Non-destructive testing method and system for biological nodules based on flexible electronic sensors

Abstract

The application discloses a biological body nodule nondestructive detection method and system based on a flexible electronic sensor, and comprises the following steps: pressing the flexible electronic sensor on the surface of a biological tissue, and synchronously collecting the flexible antenna S parameter and the piezoelectric pressing force value; performing difference operation on the reference S parameter, and obtaining electromagnetic response reflecting the coupling change of antenna deformation and tissue electromagnetic characteristics; and respectively reducing the dimensions of the S parameter and the force signal into tokens, performing fusion coding through a Transformer, and outputting the position and size parameters of the internal nodule of the tissue. The application can realize intelligent detection of nodule space parameters under noninvasive and non-radiation conditions, reduce detection cost and dependence on artificial experience, and improve the convenience and accuracy of nondestructive screening of biological body nodules.

Description

Technical Field

[0001] This invention relates to the field of nondestructive testing technology, specifically to a method and system for nondestructive testing of biological nodules based on flexible electronic sensors. Background Technology

[0002] In clinical medicine, early detection and accurate diagnosis of nodules in organs such as the lungs, liver, breast, and thyroid are crucial for improving treatment outcomes and reducing mortality. Taking lung nodules as an example, early and accurate detection directly impacts the five-year survival rate of lung cancer patients. Traditional detection methods mainly include imaging examinations (such as CT, ultrasound, and MRI) and invasive biopsies. CT scans pose a risk of ionizing radiation and are limited in frequency; ultrasound results are highly dependent on the operator's experience and have limited ability to detect deep nodules or air-containing organs (such as the lungs); while invasive biopsies are the gold standard for diagnosis, they carry risks of complications such as pneumothorax, bleeding, and infection, and sampling of small nodules is difficult. Existing technologies generally suffer from radiation hazards, invasive procedures, and the inability to simultaneously obtain information on nodule depth and size.

[0003] In recent years, microwave detection technology based on electromagnetic field theory has received widespread attention in the field of biomedical detection due to its advantages such as being non-invasive, radiation-free, and low-cost. Its basic principle is to use an antenna to emit electromagnetic waves to irradiate the tissue under test, and then invert the distribution of dielectric properties within the tissue by receiving the reflected or transmitted S-parameters. Since normal tissue and diseased tissue differ in electromagnetic parameters such as dielectric constant and conductivity, changes in the S-parameters can indirectly reflect internal abnormalities. Currently, microwave biosensors have been widely used in areas such as blood glucose concentration detection, breast imaging, and stroke detection. For example, Hagness et al. proposed a breast cancer detection system based on confocal microwave imaging, using an ultra-wideband antenna array to collect backscattered signals to achieve tumor detection; Ren Zhongyue et al. designed a miniaturized, ultra-wideband slotted antenna for early breast tumor microwave detection, and the imaging results can show the area of ​​breast lesions; Wang Mingsai et al. proposed a miniaturized, ultra-wideband wearable antenna for stroke microwave detection, and its detection system can achieve imaging of simple models.

[0004] The rapid development of flexible electronics and electronic skin technologies has brought new breakthroughs to microwave biosensors. Flexible microstrip antennas, due to their advantages such as lightweight, good skin-fitting properties, flexibility, and low cost, have become a core component of wearable microwave sensing systems, and have been validated in scenarios such as blood glucose monitoring and breast imaging. Rezaeieh et al. investigated the feasibility of a broadband microwave system for non-invasive detection and monitoring of pulmonary edema. However, the aforementioned microwave detection methods typically rely on complex imaging algorithms or multi-antenna array systems, resulting in high hardware costs, complex signal processing, and difficulty in simultaneously acquiring nodule depth and size information. Furthermore, existing research mainly utilizes differences in electromagnetic parameters for detection, and has not yet integrated mechanical deformation information (such as the deformation distribution on the sensor surface during pressure) with the electromagnetic response, missing the important pathological characteristic dimension of tissue stiffness.

[0005] In summary, existing nodule detection technologies generally suffer from limitations such as radiation hazards, invasive procedures, high dependence on operator experience, and the inability to simultaneously acquire nodule depth and size information. While flexible microwave sensors have shown potential in the biomedical detection field, there is currently no research combining the mechanical deformation information of the flexible antenna during pressing with S-parameter changes for nodule detection. Therefore, there is an urgent need to propose a detection method that integrates mechanical deformation and electromagnetic response information, is non-invasive, radiation-free, easy to operate, and can simultaneously acquire nodule location, depth, and size. Summary of the Invention

[0006] To overcome the aforementioned problems in existing nodule detection technologies, this invention proposes an artificial intelligence non-destructive testing method and system based on flexible electronic sensors. By utilizing the S-parameter differences generated by small flexible sensors during the pressing process, combined with the pressing pressure information, and through a Transformer multimodal fusion network, intelligent detection and localization of foreign objects inside organs can be achieved.

[0007] To achieve the above objectives, the present invention provides a non-destructive testing method for biological nodules based on flexible electronic sensors, comprising: The flexible electronic sensor is pressed onto the surface of the biological tissue to be tested, and the S-parameter sequence of the flexible antenna in the flexible electronic sensor and the pressing force value of the piezoelectric sensor are collected simultaneously. The S-parameter sequence includes antenna port reflection parameters at at least multiple frequency points. The S-parameter difference sequence is obtained by performing a difference operation on the S-parameter sequence corresponding to the test point based on the reference S-parameter sequence. The S-parameter difference sequence is an electromagnetic measurement result, used to characterize the change in antenna-tissue coupled electromagnetic response caused by the geometric deformation of the flexible antenna and the difference in electromagnetic properties of the biological tissue under the pressure loading state. Linear projection dimensionality reduction is performed on the S-parameter difference sequence to obtain the S-parameter feature vector, and linear projection is performed based on the pressing force value to obtain the force feature vector; The S-parameter feature vector and the force feature vector are used as S-parameter tokens and force tokens, respectively, to form a token sequence. This sequence is then input into the Transformer encoder for interactive encoding to obtain the fused features. The nodule parameters of the nodules inside the tested biological tissue are output based on the fusion features. These nodule parameters are used to characterize the spatial location and size information of the nodules within the tested biological tissue.

[0008] As a further improvement to the above technical solution, the flexible electronic sensor is a multi-layer composite structure, including a flexible serpentine antenna layer, a PDMS dielectric substrate layer, and a piezoelectric sensor layer; the flexible serpentine antenna layer is disposed on the side of the PDMS dielectric substrate layer that contacts the biological tissue to be tested and forms a serpentine electrode trace, and the piezoelectric sensor layer is encapsulated in the PDMS dielectric substrate layer and located on the side opposite to the flexible serpentine antenna layer; When the flexible electronic sensor is pressed onto the surface of the biological tissue to be tested, the flexible serpentine antenna layer deforms with the surface of the biological tissue to cause changes in the reflection parameters of the antenna port, while the piezoelectric sensor layer collects the pressing force value in real time.

[0009] As a further improvement to the above technical solution, the synchronous acquisition of the S-parameter sequence of the flexible antenna in the flexible electronic sensor and the pressing force value of the piezoelectric sensor includes: connecting the flexible antenna to a vector network analyzer, performing frequency sweep measurement on the flexible antenna within a preset frequency sweep range using the vector network analyzer to obtain the reflection parameter S11 curve; forming an S-parameter sequence from the reflection parameter S11 curve according to multiple frequency points within the preset frequency sweep range; and acquiring the current pressing force value through the piezoelectric sensor during the same pressing loading process, so that the S-parameter sequence and the pressing force value correspond to the same test point and the same pressing loading state.

[0010] As a further improvement to the above technical solution, the step of performing a difference operation on the S-parameter sequence corresponding to the test site based on the reference S-parameter sequence to obtain the S-parameter difference sequence includes: pressing a flexible electronic sensor onto a nodule-free region or a normal tissue region to collect the corresponding S-parameter sequence as a reference S-parameter sequence; pressing a flexible electronic sensor onto the test site to collect the corresponding S-parameter sequence as the test S-parameter sequence; performing a difference operation on the parameter values ​​corresponding to each frequency point in the test S-parameter sequence and the parameter values ​​corresponding to the same frequency points in the reference S-parameter sequence one by one, and arranging the difference operation results according to the frequency point order to obtain the S-parameter difference sequence; the difference operation is used to eliminate the tissue background signal of the test biological tissue and highlight the S-parameter changes caused by nodules.

[0011] As a further improvement to the above technical solution, the step of performing linear projection dimensionality reduction on the S-parameter difference sequence to obtain the S-parameter feature vector includes: passing the S-parameter difference sequence through a first fully connected layer, a ReLU activation function, a Dropout layer, and a second fully connected layer in sequence to map the 401-dimensional S-parameter difference sequence into a 24-dimensional S-parameter feature vector. The process of linearly projecting the pressure value to obtain the force feature vector includes: passing the pressure value through a first fully connected layer and a second fully connected layer in sequence to map the 1-dimensional pressure value into a 24-dimensional force feature vector.

[0012] As a further improvement to the above technical solution, the generation process of the fusion feature includes: concatenating the S-parameter token and the force token in the sequence dimension to form a token sequence with a length of 2 and a dimension of 24 for each token; inputting the token sequence into a Transformer encoder, which includes a multi-head self-attention mechanism, residual connections, layer normalization, and a feedforward network; realizing information interaction between the S-parameter token and the force token through the multi-head self-attention mechanism; and using the encoded features output by the Transformer encoder as the fusion feature.

[0013] As a further improvement to the above technical solution, the step of outputting nodule parameters of the internal nodules of the tested biological tissue based on fusion features includes: separating the encoded features output by the Transformer encoder into encoded S-parameter tokens and encoded force tokens and concatenating them; obtaining shared features through dimensionality reduction using a fusion layer; outputting nodule parameters through four output heads based on the shared features; reconcatenating the original pressing force values ​​from each output head and performing regression prediction; the nodule parameters include the two-dimensional position, depth, and diameter of the nodule; the multimodal fusion model is trained using training samples with real nodule parameter labels; and the loss function of the multimodal fusion model is Huber loss.

[0014] As a further improvement to the above technical solution, in the Transformer encoder, the number of heads in the multi-head self-attention mechanism is 2, the hidden layer dimension of the feedforward network is 64, and the number of Transformer encoder layers is 1.

[0015] As a further improvement to the above technical solution, the fusion layer includes: a first linear layer that reduces the 48-dimensional input to 20-dimensionality, followed by a ReLU activation function and a Dropout layer; and a second linear layer that reduces the 20-dimensionality input to 10-dimensionality. The network structure of each of the four output heads is as follows: the input layer maps the 11-dimensional features (after concatenating the shared features with the original pressure values) to 10-dimensionality through a linear layer, followed by a ReLU activation function, and then maps it to 1-dimensionality through another linear layer. The X and Y heads use the Tanh activation function, and the Z and D heads use the Sigmoid activation function.

[0016] This invention also provides a non-destructive testing system for biological nodules based on flexible electronic sensors, comprising: A flexible electronic sensor includes a flexible antenna and a piezoelectric sensor. The flexible antenna is used to generate an S-parameter response when the surface of the biological tissue to be tested is pressed, and the piezoelectric sensor is used to synchronously acquire the pressing force value. The press-load module is used to press the flexible electronic sensor onto the surface of the biological tissue to be tested. The S-parameter acquisition module is used to acquire the S-parameter sequence of the flexible antenna, wherein the S-parameter sequence includes at least antenna port reflection parameters at multiple frequency points; The data preprocessing module is used to perform difference calculations on the S-parameter sequences corresponding to the test points based on the reference S-parameter sequence to obtain the S-parameter difference sequence. The feature extraction module is used to perform linear projection dimensionality reduction on the S-parameter difference sequence to obtain the S-parameter feature vector, and to perform linear projection on the pressure value to obtain the force feature vector; The multimodal fusion prediction module is used to construct a token sequence by taking the S-parameter feature vector and the force feature vector as S-parameter token and force token respectively, inputting them into the Transformer encoder for interactive encoding to obtain fusion features, and outputting nodule parameters of nodules inside the tested biological tissue based on the fusion features. The nodule parameters are used to characterize the spatial location and size information of nodules in the tested biological tissue. The results display module is used to display the S-parameter sequence, pressure value, and nodule parameters.

[0017] The beneficial effects of the non-destructive testing method and system for biological nodules proposed in this invention are as follows: This study utilizes the difference in S-parameters during pressure application using a flexible snake-shaped antenna sensor to detect nodules inside organs. It reveals the coupling mechanism between deformation and electromagnetic response during sensor pressure on the organ surface. Based on a Transformer multimodal fusion model, the antenna S-parameters and pressure values ​​are input into a Transformer encoder for analysis, obtaining key nodule features including size and depth, thus achieving intelligent detection of organ nodules. This approach solves the problems of radiation risks, invasive procedures, and reliance on physician experience in existing nodule detection technologies, offering advantages such as non-invasiveness, radiation-free operation, low cost, and high-precision screening in clinical practice. Attached Figure Description

[0018] Figure 1 This is a flowchart of the non-destructive testing method for biological nodules in an embodiment of the present invention; Figure 2 This is a schematic diagram of the structure of the non-destructive testing system for biological nodules in an embodiment of the present invention; Figure 3 This is a schematic diagram of the flexible electronic sensor structure of the present invention; Figure 4 This is a planar schematic diagram of the serpentine antenna in the flexible electronic sensor of the present invention; Figure 5 A schematic diagram illustrating the detection principle of the non-destructive testing method for nodules in organisms according to the present invention; Figure 6 This is a schematic diagram of the Transformer Encoder network structure in an embodiment of the present invention; Figure 7 This is a schematic diagram of S-parameter data processing in an embodiment of the present invention; Figure 8 This is a schematic diagram of force parameter data processing in an embodiment of the present invention; Figure 9 This is a schematic diagram of the Transformer encoder structure in an embodiment of the present invention; Figure 10 This is a schematic diagram comparing the S-parameter response curves of nodules at different locations in the simulation results of an embodiment of the present invention; Figure 11 This is a schematic diagram comparing the S-parameter response curves of nodules at different locations in the actual measurement results of an embodiment of the present invention; Figure 12 This is a schematic diagram of a GUI interface for real-time display of prediction results provided in an embodiment of the present invention. Detailed Implementation

[0019] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0020] This invention provides a non-destructive detection method for biological nodules based on flexible electronic sensors, such as... Figure 1 As shown, the method includes: pressing a flexible electronic sensor onto the surface of the biological tissue to be tested, simultaneously acquiring the S-parameter sequence of the flexible antenna in the flexible electronic sensor and the pressing force value of the piezoelectric sensor, wherein the S-parameter sequence includes at least antenna port reflection parameters at multiple frequency points; performing a difference operation on the S-parameter sequence corresponding to the test point based on the reference S-parameter sequence to obtain an S-parameter difference sequence; performing linear projection dimensionality reduction on the S-parameter difference sequence to obtain an S-parameter feature vector, and performing linear projection on the pressing force value to obtain a force feature vector; using the S-parameter feature vector and the force feature vector as S-parameter token and force token respectively to form a token sequence, inputting it into a Transformer encoder for interactive encoding to obtain a fused feature; and outputting nodule parameters of the nodules inside the biological tissue to be tested based on the fused feature, wherein the nodule parameters are used to characterize the spatial location and size information of the nodules within the biological tissue to be tested.

[0021] In this invention, the S-parameter difference sequence is an electromagnetic measurement result used to characterize the antenna-tissue coupling electromagnetic response change caused by the geometric deformation of the flexible serpentine antenna and the difference in electromagnetic properties of the biological tissue under test under a pressing loading state.

[0022] Example 1 This embodiment provides a nodule detection method based on the coupling of flexible sensor deformation and electromagnetic parameters, and Transformer multimodal feature fusion. The method uses the S-parameter response generated by a flexible serpentine antenna during the pressing of the surface of the biological tissue as the primary detection signal, and simultaneously acquires the pressing pressure value. The S-parameter sequence and the pressing pressure value are used together for subsequent multimodal fusion prediction to output the spatial location and size information of the nodule within the biological tissue.

[0023] This invention is based on the inherent differences in mechanical properties between normal tissues and nodules in a biological organism. It utilizes the changes in S-parameters caused by the deformation of the antenna itself during sensor pressing to achieve non-destructive testing of nodules.

[0024] From a mechanical property perspective, there are significant differences between normal tissue and nodules in terms of Young's modulus and Poisson's ratio. When the sensor is pressed onto the tissue surface with the same pressure, normal tissue, being soft, is prone to larger deformation, while nodule areas, being harder and less compressible, exhibit smaller deformation. This difference in deformation response directly alters the geometry of the flexible serpentine antenna, thereby causing changes in the antenna's radiation characteristics.

[0025] Electromagnetic property analysis reveals differences between normal tissue and nodules in electromagnetic parameters such as dielectric constant and conductivity. However, due to the smaller volume of nodules, their disturbance to the electromagnetic field is limited, thus the contribution of electromagnetic differences to the S-parameter difference is relatively weak. This invention uses the difference in S-parameters measured by pressing normal tissue and nodular tissue as a detection signal. This difference simultaneously integrates information from both mechanical deformation and electromagnetic response, providing a physical basis for the subsequent intelligent nodule identification using an artificial intelligence model.

[0026] In terms of algorithm flow, this embodiment includes the following steps: like Figure 6 As shown, the 401-dimensional S-parameters are normalized by difference and then formed into a Stoken through Linear401→24. The 1-dimensional force value is normalized and then formed into an Ftoken through Linear1→24. The Stoken and Ftoken are concatenated and then input into the Transformer encoder. The encoding result is fused by features and then output through four output heads to output the X coordinate, Y coordinate, depth Z and diameter D respectively.

[0027] S1. Data Acquisition: Press the flexible sensor onto the surface of the organ to be tested and simultaneously acquire two types of data: antenna S-parameters (401 frequency points) and pressure value (1-dimensional).

[0028] S2. Data Preprocessing: First, the S-parameters are subtracted by comparing the S-parameters corresponding to the test points with those of the normal tissue area to eliminate background signals and highlight the subtle changes caused by nodules. At the same time, the pressure values ​​collected by the piezoelectric sensor are standardized. Then, the S-parameters after the subtraction are reduced by linear projection, and the 401-dimensional data is mapped to a 24-dimensional feature representation through a fully connected layer.

[0029] S3. Feature Projection and Token Construction: The dimensionality-reduced S-parameter features are used as S-parameter tokens (24-dimensional). The standardized pressing force values ​​are mapped to force tokens (24-dimensional) through a linear projection layer. The S-parameter tokens and force tokens are concatenated in the sequence dimension to form a token sequence with a length of 2 and each token having a dimension of 24.

[0030] S4. Transformer Encoding: The token sequence is input into the Transformer encoder for interactive encoding. The Transformer encoder contains one encoding layer, each layer including a two-head self-attention mechanism, residual connections and layer normalization, and a feedforward network (64 hidden layer dimensions) to realize information interaction between the S-parameter token and the force token.

[0031] S5. Feature Separation and Fusion: The encoded features output by the Transformer encoder are separated into encoded S-parameter tokens and encoded force tokens and concatenated to obtain 48-dimensional features; dimensionality reduction is achieved through a fusion layer (Linear 48→20→10) to obtain 10-dimensional shared features; the shared features are concatenated with the original pressing force value to obtain 11-dimensional features.

[0032] S6. Multi-task output: The four output heads share the aforementioned 11-dimensional features. Each output head uses a two-layer fully connected network (11→10→1) to output the two-dimensional position (X, Y), depth (Z), and diameter (D) of the nodule. The X and Y heads are multiplied by 16 and 8.5 respectively after applying the Tanh activation function, with output ranges of [-16, 16] mm and [-8.5, 8.5] mm; the Z and D heads are multiplied by 50 and 30 respectively after applying the Sigmoid activation function, with output ranges of [0, 50] mm and [0, 30] mm.

[0033] S7. Model Training and Evaluation: The Huber loss function is used, assigning the same weight (1.0) to all output tasks (X, Y, Z, D). The Adam optimizer (learning rate 0.001, weight decay 0.01) and ReduceLROnPlateau learning rate scheduling are used, along with an early stopping mechanism to prevent overfitting. Noise injection and random scaling are used as data augmentation during training to improve model robustness. The model's predictive accuracy is measured by calculating the mean absolute error (MAE), root mean square error (RMSE), and goodness of fit (R²) between the predicted and true values.

[0034] Example 2: Flexible Electronic Sensor Structure This embodiment further illustrates the flexible electronic sensor used in the above-described non-destructive testing method for biological nodules. The flexible electronic sensor has a multilayer composite structure, including a flexible serpentine antenna layer, a PDMS dielectric substrate layer, and a piezoelectric sensor layer. The flexible serpentine antenna layer generates an S-parameter response when the surface of the biological tissue to be tested is pressed, and the piezoelectric sensor layer is used to synchronously acquire the pressure value.

[0035] like Figure 3 As shown, the flexible sensor of this invention consists of a flexible serpentine antenna layer, a PDMS dielectric substrate layer, and a piezoelectric sensor layer. Figure 4The serpentine antenna layer is made of nickel metal patches and processed into a serpentine electrode shape, which is then attached to the bottom of the PDMS flexible substrate. The serpentine traces are used to increase the effective current path length to achieve antenna miniaturization. The piezoelectric sensor is encapsulated on the top of the substrate on the side opposite to the antenna and is used to monitor the pressing pressure in real time.

[0036] When the sensor is pressed on the surface of the organ, the flexible serpentine antenna layer deforms along with the surface of the biological tissue being tested, thereby causing changes in the antenna radiation characteristics and the shift of the S-parameter. At the same time, the piezoelectric sensor records the pressing force value, realizing the synchronous acquisition of mechanical deformation information and pressure information.

[0037] In this embodiment, the S-parameter sequence includes at least antenna port reflection parameters at multiple frequency points; in a specific experiment, the S-parameter sequence can be represented as the reflection parameter S11 curve obtained by a vector network analyzer.

[0038] This embodiment only uses the amplitude (dB) of S11, without using the phase or the complex form.

[0039] Example 3 This embodiment illustrates the simulation generation method for training and validation samples. (Reference) Figure 5 As shown in the principle, in order to establish the mapping relationship between compression deformation and electromagnetic response, this embodiment constructs a "mechanics-electromagnetism" multiphysics coupling simulation process, and obtains the S-parameter response and compression value corresponding to different nodule parameters through this process.

[0040] First, a finite element model of lung tissue was established in mechanical deformation simulation (Abaqus), with the following material mechanical parameters set: Young's modulus of normal lung tissue was 20 kPa and Poisson's ratio was 0.3; Young's modulus of nodules (simulated by clay balls) was 20,000 kPa and Poisson's ratio was 0.48; Young's modulus of nickel metal (used for sensor electrodes) was 190 GPa and Poisson's ratio was 0.31.

[0041] Spherical nodules were pre-defined within the lung tissue model, with diameters ranging from 4-28 mm (6 mm intervals), depths ranging from 10-31 mm (3 mm intervals), X-coordinates ranging from -15 mm to 15 mm (5 mm intervals), and Y-coordinates ranging from -8 mm to 8 mm (4 mm intervals). An elliptical compression area was set up to press the sensor with 1 N of pressure. The sensor deformation distribution and the deformed geometric model of the lung tissue were extracted, resulting in approximately 50 sets of mechanical simulation samples.

[0042] Subsequently, electromagnetic simulation (HFSS) was performed based on the mechanical simulation: the deformed sensor geometric model and the deformed lung geometric model calculated by Abaqus were imported into HFSS to construct a complete electromagnetic simulation model of the compression state. The material electromagnetic parameters were set as follows: the relative permittivity of normal lung tissue was 45, the conductivity was 5.54 S / m, and the dielectric loss tangent was 0.348; the relative permittivity of nodules was 48.1, the conductivity was 6.21 S / m, and the dielectric loss tangent was 0.3986; the relative permittivity of the PDMS substrate was 2.7. The sweep frequency range was set from 1 to 7 GHz to obtain the S11 parameters under different conditions.

[0043] Through the aforementioned joint simulation, the system collected S-parameter responses and pressure values ​​corresponding to different nodule parameters (X, Y, Z, D), where S represents the S11 curve (401 frequency points), and F represents the pressure value. Each data sample can be represented as (N, S, F). A total of 50 sets of simulation data were obtained, which were divided into a training set (40 sets), a validation set (5 sets), and a test set (5 sets) in an 8:1:1 ratio for subsequent model training and evaluation. Some simulation results are shown below. Figure 10 As shown in the figure. The horizontal axis represents frequency in GHz; the vertical axis represents the S-parameter in dB (S(1,1)), in dB, used to represent the decibel value of the reflection parameter S11 at the port of the flexible serpentine antenna; the blue curve represents normal lung tissue with a pressure of 0.9 N; the red curve represents the embedded nodule coordinates (0, 0, 1 cm), diameter 0.8 cm, and pressure of 0.9 N; the green curve represents the embedded nodule coordinates (0.8 cm, 0, 2.5 cm), diameter 0.8 cm, and pressure of 0.9 N.

[0044] like Figure 10 As shown, the curves, with Frequency (GHz) on the horizontal axis and dB(S(1,1)) (dB) on the vertical axis, compare the S11 amplitude response of normal lung tissue and nodules at different locations and depths under the same pressure of 0.9N. The curves in the figure show differences at the resonance valley and the high-frequency amplitude.

[0045] Example 4 This embodiment illustrates a data acquisition method in a real physical scenario. Corresponding to the method described above, which involves pressing a flexible electronic sensor onto the surface of the biological tissue to be tested and simultaneously acquiring the S-parameter sequence of the flexible serpentine antenna and the pressing force value of the piezoelectric sensor, this embodiment uses an in vitro porcine lung experiment to obtain the measured dataset.

[0046] To verify the validity of the simulation data and obtain detection data under real physical scenarios, this invention further conducted in vitro experiments on pig lungs. Fresh isolated pig lungs were used as experimental subjects, and clay balls of different diameters were embedded in different locations inside the lung tissue to simulate lung nodules. The parameters of the clay balls were set in accordance with the simulation: diameter 5-20 mm (step size 5 mm), depth 5-25 mm (step size 5 mm), and two-dimensional positions randomly distributed within a range of 50 mm × 50 mm. Three samples were prepared for each parameter combination.

[0047] The experimental setup includes: a flexible snake antenna sensor, attached to a PDMS substrate with a center frequency of 5.8 GHz; a piezoelectric thin film sensor, encapsulated on the top of the PDMS substrate; a vector network analyzer, used to acquire S11 parameters in the 1-7.5 GHz frequency band; and a pressure acquisition module, used to synchronously record the pressure value.

[0048] During data acquisition, a flexible sensor was attached to predetermined detection points on the surface of the pig lung. Pressure was applied using a constant-force pressing device, and data was collected simultaneously. A vector network analyzer acquired the S11 parameters, and a piezoelectric sensor acquired the current pressing force value. Each detection point was pressed three times, with each press lasting one second. The average value was taken as the measurement data for that point, and the actual nodule parameters (X-coordinate, Y-coordinate, depth, and diameter) at that point were recorded as labels. A total of 50 valid measured samples were collected, with 40 sets used for training and 10 sets used for testing. Some measured results are shown below. Figure 11 As shown in the figure. The horizontal axis represents frequency in GHz; the vertical axis represents S-parameters in dB, used to indicate the decibel value of the reflection parameter S11 at the port of the flexible serpentine antenna; the blue curve represents the embedded nodal coordinates (0, 0, 2cm), diameter 0.5cm, and pressing force 2.5N; the red curve represents the embedded nodal coordinates (0, 0, 2cm), diameter 0.5cm, and pressing force 3.0N; the green curve represents the embedded nodal coordinates (0, 0, 3cm), diameter 0.5cm, and pressing force 2.4N.

[0049] In this embodiment, the S-parameter sequence and the pressing force value correspond to the same test point and the same pressing loading state, thereby ensuring that the subsequent model can perform multimodal fusion analysis based on the electromagnetic response and pressure information under the same loading state.

[0050] Example 5 This embodiment illustrates the method of performing interpolation on the S-parameter sequence and extracting features from the S-parameter difference sequence. Corresponding to the detection method described above, the S-parameter sequence corresponding to the test site is interpolated based on the reference S-parameter sequence to obtain the S-parameter difference sequence; this S-parameter difference sequence is used to eliminate the tissue background signal of the biological tissue under test and highlight the weak S-parameter changes caused by nodules.

[0051] The same preprocessing operations were performed on the collected simulation data and measured data. First, S-parameter difference calculation was performed, using the S-parameter of the nodule-free area as the reference value S_normal for normal tissue, and the S-parameter difference ΔS = S_nodule - S_normal for each detection point was calculated. This difference calculation effectively eliminated the tissue background signal and highlighted the subtle changes caused by nodules.

[0052] In an embodiment corresponding to the above claims, a flexible electronic sensor can be pressed onto a nodule-free region or a normal tissue region to collect the corresponding S-parameter sequence as a reference S-parameter sequence; then, the flexible electronic sensor can be pressed onto the test point to collect the corresponding S-parameter sequence as the test S-parameter sequence; then, the parameter values ​​corresponding to each frequency point in the test S-parameter sequence are compared with the parameter values ​​corresponding to the same frequency points in the reference S-parameter sequence one by one, and the difference calculation results are arranged according to the frequency point order to obtain the S-parameter difference sequence.

[0053] The reference S-parameters are derived from data measured by pressing on known nodular regions. Both the reference S-parameters and the S-parameters to be measured must be acquired under the same or similar pressing pressure. Reference S-parameters at different pressure values ​​are acquired first, and then reference S-parameters with similar pressure values ​​are automatically matched based on the pressure value of the S-parameter to be measured.

[0054] Subsequently, linear projection dimensionality reduction is performed on the S-parameter difference sequence. A fully connected layer maps the 401-dimensional S-parameter difference sequence into a 24-dimensional S-parameter feature vector, serving as the S-parameter token. Simultaneously, the pressure value F collected by the piezoelectric sensor is standardized and mapped into a 24-dimensional force feature vector through a linear projection layer, serving as the force token. The S-parameter token and force token are concatenated along the sequence dimension to form a token sequence of length 2, with each token having a dimension of 24. This sequence is then input into the Transformer encoder for interactive encoding.

[0055] like Figure 7 As shown, the feature extraction includes the following steps: First, the difference between the original S-parameter nodules (401-dimensional) and the original S-parameter normal (401-dimensional) is calculated to obtain ΔS = S_nodules - S_normal; then, the difference result is standardized for S-parameters; finally, the standardized S-parameters are reduced to 24-dimensional S-parameter tokens through a Linear 401→64→24 linear projection layer. The S-parameters can be obtained by standard convolution + linear projection processing.

[0056] The S-parameter tokens are used as S-parameter feature points, and together with the force tokens, they form a token sequence. These S-parameter tokens characterize the compressed feature representation of the S-parameter difference sequence and provide input for subsequent interactive encoding by the Transformer encoder.

[0057] Example 6 This embodiment illustrates the process of generating a force feature vector based on the pressure value. Corresponding to the detection method described above, the pressure value is synchronously acquired by a piezoelectric sensor to characterize the loading state when the flexible electronic sensor presses against the surface of the biological tissue under test, and is input into the Transformer encoder along with the S-parameter feature vector.

[0058] The pressure value F collected by the piezoelectric sensor is standardized and then mapped into a 24-dimensional force feature vector through a linear projection layer, which serves as the force token.

[0059] Specifically, the pressure values ​​collected by the piezoelectric sensor can be standardized to obtain normalized force values; then, the 1-dimensional pressure values ​​can be mapped to 24-dimensional features through the first fully connected layer, and the 24-dimensional features can be maintained through the second fully connected layer to obtain the force feature vector.

[0060] The pressure value is a fixed-dimensional linear mapping of the single-point force value.

[0061] The force feature vector, used as the force token, is concatenated with the S-parameter token along the sequence dimension and then input into the Transformer encoder, enabling the model to analyze the S-parameter difference sequence in conjunction with the pressing pressure condition.

[0062] like Figure 8 As shown, the force value feature extraction includes the following steps: first, the original force value (1-dimensional) is standardized; then, through a Linear 1→24 linear projection layer, the standardized force value is mapped to a 24-dimensional force value token.

[0063] The force token is used to represent the compression feature of the pressing force value. After being concatenated with the S-parameter token in the sequence dimension, they are jointly input into the Transformer encoder for interactive encoding.

[0064] Example 7 This embodiment illustrates the process of generating loading state characteristics based on the pressure value. Corresponding to the detection method described above, the pressure value is synchronously acquired by a piezoelectric sensor to characterize the loading state when a flexible electronic sensor presses against the surface of the biological tissue under test, and is input into a multimodal fusion model along with the electromagnetic response characteristics.

[0065] The pressure value F collected by the piezoelectric sensor is linearly projected to reduce its dimensionality, and mapped to a low-dimensional feature space (24 dimensions) to obtain the force feature vector.

[0066] Specifically, the pressure values ​​collected by the piezoelectric sensor can be standardized, and then the 1-dimensional pressure values ​​can be mapped to a 24-dimensional force feature vector through a linear projection layer.

[0067] This loading state feature is used to characterize the loading state when the flexible electronic sensor presses on the surface of the biological tissue to be tested, and is used together with the electromagnetic response feature to input into the multimodal fusion model, so that the model can analyze the S-parameter difference sequence in combination with the pressing pressure condition.

[0068] Example 8 This embodiment illustrates the process of inputting electromagnetic response features and loading state features into a multimodal fusion model for interactive encoding to obtain fused features. The multimodal fusion model includes a Transformer encoder, which uses a multi-head self-attention mechanism to achieve information interaction between S-parameter tokens and force tokens.

[0069] After data preprocessing, the Transformer multimodal fusion network is trained. First, the S-parameter feature vector (24-dimensional) is used as the S-parameter token, and the force feature vector (24-dimensional) is used as the force token. The two are concatenated in the sequence dimension to form a token sequence of length 2 and each token has a dimension of 24, which is then input into the Transformer encoder.

[0070] In this embodiment, the electromagnetic response characteristics are used as S-parameter tokens, the loading state characteristics are used as force tokens, and the S-parameter tokens and force tokens are used together to form a token sequence; this token sequence is then input into the Transformer encoder.

[0071] like Figure 9 As shown, the single-layer Transformer Encoder structure comprises the following components: the input is a token sequence (2×24). First, a multi-head self-attention mechanism (nhead=2) is used to facilitate information exchange between the S-parameter tokens and the force tokens. Then, after residual connections and layer normalization, the output of the attention mechanism is added to the original input and normalized. Next, a feedforward network (64 hidden layer dimensions, GELU activation function) is used to enhance the model's non-linear expressive power. Finally, after another residual connection and layer normalization, the output is a token sequence (2×24) with the same dimensions as the input. This invention employs a single-layer Transformer Encoder structure as described above.

[0072] After Transformer encoding, the force token absorbs the contextual information of the S-parameters, while the S-parameter token also gains the influence of the force, achieving a deep fusion of mechanical and electromagnetic information. The encoded features output by the Transformer encoder are used as fused features for subsequent nodal parameter output.

[0073] Example 9 This embodiment illustrates the process of outputting nodule parameters of nodules within the tested biological tissue based on fusion features. The nodule parameters characterize the spatial location and size information of the nodules within the tested biological tissue; in this embodiment, they specifically include the two-dimensional location, depth, and diameter of the nodules.

[0074] After the Transformer encoder outputs the encoded features, feature separation and fusion are performed: the encoded S-parameter token and the encoded force token are concatenated to obtain 48-dimensional features, which are then reduced to 10 dimensions through a fusion layer (Linear 48→20→10) to form a unified shared feature representation. Finally, the shared features are concatenated with the original pressing force value (1-dimensional) to obtain 11-dimensional features.

[0075] In the multi-task output stage, the four output heads share the aforementioned 11-dimensional features. Each output head uses a two-layer fully connected network (11-dimensional → 10-dimensional → 1-dimensional) to output the two-dimensional position (X, Y), depth (Z), and diameter (D) of the nodule. The X and Y heads are multiplied by 16 and 8.5 respectively after applying the Tanh activation function, resulting in output ranges of [-16, 16] mm and [-8.5, 8.5] mm. The Z and D heads are multiplied by 50 and 30 respectively after applying the Sigmoid activation function, resulting in output ranges of [0, 50] mm and [0, 30] mm.

[0076] During model training, the multimodal fusion model is trained using training samples with real nodule parameter labels, including the nodule's two-dimensional location, depth, and diameter. For the loss function, the Huber loss function is used. = `_Huber(y_true, y_pred)`. In this formula, `y_true` represents the true values, i.e., the pre-recorded true nodule parameters in the training samples; `y_pred` represents the predicted values, i.e., the nodule prediction parameters output by the Transformer multimodal fusion model based on S-parameter features and pressure features. The Adam optimizer (learning rate 0.001, weight decay 0.01) and ReduceLROnPlateau learning rate scheduling are used, along with an early stopping mechanism to prevent overfitting. Noise injection is used as data augmentation during training to improve model robustness.

[0077] Example 10 This embodiment illustrates the model performance evaluation method and result analysis approach. After the model completes training, the prediction accuracy is measured by calculating the mean absolute error (MAE), root mean square error (RMSE), and goodness of fit (R²) between the predicted and actual values. The mean absolute error is: ; The root mean square error is: ; The goodness of fit is: ; The accuracy of a model's predictions is measured by calculating the mean absolute error, root mean square error, and goodness of fit between the predicted and actual values. This represents the true value of the i-th sample. This represents the predicted value of the i-th sample. represents the average of the true values, and n represents the sample size.

[0078] The model performance was evaluated on the simulation test set, and the results are as follows:

[0079] Experimental results show that the S-parameter curves corresponding to nodules of different diameters and depths have distinguishable differences in resonance point and amplitude, verifying the effectiveness of the force-electromagnetic joint simulation method and the accuracy of the proposed Transformer multimodal fusion model in predicting nodule location, depth and diameter.

[0080] Example 11 This embodiment illustrates a detection system that implements the above method. For example... Figure 2 As shown, the system includes a flexible electronic sensor, a pressure loading module, an S-parameter acquisition module, a pressure acquisition module, a data preprocessing module, a feature extraction module, a multimodal fusion prediction module, and a result display module.

[0081] The flexible electronic sensor comprises a flexible serpentine antenna, a PDMS dielectric substrate, and a piezoelectric sensor. The flexible serpentine antenna generates an S-parameter response when pressed against the surface of the biological tissue to be tested, and the piezoelectric sensor synchronously acquires the pressure value. A pressure loading module presses the flexible electronic sensor onto the surface of the biological tissue. An S-parameter acquisition module, connected to the flexible serpentine antenna, acquires the S-parameter sequence of the antenna, which includes antenna port reflection parameters at at least multiple frequency points. A pressure acquisition module, connected to the piezoelectric sensor, acquires the pressure value synchronized with the S-parameter sequence.

[0082] The data preprocessing module is used to perform difference operations on the S-parameter sequences corresponding to the test points based on the reference S-parameter sequence to obtain the S-parameter difference sequence.

[0083] The feature extraction module performs linear projection dimensionality reduction on the S-parameter difference sequence to obtain the S-parameter feature vector, and performs linear projection on the pressure value to obtain the force feature vector. The multimodal fusion prediction module uses the S-parameter feature vector and the force feature vector as S-parameter tokens and force tokens, respectively, to construct a token sequence. This sequence is input into a Transformer encoder for interactive encoding to obtain fused features. Based on these fused features, the module outputs the nodule parameters of the nodules within the tested biological tissue, including the nodule's two-dimensional location, depth, and diameter. The results display module displays the S-parameter sequence, pressure value, and nodule parameters.

[0084] To facilitate convenient operation of the detection process and visualize the results, this invention designs a graphical user interface. For example... Figure 12 As shown, the GUI interface integrates the following functional modules: a data acquisition control module for starting and stopping data acquisition; a waveform display module for real-time display of S-parameter waveforms and pressure values; a prediction result visualization module that presents nodule locations in the form of a lung thermogram and includes size annotations; and a data saving and playback module for storing historical test data and reproducing the test process. This interface effectively improves the convenience and interactivity of experimental operations and clinical applications.

[0085] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Those skilled in the art can make various modifications and improvements without departing from the inventive concept, and all such modifications and improvements should fall within the protection scope of the present invention.

Claims

1. A non-destructive testing method for biological nodules based on flexible electronic sensors, characterized in that, include: The flexible electronic sensor is pressed onto the surface of the biological tissue to be tested, and the S-parameter sequence of the flexible antenna in the flexible electronic sensor and the pressing force value of the piezoelectric sensor are collected simultaneously. The S-parameter sequence includes antenna port reflection parameters at at least multiple frequency points. The S-parameter difference sequence is obtained by performing a difference operation on the S-parameter sequence corresponding to the test point based on the reference S-parameter sequence. The S-parameter difference sequence is an electromagnetic measurement result, used to characterize the change in antenna-tissue coupled electromagnetic response caused by the geometric deformation of the flexible antenna and the difference in electromagnetic properties of the biological tissue under the pressure loading state. Linear projection dimensionality reduction is performed on the S-parameter difference sequence to obtain the S-parameter feature vector, and linear projection is performed based on the pressing force value to obtain the force feature vector; The S-parameter feature vector and the force feature vector are used as S-parameter tokens and force tokens, respectively, to form a token sequence. This sequence is then input into the Transformer encoder for interactive encoding to obtain the fused features. The nodule parameters of the nodules inside the tested biological tissue are output based on the fusion features. These nodule parameters are used to characterize the spatial location and size information of the nodules within the tested biological tissue.

2. The method for non-destructive testing of biological nodules based on flexible electronic sensors according to claim 1, characterized in that, The flexible electronic sensor is a multi-layer composite structure, including a flexible serpentine antenna layer, a PDMS dielectric substrate layer, and a piezoelectric sensor layer. The flexible serpentine antenna layer is disposed on the side of the PDMS dielectric substrate layer that contacts the biological tissue to be tested and forms a serpentine electrode trace. The piezoelectric sensor layer is encapsulated in the PDMS dielectric substrate layer and is located on the side opposite to the flexible serpentine antenna layer. When the flexible electronic sensor is pressed onto the surface of the biological tissue to be tested, the flexible serpentine antenna layer deforms with the surface of the biological tissue to cause changes in the reflection parameters of the antenna port, while the piezoelectric sensor layer collects the pressing force value in real time.

3. The method for non-destructive testing of biological nodules based on flexible electronic sensors according to claim 1, characterized in that, The synchronous acquisition of the S-parameter sequence of the flexible antenna in the flexible electronic sensor and the pressing force value of the piezoelectric sensor includes: connecting the flexible antenna to a vector network analyzer, performing frequency sweep measurement on the flexible antenna within a preset frequency sweep range using the vector network analyzer to obtain the reflection parameter S11 curve; forming the S-parameter sequence from the reflection parameter S11 curve according to multiple frequency points within the preset frequency sweep range; and acquiring the current pressing force value through the piezoelectric sensor during the same pressing loading process, so that the S-parameter sequence and the pressing force value correspond to the same test point and the same pressing loading state.

4. The method for non-destructive testing of biological nodules based on flexible electronic sensors according to claim 1, characterized in that, The step of performing a difference operation on the S-parameter sequence corresponding to the test site based on the reference S-parameter sequence to obtain the S-parameter difference sequence includes: pressing a flexible electronic sensor onto a nodule-free region or a normal tissue region to collect the corresponding S-parameter sequence as a reference S-parameter sequence; pressing the flexible electronic sensor onto the test site to collect the corresponding S-parameter sequence as the test S-parameter sequence; performing a difference operation on the parameter values ​​corresponding to each frequency point in the test S-parameter sequence and the parameter values ​​corresponding to the same frequency points in the reference S-parameter sequence one by one, and arranging the difference operation results according to the frequency point order to obtain the S-parameter difference sequence; the difference operation is used to eliminate the tissue background signal of the test biological tissue and highlight the S-parameter changes caused by nodules.

5. The method for non-destructive testing of biological nodules based on flexible electronic sensors according to claim 1, characterized in that, The step of performing linear projection dimensionality reduction on the S-parameter difference sequence to obtain the S-parameter feature vector includes: passing the S-parameter difference sequence through a first fully connected layer, a ReLU activation function, a Dropout layer, and a second fully connected layer in sequence to map the 401-dimensional S-parameter difference sequence into a 24-dimensional S-parameter feature vector. The process of linearly projecting the pressure value to obtain the force feature vector includes: passing the pressure value through a first fully connected layer and a second fully connected layer in sequence to map the 1-dimensional pressure value into a 24-dimensional force feature vector.

6. The method for non-destructive testing of biological nodules based on flexible electronic sensors according to claim 1, characterized in that, The process of generating the fusion feature includes: concatenating the S-parameter token and the force token along the sequence dimension to form a token sequence of length 2 and dimension 24 for each token; inputting the token sequence into a Transformer encoder, which includes a multi-head self-attention mechanism, residual connections, layer normalization, and a feedforward network; realizing information interaction between the S-parameter token and the force token through the multi-head self-attention mechanism; and using the encoded features output by the Transformer encoder as the fusion feature.

7. The method for non-destructive testing of biological nodules based on flexible electronic sensors according to claim 6, characterized in that, The step of outputting nodule parameters of the internal nodules of the tested biological tissue based on fusion features includes: separating the encoded features output by the Transformer encoder into encoded S-parameter tokens and encoded force tokens, concatenating them, and obtaining shared features through dimensionality reduction using a fusion layer; outputting nodule parameters through four output heads based on the shared features, and reconcatenating the original pressing force values ​​from each output head for regression prediction; the nodule parameters include the two-dimensional location, depth, and diameter of the nodule; the multimodal fusion model is trained using training samples with real nodule parameter labels; and the loss function of the multimodal fusion model is Huber loss.

8. The method for non-destructive testing of biological nodules based on flexible electronic sensors according to claim 1, characterized in that, In the Transformer encoder, the number of heads in the multi-head self-attention mechanism is 2, the hidden layer dimension of the feedforward network is 64, and the number of Transformer encoder layers is 1.

9. The method for non-destructive testing of biological nodules based on flexible electronic sensors according to claim 7, characterized in that, The fusion layer includes: a first linear layer that reduces the 48-dimensional input to 20-dimensionality, followed by a ReLU activation function and a Dropout layer; and a second linear layer that reduces the 20-dimensionality input to 10-dimensionality. The network structure of each of the four output heads is as follows: the input layer maps the 11-dimensional features (after concatenating the shared features with the original pressure values) to 10-dimensionality through a linear layer, followed by a ReLU activation function, and then maps it to 1-dimensionality through another linear layer. The X and Y heads use the Tanh activation function, and the Z and D heads use the Sigmoid activation function.

10. A non-destructive testing system for biological nodules based on flexible electronic sensors, characterized in that, include: A flexible electronic sensor includes a flexible antenna and a piezoelectric sensor. The flexible antenna is used to generate an S-parameter response when the surface of the biological tissue to be tested is pressed, and the piezoelectric sensor is used to synchronously acquire the pressing force value. The press-load module is used to press the flexible electronic sensor onto the surface of the biological tissue to be tested. The S-parameter acquisition module is used to acquire the S-parameter sequence of the flexible antenna, wherein the S-parameter sequence includes at least antenna port reflection parameters at multiple frequency points; The data preprocessing module is used to perform difference calculations on the S-parameter sequences corresponding to the test points based on the reference S-parameter sequence to obtain the S-parameter difference sequence. The feature extraction module is used to perform linear projection dimensionality reduction on the S-parameter difference sequence to obtain the S-parameter feature vector, and to perform linear projection on the pressure value to obtain the force feature vector; The multimodal fusion prediction module is used to construct a token sequence by taking the S-parameter feature vector and the force feature vector as S-parameter token and force token respectively, inputting them into the Transformer encoder for interactive encoding to obtain fusion features, and outputting nodule parameters of nodules inside the tested biological tissue based on the fusion features. The nodule parameters are used to characterize the spatial location and size information of nodules in the tested biological tissue. The results display module is used to display the S-parameter sequence, pressure value, and nodule parameters.

Images