A method, system, device and medium for predicting the trend of reimplantation finger vascular crisis

Abstract

This application relates to a method, system, device, and medium for predicting the trend of vascular crisis in replanted fingers. The method includes: acquiring clinical pathological index data, infrared monitoring image sequences, and visual monitoring image sequences; inputting the visual monitoring image sequences into a neural network model for appearance image feature extraction to extract a feature sequence of the replanted finger's filling state and a feature sequence of the replanted finger's swelling progression; inputting the infrared monitoring image sequences into a neural network model for vascular imaging feature extraction to extract a feature sequence of blood supply patency; concatenating the blood supply patency feature sequence, the replanted finger's filling state feature sequence, and the replanted finger's swelling progression feature sequence to obtain a multi-source image monitoring feature sequence; and inputting the multi-source image monitoring feature sequence and clinical pathological index data into a vascular crisis trend prediction model to generate a vascular crisis level prediction sequence. This method can identify the filling state and swelling progression of the replanted finger, enabling dynamic trend prediction of the risk of crisis.

Description

Technical Field

[0001] This application belongs to the field of vascular crisis trend prediction, and in particular relates to a method, system, device and medium for predicting vascular crisis trends in replanted fingers. Background Technology

[0002] Finger replantation is a landmark microsurgical technique for the clinical treatment of finger amputations. While surgical anastomosis techniques have become increasingly sophisticated, the prevention and control of postoperative vascular crisis remains a core bottleneck restricting the survival rate of replanted fingers. Traditional postoperative vascular crisis monitoring relies on regular visual observation, manual palpation, and single-point blood oxygenation of the fingertip by medical staff. It is primarily based on the clinical experience of medical staff and has inherent drawbacks such as strong subjectivity, long monitoring intervals, inability to continuously track dynamic events, easy omission of early hidden pathological signals, and significant delay in early warning.

[0003] Although auxiliary equipment such as laser Doppler flowmeters and single-point infrared thermometers have emerged since then, they can only achieve intermittent detection of a single indicator and cannot accurately capture the hidden dynamic changes in the early stage of vascular crisis. Manual assessment is highly subjective and the monitoring interval has a significant lag.

[0004] Most existing technologies can only achieve immediate diagnosis of vascular crises, and are unable to complete the dynamic trend prediction of the risk of crisis. Their early warning capabilities are extremely weak, and they cannot adapt to the progressive and dynamic pathological development of vascular crises. They also cannot meet the need for accurate early warning during the high-risk window period after replantation of fingers, thus restricting the improvement of the survival rate of replanted fingers. Summary of the Invention

[0005] Therefore, it is necessary to provide a method, system, device, and medium for predicting the vascular crisis trend of replanted fingers, which can identify the filling status and swelling progression of replanted fingers and realize the dynamic trend extrapolation of the risk of crisis.

[0006] In a first aspect, this application provides a method for predicting the trend of vascular crisis in replanted fingers, including:

[0007] Acquire clinical pathological index data, infrared monitoring image sequences of blood vessels in the replanted finger, and visual monitoring image sequences of the replanted finger.

[0008] The visual monitoring image sequence is input into the appearance image feature extraction neural network model to extract the feature sequence of the replanted finger body filling state and the feature sequence of the replanted finger body swelling progression;

[0009] The infrared monitoring image sequence is input into the vascular imaging feature extraction neural network model to extract the blood supply patency feature sequence. The blood supply patency feature sequence, the replanted finger filling status feature sequence, and the replanted finger swelling progression feature sequence are spliced ​​according to the time sequence to obtain the multi-source image monitoring feature sequence.

[0010] Multi-source image monitoring feature sequences and clinicopathological index data are input into the vascular crisis trend prediction model to generate a vascular crisis level prediction sequence.

[0011] In one embodiment, a sequence of visual monitoring images is input into a neural network model for extracting appearance image features to extract a feature sequence of the replanted finger's fullness state and a feature sequence of the replanted finger's swelling progression, including:

[0012] Each visual monitoring image in the visual monitoring image sequence is input into the multi-scale convolutional feature extraction module in the appearance image feature extraction neural network model to extract the multi-scale image feature set of the visual monitoring image;

[0013] The multi-scale image feature set of the visual monitoring image is input into the multi-scale feature fusion module of the fullness state extraction branch in the appearance image feature extraction neural network model. The single-scale image features of the multi-scale image feature set are fused to obtain the fullness state fused image features. The fullness state fused image features are then input into the fullness state feature decoding module of the fullness state extraction branch to generate the fullness state features of the replanted finger in each visual monitoring image.

[0014] The multi-scale image feature set of the visual monitoring image is input into the swelling progression multi-scale feature fusion module in the swelling progression extraction branch of the appearance image feature extraction neural network model. The single-scale image features in the multi-scale image feature set are fused to obtain the swelling progression fused image features. The swelling progression fused image features are then input into the swelling progression feature decoding module in the swelling progression extraction branch to generate the replanted finger swelling progression features of each visual monitoring image.

[0015] Based on the temporal splicing of the visual monitoring images in the visual monitoring image sequence, the filling state characteristics of the replanted finger body are obtained, and based on the temporal splicing of the replanted finger body swelling progression characteristics of the visual monitoring images, the swelling progression characteristics of the replanted finger body are obtained.

[0016] In one embodiment, the multi-scale image feature set includes small-scale image features and medium-scale image features. The multi-scale image feature set of the visual surveillance image is input into the filling state multi-scale feature fusion module in the filling state extraction branch of the appearance image feature extraction neural network model. The single-scale image features in the multi-scale image feature set are fused to obtain the filling state fused image features, including:

[0017] The small-scale image features in the multi-scale image feature set are input into the small-scale convolutional feature enhancement module in the full-state multi-scale feature fusion module to generate enhanced small-scale image features;

[0018] The mesoscale image features in the multi-scale image feature set are input into the full-state mesoscale convolutional feature enhancement module in the full-state multi-scale feature fusion module to generate full-state enhanced mesoscale image features. The full-state enhanced mesoscale image features are then upsampled to obtain upsampled full-state enhanced image features.

[0019] The enhanced small-scale image features and the upsampled full-state enhanced image features are fused to obtain the preliminary fused image features of the full-state. The preliminary fused image features of the full-state are then input into the full-state upsampled multi-layer convolution module in the full-state multi-scale feature fusion module to generate the full-state fused image features.

[0020] In one embodiment, the multi-scale image feature set further includes large-scale image features. The multi-scale image feature set of the visual monitoring image is input into the swelling progression multi-scale feature fusion module in the swelling progression extraction branch of the appearance image feature extraction neural network model. The single-scale image features in the multi-scale image feature set are fused to obtain the swelling progression fused image features, including:

[0021] The mesoscale image features from the multi-scale image feature set are input into the mesoscale convolutional feature enhancement module for swelling progression in the multi-scale feature fusion module for swelling progression, to generate enhanced mesoscale image features for swelling progression.

[0022] The large-scale image features in the multi-scale image feature set are input into the large-scale convolutional feature enhancement module in the swelling progression multi-scale feature fusion module to generate enhanced large-scale image features. The enhanced large-scale image features are then upsampled to obtain upsampled enhanced large-scale image features.

[0023] The swelling progression enhancement mid-scale image features and upsampling enhancement large-scale image features are fused to obtain preliminary fused image features of swelling progression. These preliminary fused image features are then input into the swelling progression downsampling multi-layer convolution module in the swelling progression multi-scale feature fusion module to generate fused image features of swelling progression.

[0024] In one embodiment, the replanted finger body filling status characteristics include normal filling level characteristics, mild abnormal filling level characteristics, moderate abnormal filling level characteristics, severe abnormal filling level characteristics, and perfusion disappearance level characteristics, and the replanted finger body swelling progression characteristics include normal swelling level, mild progressive swelling level characteristics, moderate progressive swelling level characteristics, and severe progressive swelling level characteristics.

[0025] The loss functions of the neural network model for extracting features from appearance images include the fullness level loss function and the swelling level loss function. The expressions for the fullness level loss function and the swelling level loss function are as follows:

[0026]

[0027]

[0028] In the formula, and These are the fullness level loss function and the swelling level loss function, respectively. and These are the total number of grades for the filling status characteristics of the replanted finger and the total number of grades for the swelling progression characteristics of the replanted finger, respectively. For the first The graded replantation refers to the true filling state of the replanted tissue, and its graded unique heat encoding. and The first one predicted by the full state feature decoding module is the first one. The probability of graded replantation with full body filling and the first The probability of graded replantation with the tissue fully engorged. and These are the error spacing weighting coefficients for the fullness status level and the swelling progression level, respectively. For the first Graded replantation refers to the true swelling progression grade coded by a unique thermal code. and The first one predicted by the swelling progression feature decoding module is the first one. The probability of swelling progression in replanted finger tissue and the first The probability of graded replanted finger swelling progression.

[0029] In one embodiment, the blood supply patency feature sequence includes a vascular continuity feature sequence and a systolic / diastolic rhythm feature sequence. The infrared monitoring image sequence is input into a vascular imaging feature extraction neural network model to extract the blood supply patency feature sequence, which includes:

[0030] Each infrared monitoring image in the infrared monitoring image sequence is input into the vascular spatial feature extraction module in the vascular imaging feature extraction neural network model to extract the vascular spatial features of each infrared monitoring image.

[0031] The spatial features of blood vessels are input into the vasomotor value quantification module in the vasomotor imaging feature extraction neural network model to extract the initial vasomotor values ​​of each infrared monitoring image. The spatial features of blood vessels are also input into the vasculature continuity quantification module in the vasculature imaging feature extraction neural network model to extract the initial vasculature continuity values ​​of each infrared monitoring image.

[0032] Using the sampling time of each infrared monitoring image as the center time of the vasomotor sliding window, the vasomotor time window analysis sequence corresponding to each infrared monitoring image is constructed based on the initial vasomotor values ​​within the vasomotor sliding window. The vasomotor time window analysis sequence is then input into the vasomotor rhythm feature time sequence analysis module in the vasomotor imaging feature extraction neural network model to extract the vasomotor rhythm features of each infrared monitoring image.

[0033] Using the sampling time of each infrared monitoring image as the center time of the blood vessel continuous sliding window, the blood vessel continuity time window analysis sequence corresponding to each infrared monitoring image is constructed based on the initial blood vessel continuity value within the blood vessel continuous sliding window. The blood vessel continuity time window analysis sequence is then input into the blood vessel continuity feature time sequence analysis module in the blood vessel imaging feature extraction neural network model to extract the blood vessel continuity features of each infrared monitoring image.

[0034] Based on the temporal splicing of the infrared monitoring images of each infrared monitoring image, a vasomotor rhythm feature sequence is obtained, and based on the temporal splicing of the vascular continuity features of each infrared monitoring image, a vascular continuity feature sequence is obtained.

[0035] In one embodiment, multi-source image monitoring feature sequences and clinicopathological index data are input into a vascular crisis trend prediction model to generate a vascular crisis level prediction sequence, including:

[0036] The feature sequences from multi-source image monitoring are input into the long short-term memory network encoder in the vascular crisis trend prediction model to generate encoder hidden state sequences.

[0037] Clinical pathological index data are input into the fully connected coding network in the vascular crisis trend prediction model to generate global pathological condition features.

[0038] The encoder hidden state sequence and global pathological condition features are input into the global pathological condition gating fusion module in the vascular crisis trend prediction model to fuse the hidden state sequence and global pathological condition features to generate an enhanced encoder hidden state sequence.

[0039] The enhanced encoder hidden state sequence is input into the long short-term memory network decoder in the vascular crisis trend prediction model to generate the decoder hidden state sequence. The decoder hidden state sequence is then input into the vascular crisis level prediction module in the vascular crisis trend prediction model to generate the vascular crisis level prediction sequence. The vascular crisis level prediction module includes a fully connected layer and an activation function layer.

[0040] Secondly, this application also provides a replanted finger vascular crisis trend prediction system, including:

[0041] The pathological monitoring data acquisition module is used to acquire clinical pathological index data, infrared monitoring image sequences of blood vessels in the replanted finger, and visual monitoring image sequences of the replanted finger.

[0042] The appearance image feature extraction module is used to input the visual monitoring image sequence into the appearance image feature extraction neural network model to extract the feature sequence of the replanted finger body filling state and the feature sequence of the replanted finger body swelling progression.

[0043] The vascular imaging feature extraction module is used to input the infrared monitoring image sequence into the vascular imaging feature extraction neural network model to extract the blood supply patency feature sequence, and splice the blood supply patency feature sequence, the replanted finger filling status feature sequence, and the replanted finger swelling progression feature sequence according to the time sequence to obtain the multi-source image monitoring feature sequence.

[0044] The vascular crisis level prediction module is used to input multi-source image monitoring feature sequences and clinicopathological index data into the vascular crisis trend prediction model to generate a vascular crisis level prediction sequence.

[0045] Thirdly, this application also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of the method as described in any of the first aspects of this application.

[0046] Fourthly, this application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the method described in any of the first aspects of this application.

[0047] The aforementioned method, system, equipment, and medium for predicting vascular crisis trends in replanted fingers integrate clinical pathological indicators, infrared monitoring images, and visual monitoring images. Based on a neural network model, they extract key features such as filling status, swelling progression, and blood supply patency, overcoming the limitations of traditional single-modal monitoring and improving the generalization ability and robustness of the prediction model. By employing a dual neural network model to process external images and infrared images separately and extracting quantitative features, they can eliminate reliance on physician subjective judgment and reveal the contribution of different factors to crisis risk. Through global pathological condition-guided temporal prediction of vascular crisis trends, they can achieve deep encoding of long-term temporal dependencies in multi-source image monitoring feature sequences, accurately predict the dynamic development trend of vascular crises, and capture latent pathological warning signals before the occurrence of crises, effectively improving the foresight and accuracy of vascular crisis prediction. Attached Figure Description

[0048] To more clearly illustrate the technical solutions in the embodiments or related technologies of this application, the accompanying drawings used in the description of the embodiments or related technologies will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0049] Figure 1 A flowchart illustrating a method for predicting the trend of vascular crisis in a replanted finger, provided as an embodiment of this application. Figure 1 ;

[0050] Figure 2 A flowchart illustrating a method for predicting the trend of vascular crisis in a replanted finger, provided as an embodiment of this application. Figure 2 ;

[0051] Figure 3 This is a schematic diagram of the structure of a replanted finger vascular crisis trend prediction system provided in one embodiment of this application. Detailed Implementation

[0052] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0053] In one exemplary embodiment of this application, such as Figure 1As shown, a method for predicting vascular crisis trends in replanted fingers is provided. This embodiment illustrates the application of this method to a vascular crisis prediction terminal. It is understood that this method can also be applied to a vascular crisis prediction server, and further to a vascular crisis prediction system including both a vascular crisis prediction terminal and a vascular crisis prediction server, and is implemented through the interaction between the two. In this embodiment, the method includes the following steps:

[0054] Step S101: Obtain clinical pathological index data, infrared monitoring image sequence of blood vessels in the replanted finger, and visual monitoring image sequence of the replanted finger.

[0055] Optionally, after authorization, the vascular crisis prediction terminal can acquire clinical pathological index data of the target patient, infrared monitoring image sequence of the replanted finger blood vessels of the target patient, and visual monitoring image sequence of the replanted finger of the target patient from different data sources such as the hospital's electronic medical record system, clinical laboratory testing system, and surgical monitoring equipment terminal, based on a multi-source data interaction interface.

[0056] For example, clinical pathological index data may include, but is not limited to, patient baseline physical index data, surgery-related index data, and postoperative baseline pathological index data.

[0057] Optionally, the vascular crisis prediction terminal can acquire infrared monitoring image sequences of the replanted finger vessels collected by an infrared vascular imaging monitoring device. The infrared monitoring images in the sequence may include sampling timestamps. These infrared monitoring images can be used to characterize blood perfusion within the vessels.

[0058] Optionally, the vascular crisis prediction terminal can acquire a sequence of visual monitoring images of the replanted finger, collected by a visual imaging monitoring device. The visual monitoring images in the sequence may include sampling timestamps. These visual monitoring images can be used to characterize visual features such as the appearance, skin color, and degree of tissue swelling of the replanted finger.

[0059] Schematic illustration: The vascular crisis prediction terminal can perform image enhancement preprocessing on the acquired infrared monitoring image sequence and visual monitoring image sequence. Image enhancement preprocessing may include, but is not limited to, image noise reduction, image size normalization, and image grayscale correction. The vascular crisis prediction terminal can also perform time-series alignment of the infrared monitoring image sequence and visual monitoring image sequence based on the sampling timestamp.

[0060] Step S102: Input the visual monitoring image sequence into the appearance image feature extraction neural network model to extract the replanted finger body filling state feature sequence and the replanted finger body swelling progression feature sequence.

[0061] Specifically, the vascular crisis prediction terminal can input the visual monitoring image sequence frame by frame into the appearance image feature extraction neural network model to extract the replanted finger body filling status feature sequence and the replanted finger body swelling progression feature sequence, respectively.

[0062] Optionally, the neural network model for appearance image feature extraction can be a deep learning model. This model may include a multi-scale convolutional feature extraction module, a fullness state extraction branch, and a swelling progression extraction branch. The fullness state extraction branch may include a fullness state multi-scale feature fusion module and a fullness state feature decoding module. The swelling progression extraction branch may include a swelling progression multi-scale feature fusion module and a swelling progression feature decoding module. The neural network model for appearance image feature extraction can use labeled visual monitoring images of replanted fingers and their corresponding fullness state and swelling progression levels as a training set. It can use the fullness level loss function and the swelling level loss function as the loss functions for model training, and iteratively optimize the model parameters through a backpropagation algorithm.

[0063] To illustrate, the neural network model for extracting features from appearance images can be mounted on a vascular crisis prediction terminal, a vascular crisis prediction server, or a vascular crisis prediction system; no limitations are imposed here.

[0064] Schematic, after extracting the filling state features and swelling progression features of the replanted finger from each time frame of the visual monitoring image sequence, the vascular crisis prediction terminal can splice the filling state features of the replanted finger to obtain a filling state feature sequence, and splice the swelling progression features of the replanted finger to obtain a swelling progression feature sequence.

[0065] Step S103: The infrared monitoring image sequence is input into the vascular imaging feature extraction neural network model to extract the blood supply patency feature sequence. The blood supply patency feature sequence, the replanted finger filling status feature sequence, and the replanted finger swelling progression feature sequence are spliced ​​according to the time sequence to obtain the multi-source image monitoring feature sequence.

[0066] Specifically, the vascular crisis prediction terminal can input the infrared monitoring image sequence frame by frame into the vascular imaging feature extraction neural network model to extract the blood supply patency feature sequence. The temporal sequence of the blood supply patency feature sequence can be consistent with the infrared monitoring image sequence.

[0067] Optionally, the blood supply patency characteristic sequence may include a vascular continuity characteristic sequence and a vasomotor rhythm characteristic sequence. The vascular continuity characteristic sequence can be used to reflect the patency status of the replanted finger's blood vessels, and the vasomotor rhythm characteristic sequence can be used to characterize the vasomotor rhythm status of vascular smooth muscle.

[0068] For example, the vascular imaging feature extraction neural network model can be a deep learning model, which may include a vascular spatial feature extraction module, a vascular vasoconstriction value quantification module, a vascular continuity quantification module, a vasoconstriction rhythm feature time-series analysis module, and a vascular continuity feature time-series analysis module. The vascular imaging feature extraction neural network model can use labeled infrared monitoring images of replanted fingers, the vascular continuity corresponding to the infrared monitoring images of replanted fingers, and the vasoconstriction rhythm corresponding to the infrared monitoring images of replanted fingers as training sets, and perform iterative optimization of model parameters through a regression loss function.

[0069] Schematic illustration: The vascular crisis prediction terminal can match time steps of the blood supply patency feature sequence, the replanted finger filling status feature sequence, and the replanted finger swelling progression feature sequence. The terminal can splice these sequences according to a unified monitoring time sequence, fusing the blood supply patency, replanted finger filling status, and replanted finger swelling progression features at the same time step into a high-dimensional feature vector. The high-dimensional feature vectors from each time step are then arranged chronologically to obtain a multi-source image monitoring feature sequence. This multi-source image monitoring feature sequence can be used to characterize the temporal features of the three core pathological dimensions of the replanted finger: blood supply, filling, and swelling.

[0070] Step S104: Input the multi-source image monitoring feature sequence and clinical pathological index data into the vascular crisis trend prediction model to generate a vascular crisis level prediction sequence.

[0071] Specifically, the vascular crisis prediction terminal can input multi-source image monitoring feature sequences and clinicopathological index data into the vascular crisis trend prediction model to generate a vascular crisis level prediction sequence for the target patient. The vascular crisis level prediction sequence can be a time-series prediction result of the vascular crisis risk level, which can reflect the trend of the risk level of vascular crisis in the replanted finger over several monitoring time steps.

[0072] For example, the vascular crisis trend prediction model can be a deep learning model based on a long short-term memory network and a gating fusion mechanism. This model may include a long short-term memory network encoder, a fully connected encoding network, a global pathological condition gating fusion module, a long short-term memory network decoder, and a vascular crisis level prediction module. The vascular crisis level prediction module may include fully connected layers and activation function layers. The vascular crisis trend prediction model can use labeled multi-source image monitoring feature sequences, clinical pathological index data, and vascular crisis level time-series data as training sets. It uses the classification cross-entropy loss function as the basis for model training loss calculation and iteratively optimizes the model parameters through a backpropagation algorithm.

[0073] To illustrate, the time series data of vascular crisis levels used for training can first be manually labeled as sparse time series data of vascular crisis levels, and then expanded based on the sparse time series data of vascular crisis levels to obtain the time series data of vascular crisis levels used for training.

[0074] In the aforementioned method for predicting the trend of vascular crisis in replanted fingers, the accuracy of extracting external pathological features is improved by capturing pathological change information in visual monitoring images. By accurately extracting the spatial morphological features of subcutaneous vessels in infrared monitoring images, the patency of blood supply can be quantified, and pathological changes in vascular patency can be accurately identified. This solves the technical problem that traditional manual interpretation cannot accurately capture subtle dynamic changes in blood vessels and is prone to missing early blood supply abnormality signals, thus improving the temporal representation ability of blood supply patency features and the sensitivity of early abnormality identification. Through the temporal prediction of vascular crisis trends guided by global pathological conditions, the long-term temporal dependence of multi-source image monitoring feature sequences can be deeply encoded, accurately predicting the dynamic development trend of vascular crisis, and capturing hidden pathological warning signals before the occurrence of crisis in advance. This effectively improves the foresight and accuracy of vascular crisis prediction, thereby avoiding subjective errors caused by manual interpretation, enhancing the risk control ability during the high-risk window period after replantation of fingers, and improving the overall success rate of replantation of fingers.

[0075] In an optional embodiment of this application, please refer to Figure 1 and Figure 2 Step S102 involves inputting the visual monitoring image sequence into the appearance image feature extraction neural network model to extract the replanted finger body filling state feature sequence and the replanted finger body swelling progression feature sequence, which may include:

[0076] Step S202: Input each visual monitoring image in the visual monitoring image sequence into the multi-scale convolutional feature extraction module in the appearance image feature extraction neural network model to extract the multi-scale image feature set of the visual monitoring image.

[0077] Schematic, a multi-scale convolutional feature extraction module may include multiple concatenated convolutional modules.

[0078] Step S203: Input the multi-scale image feature set of the visual monitoring image into the filling state multi-scale feature fusion module in the filling state extraction branch of the appearance image feature extraction neural network model, fuse the single-scale image features in the multi-scale image feature set to obtain the filling state fused image features, and input the filling state fused image features into the filling state feature decoding module in the filling state extraction branch to generate the replanted finger filling state features of each visual monitoring image.

[0079] Step S204: Input the multi-scale image feature set of the visual monitoring image into the swelling progression multi-scale feature fusion module in the swelling progression extraction branch of the appearance image feature extraction neural network model, fuse the single-scale image features in the multi-scale image feature set to obtain the swelling progression fused image features, and input the swelling progression fused image features into the swelling progression feature decoding module in the swelling progression extraction branch to generate the replanted finger swelling progression features of each visual monitoring image.

[0080] Step S205: Based on the time sequence of visual monitoring images in the visual monitoring image sequence, the replanted finger body filling state feature sequence is obtained by splicing the visual monitoring images of each visual monitoring image in the visual monitoring image sequence. Based on the time sequence of visual monitoring images, the replanted finger body swelling progression feature sequence is obtained by splicing the visual monitoring images of each visual monitoring image in the visual monitoring image sequence.

[0081] In the aforementioned method for predicting the vascular crisis trend of replanted fingers, multi-scale convolutional feature extraction and dedicated feature extraction with independent decoding of dual pathological branches are used to extract the appearance visual monitoring images of replanted fingers. This method can fully capture pathological change information at different scales in the visual monitoring images, focusing on the two core pathological dimensions of the replanted finger's filling state and swelling progression. Furthermore, the method uses a branch-specific multi-scale feature fusion module to enhance and adapt the fusion of features at different scales, generating highly recognizable fusion features that conform to the changing patterns of the pathological dimensions. Finally, the method uses a branch-specific feature decoding module to output the pathological features in a hierarchical and quantitative manner, which can improve the targeting of the extraction of the two types of core pathological features and improve the extraction accuracy of the appearance pathological features.

[0082] In an optional embodiment of this application, the multi-scale image feature set may include small-scale image features and medium-scale image features. The multi-scale image feature set of the visual surveillance image is input into the filling state extraction branch of the appearance image feature extraction neural network model, where the filling state multi-scale feature fusion module fuses the single-scale image features from the multi-scale image feature set to obtain the filling state fused image features. This process may include:

[0083] Specifically, the vascular crisis prediction terminal can input small-scale image features from the multi-scale image feature set into the small-scale convolutional feature enhancement module in the multi-scale feature fusion module of the full state to generate enhanced small-scale image features.

[0084] Specifically, the vascular crisis prediction terminal can input the mesoscale image features from the multi-scale image feature set into the filling state mesoscale convolutional feature enhancement module in the filling state multi-scale feature fusion module to generate filling state enhanced mesoscale image features, and upsample the filling state enhanced mesoscale image features to obtain upsampled filling state enhanced image features.

[0085] Specifically, the vascular crisis prediction terminal can fuse enhanced small-scale image features and upsampled filling state enhanced image features to obtain preliminary fused image features of the filling state. These preliminary fused image features of the filling state are then input into the filling state upsampled multi-layer convolution module in the filling state multi-scale feature fusion module to generate the filling state fused image features.

[0086] In an optional embodiment of this application, the multi-scale image feature set may further include large-scale image features. The multi-scale image feature set of the visual monitoring image is input into the swelling progression extraction branch of the appearance image feature extraction neural network model, where the swelling progression multi-scale feature fusion module fuses the single-scale image features from the multi-scale image feature set to obtain the swelling progression fused image features. This may include:

[0087] Specifically, the vascular crisis prediction terminal can input the mesoscale image features from the multi-scale image feature set into the mesoscale convolutional feature enhancement module for swelling progression in the multi-scale feature fusion module for swelling progression, and generate enhanced mesoscale image features for swelling progression.

[0088] Specifically, the vascular crisis prediction terminal can input large-scale image features from the multi-scale image feature set into the large-scale convolutional feature enhancement module in the swelling progression multi-scale feature fusion module to generate enhanced large-scale image features, and then upsample the enhanced large-scale image features to obtain upsampled enhanced large-scale image features.

[0089] Specifically, the vascular crisis prediction terminal can fuse the swelling progression enhanced mesoscale image features and the upsampling enhanced large-scale image features to obtain the preliminary fused image features of swelling progression. The preliminary fused image features of swelling progression are then input into the swelling progression downsampling multi-layer convolution module in the swelling progression multi-scale feature fusion module to generate the swelling progression fused image features.

[0090] In an optional embodiment of this application, the filling status characteristics of the replanted finger may include normal filling level characteristics, mild abnormal filling level characteristics, moderate abnormal filling level characteristics, severe abnormal filling level characteristics, and perfusion disappearance level characteristics, and the swelling progression characteristics of the replanted finger may include normal swelling level, mild progressive swelling level characteristics, moderate progressive swelling level characteristics, and severe progressive swelling level characteristics.

[0091] Optionally, the loss function of the neural network model for appearance image feature extraction may include a fullness level loss function and a swelling level loss function, the expressions of which can be:

[0092]

[0093]

[0094] In the formula, and These are the fullness level loss function and the swelling level loss function, respectively. and These are the total number of grades for the filling status characteristics of the replanted finger and the total number of grades for the swelling progression characteristics of the replanted finger, respectively. For the first The graded replantation refers to the true filling state of the replanted tissue, and its graded unique heat encoding. and The first one predicted by the full state feature decoding module is the first one. The probability of graded replantation with full body filling and the first The probability of graded replantation with the tissue fully engorged. and These are the error spacing weighting coefficients for the fullness status level and the swelling progression level, respectively. For the first Graded replantation refers to the true swelling progression grade coded by a unique thermal code. and The first one predicted by the swelling progression feature decoding module is the first one. The probability of swelling progression in replanted finger tissue and the first The probability of graded replanted finger swelling progression.

[0095] For example, when the replanted finger's filling status characteristics include normal filling level characteristics, mild filling abnormality level characteristics, moderate filling abnormality level characteristics, severe filling abnormality level characteristics, and perfusion disappearance level characteristics, the total number of levels of the replanted finger's filling status characteristics is... The total number of levels can be 5. Normal filling grade characteristics, mild abnormal filling grade characteristics, moderate abnormal filling grade characteristics, severe abnormal filling grade characteristics, and perfusion disappearance grade characteristics can correspond to the first, second, third, fourth, and fifth levels of replanted finger filling states, respectively. When the replanted finger swelling progression characteristics include normal swelling grade, mild progressive swelling grade characteristics, moderate progressive swelling grade characteristics, and severe progressive swelling grade characteristics, the total number of levels of replanted finger swelling progression characteristics is... It can be 4. The characteristics of normal swelling grade, mild progressive swelling grade, moderate progressive swelling grade, and severe progressive swelling grade can correspond to the swelling progression of replanted finger body of grade 1, grade 2, grade 3, and grade 4, respectively.

[0096] Schematic illustration: the characteristics of normal filling level, mild abnormal filling level, moderate abnormal filling level, severe abnormal filling level, and loss of perfusion level can also correspond to the filling state of replanted finger body at level 5, level 4, level 3, level 2, and level 1, respectively. Similarly, the characteristics of normal swelling level, mild progressive swelling level, moderate progressive swelling level, and severe progressive swelling level can also correspond to the swelling progression of replanted finger body at level 4, level 3, level 2, and level 1, respectively. No specific limitations are imposed here.

[0097] In an optional embodiment of this application, the blood supply patency feature sequence may include a vascular continuity feature sequence and a systolic-diastolic rhythm feature sequence. The infrared monitoring image sequence is input into a vascular imaging feature extraction neural network model to extract the blood supply patency feature sequence, which may include:

[0098] Specifically, the vascular crisis prediction terminal can input each infrared monitoring image in the infrared monitoring image sequence into the vascular spatial feature extraction module in the vascular imaging feature extraction neural network model to extract the vascular spatial features of each infrared monitoring image.

[0099] Specifically, the vascular crisis prediction terminal can input vascular spatial features into the vascular vasoconstriction value quantification module in the vascular imaging feature extraction neural network model to extract the initial vascular vasoconstriction values ​​of each infrared monitoring image, and input vascular spatial features into the vascular continuity quantification module in the vascular imaging feature extraction neural network model to extract the initial vascular continuity value of each infrared monitoring image.

[0100] Specifically, the vascular crisis prediction terminal can use the sampling time of each infrared monitoring image as the center time of the vasomotor sliding window. Based on the initial vasomotor values ​​within the vasomotor sliding window, it can construct the vasomotor time window analysis sequence corresponding to each infrared monitoring image. The vasomotor time window analysis sequence is then input into the vasomotor rhythm feature time series analysis module in the vascular imaging feature extraction neural network model to extract the vasomotor rhythm features of each infrared monitoring image.

[0101] Specifically, the vascular crisis prediction terminal can use the sampling time of each infrared monitoring image as the center time of the continuous vascular sliding window. Based on the initial vascular continuity value within the continuous vascular sliding window, it can construct the vascular continuity time window analysis sequence corresponding to each infrared monitoring image. The vascular continuity time window analysis sequence is then input into the vascular continuity feature time sequence analysis module in the vascular imaging feature extraction neural network model to extract the vascular continuity features of each infrared monitoring image.

[0102] Specifically, the vascular crisis prediction terminal can obtain a vasomotor rhythm feature sequence by splicing the vasomotor rhythm features of each infrared monitoring image in time sequence, and obtain a vascular continuity feature sequence by splicing the vascular continuity features of each infrared monitoring image in time sequence.

[0103] In an optional embodiment of this application, please refer to Figure 1 and Figure 2 Step S104 involves inputting the multi-source image monitoring feature sequence and clinicopathological index data into the vascular crisis trend prediction model to generate a vascular crisis level prediction sequence, which may include:

[0104] Step S207: Input the multi-source image monitoring feature sequence into the long short-term memory network encoder in the vascular crisis trend prediction model to generate the encoder hidden state sequence.

[0105] Step S208: Input clinical pathological index data into the fully connected coding network in the vascular crisis trend prediction model to generate global pathological condition features.

[0106] Step S209: Input the encoder hidden state sequence and global pathological condition features into the global pathological condition gating fusion module in the vascular crisis trend prediction model, fuse the hidden state sequence and global pathological condition features, and generate an enhanced encoder hidden state sequence.

[0107] For example, the expression for the hidden state sequence of the augmented encoder can be:

[0108]

[0109]

[0110]

[0111] In the formula, For the first An adaptive gating vector for each time step. This is a learnable weight matrix for gating calculations in the global pathological condition gating fusion module. For the first hidden state in the encoder's hidden state sequence The encoder's hidden state at each time step. This refers to the overall pathological condition characteristics. This refers to the learnable bias term for the gating calculation of the global pathological condition gating fusion module. To enhance the first hidden state sequence of the encoder Conditional enhancement of hidden states at each time step. This is a learnable weight matrix for the global pathological condition gating fusion module, which maps global pathological condition features to hidden state dimensions. This is a learnable bias term for the global pathological condition gating fusion module, which maps global pathological condition features to the hidden state dimension. This is the Hadamard product, which is an element-wise product. To enhance the encoder's hidden state sequence, The total number of time steps in the encoder's hidden state sequence. To enhance the hidden state dimension of the encoder's hidden state sequence, For the set of real numbers, This is the activation function.

[0112] Step S210: Input the enhanced encoder hidden state sequence into the long short-term memory network decoder in the vascular crisis trend prediction model to generate the decoder hidden state sequence, and input the decoder hidden state sequence into the vascular crisis level prediction module in the vascular crisis trend prediction model to generate the vascular crisis level prediction sequence.

[0113] Optionally, the vascular crisis severity prediction module includes a fully connected layer and an activation function layer.

[0114] Schematic, the vascular crisis level in the vascular crisis level prediction sequence can include normal level, low risk level, medium risk level and high risk level.

[0115] In one exemplary embodiment of this application, such as Figure 2 As shown, a method for predicting the trend of vascular crisis in replanted fingers is provided, including:

[0116] Step S201: Obtain clinical pathological index data, infrared monitoring image sequence of blood vessels in the replanted finger, and visual monitoring image sequence of the replanted finger.

[0117] Step S202: Input each visual monitoring image in the visual monitoring image sequence into the multi-scale convolutional feature extraction module in the appearance image feature extraction neural network model to extract the multi-scale image feature set of the visual monitoring image.

[0118] Step S203: Input the multi-scale image feature set of the visual monitoring image into the filling state multi-scale feature fusion module in the filling state extraction branch of the appearance image feature extraction neural network model, fuse the single-scale image features in the multi-scale image feature set to obtain the filling state fused image features, and input the filling state fused image features into the filling state feature decoding module in the filling state extraction branch to generate the replanted finger filling state features of each visual monitoring image.

[0119] Step S204: Input the multi-scale image feature set of the visual monitoring image into the swelling progression multi-scale feature fusion module in the swelling progression extraction branch of the appearance image feature extraction neural network model, fuse the single-scale image features in the multi-scale image feature set to obtain the swelling progression fused image features, and input the swelling progression fused image features into the swelling progression feature decoding module in the swelling progression extraction branch to generate the replanted finger swelling progression features of each visual monitoring image.

[0120] Step S205: Based on the time sequence of visual monitoring images in the visual monitoring image sequence, the replanted finger body filling state feature sequence is obtained by splicing the visual monitoring images of each visual monitoring image in the visual monitoring image sequence. Based on the time sequence of visual monitoring images, the replanted finger body swelling progression feature sequence is obtained by splicing the visual monitoring images of each visual monitoring image in the visual monitoring image sequence.

[0121] Step S206: The infrared monitoring image sequence is input into the vascular imaging feature extraction neural network model to extract the blood supply patency feature sequence. The blood supply patency feature sequence, the replanted finger filling status feature sequence, and the replanted finger swelling progression feature sequence are spliced ​​according to the time sequence to obtain the multi-source image monitoring feature sequence.

[0122] Step S207: Input the multi-source image monitoring feature sequence into the long short-term memory network encoder in the vascular crisis trend prediction model to generate the encoder hidden state sequence.

[0123] Step S208: Input clinical pathological index data into the fully connected coding network in the vascular crisis trend prediction model to generate global pathological condition features.

[0124] Step S209: Input the encoder hidden state sequence and global pathological condition features into the global pathological condition gating fusion module in the vascular crisis trend prediction model, fuse the hidden state sequence and global pathological condition features, and generate an enhanced encoder hidden state sequence.

[0125] Step S210: Input the enhanced encoder hidden state sequence into the long short-term memory network decoder in the vascular crisis trend prediction model to generate the decoder hidden state sequence, and input the decoder hidden state sequence into the vascular crisis level prediction module in the vascular crisis trend prediction model to generate the vascular crisis level prediction sequence.

[0126] The aforementioned method for predicting vascular crises in replanted fingers achieves accurate prediction and early warning of vascular crises through multimodal data fusion, deep feature extraction, and dynamic temporal modeling, addressing the issues of strong subjectivity, poor timeliness, and limited information dimensions inherent in traditional monitoring methods. Specifically, a neural network model is used to extract multidimensional dynamic features such as filling status, swelling progression, and blood supply patency. These features are then temporally spliced ​​and fused to obtain a comprehensive monitoring feature sequence. Furthermore, by combining a long short-term memory network and a global pathological condition fusion mechanism, a quantitative crisis level prediction sequence is generated. This overcomes the limitations of single-modal data and breaks through the lag of static assessment, providing clinicians with real-time and objective risk quantification, significantly improving the timeliness of intervention and the scientific nature of decision-making. Ultimately, this reduces the replantation failure rate and ensures the quality of patient prognosis.

[0127] It should be understood that although the steps in the flowcharts of the embodiments described above are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the embodiments described above may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0128] Based on the same inventive concept, this application also provides a replanted finger vascular crisis trend prediction system for implementing the above-described method for predicting the trend of vascular crisis in a replanted finger. The solution provided by this system is similar to the solution described in the above method. Therefore, the specific limitations of one or more embodiments of the replanted finger vascular crisis trend prediction system provided below can be found in the limitations of the replanted finger vascular crisis trend prediction method described above, and will not be repeated here.

[0129] In one exemplary embodiment, such as Figure 3 As shown, a replanted finger vascular crisis trend prediction system 300 is provided, including:

[0130] The pathological monitoring data acquisition module 301 can be used to acquire clinical pathological index data, infrared monitoring image sequences of blood vessels in the replanted finger, and visual monitoring image sequences of the replanted finger.

[0131] The appearance image feature extraction module 302 can be used to input the visual monitoring image sequence into the appearance image feature extraction neural network model to extract the replanted finger body filling state feature sequence and the replanted finger body swelling progression feature sequence.

[0132] The vascular imaging feature extraction module 303 can be used to input infrared monitoring image sequences into the vascular imaging feature extraction neural network model to extract blood supply patency feature sequences, and splice the blood supply patency feature sequences, replanted finger filling status feature sequences, and replanted finger swelling progression feature sequences according to the time sequence to obtain multi-source image monitoring feature sequences.

[0133] The vascular crisis level prediction module 304 can be used to input multi-source image monitoring feature sequences and clinicopathological index data into the vascular crisis trend prediction model to generate a vascular crisis level prediction sequence.

[0134] In an optional embodiment of this application, the appearance image feature extraction module 302 can also be used for:

[0135] Each visual monitoring image in the visual monitoring image sequence is input into the multi-scale convolutional feature extraction module in the appearance image feature extraction neural network model to extract the multi-scale image feature set of the visual monitoring image.

[0136] The multi-scale image feature set of the visual monitoring image is input into the multi-scale feature fusion module of the fullness state extraction branch in the appearance image feature extraction neural network model. The single-scale image features of each multi-scale image feature set are fused to obtain the fullness state fused image features. The fullness state fused image features are then input into the fullness state feature decoding module of the fullness state extraction branch to generate the replanted finger fullness state features of each visual monitoring image.

[0137] The multi-scale image feature set of the visual monitoring image is input into the swelling progression multi-scale feature fusion module in the swelling progression extraction branch of the appearance image feature extraction neural network model. The single-scale image features in the multi-scale image feature set are fused to obtain the swelling progression fused image features. The swelling progression fused image features are then input into the swelling progression feature decoding module in the swelling progression extraction branch to generate the replanted finger swelling progression features of each visual monitoring image.

[0138] Based on the temporal splicing of the visual monitoring images in the visual monitoring image sequence, the filling state characteristics of the replanted finger body are obtained, and based on the temporal splicing of the replanted finger body swelling progression characteristics of the visual monitoring images, the swelling progression characteristics of the replanted finger body are obtained.

[0139] In an optional embodiment of this application, the appearance image feature extraction module 302 can also be used for:

[0140] The small-scale image features in the multi-scale image feature set are input into the small-scale convolutional feature enhancement module in the full-state multi-scale feature fusion module to generate enhanced small-scale image features.

[0141] The mesoscale image features from the multi-scale image feature set are input into the full-state mesoscale convolutional feature enhancement module in the full-state multi-scale feature fusion module to generate full-state enhanced mesoscale image features. The full-state enhanced mesoscale image features are then upsampled to obtain upsampled full-state enhanced image features.

[0142] The enhanced small-scale image features and the upsampled full-state enhanced image features are fused to obtain the preliminary fused image features of the full-state. The preliminary fused image features of the full-state are then input into the full-state upsampled multi-layer convolution module in the full-state multi-scale feature fusion module to generate the full-state fused image features.

[0143] In an optional embodiment of this application, the appearance image feature extraction module 302 can also be used for:

[0144] The mesoscale image features from the multi-scale image feature set are input into the mesoscale convolutional feature enhancement module for swelling progression in the multi-scale feature fusion module for swelling progression, generating enhanced mesoscale image features for swelling progression.

[0145] The large-scale image features in the multi-scale image feature set are input into the large-scale convolutional feature enhancement module in the swelling progression multi-scale feature fusion module to generate enhanced large-scale image features. The enhanced large-scale image features are then upsampled to obtain upsampled enhanced large-scale image features.

[0146] The swelling progression enhancement mid-scale image features and upsampling enhancement large-scale image features are fused to obtain preliminary fused image features of swelling progression. These preliminary fused image features are then input into the swelling progression downsampling multi-layer convolution module in the swelling progression multi-scale feature fusion module to generate fused image features of swelling progression.

[0147] In an optional embodiment of this application, the vascular imaging feature extraction module 303 can also be used for:

[0148] Each infrared monitoring image in the infrared monitoring image sequence is input into the vascular spatial feature extraction module in the vascular imaging feature extraction neural network model to extract the vascular spatial features of each infrared monitoring image.

[0149] The spatial features of blood vessels are input into the vasomotor value quantification module in the vasomotor feature extraction neural network model to extract the initial vasomotor values ​​of each infrared monitoring image. The spatial features of blood vessels are also input into the vasculature continuity quantification module in the vasculature feature extraction neural network model to extract the initial vasculature continuity values ​​of each infrared monitoring image.

[0150] Using the sampling time of each infrared monitoring image as the center time of the vasomotor sliding window, the vasomotor time window analysis sequence corresponding to each infrared monitoring image is constructed based on the initial vasomotor values ​​within the vasomotor sliding window. The vasomotor time window analysis sequence is then input into the vasomotor rhythm feature time series analysis module in the vasomotor imaging feature extraction neural network model to extract the vasomotor rhythm features of each infrared monitoring image.

[0151] Using the sampling time of each infrared monitoring image as the center time of the blood vessel continuous sliding window, the blood vessel continuity time window analysis sequence corresponding to each infrared monitoring image is constructed based on the initial blood vessel continuity value within the blood vessel continuous sliding window. The blood vessel continuity time window analysis sequence is then input into the blood vessel continuity feature time sequence analysis module in the blood vessel imaging feature extraction neural network model to extract the blood vessel continuity features of each infrared monitoring image.

[0152] Based on the temporal splicing of the infrared monitoring images of each infrared monitoring image, a vasomotor rhythm feature sequence is obtained, and based on the temporal splicing of the vascular continuity features of each infrared monitoring image, a vascular continuity feature sequence is obtained.

[0153] In an optional embodiment of this application, the vascular crisis level prediction module 304 can also be used for:

[0154] The feature sequences from multi-source image monitoring are input into the long short-term memory network encoder in the vascular crisis trend prediction model to generate the encoder hidden state sequence.

[0155] Clinical pathological index data are input into the fully connected coding network in the vascular crisis trend prediction model to generate global pathological condition features.

[0156] The encoder hidden state sequence and global pathological condition features are input into the global pathological condition gating fusion module in the vascular crisis trend prediction model to fuse the hidden state sequence and global pathological condition features and generate an enhanced encoder hidden state sequence.

[0157] The enhanced encoder hidden state sequence is input into the long short-term memory network decoder in the vascular crisis trend prediction model to generate the decoder hidden state sequence. The decoder hidden state sequence is then input into the vascular crisis level prediction module in the vascular crisis trend prediction model to generate the vascular crisis level prediction sequence. The vascular crisis level prediction module includes a fully connected layer and an activation function layer.

[0158] In one embodiment, a computer device is provided, including a memory and a processor, the memory storing a computer program, the processor executing the computer program to implement the steps of the previously described method for predicting vascular crises in replanted fingers.

[0159] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the steps in the above method embodiments.

[0160] For the device embodiments, since they basically correspond to the method embodiments, the relevant parts can be referred to in the description of the method embodiments. The device embodiments described above are merely illustrative. The components described as separate parts may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this disclosure according to actual needs. Those skilled in the art can understand and implement this without creative effort.

[0161] The above-described embodiments are merely illustrative of several implementation methods of the embodiments of this application, and their descriptions are relatively specific and detailed. However, they should not be construed as limiting the scope of the patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the embodiments of this application, and these modifications and improvements all fall within the protection scope of the embodiments of this application.

Claims

1. A method for predicting the trend of vascular crisis in replanted fingers, characterized in that, The method includes: Acquire clinical pathological index data, infrared monitoring image sequences of blood vessels in the replanted finger, and visual monitoring image sequences of the replanted finger. The visual monitoring image sequence is input into the appearance image feature extraction neural network model to extract the replanted finger body filling state feature sequence and the replanted finger body swelling progression feature sequence; The infrared monitoring image sequence is input into the vascular imaging feature extraction neural network model to extract the blood supply patency feature sequence. The blood supply patency feature sequence, the replanted finger filling state feature sequence, and the replanted finger swelling progression feature sequence are spliced ​​together according to the time sequence to obtain the multi-source image monitoring feature sequence. The multi-source image monitoring feature sequence and the clinical pathological index data are input into the vascular crisis trend prediction model to generate a vascular crisis level prediction sequence.

2. The method according to claim 1, characterized in that, The step of inputting the visual monitoring image sequence into the appearance image feature extraction neural network model to extract the replanted finger body filling state feature sequence and the replanted finger body swelling progression feature sequence includes: Each visual monitoring image in the visual monitoring image sequence is input into the multi-scale convolutional feature extraction module in the appearance image feature extraction neural network model to extract the multi-scale image feature set of the visual monitoring image; The multi-scale image feature set of the visual monitoring image is input into the multi-scale feature fusion module of the fullness state extraction branch in the appearance image feature extraction neural network model. The single-scale image features of the multi-scale image feature set are fused to obtain the fullness state fused image features. The fullness state fused image features are then input into the fullness state feature decoding module of the fullness state extraction branch to generate the replanted finger fullness state features of each visual monitoring image. The multi-scale image feature set of the visual monitoring image is input into the swelling progression multi-scale feature fusion module in the swelling progression extraction branch of the appearance image feature extraction neural network model. The single-scale image features in the multi-scale image feature set are fused to obtain the swelling progression fused image features. The swelling progression fused image features are then input into the swelling progression feature decoding module in the swelling progression extraction branch to generate the replanted finger swelling progression features of each visual monitoring image. The replanted finger fill-up state feature sequence is obtained by splicing the visual monitoring images of each visual monitoring image in the visual monitoring image sequence in a temporal sequence, and the replanted finger swelling progression feature sequence is obtained by splicing the replanted finger swelling progression feature of each visual monitoring image in a temporal sequence.

3. The method according to claim 2, characterized in that, The multi-scale image feature set includes small-scale image features and medium-scale image features. The step of inputting the multi-scale image feature set of the visual monitoring image into the fullness state extraction branch of the appearance image feature extraction neural network model, and fusing the single-scale image features in the multi-scale image feature set to obtain fullness state fused image features, includes: The small-scale image features in the multi-scale image feature set are input into the small-scale convolutional feature enhancement module in the full-state multi-scale feature fusion module to generate enhanced small-scale image features. The mesoscale image features in the multi-scale image feature set are input into the full-state mesoscale convolutional feature enhancement module in the full-state multi-scale feature fusion module to generate full-state enhanced mesoscale image features, and the full-state enhanced mesoscale image features are upsampled to obtain upsampled full-state enhanced image features. The enhanced small-scale image features and the upsampled full-state enhanced image features are fused to obtain the full-state preliminary fused image features. The full-state preliminary fused image features are then input into the full-state downsampled multi-layer convolution module in the full-state multi-scale feature fusion module to generate the full-state fused image features.

4. The method according to claim 3, characterized in that, The multi-scale image feature set also includes large-scale image features. The step of inputting the multi-scale image feature set of the visual monitoring image into the swelling progression extraction branch of the appearance image feature extraction neural network model, and fusing the single-scale image features in the multi-scale image feature set to obtain fused swelling progression image features, includes: The mesoscale image features in the multi-scale image feature set are input into the mesoscale convolutional feature enhancement module for swelling progression in the multi-scale feature fusion module for swelling progression, to generate enhanced mesoscale image features for swelling progression. The large-scale image features in the multi-scale image feature set are input into the large-scale convolutional feature enhancement module in the swelling progression multi-scale feature fusion module to generate enhanced large-scale image features, and the enhanced large-scale image features are upsampled to obtain upsampled enhanced large-scale image features. The swelling progression enhanced mesoscale image features and the upsampling enhanced large-scale image features are fused to obtain preliminary fused image features of swelling progression. These preliminary fused image features of swelling progression are then input into the swelling progression downsampling multi-layer convolution module in the swelling progression multi-scale feature fusion module to generate the swelling progression fused image features.

5. The method according to any one of claims 2 to 4, characterized in that: The replanted finger's filling status characteristics include normal filling level characteristics, mild abnormal filling level characteristics, moderate abnormal filling level characteristics, severe abnormal filling level characteristics, and perfusion disappearance level characteristics. The replanted finger's swelling progression characteristics include normal swelling level, mild progressive swelling level characteristics, moderate progressive swelling level characteristics, and severe progressive swelling level characteristics. The loss function of the neural network model for extracting appearance image features includes a fullness level loss function and a swelling level loss function, the expressions of which are: In the formula, and These are the fullness level loss function and the swelling level loss function, respectively. and These are the total number of levels for the filling state characteristics of the replanted finger and the total number of levels for the swelling progression characteristics of the replanted finger, respectively. For the first The graded replantation refers to the true filling state of the replanted tissue, and its graded unique heat encoding. and The first number predicted by the fullness state feature decoding module is respectively The probability of graded replantation with full body filling and the first The probability of graded replantation with the tissue fully engorged. and These are the error spacing weighting coefficients for the fullness status level and the swelling progression level, respectively. For the first Graded replantation refers to the true swelling progression grade coded by a unique thermal code. and The first number predicted by the swelling progression feature decoding module is respectively The probability of swelling progression in replanted finger tissue and the first The probability of graded replanted finger swelling progression.

6. The method according to claim 1, characterized in that, The blood supply patency feature sequence includes a vascular continuity feature sequence and a systolic-diastolic rhythm feature sequence. The step of inputting the infrared monitoring image sequence into a vascular imaging feature extraction neural network model to extract the blood supply patency feature sequence includes: Each infrared monitoring image in the infrared monitoring image sequence is input into the vascular spatial feature extraction module in the vascular imaging feature extraction neural network model to extract the vascular spatial features of each infrared monitoring image; The vascular spatial features are input into the vascular contraction and dilation value quantification module in the vascular imaging feature extraction neural network model to extract the initial vascular contraction and dilation values ​​of each infrared monitoring image. The vascular spatial features are also input into the vascular continuity quantification module in the vascular imaging feature extraction neural network model to extract the initial vascular continuity value of each infrared monitoring image. Using the sampling time of each infrared monitoring image as the center time of the vasomotor sliding window, a vasomotor time window analysis sequence corresponding to each infrared monitoring image is constructed based on the initial vasomotor values ​​within the vasomotor sliding window. The vasomotor time window analysis sequence is then input into the vasomotor rhythm feature time sequence analysis module in the vasomotor imaging feature extraction neural network model to extract the vasomotor rhythm features of each infrared monitoring image. Using the sampling time of each infrared monitoring image as the center time of the blood vessel continuous sliding window, a blood vessel continuity time window analysis sequence corresponding to each infrared monitoring image is constructed based on the initial blood vessel continuity value within the blood vessel continuous sliding window. The blood vessel continuity time window analysis sequence is then input into the blood vessel continuity feature time sequence analysis module in the blood vessel imaging feature extraction neural network model to extract the blood vessel continuity features of each infrared monitoring image. The systolic and diastolic rhythm features are spliced ​​together from the infrared monitoring images according to their temporal sequence to obtain the systolic and diastolic rhythm feature sequence, and the vascular continuity features are spliced ​​together from the infrared monitoring images according to their temporal sequence to obtain the vascular continuity feature sequence.

7. The method according to claim 1, characterized in that, The step of inputting the multi-source image monitoring feature sequence and the clinicopathological index data into the vascular crisis trend prediction model to generate a vascular crisis level prediction sequence includes: The multi-source image monitoring feature sequence is input into the long short-term memory network encoder in the vascular crisis trend prediction model to generate the encoder hidden state sequence. The clinical pathological index data are input into the fully connected coding network in the vascular crisis trend prediction model to generate global pathological condition features. The encoder hidden state sequence and the global pathological condition features are input into the global pathological condition gating fusion module in the vascular crisis trend prediction model to fuse the hidden state sequence and the global pathological condition features to generate an enhanced encoder hidden state sequence. The enhanced encoder hidden state sequence is input into the long short-term memory network decoder in the vascular crisis trend prediction model to generate the decoder hidden state sequence. The decoder hidden state sequence is then input into the vascular crisis level prediction module in the vascular crisis trend prediction model to generate the vascular crisis level prediction sequence. The vascular crisis level prediction module includes a fully connected layer and an activation function layer.

8. A system for predicting the trend of vascular crisis in replanted fingers, characterized in that, The system includes: The pathological monitoring data acquisition module is used to acquire clinical pathological index data, infrared monitoring image sequences of blood vessels in the replanted finger, and visual monitoring image sequences of the replanted finger. The appearance image feature extraction module is used to input the visual monitoring image sequence into the appearance image feature extraction neural network model to extract the replanted finger body filling state feature sequence and the replanted finger body swelling progress feature sequence; The vascular imaging feature extraction module is used to input the infrared monitoring image sequence into the vascular imaging feature extraction neural network model, extract the blood supply patency feature sequence, and splice the blood supply patency feature sequence, the replanted finger filling state feature sequence, and the replanted finger swelling progression feature sequence according to the time sequence to obtain the multi-source image monitoring feature sequence. The vascular crisis level prediction module is used to input the multi-source image monitoring feature sequence and the clinical pathological index data into the vascular crisis trend prediction model to generate a vascular crisis level prediction sequence.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the method of any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the method of any one of claims 1 to 7.

Images