Artificial intelligence-based postoperative wound rehabilitation evaluation method for urinary surgery, electronic equipment and storage medium

Abstract

The invention belongs to the technical field of medical rehabilitation assessment, and provides a urinary surgery postoperative wound rehabilitation assessment method based on artificial intelligence, electronic equipment and a storage medium. The method comprises the steps of multi-source data acquisition, data preprocessing, spatial alignment, normalization processing, three-branch network feature extraction, feature integration, feature compression, feature mapping and wound rehabilitation grade matching. Through three-mode objective data of microwave imaging, electrical impedance imaging and bioelectrical impedance, the whole data acquisition process is free of invasion and interference, and errors caused by individual differences are reduced; through a three-branch feature extraction network and a deep fusion process, different types of data characteristics are adapted, cross-modal information complementation is realized, and wound rehabilitation multi-dimensional key signals are captured.

Description

Technical Field

[0001] This invention relates to the field of medical rehabilitation assessment technology, and in particular to an artificial intelligence-based method, electronic device and storage medium for postoperative wound rehabilitation assessment in urology. Background Technology

[0002] Current surgical wound rehabilitation assessment techniques are evolving from traditional subjective assessments to modern objective and quantitative methods, forming a multi-dimensional integrated assessment system. Traditional methods primarily rely on visual and tactile examinations by healthcare professionals combined with patient subjective reports, such as observing wound redness and swelling, exudation, and using VAS pain scores. However, these methods suffer from limitations such as strong subjectivity and insufficient real-time data. Modern technologies utilize imaging and sensing technologies for precise monitoring. Ultrasound can assess the healing layers of wound tissue, infrared thermography can determine the state of inflammation through temperature changes, and wearable devices can capture physiological signals such as heart rate variability and skin conductance in real time, correlated with the rehabilitation process. Further breakthroughs in artificial intelligence technology have enabled multimodal neural networks, combining wound images and patient report data, to predict surgical site infections with accuracy approaching that of clinicians, while significantly reducing the workload of healthcare professionals.

[0003] Traditional rehabilitation assessments rely on visual and tactile examinations by medical staff and subjective reports from patients, lacking objective data support and easily influenced by experience and individual feelings. Different types of imaging data are spatially misaligned, and existing methods lack effective feature interaction mechanisms, resulting in inefficient fusion effects. Traditional network models are difficult to adapt to multiple types of data such as microwave imaging and bioelectrical impedance analysis, and incomplete capture of key information leads to insufficient accuracy in feature extraction. Summary of the Invention

[0004] To overcome the shortcomings of existing technologies, the purpose of this invention is to provide an artificial intelligence-based method, electronic device, and storage medium for assessing postoperative wound rehabilitation in urology, thereby solving the problems of traditional methods being easily influenced by experience and individual feelings, lacking effective feature interaction mechanisms, and failing to capture key information comprehensively.

[0005] To achieve the above objectives, the present invention provides the following solution:

[0006] An artificial intelligence-based method for assessing postoperative wound rehabilitation in urology includes:

[0007] Collect raw microwave imaging data, raw electrical impedance imaging data, and raw bioelectrical impedance data from the target postoperative patient;

[0008] The microwave imaging data is subjected to adaptive Wiener filtering and median filtering in sequence to obtain preprocessed microwave imaging data. The electrical impedance imaging data is subjected to Tikhonov regularization to obtain preprocessed electrical impedance imaging data. The bioelectrical impedance data is filtered using Grubbs criterion and moving average filtering method to obtain preprocessed bioelectrical impedance data.

[0009] The preprocessed microwave imaging data and the preprocessed electrical impedance imaging data were spatially aligned using the SIFT algorithm and the least squares method.

[0010] The preprocessed microwave imaging data, the preprocessed electrical impedance imaging data, and the preprocessed bioelectrical impedance data are normalized.

[0011] The preprocessed microwave imaging data is encoded using an improved U-Net network to obtain microwave imaging features.

[0012] The preprocessed electrical impedance imaging data is encoded using an improved ResNet-34 network to obtain electrical impedance imaging features.

[0013] The preprocessed bioelectrical impedance data is encoded using an improved GRU network to obtain bioelectrical impedance features.

[0014] The microwave imaging features, the electrical impedance imaging features, and the bioelectrical impedance features are interactively processed and integrated to obtain three-modal fusion features;

[0015] The three-modal fused features are input into an improved Transformer network for computation and feature compression to obtain a global fused vector;

[0016] The global fusion vector is feature-mapped using a fully connected network to obtain the network prediction value;

[0017] The wound rehabilitation assessment result is obtained by matching the level corresponding to the network prediction value with the preset rehabilitation assessment rules.

[0018] Preferably, spatial alignment of the preprocessed microwave imaging data and the preprocessed electrical impedance imaging data is performed using the SIFT algorithm and the least squares method, including:

[0019] The CLAHE algorithm is used to enhance the local contrast of the preprocessed microwave imaging data and the preprocessed electrical impedance imaging data.

[0020] The SIFT algorithm is used to perform scale-space extremum detection, orientation assignment, and descriptor generation on the enhanced preprocessed microwave imaging data and the preprocessed electrical impedance imaging data, respectively, to obtain image descriptors.

[0021] The image descriptors corresponding to the preprocessed microwave imaging data and the preprocessed electrical impedance imaging data are matched and filtered according to Euclidean distance and RANSAC algorithm to obtain matching pairs.

[0022] The least squares method is used to solve the transformation matrix of the matching pair to obtain the alignment matrix, and the alignment matrix is ​​used to spatially align the preprocessed microwave imaging data and the preprocessed electrical impedance imaging data.

[0023] Preferably, the improved U-Net network comprises: an input layer, an encoder, a bottleneck layer, a decoder, and an output layer connected in sequence; the decoder embeds a spatial attention gate.

[0024] Preferably, the improved ResNet-34 network comprises: an input layer, an initial convolutional layer, a max-pooling layer, a residual block group, a global average pooling layer, and an output layer connected in sequence; each residual block in the residual block group embeds two frequency-adaptive convolutions; the expression for the frequency-adaptive convolution is:

[0025] ;in, It is a dynamic kernel; Based on the convolutional kernel; Frequency-shifted convolution kernel; For the first Each frequency adaptive weight; This represents the number of convolution kernels.

[0026] Preferably, the improved GRU network comprises: an input layer, a bidirectional GRU layer, a temporal attention layer, a feature fusion layer, and an output layer connected in sequence.

[0027] Preferably, the microwave imaging features, the electrical impedance imaging features, and the bioelectrical impedance features are interactively processed and integrated to obtain a three-modal fusion feature, including:

[0028] The microwave imaging features and the electrical impedance imaging features are flattened, and the bioelectrical impedance features are expanded.

[0029] A bidirectional cross-attention calculation is performed on the flattened microwave imaging features and the electrical impedance imaging features to obtain a first interactive feature and a second interactive feature; the expression for the first interactive feature is: The expression for the second interactive feature is: ;in, , These are the first interaction feature and the second interaction feature, respectively; , These are the microwave imaging features and the electrical impedance imaging features after being flattened, respectively. Attention weights for microwave-to-resistivity; Attention weights for impedance to microwave; , These are the first-value matrix and the second-value matrix, respectively.

[0030] Temporal-spatial attention is performed on the first interaction feature, the second interaction feature, and the extended bioelectrical impedance feature to obtain the first temporal-spatial interaction feature, the second temporal-spatial interaction feature, and the third temporal-spatial interaction feature; the expression of the first temporal-spatial interaction feature is:

[0031] The expression for the second temporal-spatial interaction feature is:

[0032] The expression for the third temporal-spatial interaction feature is:

[0033] ;in, , , These are the first temporal-spatial interaction feature, the second temporal-spatial interaction feature, and the third temporal-spatial interaction feature, respectively. , These are the attention weights of temporal features on spatial location and the feedback attention weights of spatial features on temporal features, respectively. , These are the third-valued matrix and the fourth-valued matrix, respectively. For the extended bioelectrical impedance characteristics; This indicates mean pooling.

[0034] The first temporal-spatial interaction feature, the second temporal-spatial interaction feature, and the third temporal-spatial interaction feature are aligned and spliced ​​to obtain the trimodal fusion feature.

[0035] Preferably, the method of using a fully connected network to perform feature mapping on the global fusion vector to obtain network prediction values ​​further includes:

[0036] The three-modal fusion features are input into the two-layer fully connected network and forward propagated several times to obtain multiple network prediction values;

[0037] The prediction mean, prediction entropy, and temporal consistency score of the network predictions are statistically analyzed.

[0038] When the predicted entropy is greater than the entropy threshold or the temporal consistency score is less than the consistency threshold, the network predicted value is marked as low confidence.

[0039] Preferably, the three-modal fused features are input into an improved Transformer network for computation and feature compression to obtain a global fused vector, including:

[0040] Spatial self-attention calculation is performed on the three-modal fusion features to obtain spatial correlation features;

[0041] The third temporal-spatial interaction feature is extended, and the spatial sequence of the extended third temporal-spatial interaction feature is spliced ​​with the spatial association feature to obtain a spatiotemporal sequence.

[0042] Joint self-attention calculation is performed on the spatiotemporal sequence to obtain global correlation features;

[0043] The global correlation features are compressed using a multilayer perceptron to obtain the global fusion vector.

[0044] Preferably, an electronic device includes: at least one processor and a memory communicatively connected to the processor; wherein the memory stores instructions executable by the processor, the instructions being executed by the processor to enable the processor to perform the aforementioned artificial intelligence-based urological postoperative wound rehabilitation assessment method.

[0045] Preferably, a non-transitory computer-readable storage medium stores computer instructions for causing a computer to execute the aforementioned artificial intelligence-based postoperative wound rehabilitation assessment method for urological surgery.

[0046] The present invention discloses the following technical effects:

[0047] This invention provides an artificial intelligence-based method, electronic device, and storage medium for assessing postoperative wound rehabilitation in urology. By using objective data from three modalities—microwave imaging, electrical impedance imaging, and bioelectrical impedance—it overcomes the shortcomings of traditional methods that are easily influenced by experience and individual perceptions, achieving non-invasive and interference-free data acquisition. Through a three-branch feature extraction network and a deep fusion process, it solves the problems of existing methods lacking an effective feature interaction mechanism and incomplete capture of key information, achieving cross-modal information complementarity. Attached Figure Description

[0048] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, 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.

[0049] Figure 1 A schematic diagram of an AI-based urological postoperative wound rehabilitation assessment process provided in an embodiment of the present invention;

[0050] Figure 2 This is a schematic diagram of the spatial alignment process provided in an embodiment of the present invention;

[0051] Figure 3 This is a schematic diagram of the interactive processing and feature integration process provided in an embodiment of the present invention;

[0052] Figure 4 This is a schematic diagram of the feature mapping process provided in an embodiment of the present invention;

[0053] Figure 5 This is a schematic diagram of the feature calculation and feature compression process provided in an embodiment of the present invention. Detailed Implementation

[0054] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0055] The purpose of this invention is to provide an artificial intelligence-based method, electronic device and storage medium for assessing postoperative wound rehabilitation in urology, which solves the problems of traditional methods being easily affected by experience and individual feelings, lacking effective feature interaction mechanisms and incomplete capture of key information.

[0056] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0057] Figure 1 This is a schematic diagram of an artificial intelligence-based urological postoperative wound rehabilitation assessment process provided in an embodiment of the present invention, such as... Figure 1 As shown, this invention provides an artificial intelligence-based method for assessing postoperative wound rehabilitation in urology, including:

[0058] Step 100: Collect raw microwave imaging data, raw electrical impedance imaging data, and raw bioelectrical impedance data of the target postoperative patient;

[0059] Step 200: Perform adaptive Wiener filtering and median filtering on the microwave imaging data in sequence to obtain preprocessed microwave imaging data; perform Tikhonov regularization on the electrical impedance imaging data to obtain preprocessed electrical impedance imaging data; and filter the bioelectrical impedance data using the Grubbs criterion and moving average filtering method to obtain preprocessed bioelectrical impedance data.

[0060] Step 300: Spatial alignment of the preprocessed microwave imaging data and the preprocessed electrical impedance imaging data is performed using the SIFT algorithm and the least squares method;

[0061] Step 400: Normalize the preprocessed microwave imaging data, the preprocessed electrical impedance imaging data, and the preprocessed bioelectrical impedance data;

[0062] Step 500: Encode the preprocessed microwave imaging data using an improved U-Net network to obtain microwave imaging features;

[0063] Step 600: Encode the preprocessed electrical impedance imaging data using an improved ResNet-34 network to obtain electrical impedance imaging features;

[0064] Step 700: Encode the preprocessed bioelectrical impedance data using an improved GRU network to obtain bioelectrical impedance features;

[0065] Step 800: Perform interactive processing and feature integration on the microwave imaging features, the electrical impedance imaging features, and the bioelectrical impedance features to obtain three-modal fusion features;

[0066] Step 900: Input the three-modal fused features into the improved Transformer network for calculation and feature compression to obtain the global fused vector;

[0067] Step 1000: Use a fully connected network to perform feature mapping on the global fusion vector to obtain the network prediction value;

[0068] Step 1100: Match the level corresponding to the network prediction value with the preset rehabilitation assessment rules to obtain the wound rehabilitation assessment result.

[0069] refer to Figure 2 Spatial alignment of the preprocessed microwave imaging data and the preprocessed electrical impedance imaging data is performed using the SIFT algorithm and the least squares method, including:

[0070] Step 301: Use the CLAHE algorithm to perform local contrast enhancement on the preprocessed microwave imaging data and the preprocessed electrical impedance imaging data;

[0071] Step 302: Using the SIFT algorithm, scale-space extremum detection, orientation assignment, and descriptor generation are performed on the enhanced preprocessed microwave imaging data and the preprocessed electrical impedance imaging data respectively to obtain image descriptors;

[0072] Step 303: Match and filter the image descriptors corresponding to the preprocessed microwave imaging data and the preprocessed electrical impedance imaging data according to Euclidean distance and RANSAC algorithm to obtain matching pairs;

[0073] Step 304: Solve the transformation matrix of the matching pair using the least squares method to obtain the alignment matrix, and use the alignment matrix to spatially align the preprocessed microwave imaging data and the preprocessed electrical impedance imaging data.

[0074] Preferably, the improved U-Net network comprises: an input layer, an encoder, a bottleneck layer, a decoder, and an output layer connected in sequence; the decoder embeds a spatial attention gate.

[0075] Preferably, the improved ResNet-34 network comprises: an input layer, an initial convolutional layer, a max-pooling layer, a residual block group, a global average pooling layer, and an output layer connected in sequence; each residual block in the residual block group embeds two frequency-adaptive convolutions; the expression for the frequency-adaptive convolution is:

[0076] ;in, It is a dynamic kernel; Based on the convolutional kernel; Frequency-shifted convolution kernel; For the first Each frequency adaptive weight; This represents the number of convolution kernels.

[0077] Preferably, the improved GRU network comprises: an input layer, a bidirectional GRU layer, a temporal attention layer, a feature fusion layer, and an output layer connected in sequence.

[0078] refer to Figure 3 The microwave imaging features, the electrical impedance imaging features, and the bioelectrical impedance features are interactively processed and integrated to obtain three-modal fusion features, including:

[0079] Step 801: Flatten the microwave imaging features and the electrical impedance imaging features, and expand the bioelectrical impedance features;

[0080] Step 802: Perform bidirectional cross-attention calculation on the flattened microwave imaging features and the electrical impedance imaging features to obtain the first interactive feature and the second interactive feature; the expression of the first interactive feature is: The expression for the second interactive feature is: ;in, , These are the first interaction feature and the second interaction feature, respectively; , These are the microwave imaging features and the electrical impedance imaging features after being flattened, respectively. Attention weights for microwave-to-resistivity; Attention weights for impedance to microwave; , These are the first-value matrix and the second-value matrix, respectively.

[0081] Step 803: Perform temporal-spatial attention calculation on the first interaction feature, the second interaction feature, and the extended bioelectrical impedance feature to obtain the first temporal-spatial interaction feature, the second temporal-spatial interaction feature, and the third temporal-spatial interaction feature; the expression of the first temporal-spatial interaction feature is:

[0082] The expression for the second temporal-spatial interaction feature is:

[0083] The expression for the third temporal-spatial interaction feature is:

[0084] ;in, , , These are the first temporal-spatial interaction feature, the second temporal-spatial interaction feature, and the third temporal-spatial interaction feature, respectively. , These are the attention weights of temporal features on spatial location and the feedback attention weights of spatial features on temporal features, respectively. , These are the third-valued matrix and the fourth-valued matrix, respectively. For the extended bioelectrical impedance characteristics; This indicates mean pooling.

[0085] Step 804: Align and splice the first temporal-spatial interaction feature, the second temporal-spatial interaction feature, and the third temporal-spatial interaction feature to obtain the trimodal fusion feature.

[0086] refer to Figure 4 The method further includes using a fully connected network to perform feature mapping on the global fusion vector to obtain network prediction values, and also includes:

[0087] Step 1001: Input the three-modal fusion features into the two-layer fully connected network and perform several forward propagations to obtain multiple network prediction values;

[0088] Step 1002: Calculate the prediction mean, prediction entropy, and temporal consistency score of the network prediction values;

[0089] Step 1003: When the predicted entropy is greater than the entropy threshold or the temporal consistency score is less than the consistency threshold, the network prediction value is marked as low confidence.

[0090] refer to Figure 5 The three-modal fused features are input into an improved Transformer network for computation and feature compression to obtain a global fused vector, including:

[0091] Step 901: Perform spatial self-attention calculation on the three-modal fusion features to obtain spatial correlation features;

[0092] Step 902: Extend the third temporal-spatial interaction feature, and concatenate the extended third temporal-spatial interaction feature with the spatial sequence of the spatial association feature to obtain a spatiotemporal sequence;

[0093] Step 903: Perform joint self-attention calculation on the spatiotemporal sequence to obtain global correlation features;

[0094] Step 904: Use a multilayer perceptron to compress the global correlation features to obtain the global fusion vector.

[0095] Specifically, multimodal data is acquired synchronously. The system first performs multimodal data acquisition synchronously on the urological postoperative wound area to provide raw data support for subsequent evaluation and ensure data consistency to guarantee the accuracy of subsequent processing: Microwave imaging data acquisition: A microwave imaging device is used to scan the wound area to obtain images of the dielectric properties distribution of the wound surface and subcutaneous tissue. It can capture subtle features such as skin edges and irregularities and can penetrate dressings without removal. Electrical impedance imaging data acquisition: An array of Ag / AgCl electrodes is arranged around the wound area. An alternating current is injected through a constant current source to measure the voltage difference between different electrodes to reconstruct the tissue conductivity distribution image. Bioelectrical impedance data acquisition: A four-electrode configuration is used to continuously acquire data at fixed intervals, measuring the resistance, reactance, and phase angle parameters of the wound area to form a time-series dataset that reflects the dynamic changes in extracellular fluid content, cell membrane integrity, and cell density.

[0096] Preferably, multimodal data preprocessing involves standardizing the collected raw data to eliminate noise, spatial misalignment, and dimensional differences, outputting high-quality, spatially aligned standardized data to provide reliable input for the neural network model. Each step is progressive and interconnected.

[0097] 1) Data cleaning and noise suppression:

[0098] Microwave images: Adaptive Wiener filtering is used to suppress Gaussian noise and motion artifacts. The formula is as follows:

[0099]

[0100] in, These are the filtered pixel values. These are the original pixel values. For the signal power spectral density, This represents the noise power spectral density.

[0101] Subsequent 3×3 median filtering further eliminates salt-and-pepper noise while preserving edge features.

[0102] 2) Electrical impedance image: Tikhonov regularization was used to optimize reconstruction accuracy.

[0103] 3) Bioelectrical impedance data: Outliers are removed using the Grubbs criterion, and then the time series data is smoothed by moving average filtering to finally form a time series.

[0104] Furthermore, spatial registration and alignment: The CLAHE algorithm was used to enhance the local contrast of the microwave and electrical impedance images, highlighting anatomical features such as wound edges, drainage tube interfaces, and skin texture; feature points were extracted using the Scale Invariant Feature Transform (SIFT) algorithm.

[0105] 1) Scale-space extremum detection: Locate potential feature points at different scales using the difference of Gaussian pyramid, and filter out unstable points with low contrast and edge response;

[0106] 2) Direction assignment: Calculate the gradient direction histogram of the neighborhood of the feature point, and take the main direction and the secondary direction within 180° away to ensure the rotation invariance of the feature point;

[0107] 3) Descriptor generation: The neighborhood of the feature point is divided into 4×4 sub-regions, and an 8-directional gradient histogram is generated for each sub-region to obtain the descriptor.

[0108] Feature point similarity is calculated using Euclidean distance, matching point pairs are selected, and false matches are eliminated using the RANSAC algorithm. Based on the matching point pairs, the affine transformation matrix is ​​solved using the least squares method to align the impedance image with the reference image. The transformation formula is as follows:

[0109]

[0110] in, Let be the affine transformation matrix. , These are the transformed x and y coordinates. , These are the original x-coordinates and original y-coordinates.

[0111] The impedance image is transformed using the transformation formula obtained from the solution.

[0112] Optionally, data standardization and normalization are performed as follows: Image data: Linear normalization is used to map microwave images and bioelectrical impedance images to unit intervals; microwave images are further optimized for contrast using gamma correction. Bioelectrical impedance data: Z-score standardization is used to eliminate the influence of dimensions; the standardized data have a mean of 0 and a standard deviation of 1, while retaining the time-series dimension.

[0113] Specifically, improved neural network feature extraction and fusion. 1) The improved U-Net network includes: an input layer, an encoder, a bottleneck layer, a decoder, and an output layer connected in sequence; the decoder embeds a spatial attention gate (SAG). The SAG structure is as follows:

[0114]

[0115] in, For spatial attention gate output, For network input, Use the Sigmoid activation function; It is a 1×1 convolution; , These are max pooling and average pooling, respectively. for .

[0116] The input preprocessed microwave image is downsampled at 4 levels by the encoder, the bottleneck layer further convolves the highest dimension feature, the decoder performs upsampling at 4 levels, and the output layer is compressed through a 1×1 convolution channel to obtain the microwave image spatial feature map.

[0117] 2) The improved ResNet-34 network comprises: an input layer, an initial convolutional layer, a max-pooling layer, a residual block group, a global average pooling layer, and an output layer connected in sequence; each residual block in the residual block group embeds two frequency-adaptive convolutions; the expression for the frequency-adaptive convolution (FAC) is:

[0118] ;in, It is a dynamic kernel; Based on the convolutional kernel; Frequency-shifted convolution kernel; For the first Each frequency adaptive weight; This represents the number of convolution kernels.

[0119] The preprocessed electrical impedance image is input, and its dimensionality is reduced and basic features are extracted by the initial convolutional layer. The max pooling layer further compresses the spatial size and focuses on the core region features. Each group of residual blocks is dynamically adapted to the current frequency through FAC to capture multi-frequency conductivity features. The global average pooling layer compresses the spatial features to a 1×1 dimension. The output layer adjusts the number of channels through 1×1 convolution to obtain the spatial feature map of the electrical impedance image.

[0120] 3) The improved GRU network consists of an input layer, a bidirectional GRU layer, a temporal attention layer, a feature fusion layer, and an output layer connected sequentially. The input is a bioelectrical impedance temporal sequence. The forward GRU extracts inflammation development features from the first time step to the final time step, and the backward GRU extracts repair regression features from the final time step to the first time step, outputting a bidirectional hidden state matrix. Attention scores are calculated for the bidirectional hidden states at each time step, and weights are obtained through Softmax. All bidirectional hidden states are weighted and summed according to their attention weights to obtain the temporal feature vector. The dimensions are adjusted through a fully connected layer, outputting the bioelectrical impedance temporal feature vector.

[0121] Furthermore, the three-modal cross-feature interaction comprehensively correlates the three-modal features of microwave (M), electrical impedance (E), and bioelectrical impedance (B):

[0122] 1) Spatial feature flattening and alignment:

[0123] Microwave and electrical impedance features are flattened into sequence features (preserving spatial location information); to adapt to temporal-spatial interaction, the bioelectrical impedance temporal features are expanded into a temporal-spatial correlation dimension.

[0124] 2) Spatial interaction between microwave and electrical impedance:

[0125] Calculate the attention weights for microwave-to-resistance separately. Attention weights from impedance to microwave :

[0126]

[0127]

[0128] in, ; ; ; ; It is the softmax function; For query matrix; The key matrix; For attention interaction dimension; Indicates microwave; Indicates electrical impedance; This is the weight matrix.

[0129] Output Space - Spatial Interaction Features:

[0130]

[0131]

[0132] in, , These are the first interaction feature and the second interaction feature, respectively; , These are the microwave imaging features and the electrical impedance imaging features after being flattened, respectively. , These are the first-value matrix and the second-value matrix, respectively.

[0133] 3) Temporal-spatial interaction of bioelectrical impedance:

[0134] Design temporal-spatial attention to establish the correlation between the temporal dynamics of bioelectrical impedance and the spatial region of the wound: calculate the attention weight of temporal features for each spatial location. :

[0135]

[0136] in, , These represent temporal and spatial aspects, respectively.

[0137] Calculate the feedback attention weights of spatial features on temporal features:

[0138]

[0139] Output temporal-spatial interaction features:

[0140]

[0141]

[0142]

[0143] in, , , These are the first temporal-spatial interaction feature, the second temporal-spatial interaction feature, and the third temporal-spatial interaction feature, respectively. , These are the third-valued matrix and the fourth-valued matrix, respectively. For the extended bioelectrical impedance characteristics; This indicates mean pooling.

[0144] 4) Integration of three-modal interaction features: unify the dimensions of the interactive features to form the basic features for three-modal fusion.

[0145] Preferably, the temporal-spatial global fusion module improves the original spatiotemporal Transformer (STT) into a three-modal spatiotemporal fusion Transformer, which simultaneously handles spatial location association, temporal dynamic association, and intermodal association: Input layer: accesses three-modal interaction features and temporal features.

[0146] Spatial self-attention layer: Apply self-attention to the trimodal interaction features to capture the correlation between different spatial regions of the wound (such as the trend of inflammation spreading from the center to the edge) and output spatial correlation features.

[0147] Temporal-spatial joint self-attention layer (new): Extends temporal features and splices them with spatially related features into a spatiotemporal sequence, applies joint self-attention, and captures the global correlation between spatial regional changes and temporal stages.

[0148] Feature compression and output: The features output by joint self-attention are compressed into a global fusion vector through a multilayer perceptron (MLP), which contains spatial, temporal and correlation information of the three modes.

[0149] Furthermore, a two-layer fully connected network is connected after the global fusion vector to perform several forward propagations on the same input (randomly shutting down some neurons in each propagation) to obtain multiple prediction results. The core evaluation metrics are calculated as: prediction mean, prediction entropy, and temporal consistency score. The credibility determination rule is: when the prediction entropy is greater than the entropy threshold or the temporal consistency score is less than the consistency threshold, the result is marked as low credibility, triggering manual review.

[0150] As an optional implementation, this embodiment also provides an electronic device, including: at least one processor, and a memory communicatively connected to the processor; wherein the memory stores instructions executable by the processor, the instructions being executed by the processor to enable the processor to perform the aforementioned artificial intelligence-based urological postoperative wound rehabilitation assessment method.

[0151] As an optional implementation, this embodiment also provides a non-transitory computer-readable storage medium storing computer instructions for causing a computer to execute the aforementioned artificial intelligence-based postoperative wound rehabilitation assessment method for urological surgery.

[0152] The beneficial effects of this invention are as follows:

[0153] This invention utilizes objective data from three modalities—microwave imaging, electrical impedance imaging, and bioelectrical impedance—in a non-invasive and interference-free manner throughout the data acquisition process, reducing errors caused by individual differences. Through a three-branch feature extraction network and a deep fusion process, it adapts to the characteristics of different types of data, achieves cross-modal information complementarity, and captures key signals for multi-dimensional wound recovery.

[0154] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other.

[0155] This document uses specific examples to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of the present invention. Furthermore, those skilled in the art will recognize that, based on the ideas of the present invention, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A method for assessing postoperative wound rehabilitation in urology based on artificial intelligence, characterized in that, include: Collect raw microwave imaging data, raw electrical impedance imaging data, and raw bioelectrical impedance data from the target postoperative patient; The microwave imaging data is subjected to adaptive Wiener filtering and median filtering in sequence to obtain preprocessed microwave imaging data. The electrical impedance imaging data is subjected to Tikhonov regularization to obtain preprocessed electrical impedance imaging data. The bioelectrical impedance data is filtered using Grubbs criterion and moving average filtering method to obtain preprocessed bioelectrical impedance data. The preprocessed microwave imaging data and the preprocessed electrical impedance imaging data were spatially aligned using the SIFT algorithm and the least squares method. The preprocessed microwave imaging data, the preprocessed electrical impedance imaging data, and the preprocessed bioelectrical impedance data are normalized. The preprocessed microwave imaging data is encoded using an improved U-Net network to obtain microwave imaging features. The preprocessed electrical impedance imaging data is encoded using an improved ResNet-34 network to obtain electrical impedance imaging features. The preprocessed bioelectrical impedance data is encoded using an improved GRU network to obtain bioelectrical impedance features. The microwave imaging features, the electrical impedance imaging features, and the bioelectrical impedance features are interactively processed and integrated to obtain three-modal fusion features; The three-modal fused features are input into an improved Transformer network for computation and feature compression to obtain a global fused vector; The global fusion vector is feature-mapped using a fully connected network to obtain the network prediction value; The wound rehabilitation assessment result is obtained by matching the level corresponding to the network prediction value with the preset rehabilitation assessment rules.

2. The method for assessing postoperative wound rehabilitation in urology based on artificial intelligence according to claim 1, characterized in that, Spatial alignment of the preprocessed microwave imaging data and the preprocessed electrical impedance imaging data is performed using the SIFT algorithm and the least squares method, including: The CLAHE algorithm is used to enhance the local contrast of the preprocessed microwave imaging data and the preprocessed electrical impedance imaging data. The SIFT algorithm is used to perform scale-space extremum detection, orientation assignment, and descriptor generation on the enhanced preprocessed microwave imaging data and the preprocessed electrical impedance imaging data, respectively, to obtain image descriptors. The image descriptors corresponding to the preprocessed microwave imaging data and the preprocessed electrical impedance imaging data are matched and filtered according to Euclidean distance and RANSAC algorithm to obtain matching pairs. The least squares method is used to solve the transformation matrix of the matching pair to obtain the alignment matrix, and the alignment matrix is ​​used to spatially align the preprocessed microwave imaging data and the preprocessed electrical impedance imaging data.

3. The method for assessing postoperative wound rehabilitation in urology based on artificial intelligence according to claim 1, characterized in that, The improved U-Net network includes: an input layer, an encoder, a bottleneck layer, a decoder, and an output layer connected in sequence; the decoder embeds a spatial attention gate.

4. The method for assessing postoperative wound rehabilitation in urology based on artificial intelligence according to claim 1, characterized in that, The improved ResNet-34 network comprises: an input layer, an initial convolutional layer, a max-pooling layer, a residual block group, a global average pooling layer, and an output layer, connected in sequence; each residual block in the residual block group embeds two frequency-adaptive convolutions; the expression for the frequency-adaptive convolution is: ;in, It is a dynamic kernel; Based on the convolutional kernel; Frequency-shifted convolution kernel; For the first Each frequency adaptive weight; This represents the number of convolution kernels.

5. The method for assessing postoperative wound rehabilitation in urology based on artificial intelligence according to claim 1, characterized in that, The improved GRU network comprises: an input layer, a bidirectional GRU layer, a temporal attention layer, a feature fusion layer, and an output layer connected in sequence.

6. The method for assessing postoperative wound rehabilitation in urology based on artificial intelligence according to claim 1, characterized in that, Interactive processing and feature integration are performed on the microwave imaging features, the electrical impedance imaging features, and the bioelectrical impedance features to obtain trimodal fusion features, including: The microwave imaging features and the electrical impedance imaging features are flattened, and the bioelectrical impedance features are expanded. A bidirectional cross-attention calculation is performed on the flattened microwave imaging features and the electrical impedance imaging features to obtain a first interactive feature and a second interactive feature; the expression for the first interactive feature is: The expression for the second interactive feature is: ;in, , These are the first interaction feature and the second interaction feature, respectively; , These are the microwave imaging features and the electrical impedance imaging features after being flattened, respectively. Attention weights for microwave-to-resistivity; Attention weights for impedance to microwave; , These are the first-value matrix and the second-value matrix, respectively. Temporal-spatial attention is performed on the first interaction feature, the second interaction feature, and the extended bioelectrical impedance feature to obtain the first temporal-spatial interaction feature, the second temporal-spatial interaction feature, and the third temporal-spatial interaction feature; the expression of the first temporal-spatial interaction feature is: The expression for the second temporal-spatial interaction feature is: The expression for the third temporal-spatial interaction feature is: ;in, , , These are the first temporal-spatial interaction feature, the second temporal-spatial interaction feature, and the third temporal-spatial interaction feature, respectively. , These are the attention weights of temporal features on spatial location and the feedback attention weights of spatial features on temporal features, respectively. , These are the third-valued matrix and the fourth-valued matrix, respectively. For the extended bioelectrical impedance characteristics; This indicates mean pooling. The first temporal-spatial interaction feature, the second temporal-spatial interaction feature, and the third temporal-spatial interaction feature are aligned and spliced ​​to obtain the trimodal fusion feature.

7. The method for assessing postoperative wound rehabilitation in urology based on artificial intelligence according to claim 1, characterized in that, The method further includes using a fully connected network to perform feature mapping on the global fusion vector to obtain network prediction values, and also includes: The three-modal fusion features are input into the two-layer fully connected network and forward propagated several times to obtain multiple network prediction values; The prediction mean, prediction entropy, and temporal consistency score of the network predictions are statistically analyzed. When the predicted entropy is greater than the entropy threshold or the temporal consistency score is less than the consistency threshold, the network predicted value is marked as low confidence.

8. The method for assessing postoperative wound rehabilitation in urology based on artificial intelligence according to claim 6, characterized in that, The three-modal fused features are input into an improved Transformer network for computation and feature compression to obtain a global fused vector, including: Spatial self-attention calculation is performed on the three-modal fusion features to obtain spatial correlation features; The third temporal-spatial interaction feature is extended, and the spatial sequence of the extended third temporal-spatial interaction feature is spliced ​​with the spatial association feature to obtain a spatiotemporal sequence. Joint self-attention calculation is performed on the spatiotemporal sequence to obtain global correlation features; The global correlation features are compressed using a multilayer perceptron to obtain the global fusion vector.

9. An electronic device, characterized in that, include: At least one processor, and a memory communicatively connected to the processor; wherein the memory stores instructions executable by the processor, the instructions being executed by the processor to enable the processor to perform an artificial intelligence-based method for assessing postoperative wound rehabilitation in urology, as described in any one of claims 1 to 8.

10. A non-transitory computer-readable storage medium storing computer instructions, characterized in that, The computer instructions are used to cause the computer to execute any one of claims 1 to 8, a method for assessing postoperative wound rehabilitation in urology based on artificial intelligence.

Images