Bioelectrical impedance tumor detection method based on multiple features

Abstract

The invention discloses a bioelectrical impedance tumor detection method based on multiple features, and the method comprises the following steps: S1, obtaining bioelectrical impedance data of a tumor tissue and a normal tissue sample, the bioelectrical impedance data comprising electrical impedance values and amplitudes under different frequencies; s2, performing data cleaning and normalization processing on the electrical impedance values and the amplitudes, constructing a two-dimensional feature matrix, and converting the two-dimensional feature matrix into two-dimensional image features; s3, constructing a multi-feature set containing the original electrical impedance value, the original amplitude and the two-dimensional image features; and S4, inputting the multi-feature set into a multi-feature fusion detection network, and outputting a tumor detection result. The bioelectrical impedance data is converted into the two-dimensional image features, the two-dimensional image features are fused with the original electrical signal features to form the multi-feature set, and the multi-features are dynamically weighted and fused through the two-channel network of the multi-feature fusion detection network in combination with the attention mechanism, so that the accuracy and robustness of tumor detection are effectively improved.

Description

Technical Field

[0001] This invention relates to the field of tumor detection technology, and in particular to a multi-feature-based bioelectrical impedance analysis method for tumor detection. Background Technology

[0002] Malignant tumors are the leading cause of death worldwide, with their incidence rate increasing annually. The 5-year survival rate for late-stage patients is generally below 20%, while early diagnosis can improve treatment effectiveness by more than 50%. Therefore, achieving early and accurate tumor detection is a key breakthrough in reducing mortality. Currently used clinical tumor detection technologies have significant limitations. Bioelectrical impedance analysis (BIA) technology offers an innovative approach to tumor detection: due to abnormal cell proliferation, altered cell membrane permeability, and differences in vascular distribution, tumor tissue exhibits significant differences in impedance, conductivity, and phase angle compared to normal tissue over a wide frequency range (10Hz-10MHz).

[0003] Chinese patent CN120105251B discloses "a bioelectrical impedance tumor detection method based on pattern recognition". By acquiring current, voltage and impedance values ​​at different depths and frequencies, key features are extracted from the impedance data at different frequencies of each channel. The extracted feature data is input into a multi-level feature fusion module to form new fused features. The fused features are then input into an adaptive classification module for classification.

[0004] However, the above methods rely only on single-dimensional information such as images or biochemical indicators, which cannot reflect the correlation between the electrophysiological characteristics and structural features of tumors. Other bioelectrical impedance-based studies are limited to fixed-frequency point analysis, ignoring the dynamic changes in the electrical properties of tumor tissue under different physiological states, resulting in low detection accuracy. Therefore, it is urgent to propose a multi-feature bioelectrical impedance tumor detection method to solve the problems of the existing technologies. Summary of the Invention

[0005] To address the aforementioned deficiencies or improvement needs of existing technologies, this invention provides a multi-feature-based bioelectrical impedance analysis (BIA) method for tumor detection. By converting bioelectrical impedance data into two-dimensional image features and fusing them with the original electrical signal features to form a multi-feature set, and by combining a dual-channel network of the multi-feature fusion detection network with an attention mechanism, the multi-features are dynamically weighted and fused, effectively improving the accuracy and robustness of tumor detection.

[0006] To achieve the above objectives, according to one aspect of the present invention, a method for detecting tumor activity based on multi-feature bioelectrical impedance analysis is provided, the method comprising the following steps: S1: Obtain raw bioelectrical impedance data of tumor tissue and normal tissue samples, wherein the raw bioelectrical impedance data includes raw impedance values ​​and raw amplitudes at different frequencies; S2: The obtained original impedance value and original amplitude are cleaned, the cleaned impedance value and amplitude are normalized, a two-dimensional feature matrix is ​​constructed based on the normalized impedance value and amplitude, and the two-dimensional feature matrix is ​​converted into two-dimensional image features. S3: Perform feature alignment and dimension unification on the original impedance value and original amplitude, preprocess the two-dimensional image features to match the sample dimension of the original impedance value and original amplitude; establish a mapping relationship table of original impedance value, original amplitude and two-dimensional image features based on the unique identifier of the sample, construct a multi-feature matrix, standardize and verify the multi-feature matrix to form the final multi-feature set; S4: Construct a multi-feature fusion detection network model and train and optimize it. Input the multi-feature set into the trained multi-feature fusion detection network model and output the tumor detection result. The multi-feature fusion detection network model includes an electrical impedance feature processing module, an amplitude processing module, an image feature processing module, a multi-feature fusion module, and an output layer.

[0007] As an embodiment of this application, the data cleaning of the obtained original impedance value and original amplitude in step S2 specifically includes: S211: Adopted The criteria identify outlier data points. For each frequency point, based on the raw impedance and raw amplitude, the mean of all samples at that frequency point is calculated. and standard deviation The calculation formula is as follows:

[0008] in, The total number of samples, For the first The value of each sample; when the data points satisfy When this happens, it is considered an outlier; S212: Replace the identified outliers with the sample mean of the corresponding frequency points. The calculation formula is as follows:

[0009] in, This is the mean of all normal samples at this frequency point; The replaced value; S213: By scanning the impedance and amplitude data of each sample, identify missing frequency points and calculate the missing rate of each sample. If the sample missing rate is <1%, use linear interpolation to fill in the missing data. The calculation formula is as follows:

[0010] in, For missing frequency points, Missing frequency points to be filled The corresponding impedance value or amplitude, for The previous effective frequency point, Effective frequency point The corresponding impedance value or amplitude, for The next effective frequency point, Effective frequency point The corresponding impedance value or amplitude; If the sample missing rate is ≥1%, it is considered an invalid sample and needs to be collected again. S214: Gaussian filtering is used to smooth the impedance and amplitude data, removing high-frequency noise. The calculation formula is as follows:

[0011] in, For smoothing the first The impedance or amplitude at each frequency point The filter window half-width, The index of the data points within the window relative to the center. For the Gaussian function at the index The weight values ​​at each point are used for the raw data within the weighted average window. For the first in the original data The Gaussian function formula for the impedance or amplitude at a given frequency point is as follows:

[0012] in, The standard deviation of the Gaussian function is used to control the rate of weight decay. The index of the data points within the window relative to the center. It is a natural constant.

[0013] As an embodiment of this application, the normalization process of the impedance value and the original amplitude after data cleaning in step S2 specifically includes: S221: For the impedance value and amplitude after data cleaning, iterate through every frequency point of all samples. The formula for calculating the impedance value is as follows:

[0014]

[0015] in, The total number of samples, The total number of frequency points. Indicates the first The sample at the th The impedance value at each frequency point This is the global minimum impedance value. This is the global maximum impedance value; The formula for calculating amplitude is as follows:

[0016]

[0017] in, The total number of samples, The total number of frequency points. Indicates the first The sample at the th The amplitude at each frequency point The global minimum amplitude, This is the global maximum amplitude. S222: Normalize the impedance and amplitude of each sample at each frequency point. The normalization formula for the impedance is as follows:

[0018] in, For the first The first sample Normalized impedance values ​​at each frequency point Indicates the first The sample at the th The impedance value at each frequency point This is the global minimum impedance value. This is the global maximum impedance value; The amplitude normalization formula is as follows:

[0019] in, For the first The first sample Normalized amplitude at each frequency point Indicates the first The sample at the th The amplitude at each frequency point The global minimum amplitude, This is the global maximum amplitude. S223: When the maximum value of the impedance or amplitude parameter is equal to the minimum value in the global range, and the parameter is indistinguishable in all samples, its normalized value is uniformly set to 0.5.

[0020] As an embodiment of this application, the construction of the two-dimensional feature matrix in step S2 includes: S231: Create a size of Two-dimensional feature matrix ,in The total number of frequency points, matrix elements row index Frequency index corresponding to the impedance value, column index The frequency index corresponding to the amplitude, where ; S232: Calculate matrix elements using cross-product method. The formula is as follows:

[0021] in, For the first Normalized impedance values ​​at each frequency point For the first Normalized amplitude at each frequency point; For the first Normalized impedance values ​​at each frequency point For the first Normalized amplitude at each frequency point; S233: For the constructed two-dimensional feature matrix Symmetry verification and outlier correction are performed, and the asymmetry index is calculated using the following formula:

[0022] in, It is an index of the asymmetry of a two-dimensional feature matrix. matrix elements , The absolute difference, This indicates that only the upper triangular part of the two-dimensional feature matrix is ​​calculated. This represents the sum of the absolute values ​​of all elements in the two-dimensional feature matrix; if the asymmetry index exceeds the threshold of 0.01, then the asymmetry elements are averaged for correction; if the detection matrix exceeds... The outlier element of the criterion is replaced by the median of the row and column containing that element; S234: Adopted The function performs non-linear contrast enhancement on the corrected matrix to highlight feature differences, as shown in the following formula:

[0023] in, For the original matrix elements, For the enhanced matrix elements, The slope parameter is used to control the reinforcement strength. For center offset parameters, It is a natural constant.

[0024] As an embodiment of this application, step S2, which converts the two-dimensional feature matrix into two-dimensional image features, specifically includes: S241: Transform the two-dimensional feature matrix Linear mapping to grayscale image grayscale image The pixel value range is [0, 255], and the formula is as follows:

[0025] in, This indicates rounding to the nearest integer. grayscale image median coordinate The pixel values ​​at a given location are mapped to bright pixels in the two-dimensional feature matrix, with high-value regions being mapped to bright pixels and low-value regions being mapped to dark pixels, thus forming the initial grayscale contrast. S242: For grayscale images Perform histogram equalization to enhance the overall contrast of the image. The formula is as follows:

[0026] in, grayscale value In image grayscale image The probability distribution in This is the grayscale image after equalization; S243: Adopted Color mapping schemes will equalize the grayscale image. Convert to heatmap The color mapping function formula is as follows:

[0027] in, for Color mapping function, output Triples For indexing at frequency points Heatmaps of locations are used to enhance feature recognition through color coding; S244: Using bilinear interpolation for heatmaps Spatial standardization is performed using the following formula:

[0028] in, The coordinates of the four nearest pixels in the original image. The weights of the corresponding coordinate points, The image is spatially normalized.

[0029] As an embodiment of this application, step S3, which involves feature alignment and dimensional unification of the original impedance value and the original amplitude, specifically includes: S311: Convert the original impedance value vector Convert to electrical impedance matrix ,in, The number of frequency points sampled; The original magnitude vector Convert to magnitude matrix , This represents the number of frequency points sampled. S312: Linear interpolation is used to transform the impedance matrix. and magnitude matrix Time series alignment is performed by resampling one feature matrix to the frequency points of the other feature matrix, as shown in the following formula:

[0030] in, These are the target frequency points that need to be aligned. It is less than the original amplitude data The maximum frequency point, It is greater than the original amplitude data The minimum frequency point, The original amplitude at the frequency point The value at that location, The original amplitude at the frequency point The value at that location, For the target frequency points that need to be aligned The amplitude after resampling at the location; S313: The impedance matrix after time series alignment and magnitude matrix Standardize them separately, including the electrical reactance matrix. The standardized formula is as follows:

[0031] in, It is the first normalized electrical impedance matrix. line, number Column elements, It is the first normalized impedance matrix. line, number Column elements, It is the global mean of the impedance values ​​of all samples at all frequency points. The global standard deviation of the impedance values ​​of all samples at all frequency points; Amplitude matrix The standardized formula is as follows:

[0032] in, It is the first normalized amplitude matrix. line, number Column elements, It is the first normalized magnitude matrix. line, number Column elements, It is the global mean of the amplitude of all samples across all frequency points. The global standard deviation of the amplitude of all samples at all frequency points; S314: Using the Pearson correlation coefficient to standardize the electrical impedance matrix and magnitude matrix Feature selection is performed, retaining frequency features with high correlation, including the electrical impedance matrix. The feature selection formula is as follows:

[0033] in, It is the first The impedance values ​​at each frequency point and the Pearson correlation coefficient of the sample labels, For the sample size, It is the first The sample at the th Standardized impedance values ​​at each frequency point For all samples in the first The mean of the standardized electrical impedance values ​​at each frequency point For the first Sample labels for each sample; Amplitude matrix The feature selection formula is as follows:

[0034] in, It is the first The amplitude at each frequency point and the Pearson correlation coefficient of the sample label, For the sample size, It is the first The sample at the th Standardized amplitude at each frequency point For all samples in the first The mean of the standardized amplitude at each frequency point For the first Sample labels for each sample.

[0035] As an embodiment of this application, the preprocessing of the two-dimensional image features in step S3 to match the sample dimensions of the original impedance value and the original amplitude specifically includes: S321: A pre-trained convolutional neural network is used to extract deep features from two-dimensional image features to obtain a high-dimensional image feature vector, which is then input into the heatmap. The formula is as follows:

[0036] in, For high-dimensional image feature vectors, Represents the feature extraction function; S322: Principal component analysis is used to reduce the dimensionality of the extracted high-dimensional image feature vectors. A linear transformation is then used to project the high-dimensional features into a low-dimensional space, as shown in the following formula:

[0037] in, For high-dimensional image feature vectors, Principal component analysis projection matrix, These are the eigenvectors after dimensionality reduction; S323: Standardize the feature vectors of the dimension-reduced image to make them consistent with the numerical distribution of the original electrical signal features. The standardization formula is as follows:

[0038] in, It is the standardized image feature vector. It is the image feature vector after dimensionality reduction. This is the mean vector of image features after dimensionality reduction for all samples. This is the corresponding standard deviation vector.

[0039] As an embodiment of this application, step S3, which establishes a mapping table of original impedance value, original amplitude, and two-dimensional image features based on the unique identifier of the sample, and constructs a multi-feature matrix, specifically includes: S331: Based on the unique identifier of the sample, establish a mapping table of the original electrical impedance value, the original amplitude, and the two-dimensional image features. For each sample, concatenate the electrical impedance feature, the amplitude feature, and the image feature along the feature dimension to form a multi-feature vector. The concatenation formula is as follows:

[0040] in, Indicates the characteristics of electrical impedance value, Indicates amplitude characteristics, Represents the image feature vector. For a single sample with multiple feature vectors, Indicates concatenation by column; S332: Stack the multi-feature vectors of all samples row-wise to construct a multi-feature matrix, as shown in the following formula:

[0041] in, It is a multi-feature matrix. The total number of samples, For the first Multiple feature vectors of a sample.

[0042] As an embodiment of this application, step S3, which standardizes and verifies the multi-feature matrix to form the final multi-feature set, specifically includes: S341: For the multi-feature matrix Each feature dimension is standardized to unify the numerical distribution of all feature dimensions, as shown in the following formula:

[0043] in, Represents each feature dimension in the matrix. Indicates the standardized first The sample, the first The numerical values ​​of the dimensional features, Indicates the first before standardization The sample, the first The original numerical values ​​of the dimensional features. Indicates the first The global mean of the dimensional feature across all samples. Indicates the first The global standard deviation of the dimensional feature across all samples; S342: Calculate the cosine similarity between the electrical signal features and image features of each sample. The calculation formula is as follows:

[0044] in, Indicates the first Similarity between features of each sample Indicates the first The electrical impedance screening feature vector of each sample, Indicates the first The standardized image feature vector of each sample, Represents the vector dot product. express Norm, when Samples with low correlation between features are removed. S343: Calculate the Mahalanobis distance between the multiple feature vectors of each sample and the global mean vector. The calculation formula is as follows:

[0045] in, Indicates the first Mahalanobis distance of each sample Indicates the first Multiple feature vectors of a sample This represents the multi-feature mean vector of all samples. It is the inverse of the covariance matrix with multiple features; S344: Divide the validated feature set into training, validation and test sets in a ratio of 7:1.5:1.5.

[0046] As an embodiment of this application, step S4 specifically includes: S41: A multilayer perceptron is used to extract high-order features of impedance values ​​and amplitudes, and the filtered impedance features are then... The input is fed into the electrical impedance characteristic processing module, where it is processed through three fully connected layers. The calculation formula is as follows:

[0047]

[0048]

[0049] in, This represents the feature vector output by the first fully connected layer. The activation function is used to introduce a nonlinear transformation. This represents the weight matrix of the first fully connected layer. This represents the bias vector of the first fully connected layer, used to adjust the baseline of the linear transformation. This represents the feature vector output by the second fully connected layer. This represents the weight matrix of the second fully connected layer. This represents the bias vector of the second fully connected layer. This represents the feature vector output by the third fully connected layer. This represents the weight matrix of the third fully connected layer. This represents the bias vector of the third fully connected layer; Selected amplitude features The input is fed into the amplitude feature processing module, where it undergoes the same operations as the impedance feature processing module, ultimately outputting a feature vector. ; Image features The image is input to the image feature processing module and processed sequentially through three convolutional blocks of different dimensions. The calculation formula is as follows:

[0050]

[0051]

[0052] in, Indicates the first dimension is The feature map output by the convolutional block, This represents the bias vector of the first convolutional layer. For activation function, This is a max pooling operation used to reduce the size of the feature map while retaining key information; Indicates the second dimension is The feature map output by the convolutional block, This represents the bias vector of the second convolutional layer. Indicates the third dimension as The feature map output by the convolutional block, This represents the bias vector of the third convolutional layer. For flattening operation; S42: Outputs from the impedance characteristic processing module, amplitude processing module, and image feature processing module respectively. , , The input is fed into the feature fusion module, where each feature is mapped to a scalar attention level through a linear transformation. The normalization formula is as follows:

[0053]

[0054]

[0055] in, , , These represent the attention scalars for electrical impedance, amplitude, and image features, respectively. For linear layers, , , For attention weights, Represents an exponential function; After linear transformation , , The final fusion feature is obtained by weighted summation, and the calculation formula is as follows:

[0056] in, The fusion feature is obtained by weighted summation; S43: Integrate features Input is fed into the output layer to calculate the tumor probability. The formula is as follows:

[0057] in, This represents the activation function. The weight matrix represents the classification head. The scalar represents the bias of the classification head.

[0058] The beneficial effects of this invention are as follows: (1) By cleaning and normalizing the impedance and amplitude data, this invention can eliminate the interference of measurement noise and individual differences on the data and ensure signal stability. By converting the pre-processed impedance and amplitude into two-dimensional image features, the frequency correlation characteristics of one-dimensional electrical signals can be transformed into visualized spatial features. The two-dimensional structure constructed by matrix cross-product can intuitively present the coordinated change law of impedance and amplitude at different frequency points. Compared with single-dimensional signals, it is easier to be captured by visual feature extraction networks, which solves the problem of single feature dimension and weak correlation of traditional electrical signals.

[0059] (2) This invention establishes a mapping relationship table of original electrical impedance value, original amplitude and two-dimensional image features based on the unique identifier of the sample, constructs a multi-feature matrix, standardizes and verifies the multi-feature matrix, and forms the final multi-feature set. The original electrical impedance value and amplitude retain the time-series frequency characteristics of the electrical signal, while the two-dimensional image features supplement the spatial correlation information between frequencies. The two achieve information complementarity. The multi-feature set integrates the frequency response characteristics of electrophysiological signals and the spatial texture characteristics of images, makes up for the limitations of single features, and significantly improves the information density and discriminability of features.

[0060] (3) This invention adapts the feature differences between electrical signals and image features through the dual-channel structure of the multi-feature fusion detection network model. It extracts the frequency domain features of the original electrical signal through a multilayer perceptron, captures the spatial domain features of the two-dimensional image through a convolutional neural network, and then achieves weighted fusion across features through an attention mechanism. This fusion method can fully explore the correlation between the frequency difference of the impedance value, the attenuation law of the amplitude, and the spatial distribution features of the image, avoid the learning bias of the traditional single-feature model for complex tumor features, and significantly improve the ability to identify early small tumors or atypical tumors.

[0061] (4) This invention acquires multi-frequency bioelectrical impedance data, preprocesses it into two-dimensional image features, constructs a multi-feature set, and inputs it into a multi-feature fusion detection network model, thereby realizing the collaborative analysis of the temporal features of electrical signals and the spatial features of images. This method enhances feature richness through the complementarity of multi-feature information, strengthens cross-feature association learning with the help of a fusion network, and effectively utilizes the quantitative characteristics of electrical signals and the qualitative features of images, especially significantly improving the detection sensitivity of early small tumors. Attached Figure Description

[0062] Figure 1 This is a schematic flowchart of a multi-feature-based bioelectrical impedance antitumor detection method provided in an embodiment of the present invention; Figure 2 This is a schematic diagram of the two-dimensional image feature conversion process of a multi-feature bioelectrical impedance antitumor detection method provided in an embodiment of the present invention; Figure 3 This is a schematic diagram illustrating the construction of a multi-feature set for a bioelectrical impedance analysis method for tumor detection based on multiple features, provided in an embodiment of the present invention. Figure 4 This is a schematic diagram of a multi-feature fusion detection network model structure for a multi-feature bioelectrical impedance analysis method for tumor detection provided in an embodiment of the present invention. Detailed Implementation

[0063] 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 a part of the embodiments of the present invention, and not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0064] Reference Figures 1-4 The first aspect of this invention provides a method for detecting tumor activity based on multi-feature bioelectrical impedance analysis, the method comprising the following steps: S1: Obtain raw bioelectrical impedance data from tumor tissue and normal tissue samples. The raw bioelectrical impedance data includes raw impedance values ​​and raw amplitudes at different frequencies. Specifically, this invention uses a bioelectrical impedance measurement device to collect impedance values ​​and amplitudes at different frequencies. These data can reflect abnormalities in the electrophysiological characteristics of tissues, and their impedance values ​​and amplitudes differ significantly from those of normal tissues over a wide frequency range. For example, tumor tissues often exhibit increased impedance values ​​and decreased amplitudes at high frequencies, providing raw electrical signal evidence for subsequent tumor identification.

[0065] S2: The obtained original impedance value and original amplitude are cleaned, the cleaned impedance value and amplitude are normalized, a two-dimensional feature matrix is ​​constructed based on the normalized impedance value and amplitude, and the two-dimensional feature matrix is ​​converted into two-dimensional image features. S3: Perform feature alignment and dimension unification on the original impedance value and original amplitude, preprocess the two-dimensional image features to match the sample dimension of the original impedance value and original amplitude; establish a mapping relationship table of original impedance value, original amplitude and two-dimensional image features based on the unique identifier of the sample, construct a multi-feature matrix, standardize and verify the multi-feature matrix to form the final multi-feature set; S4: Construct a multi-feature fusion detection network model and train and optimize it. Input the multi-feature set into the trained multi-feature fusion detection network model and output the tumor detection result. The multi-feature fusion detection network model includes an electrical impedance feature processing module, an amplitude processing module, an image feature processing module, a multi-feature fusion module, and an output layer.

[0066] As an embodiment of this application, the data cleaning of the obtained original impedance value and original amplitude in step S2 specifically includes: S211: Adopted The criteria identify outlier data points. For each frequency point, based on the raw impedance and raw amplitude, the mean of all samples at that frequency point is calculated. and standard deviation The calculation formula is as follows:

[0067] in, The total number of samples, For the first The value of each sample; when the data points satisfy When the value is determined to be outlier, the bioelectrical impedance data follows an approximately normal distribution. The criteria can cover 99.7% of normal data, and data points outside this range are judged as outliers; it iterates through the impedance values ​​and amplitudes of all frequency points, marks abnormal data points that exceed the threshold, and records the sample ID, frequency point index, and outlier.

[0068] S212: Replace the identified outliers with the sample mean of the corresponding frequency points. The calculation formula is as follows:

[0069] in, This is the mean of all normal samples (non-outliers) at this frequency point; For the replaced value, if the number of abnormal samples at a certain frequency point is less than or equal to 5% of the total number of samples, direct replacement is performed; if the number of abnormal samples at a certain frequency point is greater than 5% of the total number of samples, the stability of the measurement equipment at that frequency point needs to be rechecked and the corresponding samples need to be re-collected.

[0070] S213: Detect missing values ​​in impedance and amplitude data, and implement an imputation strategy based on the missing value rate. Specifically, by scanning the impedance and amplitude data of each sample, identify missing frequency points, mark them as NULL, and calculate the missing value rate for each sample (number of missing frequency points / total number of frequency points). If the sample missing value rate is <1%, use linear interpolation to imput the missing values. The calculation formula is as follows:

[0071] in, For missing frequency points, Missing frequency points to be filled The corresponding impedance value or amplitude, for The previous effective frequency point, Effective frequency point The corresponding impedance value or amplitude, for The next effective frequency point, Effective frequency point The corresponding impedance value or amplitude; If the sample missing rate is ≥1%, it is considered an invalid sample and needs to be collected again. S214: Gaussian filtering is used to smooth the impedance and amplitude data, removing high-frequency noise. The calculation formula is as follows:

[0072] in, For smoothing the first The impedance or amplitude at each frequency point, i.e., the output value. The half-width of the filter window is set to 2, meaning the total window size is [value missing]. It covers the current point and two adjacent frequency points on each side; The index of the data points within the window relative to the center. For the Gaussian function at the index The weight values ​​at each point are used for the raw data within the weighted average window. For the first in the original data The impedance or amplitude at a given frequency point, i.e., the input value, can be expressed using the Gaussian function formula as follows:

[0073] in, is the standard deviation of the Gaussian function, which is set to 1.0 in this embodiment to control the weight decay rate; The index of the data points within the window relative to the center. It is a natural constant.

[0074] Specifically, this invention provides stable basic data for subsequent normalization and image conversion by processing outliers, missing values, and smoothing noise in a hierarchical manner.

[0075] As an embodiment of this application, the normalization process of the impedance value and the original amplitude after data cleaning in step S2 specifically includes: S221: For the impedance value and amplitude after data cleaning, iterate through every frequency point of all samples. The formula for calculating the impedance value is as follows:

[0076]

[0077] in, The total number of samples, The total number of frequency points. Indicates the first The sample at the th The impedance value at each frequency point This is the global minimum impedance value. This is the global maximum impedance value; The formula for calculating amplitude is as follows:

[0078]

[0079] in, The total number of samples, The total number of frequency points. Indicates the first The sample at the th The amplitude at each frequency point The global minimum amplitude, This is the global maximum amplitude. S222: Normalize the impedance and amplitude of each sample at each frequency point. The normalization formula for the impedance is as follows:

[0080] in, For the first The first sample Normalized impedance values ​​at each frequency point Indicates the first The sample at the th The impedance value at each frequency point This is the global minimum impedance value. This is the global maximum impedance value; The amplitude normalization formula is as follows:

[0081] in, For the first The first sample Normalized amplitude at each frequency point Indicates the first The sample at the th The amplitude at each frequency point The global minimum amplitude, This represents the global maximum amplitude; the normalized data range is unified as follows: .

[0082] S223: Next, special cases that occur during the normalization process are corrected. When the maximum value of the impedance or amplitude parameter in the global range equals the minimum value, that is... This indicates that the parameter is not different across all samples. Therefore, its normalized value is uniformly set to 0.5, as expressed in the formula below:

[0083] in, This represents the normalized value of this parameter.

[0084] As an embodiment of this application, the step S2 of constructing a two-dimensional feature matrix based on the normalized impedance value and amplitude specifically includes: S231: Create a size of Two-dimensional feature matrix ,in The total number of frequency points, matrix elements row index Frequency index corresponding to the impedance value, column index The frequency index corresponding to the amplitude, where ; S232: Calculate matrix elements using cross-product method. The formula is as follows:

[0085] in, For the first Normalized impedance values ​​at each frequency point For the first Normalized amplitude at each frequency point; For the first Normalized impedance values ​​at each frequency point For the first Normalized amplitude at each frequency point; S233: For the constructed two-dimensional feature matrix Symmetry verification and outlier correction are performed because of the formula. Theoretically, matrix M is a two-dimensional characteristic matrix. It should be a symmetric matrix. The asymmetry index is calculated using the following formula:

[0086] in, It is an index of the asymmetry of a two-dimensional feature matrix. matrix elements , The absolute difference, This indicates that only the upper triangular part of the two-dimensional feature matrix is ​​calculated. This represents the sum of the absolute values ​​of all elements in the two-dimensional feature matrix; if the asymmetry index exceeds the threshold of 0.01, the asymmetric elements are averaged and corrected, as shown in the following formula:

[0087] in, and This represents the pair of asymmetric matrix elements that need to be corrected. This indicates that the arithmetic mean of the two elements is used as the correction value; the detection matrix exceeds... The outlier element of the criterion is replaced by the median of the row and column containing that element, as shown in the following formula:

[0088] in, This indicates the outlier elements that need to be corrected. For the matrix of the first All elements of the row, For the matrix of the first All elements of the row, This indicates a set merging operation. This indicates that the median of the array is calculated. S234: Adopted The function performs nonlinear contrast enhancement on the corrected matrix to highlight feature differences within the matrix (such as high-frequency features of tumor tissue), as shown in the following formula:

[0089] in, For the original matrix elements, For the enhanced matrix elements, The slope parameter is used to control the reinforcement strength. This is the center offset parameter, used to control the center position. It is a natural constant; specifically, in this embodiment... , The values ​​are set to 5 and 0.5 respectively.

[0090] As an embodiment of this application, step S2, which converts the two-dimensional feature matrix into two-dimensional image features, specifically includes: S241: Transform the two-dimensional feature matrix Linear mapping to grayscale image grayscale image The pixel value range is [0, 255], and the formula is as follows:

[0091] in, This indicates rounding to the nearest integer. grayscale image median coordinate The pixel values ​​at a given location are mapped as follows: high-value regions in the two-dimensional feature matrix (such as frequency points corresponding to tumor tissue) are mapped as bright pixels, and low-value regions are mapped as dark pixels, forming an initial grayscale contrast. S242: For grayscale images Perform histogram equalization to enhance the overall contrast of the image. The formula is as follows:

[0092] in, grayscale value In image grayscale image The probability distribution in, i.e. Pixel count / total pixels To equalize the grayscale image, the grayscale histogram of the image is redistributed to make the bright parts of the image brighter and the dark parts darker, highlighting the boundary difference between tumor tissue and normal tissue. S243: Adopted Color mapping schemes will equalize the grayscale image. Convert to heatmap The color mapping function formula is as follows:

[0093] in, for Color mapping function, output Triples For indexing at frequency points Heatmaps of locations are used to enhance feature recognition through color coding; S244: Using bilinear interpolation for heatmaps Spatial standardization is performed to unify the image size to 224×224 pixels, using the following formula:

[0094] in, The coordinates of the four nearest pixels in the original image. The weights of the corresponding coordinate points, The spatially standardized images eliminate image size differences caused by varying frequency points, ensuring that all sample images have a uniform spatial resolution and providing consistent input for subsequent deep learning models.

[0095] Specifically, this invention eliminates the interference of measurement noise and individual differences on the data by removing outliers, normalizing and smoothing the impedance and amplitude values, thus ensuring signal stability. By converting the preprocessed impedance and amplitude values ​​into two-dimensional image features, the frequency correlation characteristics of the one-dimensional electrical signal can be transformed into visualized spatial features. The two-dimensional structure constructed by matrix cross-product can intuitively present the coordinated change law of impedance and amplitude at different frequency points, providing image-dimensional information support for multi-feature fusion.

[0096] As an embodiment of this application, step S3, which involves feature alignment and dimensional unification of the original impedance value and the original amplitude, specifically includes: S311: Convert the original impedance value vector Convert to electrical impedance matrix ,in, Total number of frequency points; The original magnitude vector Convert to magnitude matrix ,in, This represents the total number of frequency points. Specifically, this invention converts the one-dimensional time-series features of the original impedance value and the original amplitude into a two-dimensional matrix form, which facilitates dimensional alignment with image features, while retaining the frequency point index as a feature dimension.

[0097] S312: Verify the electrical impedance matrix and magnitude matrix Whether the frequency point sequences are consistent, that is, checking the frequency point lists of the two. and Does it meet the requirements? If the frequency point sequences are inconsistent, linear interpolation is used to transform the impedance matrix. and magnitude matrix Time series alignment is performed by resampling one feature matrix to the frequency points of the other feature matrix, as shown in the following formula:

[0098] in, These are the target frequency points that need to be aligned. It is less than the original amplitude data The maximum frequency point, It is greater than the original amplitude data The minimum frequency point, The original amplitude at the frequency point The value at that location, The original amplitude at the frequency point The value at that location, For the target frequency points that need to be aligned The amplitude after resampling at the location; S313: The impedance matrix after time series alignment and magnitude matrix Standardization is performed separately to unify the feature scale. Among these, the electrical impedance matrix... The standardized formula is as follows:

[0099] in, It is the first normalized electrical impedance matrix. line, number Column elements, It is the first unnormalized electrical impedance matrix. line, number Column elements, It is the global mean of the impedance values ​​of all samples at all frequency points. The global standard deviation of the impedance values ​​of all samples at all frequency points.

[0100] Amplitude matrix The standardized formula is as follows:

[0101] in, It is the first normalized amplitude matrix. line, number Column elements, It is the first normalized magnitude matrix. line, number Column elements, It is the global mean of the amplitude of all samples across all frequency points. The global standard deviation of the amplitude of all samples at all frequency points eliminates the dimensional differences between different features, making the model learn all features more evenly, while preserving the relative distribution information of the features.

[0102] S314: Using the Pearson correlation coefficient to standardize the electrical impedance matrix and magnitude matrix Feature selection is performed, retaining frequency features with high correlation, including the electrical impedance matrix. The feature selection formula is as follows:

[0103] in, It is the first The impedance values ​​at each frequency point and the Pearson correlation coefficient of the sample labels, For the sample size, It is the first The sample at the th Standardized impedance values ​​at each frequency point For all samples in the first The mean of the standardized electrical impedance values ​​at each frequency point For the first Sample labels for each sample; Amplitude matrix The feature selection formula is as follows:

[0104] in, It is the first The amplitude at each frequency point and the Pearson correlation coefficient of the sample label, For the sample size, It is the first The sample at the th Standardized amplitude at each frequency point For all samples in the first The mean of the standardized amplitude at each frequency point For the first Sample labels for each sample.

[0105] As an embodiment of this application, the preprocessing of the two-dimensional image features in step S3 to match the sample dimensions of the original impedance value and the original amplitude specifically includes: S321: A pre-trained convolutional neural network is used to extract deep features from two-dimensional image features to obtain a high-dimensional image feature vector, which is then input into the heatmap. The formula is as follows:

[0106] in, For high-dimensional image feature vectors, This represents a feature extraction function used to convert an image into a high-dimensional feature vector. The input heatmap has the following dimensions: The pixel value range is [0,1].

[0107] S322: Principal component analysis is used to reduce the dimensionality of the extracted high-dimensional image feature vectors, retaining more than 95% of the feature variance. A linear transformation is then used to project the high-dimensional features into a low-dimensional space, as shown in the following formula:

[0108] in, It is a high-dimensional image feature vector with dimension . , Principal component analysis projection matrix, dimension 1 , The dimension after dimensionality reduction. The feature vector after dimensionality reduction has a dimension of . The smallest value with a cumulative variance contribution of ≥95% is retained. Value, usually .

[0109] S323: Standardize the feature vectors of the dimension-reduced image to make them consistent with the numerical distribution of the original electrical signal features. The standardization formula is as follows:

[0110] in, It is the standardized image feature vector. It is the image feature vector after dimensionality reduction. This is the mean vector of image features after dimensionality reduction for all samples. The standard deviation vector is the corresponding standard deviation vector. The purpose of standardization is to ensure that the image features, electrical impedance features, and amplitude features are in the same numerical range, so as to avoid the fusion model from having a preference for a certain type of feature.

[0111] As an embodiment of this application, step S3, which establishes a mapping table of original impedance value, original amplitude, and two-dimensional image features based on the unique identifier of the sample, and constructs a multi-feature matrix, specifically includes: S331: Establish a mapping table between the original electrical impedance value, original amplitude, and two-dimensional image features based on the sample's unique identifier (ID). For example, "Tumor_001" represents a tumor sample, and "Normal_002" represents a normal sample. For each sample, concatenate the electrical impedance feature, amplitude feature, and image features along the feature dimension to form a multi-feature vector. The concatenation formula is as follows:

[0112] in, Indicates the characteristics of electrical impedance value, Indicates amplitude characteristics, Represents the image feature vector. For a single sample with multiple feature vectors, This indicates concatenation by column, i.e., the superposition of feature dimensions; S332: Stack the multi-feature vectors of all samples row-wise to construct a multi-feature matrix, as shown in the following formula:

[0113] in, It is a multi-feature matrix. The total number of samples, For the first Multiple feature vectors of a sample; Add sample labels to the multi-feature matrix to form a complete multi-feature dataset. Label definition: ,in Indicates the first The sample was tumor tissue. Indicates normal tissue.

[0114] As an embodiment of this application, step S3, which standardizes and verifies the multi-feature matrix to form the final multi-feature set, specifically includes: S341: For the multi-feature matrix Each feature dimension is standardized using Z-score to unify the numerical distribution of all feature dimensions, as shown in the following formula:

[0115] in, Represents each feature dimension in the matrix. Indicates the standardized first The sample, the first The numerical values ​​of the dimensional features, Indicates the first before standardization The sample, the first The original numerical values ​​of the dimensional features. Indicates the first The global mean of the dimensional feature across all samples. Indicates the first The global standard deviation of the dimensional feature across all samples; S342: Calculate the cosine similarity between the electrical signal features and image features of each sample. The calculation formula is as follows:

[0116] in, Indicates the first Similarity between features of each sample Indicates the first The electrical impedance screening feature vector of each sample, Indicates the first The standardized image feature vector of each sample, Represents the vector dot product. express Norm, when Samples with low correlation between features are removed. S343: Calculate the Mahalanobis distance between the multiple feature vectors of each sample and the global mean vector. The calculation formula is as follows:

[0117] in, Indicates the first Mahalanobis distance of each sample Indicates the first Multiple feature vectors of a sample This represents the multi-feature mean vector of all samples. Let the covariance matrix of a multi-feature matrix be the inverse matrix. Samples exceeding the 95th percentile threshold are considered outliers and are removed. S344: Divide the validated feature set into a training set, a validation set, and a test set in a ratio of 7:1.5:1.5. The training set (70%) is used for model parameter learning; the validation set (15%) is used for hyperparameter tuning (such as learning rate and regularization coefficient); and the test set (15%) is used to evaluate the model's generalization ability.

[0118] Specifically, this invention achieves precise alignment of multiple features by associating the original electrical impedance value, original amplitude, and two-dimensional image features with the sample ID: the original electrical impedance value and amplitude retain the temporal frequency characteristics of the electrical signal, and the two-dimensional image features supplement the frequency correlation information in the spatial domain. The two form a "temporal-spatial" information complementarity, avoiding the problem that a single electrical signal feature is insufficient to capture tumor heterogeneity, and significantly improving the richness and comprehensiveness of the feature set.

[0119] As an embodiment of this application, step S4 specifically includes: S41: A multilayer perceptron is used to extract high-order features of impedance and amplitude, adapted to one-dimensional time-series characteristics, and the filtered impedance features are then... The input is fed into the electrical impedance characteristic processing module, where it is processed through three fully connected layers. The calculation formula is as follows:

[0120]

[0121]

[0122] in, This represents the feature vector output by the first fully connected layer. The activation function is used to introduce a nonlinear transformation. This represents the weight matrix of the first fully connected layer. This represents the bias vector of the first fully connected layer, used to adjust the baseline of the linear transformation. This represents the feature vector output by the second fully connected layer. This represents the weight matrix of the second fully connected layer. This represents the bias vector of the second fully connected layer. This represents the feature vector output by the third fully connected layer. This represents the weight matrix of the third fully connected layer. This represents the bias vector of the third fully connected layer; Selected amplitude features The input is fed into the amplitude feature processing module, where it undergoes the same operations as the impedance feature processing module, ultimately outputting a feature vector. ; The image feature processing module uses a lightweight convolutional neural network to capture the spatial texture features of the image, adapting to the characteristics of two-dimensional vision, and processing the image features. The image is input to the image feature processing module and processed sequentially through three convolutional blocks of different dimensions. The calculation formula is as follows:

[0123]

[0124]

[0125] in, Indicates the first dimension is The feature map output by the convolutional block, This represents the bias vector of the first convolutional layer, with dimension 1. ; For activation function, This is a max pooling operation used to reduce the size of the feature map while retaining key information; Indicates the second dimension is The feature map output by the convolutional block, This represents the bias vector of the second convolutional layer, with dimension . ; Indicates the third dimension as The feature map output by the convolutional block, This represents the bias vector of the third convolutional layer, with dimension . ; To perform the flattening operation, the 3D feature map is converted into a 1D vector, and then compressed to 128 dimensions through a fully connected layer; S42: Outputs from the impedance characteristic processing module, amplitude processing module, and image feature processing module respectively. , , The input is fed into the feature fusion module, where each feature is mapped to a scalar attention level through a linear transformation. The normalization formula is as follows:

[0126]

[0127]

[0128] in, , , These represent attention scalars for electrical impedance, amplitude, and image features, respectively, reflecting the correlation strength between each feature and the tumor label. For linear layers, , , These are the attention weights, and their sum is 1; This represents an exponential function, mapping scalars to the non-negative interval and amplifying differences in attention.

[0129] After linear transformation , , The final fusion feature is obtained by weighted summation, and the calculation formula is as follows:

[0130] in, The fusion feature is obtained by weighted summation; S43: Integrate features Input is fed into the output layer to calculate the tumor probability. The formula is as follows:

[0131] in, This represents the activation function, which maps the linear output to the interval [0,1]. The weight matrix represents the classification head. This represents the bias scalar of the classification head, used to adjust the baseline of the linear transformation.

[0132] As an embodiment of this application, the present invention also enables the multi-feature fusion detection network model to learn the association rules between multiple features and tumor labels through training strategies and hyperparameter tuning, and uses binary classification cross-entropy loss to quantify the prediction error, as shown in the following formula:

[0133] in, This represents the average loss value of the batch of samples. Indicates batch size. For the first The true label of each sample This indicates that the multi-feature fusion detection network model is effective for the first... Tumor probability prediction value for each sample. The loss function is the natural logarithm, which imposes a greater penalty on samples with incorrect predictions. The summation and averaging operations are used to sum and average the losses of all samples within a batch to obtain the batch loss. The Adam optimizer is used with step decay, decreasing to 1 / 10 of the current value every 5 epochs. If the validation set loss does not decrease for 10 consecutive epochs, training is terminated, and the current optimal model parameters are saved.

[0134] As an embodiment of this application, step S4 further includes performance evaluation of the trained model, the specific steps of which are as follows: S44: Evaluate the trained multi-feature fusion detection network model using the test set. The accuracy calculation formula is as follows:

[0135] in, For accuracy, This means that samples that are actually tumors are correctly predicted as tumors by the model. For samples that are actually normal, the model correctly predicts them as normal. Samples that are actually normal are incorrectly predicted as tumors by the model. Samples that are actually tumors are incorrectly predicted as normal by the model; S45: Based on the prediction results of the test set, set the threshold to 0.5, and output a tumor detection report containing the sample ID, prediction category, and confidence level. Input multiple features of the sample into the model to obtain the prediction probability. ,like The threshold is used to determine whether a sample is "tumor tissue" or "normal tissue". A test report is generated, which includes the sample's unique ID, predicted category, confidence level, and model evaluation metrics.

[0136] Specifically, the multi-feature fusion detection network model in this invention adopts a two-channel processing mechanism. The multilayer perceptron extracts the frequency domain high-order features of the original impedance value and amplitude, the convolutional neural network captures the spatial texture features of the two-dimensional image, and then the attention layer calculates the weights of each feature to achieve weighted fusion. This design can automatically focus on key features that are strongly related to tumors, effectively integrate cross-feature information, and improve the detection sensitivity of early tumors and atypical lesions.

[0137] The above description is merely a selection of preferred embodiments of this disclosure and an explanation of the technical principles employed. Those skilled in the art should understand that the scope of the invention involved in the embodiments of this disclosure is not limited to technical solutions formed by specific combinations of the above-described technical features, but should also cover other technical solutions formed by arbitrary combinations of the above-described technical features or their equivalents without departing from the above-described inventive concept.

Claims

1. A multi-feature-based bioelectrical impedance method for detecting tumor activity, characterized in that, The method includes the following steps: S1: Obtain raw bioelectrical impedance data of tumor tissue and normal tissue samples, wherein the raw bioelectrical impedance data includes raw impedance values ​​and raw amplitudes at different frequencies; S2: The obtained original impedance value and original amplitude are cleaned, the cleaned impedance value and amplitude are normalized, a two-dimensional feature matrix is ​​constructed based on the normalized impedance value and amplitude, and the two-dimensional feature matrix is ​​converted into two-dimensional image features. S3: Perform feature alignment and dimension unification on the original impedance value and original amplitude, preprocess the two-dimensional image features to match the sample dimension of the original impedance value and original amplitude; establish a mapping relationship table of original impedance value, original amplitude and two-dimensional image features based on the unique identifier of the sample, construct a multi-feature matrix, standardize and verify the multi-feature matrix to form the final multi-feature set; S4: Construct a multi-feature fusion detection network model and train and optimize it. Input the multi-feature set into the trained multi-feature fusion detection network model and output the tumor detection result. The multi-feature fusion detection network model includes an electrical impedance feature processing module, an amplitude processing module, an image feature processing module, a multi-feature fusion module, and an output layer.

2. The method for detecting tumor resistance based on multi-feature bioelectrical impedance as described in claim 1, characterized in that, The data cleaning process for the obtained original impedance value and original amplitude in step S2 specifically includes: S211: Adopted The criteria identify outlier data points. For each frequency point, based on the raw impedance and raw amplitude, the mean of all samples at that frequency point is calculated. and standard deviation The calculation formula is as follows: in, The total number of samples, For the first The value of each sample; when the data points satisfy When this happens, it is considered an outlier; S212: Replace the identified outliers with the sample mean of the corresponding frequency points. The calculation formula is as follows: in, This is the mean of all normal samples at this frequency point; For the replaced value, if the number of abnormal samples at a certain frequency point is less than or equal to 5% of the total number of samples, the value is directly replaced; if the number of abnormal samples at a certain frequency point is greater than 5% of the total number of samples, the corresponding sample is re-collected. S213: By scanning the impedance and amplitude data of each sample, identify missing frequency points and calculate the missing rate of each sample. If the sample missing rate is <1%, use linear interpolation to fill in the missing data. The calculation formula is as follows: in, For missing frequency points, Missing frequency points to be filled The corresponding impedance value or amplitude, for The previous effective frequency point, Effective frequency point The corresponding impedance value or amplitude, for The next effective frequency point, Effective frequency point The corresponding impedance value or amplitude; If the sample missing rate is ≥1%, it is considered an invalid sample and needs to be collected again. S214: Gaussian filtering is used to smooth the impedance and amplitude data, removing high-frequency noise. The calculation formula is as follows: in, For smoothing the first The impedance or amplitude at each frequency point The filter window half-width, The index of the data points within the window relative to the center. Gaussian function at index The weight values ​​at each point are used for the raw data within the weighted average window. For the first in the original data The Gaussian function formula for the impedance or amplitude at a given frequency point is as follows: in, The standard deviation of the Gaussian function is used to control the rate of weight decay. The index of the data points within the window relative to the center. It is a natural constant.

3. The method for detecting tumor resistance based on multi-feature bioelectrical impedance as described in claim 1, characterized in that, The normalization process for the cleaned impedance value and the original amplitude in step S2 specifically includes: S221: For the impedance value and amplitude after data cleaning, iterate through every frequency point of all samples. The formula for calculating the impedance value is as follows: in, The total number of samples, The total number of frequency points. Indicates the first The sample at the th The impedance value at each frequency point This is the global minimum impedance value. This is the global maximum impedance value; The formula for calculating amplitude is as follows: in, The total number of samples, The total number of frequency points. Indicates the first The sample at the th The amplitude at each frequency point The global minimum amplitude, This is the global maximum amplitude. S222: Normalize the impedance and amplitude of each sample at each frequency point. The normalization formula for the impedance is as follows: in, For the first The first sample Normalized impedance values ​​at each frequency point Indicates the first The sample at the th The impedance value at each frequency point This is the global minimum impedance value. This is the global maximum impedance value; The amplitude normalization formula is as follows: in, For the first The first sample Normalized amplitude at each frequency point Indicates the first The sample at the th The amplitude at each frequency point The global minimum amplitude, This is the global maximum amplitude. S223: When the maximum value of the impedance or amplitude parameter is equal to the minimum value in the global range, and the parameter is indistinguishable in all samples, its normalized value is uniformly set to 0.

5.

4. The method for detecting tumor resistance based on multi-feature bioelectrical impedance as described in claim 1, characterized in that, The construction of the two-dimensional feature matrix based on the normalized impedance value and amplitude in step S2 specifically includes: S231: Create a size of Two-dimensional feature matrix ,in The total number of frequency points, matrix elements row index Frequency index corresponding to the impedance value, column index The frequency index corresponding to the amplitude, where ; S232: Calculate matrix elements using cross-product method. The formula is as follows: in, For the first Normalized impedance values ​​at each frequency point For the first Normalized amplitude at each frequency point; For the first Normalized impedance values ​​at each frequency point For the first Normalized amplitude at each frequency point; S233: For the constructed two-dimensional feature matrix Symmetry verification and outlier correction are performed, and the asymmetry index is calculated using the following formula: in, It is an index of the asymmetry of a two-dimensional feature matrix. matrix elements , The absolute difference This indicates that only the upper triangular part of the two-dimensional feature matrix is ​​calculated. This represents the sum of the absolute values ​​of all elements in the two-dimensional feature matrix; if the asymmetry index exceeds the threshold of 0.01, then the asymmetry elements are averaged for correction; [The last part, "detecting the matrix exceeding...", appears to be incomplete and requires further context.] The outlier element of the criterion is replaced by the median of the row and column containing that element; S234: Adopted The function performs non-linear contrast enhancement on the corrected matrix to highlight feature differences, as shown in the following formula: in, For the original matrix elements, For the enhanced matrix elements, The slope parameter is used to control the reinforcement strength. This is the center offset parameter, used to control the center position. It is a natural constant.

5. The method for detecting tumor resistance based on multi-feature bioelectrical impedance as described in claim 1, characterized in that, The specific steps in step S2, which involve converting the two-dimensional feature matrix into two-dimensional image features, include: S241: Transform the two-dimensional feature matrix Linear mapping to grayscale image grayscale image The pixel value range is [0, 255], and the formula is as follows: in, This indicates rounding to the nearest integer. grayscale image median coordinate Pixel value at; S242: For grayscale images Perform histogram equalization to enhance the overall contrast of the image. The formula is as follows: in, grayscale value In image grayscale image The probability distribution in This is the grayscale image after equalization; S243: Adopted Color mapping schemes will equalize the grayscale image. Convert to heatmap The color mapping function formula is as follows: in, for Color mapping function, output Triples For indexing at frequency points Heatmaps of locations are used to enhance feature recognition through color coding; S244: Using bilinear interpolation for heatmaps Spatial standardization is performed using the following formula: in, The coordinates of the four nearest pixels in the original image. The weights of the corresponding coordinate points, The image is spatially normalized.

6. The method for detecting tumor resistance based on multi-feature bioelectrical impedance as described in claim 1, characterized in that, The feature alignment and dimensional unification of the original impedance value and original amplitude in step S3 specifically includes: S311: Convert the original impedance value vector Convert to electrical impedance matrix ,in, Total number of frequency points; The original magnitude vector Convert to magnitude matrix ,in, This represents the total number of frequency points. S312: Linear interpolation is used to transform the impedance matrix. and magnitude matrix Time series alignment is performed by resampling one feature matrix to the frequency points of the other feature matrix, as shown in the following formula: in, These are the target frequency points that need to be aligned. It is less than the original amplitude data The maximum frequency point, It is greater than the original amplitude data The minimum frequency point, The original amplitude at the frequency point The value at that location, The original amplitude at the frequency point The value at that location, For the target frequency points that need to be aligned The amplitude after resampling at the location; S313: The impedance matrix after time series alignment and magnitude matrix Standardize them separately, including the electrical reactance matrix. The standardized formula is as follows: in, It is the first normalized electrical impedance matrix. line, number Column elements, It is the first normalized impedance matrix. line, number Column elements, It is the global mean of the impedance values ​​of all samples at all frequency points. The global standard deviation of the impedance values ​​of all samples at all frequency points; Amplitude matrix The standardized formula is as follows: in, It is the first normalized amplitude matrix. line, number Column elements, It is the first normalized magnitude matrix. line, number Column elements, It is the global mean of the amplitude of all samples across all frequency points. The global standard deviation of the amplitude of all samples at all frequency points; S314: Using the Pearson correlation coefficient to standardize the electrical impedance matrix and magnitude matrix Feature selection is performed, retaining frequency features with high correlation, including the electrical impedance matrix. The feature selection formula is as follows: in, It is the first The impedance values ​​at each frequency point and the Pearson correlation coefficient of the sample labels, For the sample size, It is the first The sample at the th Standardized impedance values ​​at each frequency point For all samples in the first The average of the standardized electrical impedance values ​​at each frequency point For the first Sample labels for each sample; Amplitude matrix The feature selection formula is as follows: in, It is the first The amplitude at each frequency point and the Pearson correlation coefficient of the sample label, For the sample size, It is the first The sample at the th Standardized amplitude at each frequency point For all samples in the first The mean of the standardized amplitude at each frequency point For the first Sample labels for each sample.

7. The method for detecting tumor resistance based on multi-feature bioelectrical impedance as described in claim 1, characterized in that, The preprocessing of the two-dimensional image features in step S3 to match the sample dimensions of the original impedance value and the original amplitude specifically includes: S321: A pre-trained convolutional neural network is used to extract deep features from two-dimensional image features to obtain a high-dimensional image feature vector, which is then input into the heatmap. The formula is as follows: in, For high-dimensional image feature vectors, Represents the feature extraction function; S322: Principal component analysis is used to reduce the dimensionality of the extracted high-dimensional image feature vectors, as shown in the following formula: in, For high-dimensional image feature vectors, Principal component analysis projection matrix, These are the eigenvectors after dimensionality reduction; S323: Standardize the feature vectors of the dimension-reduced image to make them consistent with the numerical distribution of the original electrical signal features. The standardization formula is as follows: in, It is the standardized image feature vector. It is the image feature vector after dimensionality reduction. This is the mean vector of image features after dimensionality reduction for all samples. This is the corresponding standard deviation vector.

8. The method for detecting tumor resistance based on multi-feature bioelectrical impedance as described in claim 1, characterized in that, The construction of the multi-feature matrix in step S3 specifically includes: S331: Based on the unique identifier of the sample, establish a mapping table of the original electrical impedance value, the original amplitude, and the two-dimensional image features. For each sample, concatenate the electrical impedance feature, the amplitude feature, and the image feature along the feature dimension to form a multi-feature vector. The concatenation formula is as follows: in, Indicates the characteristics of electrical impedance value, Indicates amplitude characteristics, Represents the image feature vector. For a single sample with multiple feature vectors, Indicates concatenation by column; S332: Stack the multi-feature vectors of all samples row-wise to construct a multi-feature matrix, as shown in the following formula: in, It is a multi-feature matrix. For the sample size, For the first Multiple feature vectors of a sample.

9. The method for detecting tumor resistance based on multi-feature bioelectrical impedance as described in claim 1, characterized in that, The formation of the final multi-feature set in step S3 specifically includes: S341: For the multi-feature matrix Each feature dimension is standardized to unify the numerical distribution of all feature dimensions, as shown in the following formula: in, Represents each feature dimension in the matrix. Indicates the standardized first The sample, the first The numerical values ​​of the dimensional features, Indicates the first before standardization The sample, the first The original numerical values ​​of the dimensional features. Indicates the first The global mean of the dimensional feature across all samples. Indicates the first The global standard deviation of the dimensional feature across all samples; S342: Calculate the cosine similarity between the electrical signal features and image features of each sample. The calculation formula is as follows: in, Indicates the first Similarity between features of each sample Indicates the first The electrical impedance screening feature vector of each sample, Indicates the first The standardized image feature vector of each sample, Represents the dot product of vectors. express Norm; S343: Calculate the Mahalanobis distance between the multiple feature vectors of each sample and the global mean vector. The calculation formula is as follows: in, Indicates the first Mahalanobis distance of each sample Indicates the first Multiple feature vectors of a sample This represents the multi-feature mean vector of all samples. It is the inverse of the covariance matrix with multiple features; S344: Divide the validated feature set into training, validation and test sets in a ratio of 7:1.5:1.

5.

10. The method for detecting tumor resistance based on multi-feature bioelectrical impedance as described in claim 9, characterized in that, Step S4 specifically includes: S41: A multilayer perceptron is used to extract high-order features of impedance values ​​and amplitudes, and the filtered impedance features are then... The input is fed into the electrical impedance characteristic processing module, where it is processed through three fully connected layers. The calculation formula is as follows: in, This represents the feature vector output by the first fully connected layer. The activation function is used to introduce a nonlinear transformation. This represents the weight matrix of the first fully connected layer. This represents the bias vector of the first fully connected layer, used to adjust the baseline of the linear transformation. This represents the feature vector output by the second fully connected layer. This represents the weight matrix of the second fully connected layer. This represents the bias vector of the second fully connected layer. This represents the feature vector output by the third fully connected layer. This represents the weight matrix of the third fully connected layer. This represents the bias vector of the third fully connected layer; Selected amplitude features The input is fed into the amplitude feature processing module, where it undergoes the same operations as the impedance feature processing module, ultimately outputting a feature vector. ; Image features The image is input to the image feature processing module and processed sequentially through three convolutional blocks of different dimensions. The calculation formula is as follows: in, Indicates the first dimension is The feature map output by the convolutional block, This represents the bias vector of the first convolutional layer. For activation function, This is a max pooling operation used to reduce the size of the feature map while retaining key information; Indicates the second dimension is The feature map output by the convolutional block, This represents the bias vector of the second convolutional layer. Indicates the third dimension as The feature map output by the convolutional block, This represents the bias vector of the third convolutional layer. To perform the flattening operation, the 3D feature map is converted into a 1D vector, and then compressed to 128 dimensions through a fully connected layer; S42: Outputs from the impedance characteristic processing module, amplitude processing module, and image feature processing module respectively. , , The input is fed into the feature fusion module, where each feature is mapped to a scalar attention level through a linear transformation. The normalization formula is as follows: in, , , These represent attention scalars for electrical impedance, amplitude, and image features, respectively, reflecting the correlation strength between each feature and the tumor label. For linear layers, , , For attention weights, Represents an exponential function; After linear transformation , , The final fusion feature is obtained by weighted summation, and the calculation formula is as follows: in, The fusion feature is obtained by weighted summation; S43: Integrate features Input is fed into the output layer to calculate the tumor probability. The formula is as follows: in, This represents the activation function. The weight matrix represents the classification head. The scalar represents the bias of the classification head.

Images