A Deep Learning-Based Bioelectrical Impedance Cancer Detection Method

Abstract

The present invention discloses a method for detecting cancer by bioelectrical impedance based on deep learning, comprising the following steps: S1: obtaining the electrical signal data of benign and malignant samples of cancer tissues by using an electrical impedance imaging system and combining it with physiological data; S2: extracting global features of the electrical signals by using a matrix decomposition module; S3: taking the result of global feature extraction and the original electrical signal data together as new inputs and further extracting signal features by using data dependency; S4: integrating the extracted feature maps into a final classification result. The present invention uses an electrical impedance imaging system to obtain the electrical signal data of cancer tissues, performs global feature extraction through a matrix decomposition module, and at the same time takes the result of global feature extraction and the original electrical signal data together as new inputs, and further extracts signal features by using data dependency, which can improve the utilization rate of data and further enhance the accuracy and reliability of cancer detection.

Description

Technical Field

[0001] The present invention relates to the technical field of cancer detection, and in particular to a bioelectrical impedance cancer detection method based on deep learning. Background Art

[0002] As one of the major diseases seriously threatening human health in the world today, the early and accurate detection of cancer is crucial for improving the survival rate of patients and the treatment effect. Traditional cancer detection methods mainly include tissue biopsy, imaging examination, and tumor marker detection, etc. Although tumor marker detection is relatively convenient, its specificity and accuracy are affected by various factors. A single marker often cannot achieve accurate diagnosis, and combined detection increases complexity and false positive rate.

[0003] In the field of medical health, deep learning has been tried and applied to medical image diagnosis, disease prediction, etc., and has shown advantages over traditional methods. Introducing deep learning technology into bioelectrical impedance cancer detection is expected to overcome the short - comings of traditional methods in feature extraction and pattern recognition. By learning and training a large amount of bioelectrical impedance data of different individuals, different cancer types and stages, a highly accurate cancer detection model can be constructed, so as to achieve more sensitive, accurate and non - invasive early cancer screening and diagnosis, and open up a new way for cancer prevention and treatment.

[0004] In the prior art, Chinese Patent No. CN117481630A discloses "a breast cancer detection method based on bioelectrical impedance analysis method". This method uses bioelectrical impedance analysis method to obtain the electrical signal data of breast tissue, and detects the data through a multi - dimensional feature extraction network module. At the same time, by designing a segmentation mechanism and a channel synthesis network module, the effectiveness and richness of the data can be improved. However, this method ignores the temporal features in the data and fails to effectively capture the dynamic changes during the tumor growth process, while the detection method based on deep - learning bioelectrical impedance technology provides a new solution for cancer detection.

[0005] Therefore, a bioelectrical impedance cancer detection method based on deep learning is proposed, which can more effectively extract the electrical features of cancer and overcome the limitations of the prior art. Summary of the Invention

[0006] Aiming at the above - mentioned defects or improvement requirements of the prior art, the present invention provides a bioelectrical impedance cancer detection method based on deep learning. Through a matrix decomposition module, global features of the electrical signal data of cancer tissues obtained are extracted. At the same time, the results of global feature extraction and the original electrical signal data are used as new inputs together, and the extracted feature maps are integrated into the final classification result, which can improve the utilization rate of data and further enhance the accuracy and reliability of cancer detection.

[0007] To achieve the above object, according to one aspect of the present invention, a method for detecting cancer by bioelectrical impedance based on deep learning is provided, and the method includes the following steps:

[0008] S1: Using an electrical impedance imaging system to obtain electrical signal data of benign and malignant samples of cancer tissues and combine it with physiological data; the electrical signal data includes electrical parameters, impedance data, and frequency response data;

[0009] S2: Using a matrix factorization module to extract global features from the electrical signals;

[0010] S3: Using the result of the global feature extraction and the original electrical signal data as new inputs, and further extracting electrical signal features using data dependencies;

[0011] S4: Integrating the extracted feature maps into a final classification result.

[0012] As an embodiment of the present application, the step S1 specifically includes the following steps:

[0013] S11: Associating different samples with physiological data, where the physiological data includes age, gender, personal bad habits, and body mass index;

[0014] S12: Measuring the impedance amplitude and phase at multiple frequencies, and plotting a frequency response curve to reflect the response of the tissue to currents of different frequencies.

[0015] As an embodiment of the present application, the step S2 specifically includes:

[0016] S21: First, the input electrical signal data is processed by a 1X1 convolutional layer for time series data;

[0017] S22: Decomposing the electrical signal processed by the 1X1 convolutional layer into a low-rank factor and a basis sequence;

[0018] S23: Using a temporal convolutional network to extract time-dependent features from the basis sequence and combining with the low-rank factor for global feature fusion to achieve the extraction of global features.

[0019] As an embodiment of the present application, the calculation formula for decomposing the electrical signal processed by the 1X1 convolutional layer into a low-rank factor and a basis sequence in the step S22 is as follows:

[0020]

[0021]

[0022] Among them, is the input electrical signal data, is a low-rank factor used to capture the global correlation between signals, is a basis sequence used to reflect the time pattern of the signal, is the dimension of the electrical signal, is the number of time steps.

[0023] As an embodiment of the present application, the step S23 specifically includes:

[0024] S231: Input the basis sequence into the temporal convolutional network. Through the dilated convolution and residual connection of the temporal convolutional network, capture the long-term dependencies in the time dimension. The convolution operation formula for each layer is as follows:

[0025]

[0026] where, is the layer number, is the dilation factor of the current layer, is the weight of the convolutional kernel, is the output of the previous layer;

[0027] The residual connection operation formula is as follows:

[0028]

[0029] where, is the output of the layer,

[0030] S232: Fuse the low-rank factor and the basis sequence processed by the temporal convolutional network to form a complete global feature by concatenating them in the feature dimension , and its calculation formula is as follows:

[0031]

[0032]

[0033] where, is the concatenated global feature, and its dimension is , represents the original electrical signal dimension, represents the dimension of the global feature, represents the number of time steps, which contains information from the low-rank factor and the basis sequence .

[0034] As an embodiment of the present application, the step S3 specifically includes:

[0035] S31: Concatenate and fuse the results of global feature extraction with the original electrical signal data;

[0036] S32: Add position encoding to the data after concatenation and fusion obtained in step S31 to obtain the final input data;

[0037] S33: Input the final input data obtained in step S32 into a model constructed based on a deep learning framework to further extract global features.

[0038] As an embodiment of the present application, the calculation formula for concatenating and fusing the global feature with the original electrical signal data in step S31 is as follows:

[0039]

[0040] where, is the newly input data, is the original electrical signal data, is the global feature;

[0041] Global feature and the original electrical signal data are concatenated in the channel dimension or time dimension. The newly input data after concatenation in the time dimension has a dimension of .

[0042] As an embodiment of the present application, step S32 specifically includes:

[0043] S321: Convert the input data into a suitable sequence format, concatenate according to time steps, and the number of time steps remains unchanged, and the feature dimension increases. The new input data is , where is the new feature dimension, representing the length of the feature vector for each time step;

[0044] S322: Use an absolute position encoder to add position information, generate the position information for each time step through a function, and the position encoding is calculated through sine and cosine functions. The calculation formula is as follows:

[0045]

[0046]

[0047] where, is the position encoding, is the time step, is the index of the feature dimension, is the size of the feature dimension. The positional encoding encodes each time step through sine and cosine functions.

[0048] Add the positional encoding to the new input data after fusion to integrate the time position information into the features and obtain the final input data , and its calculation formula is as follows:

[0049]

[0050] .

[0051] As an embodiment of the present application, the step S33 specifically includes:

[0052] S331: Input the processed input data into the model. First, through a linear transformation, convert the input data into queries, keys, and values, and its calculation formula is as follows:

[0053]

[0054]

[0055] Among them, is the learned weight matrix, are the query, key, and value respectively, is the dimension of the key, as the input of the self-attention mechanism to calculate the dependencies between each time step;

[0056] S332: Calculate the similarity between the query and the key, and use this similarity to weight the value. The calculation formula is as follows:

[0057]

[0058]

[0059] Among them, is the dimension of the key for scaling;

[0060] Use a function to convert the attention scores into weights, and its calculation formula is as follows:

[0061]

[0062] Among them, represents the query and the key the attention weight between them, and the final weighted sum value;

[0063] S333: Divide the query, key, and value into multiple heads for calculation, which can capture feature information in different subspaces in parallel. The calculation method for each head is the same as above. Finally, the outputs of all heads are concatenated, and the calculation formula is as follows:

[0064]

[0065]

[0066] Among them, is the number of heads, and the output of each head is , is the linear transformation matrix of the output, is the output;

[0067] S334: Use a neural network composed of two layers of linear transformations to further extract features, and use activation function after each layer of linear transformation. The calculation formula for the first layer of linear transformation is as follows:

[0068]

[0069] Among them, is the weight matrix of the first layer of linear transformation, is the bias term;

[0070] The calculation formula for the second layer of linear transformation is as follows:

[0071]

[0072] Among them, is the weight matrix of the second layer of linear transformation, is the bias term, is the output of the neural network;

[0073] S335: Perform residual connection and layer normalization on the output of the neural network to obtain the final output .

[0074] As an embodiment of the present application, the step S4 specifically includes:

[0075] After processing, the final output can be used for downstream tasks. The output formula for the classification task is as follows:

[0076]

[0077] Among them, is the output, The layer maps the output to the probability space for the classification task, The calculation formula for the layer is as follows:

[0078]

[0079] Among them, represents the output value of the model for class , represents the total number of classes, in represents the probability that the model predicts class for the given input ;

[0080] Each probability value output through the calculation of the layer represents the possibility that the input electrical signal belongs to each class; finally, the model will select the class with the highest probability as the classification result.

[0081] The beneficial effects of the present invention are as follows:

[0082] (1) The present invention uses an electrical impedance tomography system to obtain electrical signal data of cancer tissues. By measuring and analyzing the electrical signal data of cancer tissues, information about tissue electrical parameters, impedance data, and frequency response data can be obtained, and combined with the individual's physiological data (such as age, gender, body mass index, etc.), it helps to eliminate the interference caused by individual differences, can comprehensively reflect the electrical characteristics of tissues, helps the model better distinguish normal and cancerous tissues, and strengthens the generalization ability of the model by using multi-modal input. Through cross-modal feature learning, the bioelectrical impedance signal and the patient's physiological characteristics can be effectively utilized to improve the reliability and accuracy of detection.

[0083] (2) The present invention decomposes the electrical signal into low-rank factors and basis sequences through a matrix decomposition module, which can effectively extract global temporal features from the electrical signal data, capture the global correlation between time series, thereby providing rich feature information for subsequent cancer detection, and can extract the low-rank factors in the data to effectively remove noise and redundant information in the signal and improve the representation ability of features. Compared with traditional feature extraction methods, the present invention can better maintain temporal information, improve the performance of the model, and enhance its generalization ability in practical applications.

[0084] (3) The deep learning framework of the present invention has strong scalability, can process a large amount of electrical signal data and multi-dimensional physiological information, and is suitable for application in large-scale data analysis in clinical practice. Especially in early cancer screening and health monitoring, it can achieve rapid and accurate cancer detection and provide effective support for clinical diagnosis.

[0085] (4) By combining the deep learning model with the bioelectrical impedance detection technology, compared with the traditional bioelectrical impedance detection method, the present invention can not only improve the diagnostic accuracy, but also reduce the manual intervention, has strong technological innovation and application potential, and is applicable to the early screening and detection of various cancers such as cancer. BRIEF DESCRIPTION OF THE DRAWINGS

[0086] Figure 1 It is a schematic flow chart of a deep learning-based bioelectrical impedance cancer detection method provided in an embodiment of the present invention;

[0087] Figure 2 It is a schematic flow chart of a matrix decomposition module of a deep learning-based bioelectrical impedance cancer detection method provided in an embodiment of the present invention;

[0088] Figure 3 It is a schematic diagram of feature fusion of a deep learning-based bioelectrical impedance cancer detection method provided in an embodiment of the present invention;

[0089] Figure 4 It is a schematic diagram of feature extraction of a deep learning-based bioelectrical impedance cancer detection method provided in an embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0090] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without making creative efforts belong to the scope of protection of the present invention.

[0091] It should be noted that all the directional indications (such as up, down, left, right, front, back...) in the embodiments of the present invention are only used to explain the relative position relationship and movement conditions between components in a specific posture (as shown in the drawings). If the specific posture changes, the directional indications will also change accordingly.

[0092] In the present invention, unless otherwise clearly specified and limited, the terms "connection", "fixation", etc. should be understood in a broad sense. For example, "fixation" can be a fixed connection, a detachable connection, or integrated; it can be a mechanical connection or an electrical connection; it can be directly connected or indirectly connected through an intermediate medium, and can be the communication inside two components or the interaction relationship between two components, unless otherwise clearly limited. For those of ordinary skill in the art, the specific meanings of the above terms in the present invention can be understood according to specific situations.

[0093] In addition, if the descriptions such as "first" and "second" are involved in the embodiments of the present invention, the descriptions of "first", "second", etc. are only for descriptive purposes and cannot be understood as indicating or implying their relative importance or implicitly specifying the quantity of the indicated technical features. Thus, the features defined with "first" and "second" may explicitly or implicitly include at least one such feature. In addition, the meaning of "and / or" appearing throughout the text includes three parallel scenarios. Taking "A and / or B" as an example, it includes Scenario A, or Scenario B, or the scenario where both A and B are satisfied simultaneously. In addition, the technical solutions between various embodiments can be combined with each other, but it must be based on the ability of those of ordinary skill in the art to implement. When the combination of technical solutions results in contradictions or cannot be implemented, it should be considered that such a combination of technical solutions does not exist and is not within the scope of protection required by the present invention.

[0094] Referring to Figures 1 - 4 , the first aspect of the present invention provides a method for detecting cancer by bioelectrical impedance based on deep learning, and the method includes the following steps:

[0095] S1: Using an electrical impedance tomography system to obtain the electrical signal data of benign and malignant samples of cancer tissues and combine it with physiological data; the electrical signal data includes electrical parameters, impedance data, and frequency response data;

[0096] The present invention uses an electrical impedance tomography system to obtain the electrical signal data of cancer tissues. By measuring and analyzing the electrical signal data of cancer tissues, information about tissue electrical parameters, impedance data, and frequency response data can be obtained. Combining with the physiological data of an individual (such as age, gender, body mass index, etc.) helps to eliminate the interference caused by individual differences, can comprehensively reflect the electrical characteristics of tissues, and helps the model better distinguish normal and cancerous tissues.

[0097] S2: Using a matrix decomposition module to extract global features from the electrical signals. The matrix decomposition module can effectively decompose the global time-series features of the electrical signal data, capture the long-term dependence relationship and global trend in the electrical signals, so as to better understand the overall change of the electrical signals. Through the extraction of global features, the model can maintain good performance on different data sets, avoid the overfitting problem, and enhance its generalization ability in practical applications.

[0098] S3: Using the result of the global feature extraction and the original electrical signal data together as new inputs to further extract electrical signal features using data dependence relationships;

[0099] In the present invention, by combining the global features with the original signal data, the model can comprehensively consider the global features and local details, enhancing the ability to understand and analyze electrical signals. This fusion method can better capture cancer-related features, further extract electrical signal features using data dependency relationships, enabling the model to adaptively select the most important features and further improving the classification performance.

[0100] S4: Integrate the extracted feature maps into the final classification result.

[0101] In the present application, the global features provide a macroscopic understanding of the signal, while the original signal data is regarded as local information, and the extracted original signal features as local features focus on the microscopic details of the signal; the combination of the two can make up for the limitations of a single feature, and enable the model to not only consider long-term patterns during classification, but also keenly capture small, local abnormal changes, thus better identifying cancer signals.

[0102] The present invention uses an electrical impedance imaging system to obtain electrical signal data of cancer, extracts global features of the electrical signal through a matrix decomposition module, uses the result of the global feature extraction and the original electrical signal data together as a new input, further extracts electrical signal features using data dependency relationships, and integrates the extracted feature maps into the final classification result. The present invention has the advantages of high-precision data acquisition, effective signal processing, and the integration of global and local features, and is expected to play an important role in the early detection and diagnosis of cancer.

[0103] As an embodiment of the present application, the step S1 specifically includes the following steps:

[0104] S11: Associate different samples with physiological data, where the physiological data includes age, gender, personal bad habits, and body mass index;

[0105] S12: Measure the electrical impedance amplitudes and phases at multiple frequencies, and plot a frequency response curve to reflect the response of the tissue to currents at different frequencies.

[0106] Specifically, by combining the physiological data of an individual (such as age, gender, body mass index, etc.), the present application can comprehensively reflect the electrical characteristics of tissues, helping the model better distinguish normal and cancerous tissues. By associating the electrical signal data with physiological factors, it helps to remove the influence of factors such as smoking and obesity on the electrical impedance signal, enabling the model to focus more on cancer features and improving the sensitivity of detection.

[0107] As Figure 2 shown, as an embodiment of the present application, the step S2 specifically includes:

[0108] S21: First, the input electrical signal data is processed for time series data through a 1X1 convolutional layer;

[0109] S22: Decompose the electrical signal processed by the 1X1 convolutional layer into low-rank factors and a basis sequence;

[0110] S23: Use a temporal convolutional network to extract time-dependent features from the basis sequence and perform global feature fusion in combination with the low-rank factors, achieving the extraction of global features.

[0111] As an embodiment of the present application, the calculation formula for decomposing the electrical signal processed by the 1X1 convolutional layer into low-rank factors and a basis sequence in step S22 is as follows:

[0112]

[0113]

[0114] Where, is the input electrical signal data, is the low-rank factor, used to capture the global correlation between signals, is the basis sequence, used to reflect the time pattern of the signal, which can help compress the data dimension and extract more representative global features, is the dimension of the electrical signal, is the number of time steps.

[0115] Specifically, through the matrix decomposition module of the present application, the global time series features of the electrical signal data can be effectively decomposed, and the long-term dependence relationship and global trend in the signal can be captured, so as to better understand the overall change of the signal.

[0116] As an embodiment of the present application, step S23 specifically includes:

[0117] S231: Pass the basis sequence into the temporal convolutional network. Through the dilated convolution and residual connection of the temporal convolutional network, capture the long-term dependence relationship in the time dimension. The convolution operation formula for each layer is as follows:

[0118]

[0119] Where, is the number of layers, is the dilation factor of the current layer, is the weight of the convolution kernel, is the output of the previous layer;

[0120] The residual connection operation formula is as follows:

[0121]

[0122] Among them, is the output of the layer,

[0123] S232: Fuse the low-rank factor and the basis sequence processed by the temporal convolutional network to form a complete global feature by concatenating them in the feature dimension , and its calculation formula is as follows:

[0124]

[0125]

[0126] Among them, is the concatenated global feature, and its dimension is , represents the dimension of the original electrical signal, represents the dimension of the global feature, represents the number of time steps, including information from the low-rank factor and the basis sequence , and this fusion process captures the correlation of the electrical signal in channels and time.

[0127] Specifically, the matrix decomposition module can extract the low-rank factor in the data, effectively remove the noise and redundant information in the signal, improve the representation ability of the feature. Through the extraction of the global feature, the model can maintain good performance on different data sets, avoid the overfitting problem, and enhance its generalization ability in practical applications.

[0128] As an embodiment of the present application, step S3 specifically includes:

[0129] S31: Concatenate and fuse the result of the global feature extraction with the original electrical signal data;

[0130] S32: Add position encoding to the data obtained after the concatenation and fusion in step S31 to obtain the final input data;

[0131] S33: Input the input data obtained in step S32 into the model constructed based on the deep learning framework, and its role is to further extract the global feature.

[0132] Such as Figure 3As shown in the figure, as an embodiment of the present application, the calculation formula for splicing and fusing the global feature and the original electrical signal data in step S31 is as follows:

[0133]

[0134] Among them, is the newly input data, is the original electrical signal data, is the global feature;

[0135] Global feature and the original electrical signal data are spliced in the channel dimension or the time dimension. The newly input data after splicing in the time dimension, the dimension is .

[0136] Specifically, by combining the global feature with the original signal data, the model can comprehensively consider the global feature and local details, improving the ability to understand and analyze electrical signals. This fusion method can better capture cancer-related features.

[0137] As an embodiment of the present application, step S32 specifically includes:

[0138] S321: Convert the input data into a suitable sequence format, splice it according to time steps, the number of time steps remains unchanged, the feature dimension increases, and the new input data is , where is the new feature dimension, representing the length of the feature vector for each time step;

[0139] S322: Use an absolute position encoder to add position information. Generate the position information for each time step through a function. The position encoding is calculated through sine and cosine functions, and its calculation formula is as follows:

[0140]

[0141]

[0142] Among them, is the position encoding, is the time step, is the index of the feature dimension, is the size of the feature dimension. The position encoding encodes each time step through sine and cosine functions;

[0143] Add the position encoding to the fused new input data to integrate the time position information into the feature and obtain the final input data , and its calculation formula is as follows:

[0144]

[0145] .

[0146] Specifically, the global features and the original electrical signal features are integrated into a unified input format to support subsequent model processing. After feature concatenation, the local characteristics of the original electrical signal are retained, and at the same time, the global correlation information in the time series is enhanced; the time position information is incorporated into the input features, enabling the model to recognize the order of different time steps in the input sequence; the periodic features of sine and cosine enable the model to effectively capture the long-range dependence relationships between time steps.

[0147] As Figure 4 shown, as an embodiment of the present application, step S33 specifically includes:

[0148] S331: Input the processed input data into the model. First, through a linear transformation, the input data is converted into queries, keys, and values, and its calculation formula is as follows:

[0149]

[0150]

[0151] Among them, is the learned weight matrix, are the matrices of queries, keys, and values respectively, is the dimension of the key, as the input of the self-attention mechanism, is used to calculate the dependence relationships between each time step;

[0152] S332: Calculate the similarity between the query and the key, and use this similarity to weight the values. The calculation formula is as follows:

[0153]

[0154]

[0155] Among them, is the dimension of the key, used for scaling;

[0156] The function is used to convert the attention scores into weights, and its calculation formula is as follows:

[0157]

[0158] Among them, represents the query The attention weight between the query and the key, and the final weighted sum value;

[0159] S333: Divide the query, key, and value into multiple heads for calculation, which can capture feature information in different subspaces in parallel. The calculation method for each head is the same as above. Finally, concatenate the outputs of all heads. The calculation formula is as follows:

[0160]

[0161]

[0162] Among them, is the number of heads, and the output of each head is , is the linear transformation matrix of the output, is the output;

[0163] S334: Use a neural network composed of two layers of linear transformations to further extract features, and use activation function after each layer of linear transformation. The calculation formula for the first layer of linear transformation is as follows:

[0164]

[0165] Among them, is the weight matrix of the first layer of linear transformation, is the bias term;

[0166] The calculation formula for the second layer of linear transformation is as follows:

[0167]

[0168] Among them, the weight matrix of the second layer of linear transformation, is the bias term, is the output of the neural network;

[0169] S335: Perform residual connection and layer normalization on the output of the neural network to obtain the final output .

[0170] Specifically, by calculating the similarity between the query and the key, the model can understand the dependencies between different time steps. This mechanism helps the model focus on other time steps related to the current time step, thereby improving the model's ability to learn long-range dependencies; through the multi-head mechanism, the model can understand the data from multiple perspectives, integrate the information of different attention heads, and improve the expressiveness and accuracy of the final output features. Through two layers of linear transformation (with ​With the activation function incorporated, the neural network can further extract and process the features after the attention mechanism. This process helps the model non-linearly map the complex electrical signal features, making the final representation more rich and profound.

[0171] As an embodiment of the present application, step S4 specifically includes:

[0172] After processing, the final output is obtained which can be used for downstream tasks. The output formula for the classification task is as follows:

[0173]

[0174] where, is the output, The layer maps the output to the probability space for the classification task, The calculation formula of the layer is as follows:

[0175]

[0176] where, represents the output value of the model for class , represents the total number of classes, in represents the probability that the given input , the model predicts class ;

[0177] Through Each probability value calculated and output by the layer represents the possibility that the input electrical signal belongs to each class; finally, the model will select the class with the highest probability as the classification result.

[0178] Specifically, through the processing of multiple encoder layers, the model can learn more abstract and profound feature representations from the input data, capturing the complex relationships in the data. These deep features will directly affect the result of the classification task. Finally, the output of the classification task is determined by the probability distribution generated by the layer, so that it can effectively classify the input electrical signal into different classes. Finally, the output of the classification task is determined by the probability distribution generated by the layer, so that it can effectively classify the input signal into different classes.

[0179] The above description is only some preferred embodiments of the present disclosure and an explanation of the applied technical principles. Those skilled in the art should understand that the scope of the invention involved in the embodiments of the present disclosure is not limited to the technical solutions formed by the specific combination of the above technical features, and should also cover other technical solutions formed by any combination of the above technical features or their equivalent features without departing from the above inventive concept. For example, the technical solutions formed by mutually replacing the above features with the technical features (but not limited to) having similar functions disclosed in the embodiments of the present disclosure.

Claims

1. A method for detecting cancer by bioelectrical impedance based on deep learning, characterized in that, The method includes the following steps: S1: Obtain the electrical signal data of benign and malignant samples of cancer tissues using an electrical impedance tomography system and combine it with physiological data; the electrical signal data includes electrical parameters, impedance data, and frequency response data; S2: Use a matrix decomposition module to extract global features from the electrical signals; S3: Take the result of the global feature extraction and the original electrical signal data together as new inputs, and further extract electrical signal features using data dependencies; S4: Map the extracted features into a final classification result; Integrate the global features and the original electrical signal features into a unified input format; after feature concatenation, retain the local characteristics of the original electrical signal and enhance the global correlation information in the time series; incorporate the time position information into the input features so that the model can recognize the order of different time steps in the input sequence; enable the model to effectively capture the long-range dependencies between time steps through sine and cosine functions.

2. The method for detecting cancer by bioelectrical impedance based on deep learning according to claim 1, wherein, The specific steps of step S1 include the following steps: S11: Associate different samples with physiological data, where the physiological data includes age, gender, personal bad habits, and body mass index; S12: Measure the impedance amplitude and phase at multiple frequencies and plot a frequency response curve to reflect the response of the tissue to currents at different frequencies.

3. The method for detecting cancer by bioelectrical impedance based on deep learning according to claim 1, wherein The specific steps of step S2 include: S21: First, the input electrical signal data is processed through a 1X1 convolutional layer for time series data; S22: Decompose the electrical signal processed by the 1X1 convolutional layer into low-rank factors and a basis sequence; S23: Use a temporal convolutional network to extract time-dependent features from the basis sequence and perform global feature fusion in combination with the low-rank factors to achieve the extraction of global features.

4. The method for detecting cancer based on bioelectrical impedance by deep learning according to claim 3, characterized in that, The calculation formula for decomposing the electrical signal processed by the 1X1 convolutional layer into low-rank factors and a basis sequence in step S22 is as follows: Among them, is the input electrical signal data, is the low-rank factor, which is used to capture the global correlation between signals, is the basis sequence, which is used to reflect the time pattern of the signal, is the dimension of the electrical signal, is the number of time steps.

5. The method for detecting cancer by bioelectrical impedance based on deep learning according to claim 4, wherein, The specific steps of step S23 include: S231: Input the base sequence into the temporal convolutional network. Through the dilated convolution and residual connection of the temporal convolutional network, capture the long-term dependencies in the temporal dimension. The convolution operation formula for each layer is as follows: Among them, is the number of layers, is the dilation factor of the current layer, is the weight of the convolutional kernel, is the output of the previous layer; The formula for the residual connection operation is as follows: Among them, is the output of the n-th layer and is the input of the previous layer; S232: Combine the low-rank factor and the basis sequence processed by the temporal convolutional network to form a complete global feature by concatenating them in the feature dimension , and its calculation formula is as follows: Among them, is the global feature after splicing, and its dimension is , represents the dimension of the original electrical signal, represents the dimension of the global feature, represents the number of time steps, which contains information from the low-rank factor and the basis sequence .

6. The method for detecting cancer by bioelectrical impedance based on deep learning according to claim 1, wherein The specific steps of step S3 include: S31: Concatenate and fuse the result of the global feature extraction and the original electrical signal data; S32: Add position encoding to the data obtained from the concatenation and fusion in step S31 to obtain the final input data; S33: Input the final input data obtained in step S32 into a model constructed based on a deep learning framework to further extract global features.

7. The method for detecting cancer by bioelectrical impedance based on deep learning according to claim 6, wherein The calculation formula for concatenating and fusing the global features and the original electrical signal data in step S31 is as follows: Among them, is the newly input data, is the original electrical signal data, is the global feature; Global features and the original electrical signal data are concatenated in the channel dimension or the time dimension. The new input data concatenated in the time dimension , and the dimension is .

8. The method for detecting cancer by bioelectrical impedance based on deep learning according to claim 7, wherein, The specific steps of step S32 include: S321: Convert the input data into an appropriate sequence format, concatenate it according to time steps, and the number of time steps remains unchanged, the feature dimension increases, and the new input data is , where is the new feature dimension, representing the length of the feature vector for each time step; S322: Add position information using an absolute position encoder, generate the position information for each time step through a function, and calculate the position encoding through sine and cosine functions. The calculation formula is as follows: Among them, is the positional encoding, is the time step, is the index of the feature dimension, is the size of the feature dimension, and the positional encoding encodes each time step through sine and cosine functions; Add the positional encoding to the new fused input data to incorporate the temporal position information into the features and obtain the final input data , and its calculation formula is as follows: 。 9. The method for detecting cancer by bioelectrical impedance based on deep learning according to claim 8, wherein, The specific steps of step S33 include: S331: Input the processed input data into the model. First, it undergoes a linear transformation to convert the input data into queries, keys, and values. The calculation formulas are as follows: Among them, is the learned weight matrix, are the query, key, and value respectively, is the key dimension, as the input of the self-attention mechanism, used to calculate the dependencies between each time step; S332: Calculate the similarity between the query and the key and use this similarity to weight the values. The calculation formula is as follows: Among them, is the dimension of the key for scaling; Use a function to convert the attention scores into weights. The calculation formula is as follows: Among them, represents the attention weight between the query and the key and the final weighted sum value; S333: Divide the query, key, and value into multiple heads for calculation, which can capture feature information in different subspaces in parallel. The calculation method of each head is the same as above. Finally, the outputs of all heads are concatenated, and the calculation formula is as follows: Among them, is the number of heads, and the output of each head is , is the linear transformation matrix of the output, is the output; S334: Use a neural network consisting of two layers of linear transformations to further extract features, and use an activation function after each layer of linear transformation. The calculation formula for the first layer of linear transformation is as follows: Among them, is the weight matrix of the first-layer linear transformation, is the bias term of the first-layer linear transformation; The calculation formula of the second-layer linear transformation is as follows: Among them, the weight matrix of the second-layer linear transformation, is the bias term of the second-layer linear transformation, is the output of the neural network; S335: For the output of the neural network perform residual connection and layer normalization to obtain the final output .

10. A method for detecting cancer based on bioelectrical impedance using deep learning according to claim 9, characterized in that, The specific steps of step S4 include: After processing, the final output obtained For downstream tasks, the output formula for the classification task is as follows: Among them, is the output, The layer maps the output to a probability space for classification tasks, The calculation formula of the layer is as follows: Among them, represents the output value of the model for the category , represents the total number of categories, in represents the probability that, given the input , the model predicts the category . By Each probability value calculated and output by the layer represents the likelihood that the input electrical signal belongs to each category; ultimately, the model will select the category with the highest probability as the classification result.

Images