion latent vectors and probabilistic-evidence uncertainty. The early warning module is used to estimate individual baseline risk based on patient baseline characteristics, perform log-odds calibration on fusion risk, construct individualized dynamic thresholds based on sliding window statistics and trend terms, and output confidence bands by combining slope and conformal prediction.

Images