Bioelectrical impedance tumor detection method based on traceability mapping relation

Abstract

The invention discloses a bioelectrical impedance tumor detection method based on a traceability mapping relation, 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 at different frequencies; s2, performing standardization processing on the electrical impedance value and the amplitude value, and converting the electrical impedance value and the amplitude value into two-dimensional image features; s3, establishing a traceability mapping relation between two-dimensional image features and bioelectrical impedance original data, and extracting key associated frequency points; and S4, constructing a cross-modal correlation enhancement fusion network based on the traceability mapping relationship, performing fusion detection on the electrical impedance value features, the amplitude features and the two-dimensional image features, and outputting a tumor detection result. According to the method, the traceability mapping relation between the two-dimensional image features and the original bioelectrical impedance data is established, real associated features are focused, and the precision and interpretability of tumor detection are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of tumor detection, and in particular to a bioelectrical impedance tumor detection method based on traceability mapping relationship. BACKGROUND

[0002] Solid tumors such as lung cancer, liver cancer, breast cancer, etc. are one of the major diseases threatening human health. Early and accurate detection is of great significance to improve patient survival rate and reduce treatment cost. Traditional tumor detection methods have many limitations: imaging examination can provide anatomical information, but it has the problems of radiation exposure risk or high equipment cost; as the "gold standard" for diagnosis, puncture biopsy is an invasive operation that may cause complications and is difficult to achieve dynamic monitoring; serum marker detection has the defects of insufficient sensitivity and easy interference. Therefore, it is an urgent clinical need to develop a non-invasive, low-cost and high-specificity early tumor detection technology.

[0003] The tumor detection method based on bioelectrical impedance technology measures the electrical signal characteristics such as electrical impedance value and amplitude of the tissue at different frequencies, and realizes detection by using the difference in electrical physiological characteristics between tumor tissue and normal tissue. It has the advantages of non-invasiveness, real-time monitoring, and portable equipment, and is suitable for screening and diagnosis of various solid tumors. With the development of deep learning technology, its powerful feature mining capability provides a new tool for bioelectrical impedance multi-modal data analysis - by constructing a cross-modal fusion model, the correlation between the frequency regularity of electrical signals and the spatial pattern of images can be fully mined, further improving the accuracy of tumor detection. The combination of bioelectrical impedance technology and deep learning opens up a new path for early non-invasive detection of various solid tumors, and has a wide clinical application prospect.

[0004] Chinese patent CN119969995B discloses a bioelectrical impedance cancer detection method based on deep learning. The matrix decomposition module extracts global features from electrical signals, and the results of global feature extraction are used as new inputs together with original electrical signal data to further extract signal features using data dependency relationship, and the extracted signal features are mapped and integrated into the final classification results.

[0005] However, the above method focuses on the extraction and enhancement of single type of features, and fails to fully utilize the complementarity of bioelectrical impedance multi-modal features. Moreover, the multi-modal fusion algorithm in the prior art ignores the intrinsic traceability relationship between image features and original electrical signals, resulting in problems such as pseudo-association and fuzzy positioning of key features. Therefore, there is an urgent need to propose a bioelectrical impedance tumor detection method based on traceability mapping relationship to solve the above problems existing in the prior art. SUMMARY

[0006] In view of the above defects or improvement needs of the prior art, the present application provides a bioelectrical impedance tumor detection method based on a traceability mapping relationship, which focuses on real correlation features by establishing a mapping relationship between image pixels and original bioelectrical impedance frequency points, and improves the accuracy and interpretability of tumor detection.

[0007] To achieve the above object, according to one aspect of the present application, a bioelectrical impedance tumor detection method based on a traceability mapping relationship is provided, which comprises the following steps: S1: acquiring bioelectrical impedance data of tumor tissue and normal tissue samples, the bioelectrical impedance data including bioelectrical impedance values and amplitude values at different frequencies; S2: measuring the acquired effective sample bioelectrical impedance values and amplitude values to obtain a bioelectrical impedance value matrix and an amplitude value matrix, and performing standardization processing thereon, constructing a two-dimensional feature matrix based on the standardized bioelectrical impedance value matrix and amplitude value matrix, and converting the two-dimensional feature matrix into a two-dimensional image feature; S3: establishing a traceability mapping relationship between the two-dimensional image feature and the original bioelectrical impedance data; segmenting a high-value area of the two-dimensional image to extract a high-value area pixel set; and obtaining a key frequency point set through coordinate traceability conversion; S4: constructing a cross-modal correlation enhancement fusion network based on the traceability mapping relationship, and performing fusion detection on the bioelectrical impedance value feature, the amplitude value feature and the two-dimensional image feature to output a tumor detection result.

[0008] As an embodiment of the present application, the step S2 of measuring the acquired effective bioelectrical impedance values and amplitude values to obtain a bioelectrical impedance value matrix and an amplitude value matrix, and performing standardization processing thereon specifically comprises: S211: calculating a statistic quantity of each frequency point, measuring the bioelectrical impedance values and amplitude values of the effective sample set to obtain a bioelectrical impedance value matrix and an amplitude value matrix ; S212: performing standardization processing on the bioelectrical impedance value matrix and the amplitude value matrix , and the calculation formula is as follows:

[0009]

[0010] wherein, is a standardized bioelectrical impedance value of the i th effective sample at the j th frequency point, is an original bioelectrical impedance value of the i th effective sample at the j th frequency point, is a standardized bioelectrical impedance value of the i th effective sample at the j th frequency point, is an original bioelectrical impedance value of the i th effective sample at the j th frequency point, is a standardized bioelectrical impedance value of the i th effective sample at the j th frequency point, is an original bioelectrical impedance value of the i th effective sample at the j th frequency point, is a standardized bioelectrical impedance value of the i th effective sample at the j th frequency point, the minimum value of the frequency point in the resistance impedance matrix of all effective samples, is the maximum value of the frequency point in the resistance impedance matrix of all effective samples; is the standardized amplitude value of the effective sample at the frequency point, is the original amplitude value of the effective sample at the frequency point, is the minimum value of the frequency point in the amplitude matrix of all effective samples, is the maximum value of the frequency point in the amplitude matrix of all effective samples.

[0011] As an embodiment of the present application, the step S2 of constructing a two-dimensional feature matrix based on the normalized resistance impedance value matrix and the amplitude matrix and converting it into a two-dimensional image feature specifically includes: S221: Constructing a two-dimensional feature matrix based on the normalized resistance impedance value matrix and the amplitude matrix , the matrix element calculation formula is as follows:

[0012] wherein, is the element of the row and the column of the matrix, is the normalized resistance impedance value of the effective sample at the frequency point, is the normalized amplitude value of the effective sample at the frequency point; S222: Adjusting the two-dimensional feature matrix to a two-dimensional feature matrix with a size of using the bilinear interpolation method, and then calculating the pixel value and converting it into a two-dimensional image feature, the calculation formula is as follows:

[0013] wherein, indicates rounding off; is the image coordinate, is the pixel value of the row and the column of the two-dimensional image, and the pixel value matrix is stored as a gray-scale image as a two-dimensional image feature.

[0014] As an embodiment of this application, the formula for establishing the source mapping relationship between two-dimensional image features and original bioelectrical impedance data in step S3 is expressed as follows:

[0015]

[0016] in, For source mapping functions, This is a round-down operation; For the original frequency point index, ; Image coordinates Compared with the original frequency point index The corresponding relationships are stored as a dictionary, forming a source mapping table.

[0017] As an embodiment of this application, step S3, which involves segmenting the two-dimensional image into high-value regions and extracting the pixel set of high-value regions, specifically includes: S321: For two-dimensional images set of pixel values Iterate through all thresholds Calculate the inter-class variance The formula is as follows:

[0018] in, pixel value The regional proportion; pixel value The regional proportion; pixel value The regional mean; pixel value The regional mean; S322: Determine the segmentation threshold and take a value that minimizes the inter-class variance. The largest As the optimal threshold, the set of pixels in the high-value region is extracted using the following formula:

[0019] in, For the first The set of high-value pixels in the image of each sample; Indicates the first in the high value region The coordinates of a pixel. Indicates the first The image of each sample is in coordinates The pixel value at that location.

[0020] As an embodiment of this application, the step S3 of obtaining the key frequency point set through coordinate tracing transformation specifically includes: S311: Convert the pixel coordinates of high-value areas in the image into frequency point indices of the original electrical signal through coordinate source transformation. For each pixel coordinate... The corresponding frequency point index is obtained through the source mapping table. The coordinate source transformation formula is as follows:

[0021] in, Indicates the first in the high value region The coordinates of a pixel. For source mapping functions, Represents the first obtained after mapping Index of original electrical signal frequency points; S332: Next, duplicate original electrical signal frequency point indices are removed through deduplication to obtain the key frequency point set. ,in, Represents the first obtained after mapping Index of the original electrical signal frequency points, This indicates the number of frequency point index pairs after deduplication.

[0022] As an embodiment of this application, step S4 specifically includes: S41: Construct a cross-modal correlation enhancement fusion network, which includes a resistance value feature extraction branch, an amplitude feature extraction branch, and a two-dimensional image feature extraction branch, and respectively extracts features from resistance value features, amplitude features, and two-dimensional image features; S42: Weighted processing of electrical signal features and two-dimensional image features based on key frequency point set and high value region pixel set; S43: Perform cross-modal feature fusion on the weighted electrical signal features and two-dimensional image features to output tumor detection results.

[0023] As an embodiment of this application, step S41 specifically includes: S411: The impedance feature extraction branch uses a 3-layer multilayer perceptron network to process the standardized impedance value. Features are extracted sequentially through three fully connected layers, and the calculation formula is as follows:

[0024]

[0025]

[0026] in, is an output feature of the first layer fully connected layer, is a weight matrix of the first layer fully connected layer, is a bias of the first layer fully connected layer, is an activation function, is a batch normalization operation, is a random inactivation operation; is an output feature of the second layer fully connected layer, is a weight matrix of the second layer fully connected layer, is a bias of the second layer fully connected layer, is an output impedance feature vector of the third layer fully connected layer, is a weight matrix of the third layer fully connected layer, is a bias of the third layer fully connected layer; S412: The amplitude feature extraction branch adopts a 3-layer multi-layer perception network symmetrical with the electrical impedance feature extraction branch to process the standardized amplitude , sequentially passes through three fully connected layers to extract features, and finally outputs an amplitude feature vector ; S413: The image feature extraction branch adopts three convolutional blocks with residual connection to process two-dimensional image features , and outputs a feature vector after global average pooling, and the calculation formula is as follows:

[0027]

[0028]

[0029]

[0030]

[0031]

[0032] wherein, is a 1x1 convolution kernel, represents a step, represents padding, is a residual input after adjusting the channel by the 1x1 convolution kernel, is an output feature map of the first convolutional block, is a 3x3 convolution kernel, is a convolution operation, is an activation function, is a batch normalization operation, is a max-pooling operation; is Residual input after adjusting channels by 1x1 convolution, is a 1x1 convolution kernel, is the feature map output by the second convolution block, is a 3x3 convolution kernel, is Residual input after adjusting channels by 1x1 convolution, is a 1x1 convolution kernel, is the feature map output by the third convolution block, is a 3x3 convolution kernel; Then the feature map output by the third convolution block Global average pooling is performed on the image feature vector, and the formula is as follows:

[0033] wherein, is the image feature vector, is the global average pooling operation.

[0034] As an embodiment of the present application, the step S42 specifically comprises: S421: Based on the key frequency point set The electrical signal features are weighted and processed, and the formula is as follows:

[0035] wherein, is the element product, is the electrical impedance value feature vector, is the electrical impedance weight vector, is the amplitude feature vector, is the amplitude weight vector, is the weighted and processed electrical signal feature; S422: Based on the high-value area pixel set The image features are weighted and processed, and the high-value area pixel set corresponding features are focused, and the spatial attention mechanism is used to enhance the area feature weight, and the formula is as follows:

[0036] wherein, is the global average pooling, is the global maximum pooling, is the channel splicing, is the 1x1 convolution, is the activation function; is the weighted and processed two-dimensional image feature.

[0037] As an embodiment of the present application, the step S43 specifically comprises: S431: Uniformly weighted processed electrical signal features , and image features of dimensions and data types, feature fusion is achieved through element addition, and the formula is:

[0038] wherein, is the final fused feature, which comprehensively reflects the correlation between the frequency law of the electrical signal and the spatial pattern of the image; S432: Input the fused feature into the fully connected layer, and output the tumor probability through activation, and the formula is as follows:

[0039] wherein, is the output layer weight, is the bias, is the probability that the sample is a tumor tissue.

[0040] The beneficial effects of the present application are: (1) The present application establishes a traceable mapping relationship between two-dimensional image features and original bioelectrical impedance data, so that any area in the image can be traced back to the frequency point combination of the original electrical signal. By extracting the key frequency points corresponding to the high-value area of the image, the core frequency information related to the characteristics of the tumor can be locked, and the precise tracing from the "image spatial pattern" to the "electrical signal frequency law" can be realized, thereby fundamentally solving the cross-modal "pseudo-correlation" problem and providing a clear target for feature enhancement.

[0041] (2) The present application performs weighted processing on the electrical signal features and two-dimensional image features based on the key frequency point set and the high-value area pixel set, which can accurately trace the high-value area of the image to the key frequency points of the original electrical signal, and can also adaptively strengthen the real correlation features and suppress noise interference. In the scene where the modal data is slightly missing or the sample heterogeneity is high, the detection accuracy can still remain stable. Through the "space-frequency" dual-dimension focusing, hierarchical feature mining from macro image pattern to micro electrical signal anomaly is realized. BRIEF DESCRIPTION OF DRAWINGS

[0042] Figure 1 is a flowchart of a bioelectrical impedance tumor detection method based on traceable mapping relationship provided in an embodiment of the present application; Figure 2 is a two-dimensional image feature conversion flowchart of a bioelectrical impedance tumor detection method based on traceable mapping relationship provided in an embodiment of the present application; Figure 3A traceability mapping relationship schematic diagram of a multi-modal based bioelectrical impedance tumor detection method provided in an embodiment of the present application; Figure 4 A cross-modal correlation enhanced fusion network structure schematic diagram of a multi-modal based bioelectrical impedance tumor detection method provided in an embodiment of the present application. DETAILED DESCRIPTION

[0043] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0044] Reference Figures 1-4 The first aspect of the present application provides a bioelectrical impedance tumor detection method based on traceability mapping relationship, which comprises the following steps: S1: acquiring bioelectrical impedance data of tumor tissue and normal tissue samples, the bioelectrical impedance data comprising electrical impedance values and amplitude values at different frequencies; The present application collects electrical signal data of tumor and normal tissues through a multi-frequency bioelectrical impedance measurement system, and uses the differences in cell membrane permeability and ion concentration between tumor tissue and normal tissue, so that the electrical impedance values and amplitude values will show specific changes at different frequencies. These raw data contain basic information of the electrical physiological characteristics of the tissue, and are the core basis for subsequent cross-modal analysis, providing quantitative indicators for the identification of tumor and normal tissue.

[0045] S2: measuring the acquired effective sample electrical impedance values and amplitude values to obtain an electrical impedance value matrix and an amplitude value matrix, and performing standardization processing thereon, constructing a two-dimensional feature matrix based on the standardization-processed electrical impedance value matrix and amplitude value matrix, and converting the two-dimensional feature matrix into a two-dimensional image feature; Specifically, the present application can eliminate the dimensional difference and noise interference of the raw data by standardizing the electrical impedance value matrix and the amplitude value matrix, and convert the electrical signal into a two-dimensional image feature, which can map the abstract frequency domain information into an intuitive spatial domain mode, facilitating the mining of the coupling relationship of the electrical signal between different frequency points and providing a visual carrier for cross-modal correlation.

[0046] S3: establishing a traceability mapping relationship between the two-dimensional image feature and the original bioelectrical impedance data; segmenting the high-value area of the two-dimensional image to extract a high-value area pixel set; and obtaining a key frequency point set through coordinate traceability conversion; S4: constructing a cross-modal correlation enhanced fusion network based on the traceability mapping relationship, fusing and detecting the electrical impedance value features, amplitude value features and two-dimensional image features, and outputting a tumor detection result.

[0047] As an embodiment of the present application, the step S2 of measuring the obtained effective electrical impedance value and amplitude value to obtain an electrical impedance value matrix and an amplitude value matrix, and performing standardization processing thereon specifically includes: S211: calculating a statistical quantity of each frequency point, measuring the effective sample set electrical impedance value and amplitude value to obtain an electrical impedance value matrix and an amplitude value matrix ; S212: performing standardization processing on the electrical impedance value matrix and the amplitude value matrix , and the calculation formula is as follows:

[0048]

[0049] wherein, is a standardized electrical impedance value of the i-th effective sample at the j-th frequency point, is an original electrical impedance value of the i-th effective sample at the j-th frequency point, is a minimum value of the j-th frequency point in the electrical impedance matrix of all effective samples, is a maximum value of the j-th frequency point in the electrical impedance matrix of all effective samples; is a standardized amplitude value of the i-th effective sample at the j-th frequency point, is an original amplitude value of the i-th effective sample at the j-th frequency point, is a minimum value of the j-th frequency point in the amplitude value matrix of all effective samples, is a maximum value of the j-th frequency point in the amplitude value matrix of all effective samples. Specifically, the present application maps the electrical signal data to a unified interval through standardization processing, eliminates the magnitude difference between different frequency points, makes the features of each frequency point comparable, and provides stable input for subsequent two-dimensional matrix construction and image conversion.

[0050] Specifically, the present application maps the electrical signal data to a unified interval through standardization processing, eliminates the magnitude difference between different frequency points, makes the features of each frequency point comparable, and provides stable input for subsequent two-dimensional matrix construction and image conversion.

[0051] ​​​​​​​​​​​As an embodiment of this application, step S2, which involves constructing a two-dimensional feature matrix based on the standardized impedance matrix and amplitude matrix, and converting it into two-dimensional image features, specifically includes: S221: Based on the standardized electrical reactance matrix and magnitude matrix, construct... Two-dimensional feature matrix ,in, The matrix elements represent the total number of frequency points collected, and the formula for calculating the matrix elements is as follows:

[0052] in, For the matrix of the first Line number Column elements, For the first The number of valid samples in the th... Standardized impedance values ​​at each frequency point For the first The number of valid samples in the th... Standardized amplitude at each frequency point; S222: Using bilinear interpolation to transform the two-dimensional feature matrix Adjusted to Two-dimensional feature matrix To meet the feature extraction requirements of image modalities, pixel values ​​are calculated and converted into two-dimensional image features. The calculation formula is as follows:

[0053] in, This indicates rounding to the nearest integer. For image coordinates, For the two-dimensional image Line number The pixel values ​​of the columns, the pixel value matrix Stored as a grayscale image, used as a feature of the two-dimensional image.

[0054] Specifically, the two-dimensional feature matrix captures the coordinated changes in impedance and amplitude at different frequency points through cross-product. When converted into an image, this coupling relationship is presented intuitively in the form of pixel brightness, which facilitates the location of key areas through image segmentation technology and provides spatial anchors for subsequent source tracing mapping.

[0055] As an embodiment of this application, the formula for establishing the source mapping relationship between two-dimensional image features and original bioelectrical impedance data in step S3 is expressed as follows:

[0056]

[0057] in, For source mapping functions, This is a round-down operation; For the original frequency point index, ; Image coordinates Compared with the original frequency point index The corresponding relationships are stored as a dictionary, forming a source mapping table.

[0058] As an embodiment of this application, step S3, which involves segmenting the two-dimensional image into high-value regions and extracting the pixel set of high-value regions, specifically includes: S321: For two-dimensional images set of pixel values Iterate through all possible thresholds Calculate the inter-class variance The formula is as follows:

[0059] in, pixel value The regional proportion; pixel value The regional proportion, ; pixel value The regional mean; pixel value The regional mean; S322: Determine the segmentation threshold and take a value that minimizes the inter-class variance. The largest As the optimal threshold, the set of pixels in the high-value region is extracted using the following formula:

[0060] in, For the first The set of high-value pixels in the image of each sample; Indicates the first in the high value region The coordinates of a pixel. Indicates the first The image of each sample is in coordinates The pixel value at that location.

[0061] Specifically, this invention automatically identifies high-value regions by maximizing inter-class variance, ensuring that the segmented regions are the most discriminative parts of the image. This avoids the subjectivity of manually setting thresholds and provides objective feature regions for subsequent source tracing. As an embodiment of this application, the step S3 of obtaining the key frequency point set through coordinate tracing transformation specifically includes: S311: Convert the pixel coordinates of high-value areas in the image into frequency point indices of the original electrical signal through coordinate source transformation. For each pixel coordinate... The corresponding frequency point index is obtained through the source mapping table. The coordinate source transformation formula is as follows:

[0062] in, Indicates the first in the high value region The coordinates of a pixel. For source mapping functions, Represents the first obtained after mapping Index of original electrical signal frequency points; S332: Next, the index is traversed and duplicate original electrical signal frequency point indices are removed through deduplication to obtain the key frequency point set. ,in, Represents the first obtained after mapping Index of the original electrical signal frequency points, This indicates the number of frequency point index pairs after deduplication.

[0063] Specifically, this invention, through source mapping and deduplication, associates high-value regions in an image with the core frequency points of the original electrical signal. These frequency points are direct carriers of abnormal tumor electrical properties, providing precise targets for subsequent feature enhancement and preventing the model from wasting computational power on irrelevant frequency points. Furthermore, source mapping directly correlates with changes in electrical signals at specific frequency points, the physical meaning of which can be traced back to tissue electrophysiological characteristics. This addresses the pain point of "uninterpretable features" in traditional deep learning models, providing quantifiable and interpretable evidence for clinical result verification and mechanism research, and accelerating the clinical translation of AI technology in the field of tumor detection.

[0064] As an embodiment of this application, step S4 specifically includes: S41: Construct a cross-modal correlation enhancement fusion network, which includes a resistance value feature extraction branch, an amplitude feature extraction branch, and a two-dimensional image feature extraction branch, and respectively extracts features from resistance value features, amplitude features, and two-dimensional image features; S42: Weighted processing of electrical signal features and two-dimensional image features based on key frequency point set and high value region pixel set; S43: Perform cross-modal feature fusion on the weighted electrical signal features and two-dimensional image features to output tumor detection results.

[0065] Specifically, the application extracts features of each mode through a cross-modal correlation enhanced fusion network, and then uses the key frequency points and high-value areas obtained through tracing to perform correlation enhancement, so that the model focuses on real and effective cross-modal correlation features, and finally realizes the complementation of multi-dimensional information through fusion. This mechanism not only retains the frequency specificity of the electrical signal, but also integrates the spatial correlation of the image, significantly improving the accuracy and robustness of tumor detection.

[0066] As an embodiment of the present application, the step S41 specifically comprises: S411: The electrical impedance feature extraction branch adopts a 3-layer multi-layer perception network to process the standardized electrical impedance value , and extracts features through three full connection layers in turn, and the calculation formula is as follows:

[0067]

[0068]

[0069] Among them, is the output feature of the first full connection layer, and the dimension is 256; is the weight matrix of the first full connection layer, is the bias of the first full connection layer, is the activation function, is the batch normalization operation, is the random inactivation operation; is the output feature of the second full connection layer, and the dimension is 192; is the weight matrix of the second full connection layer, is the bias of the second full connection layer, is the electrical impedance feature vector output by the third full connection layer, and the output dimension is 128; is the weight matrix of the third full connection layer, is the bias of the third full connection layer; S412: The amplitude feature extraction branch adopts a 3-layer multi-layer perception network symmetrical to the electrical impedance feature extraction branch to process the standardized amplitude , and extracts features through three full connection layers in turn, and finally outputs the amplitude feature vector ; S413: The image feature extraction branch adopts three convolution blocks with residual connection to process two-dimensional image features , and outputs the feature vector through global average pooling, and the calculation formula is as follows:

[0070]

[0071]

[0072]

[0073]

[0074]

[0075] wherein, is a 1x1 convolution kernel, and the number is 32; represents a step size, represents padding, is The residual input after adjusting the channel by the 1x1 convolution kernel is matched with the channel number . is the output feature map of the first convolution block, the channel number is 32, and the size is 112x112; is a 3x3 convolution kernel, and the number is 32; is a convolution operation, is an activation function, is a batch normalization operation, is a max pooling operation; is The residual input after adjusting the channel by the 1x1 convolution kernel, is a 1x1 convolution kernel, and the number is 64; is the output feature map of the second convolution block, the channel number is 64, and the size is 56x56; is a 3x3 convolution kernel, and the number is 64; is The residual input after adjusting the channel by the 1x1 convolution kernel, is a 1x1 convolution kernel, and the number is 128; is the output feature map of the third convolution block, the channel number is 128, and the size is 56x56; is a 3x3 convolution kernel, and the number is 128; Then the output feature map of the third convolution block is globally averaged to output an image feature vector, and the formula is as follows:

[0076] wherein, is an image feature vector, and the output dimension is 128; is a global average pooling operation, which compresses a 28x28x128 feature map into a 1x128 feature vector.

[0077] Specifically, the cross-modal correlation enhanced fusion network constructed by the application is designed with different structures for different modal characteristics: the MLP branch captures the frequency domain law of the electrical signal through multi-layer nonlinear transformation, the convolution block branch extracts the spatial hierarchical features of the image through residual connection and pooling operation, and each branch is normalized through batch normalization and Inhibition of overfitting ensures the stability and effectiveness of feature extraction.

[0078] As an embodiment of the present application, the step S42 specifically comprises: S421: based on the key frequency point set The electrical signal features are weighted to highlight the key frequency point corresponding features, generate the electrical impedance weight vector and the amplitude weight vector, and the calculation formula is as follows:

[0079] Among them, is the element product, is the electrical impedance value feature vector, is the electrical impedance weight vector, is the amplitude feature vector, is the amplitude weight vector, is the weighted electrical signal feature; if the feature component corresponds to the frequency point index in the key frequency point set , , otherwise .

[0080] S422: based on the high value area pixel set The image features are weighted to focus on the high value area pixel set corresponding features, and the spatial attention mechanism is used to enhance the regional feature weight, and the calculation formula is:

[0081] Among them, is the global average pooling, is the global maximum pooling, is the channel splicing, is the 1x1 convolution, is the activation function; is the weighted two-dimensional image feature.

[0082] The application enhances the weight by correlation, so that the model automatically focuses on the key frequency points and high value area features obtained by tracing, and suppresses irrelevant information interference. This "targeted enhancement" mechanism solves the problem of feature weight averaging in traditional fusion, so that tumor-specific features dominate in fusion.

[0083] As an embodiment of this application, step S43 specifically includes: S431: Characteristics of electrical signals after unified weighting , Image features Given the dimensions and data types, feature fusion is achieved through element-wise addition, with the formula as follows:

[0084] in, The final fusion feature comprehensively reflects the correlation between the frequency law of the electrical signal and the spatial pattern of the image; S432: Merge features Input to fully connected layer, after The activation formula for outputting tumor probability is as follows:

[0085] in, For output layer weights, For bias, This represents the probability that the sample is tumor tissue.

[0086] Specifically, element-wise additive fusion allows modal features to directly complement each other in the same dimension (such as superimposing the frequency intensity of an electrical signal with the spatial location information of an image), while The function maps fused features to tumor probabilities, achieving a transformation from the feature space to the decision space. This fusion method not only preserves the core information of each modality but also ensures feature consistency through correlation enhancement, ultimately significantly improving the accuracy, sensitivity, and specificity of tumor detection.

[0087] 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. For example, technical solutions formed by substituting the above-described features with (but not limited to) technical features with similar functions disclosed in the embodiments of this disclosure.

Claims

1. A bioelectrical impedance tomography method based on a traceability mapping relationship for tumor detection, characterized in that, The method comprises the following steps: S1: acquiring bioelectrical impedance data of tumor tissue and normal tissue samples, the bioelectrical impedance data comprising electrical impedance values and amplitude values at different frequencies; S2: measuring the acquired effective sample electrical impedance values and amplitude values to obtain an electrical impedance value matrix and an amplitude value matrix, and performing standardization processing thereon, constructing a two-dimensional feature matrix based on the standardized electrical impedance value matrix and the amplitude value matrix, and converting the two-dimensional feature matrix into a two-dimensional image feature; S3: establishing a traceability mapping relationship between the two-dimensional image feature and the original bioelectrical impedance data; segmenting a high-value area of the two-dimensional image to extract a high-value area pixel set; and obtaining a key frequency point set through coordinate traceability conversion; S4: constructing a cross-modal correlation enhanced fusion network based on the traceability mapping relationship, fusing and detecting the electrical impedance value feature, the amplitude value feature and the two-dimensional image feature, and outputting a tumor detection result.

2. The bioelectrical impedance tomography method based on traceability mapping relationship for tumor detection according to claim 1, characterized in that, The step S2 of measuring the acquired effective electrical impedance values and amplitude values to obtain an electrical impedance value matrix and an amplitude value matrix, and performing standardization processing thereon specifically comprises: S211: Calculate the statistics of each frequency point, measure the resistance and impedance value and amplitude value of the effective sample set, and obtain the resistance and impedance value matrix and amplitude matrix ; S212: normalizing the electrical impedance value matrix and the amplitude matrix S213: calculating the normalized electrical impedance value matrix and the normalized amplitude matrix in, For the first The number of valid samples in the th... Standardized impedance values ​​at each frequency point For the first The number of valid samples in the th... The original impedance value at each frequency point It is the first The minimum value of a frequency point in the electrical impedance matrix of all effective samples. It is the first The maximum value of a frequency point in the electrical impedance matrix of all valid samples; For the first The number of valid samples in the th... Standardized amplitude at each frequency point For the first The number of valid samples in the th... The original amplitude at each frequency point It is the first The minimum value of a frequency point in the amplitude matrix of all valid samples. It is the first The maximum value of a frequency point in the amplitude matrix of all valid samples.

3. The bioelectrical impedance tomography method based on traceability mapping relationship for tumor detection according to claim 1, characterized in that, The step S2 of constructing a two-dimensional feature matrix based on the standardized electrical impedance value matrix and the amplitude value matrix, and converting the two-dimensional feature matrix into a two-dimensional image feature specifically comprises: S221: Construct a two-dimensional feature matrix based on the standardized resistance and impedance value matrix and the amplitude matrix The matrix element calculation formula is as follows: wherein, is the element of the matrix in the row and the column, is the normalized impedance value of the effective sample at the frequency point, is the normalized amplitude value of the effective sample at the frequency point; S222: Using bilinear interpolation to transform the two-dimensional feature matrix Adjust to size Two-dimensional feature matrix Then, the pixel values ​​are calculated and converted into two-dimensional image features. The calculation formula is as follows: wherein, denotes rounding off; is an image coordinate, is a pixel value of the two-dimensional image at the i-th row and the j-th column, is a pixel value of the two-dimensional image at the i-th row and the j-th column, is a pixel value of the two-dimensional image at the i-th row and the j-th column, is a two-dimensional image feature after the pixel value matrix is converted.

4. The bioelectrical impedance tomography method based on traceability mapping relationship for tumor detection according to claim 1, characterized in that, The formula for establishing the traceability mapping relationship between the two-dimensional image feature and the original bioelectrical impedance data in the step S3 is as follows: wherein, is a trace mapping function, is a floor operation; is an original frequency point index, ; The image coordinates correspondence with the original frequency point index are stored as a dictionary to form a traceability mapping table.

5. The bioelectrical impedance tomography method based on traceability mapping relationship for tumor detection according to claim 1, characterized in that, The step S3 of segmenting a high-value area of the two-dimensional image to extract a high-value area pixel set specifically comprises: S321: For two-dimensional images set of pixel values Iterate through all thresholds Calculate the inter-class variance The formula is as follows: wherein, a region proportion of pixel values a region proportion of pixel values a region proportion of pixel values a region proportion of pixel values a region mean of pixel values a region mean of pixel values a region mean of pixel values a region mean of pixel values S322: determine the segmentation threshold, take the inter-class variance the largest As the optimal threshold, extract the high-value region pixel set, the formula is as follows: wherein, is a set of image high value region pixels of the th sample; represents the coordinates of the th pixel in the high value region, represents the pixel value of the th sample at coordinates .

6. The bioelectrical impedance tomography method based on traceability mapping relationship for tumor detection according to claim 1, characterized in that, The step S3 of obtaining a key frequency point set through coordinate traceability conversion specifically comprises: S311: convert the image high value area pixel coordinates into the frequency point indexes of the original electric signal by coordinate tracing conversion, for each pixel coordinate , get the corresponding frequency point index through the tracing mapping table, and the coordinate tracing conversion formula is as follows: wherein, denotes the coordinates of the i-th pixel in the high-value region, is a trace mapping function, denotes the i-th original electric signal frequency point index obtained by mapping, denotes the i-th original electric signal frequency point index obtained by mapping,​ S332: Then, the repeated original electric signal frequency point indexes are removed by a deduplication processing operation to obtain a key frequency point set wherein, represents the mapped original electric signal frequency point index, represents the number of deduplicated frequency point index pairs.​ 7. The bioelectrical impedance tomography method based on traceability mapping relationship for tumor detection according to claim 1, characterized in that, The step S4 specifically comprises: S41: constructing a cross-modal correlation enhanced fusion network, the cross-modal correlation enhanced fusion network comprising an electrical impedance value feature extraction branch, an amplitude value feature extraction branch and a two-dimensional image feature extraction branch, and performing feature extraction on the electrical impedance value feature, the amplitude value feature and the two-dimensional image feature, respectively; S42: performing weighted processing on the electrical signal feature and the two-dimensional image feature based on the key frequency point set and the high-value area pixel set; S43: performing cross-modal feature fusion on the weighted electrical signal feature and the two-dimensional image feature, and outputting a tumor detection result.

8. The bioelectrical impedance tomography method based on traceability mapping relationship for tumor detection according to claim 7, characterized in that, The step S41 specifically comprises: S411: the electrical impedance feature extraction branch adopts a 3-layer multi-layer perception network to process the standardized electrical impedance values , and sequentially extracts features through three fully connected layers, and the calculation formula is as follows: wherein, is an output feature of the first layer fully connected layer, is a weight matrix of the first layer fully connected layer, is a bias of the first layer fully connected layer, is an activation function, is a batch normalization operation, is a random deactivation operation; is an output feature of the second layer fully connected layer, is a weight matrix of the second layer fully connected layer, is a bias of the second layer fully connected layer, is an electrical impedance feature vector output by the third layer fully connected layer, is a weight matrix of the third layer fully connected layer, is a bias of the third layer fully connected layer; S412: the amplitude feature extraction branch adopts a 3-layer multi-layer perception network symmetric to the electrical impedance feature extraction branch to process the standardized amplitude , sequentially passes through three fully connected layers to extract features, and finally outputs an amplitude feature vector ; S413: the image feature extraction branch adopts three convolution blocks with residual connection to process two-dimensional image features The global average pooling output feature vector is calculated according to the following formula: wherein, is a 1x1 convolution kernel, denotes a step size, denotes padding, is a residual input after adjusting channels by a 1x1 convolution kernel, is a feature map output by a first convolution block, is a 3x3 convolution kernel, is a convolution operation, is an activation function, is a batch normalization operation, is a max-pooling operation; is a residual input after adjusting channels by a 1x1 convolution, is a 1x1 convolution kernel, is a feature map output by a second convolution block, is a 3x3 convolution kernel, is a residual input after adjusting channels by a 1x1 convolution, is a 1x1 convolution kernel, is a feature map output by a third convolution block, is a 3x3 convolution kernel; Then the output feature map of the third convolutional block is The global average pooling is performed on the output image feature vector, and the formula is as follows: wherein, is an image feature vector, is a global average pooling operation.

9. The bioelectrical impedance tomography method based on traceability mapping relationship for tumor detection according to claim 7, characterized in that, The step S42 specifically comprises: S421: based on the key frequency point set The electric signal characteristics are weighted, and the calculation formula is as follows: wherein, is an element product, is an electrical impedance value feature vector, is an electrical impedance weight vector, is an amplitude value feature vector, is an amplitude weight vector, is a weighted processed electrical signal feature; S422: based on the high-value region pixel set Weighted processing of image features, focusing on high-value region pixel set Corresponding features, enhanced by spatial attention mechanism Region feature weight, the calculation formula is: wherein, is a global average pooling, is a global max pooling, is a channel concatenation, is a 1x1 convolution, is an activation function; is a two-dimensional image feature after weighting processing.

10. The bioelectrical impedance tomography method based on traceability mapping relationship for tumor detection according to claim 9, characterized in that, The step S43 specifically comprises: S431: Characteristics of electrical signals after unified weighting , Image features Given the dimensions and data types, feature fusion is achieved through element-wise addition, with the formula as follows: wherein, is the final fusion feature, which comprehensively reflects the correlation between the frequency law of the electrical signal and the image spatial pattern. S432: fuse features input fully connected layer, pass through activation output tumor probability, formula as follows: wherein, is an output layer weight, is a bias, is a probability that the sample is a tumor tissue.

Images