Lung cancer early screening electronic nose based on MEMS sensor array and hybrid algorithm

Abstract

The invention discloses a lung cancer early screening electronic nose based on an MEMS sensor array and a hybrid algorithm, and belongs to the technical field of gas detection. The device comprises an MEMS gas sensor array, an enrichment device, an electromagnetic valve, a gas chamber, a single chip microcomputer main control board, a gas pump, an upper computer, a recognition algorithm and the like. The electronic nose system focuses on lung cancer sensor array construction and algorithm design, and is small in size and high in accuracy on the basis of a traditional electronic nose. For a core sensor array of an electronic nose, a novel method of'correlation coefficient + distinguishing performance value 'dual screening is adopted, simplification of the number of sensors is realized on the premise of ensuring target gas selectivity, and technical support is provided for efficient detection of lung cancer symbolic VOCs (Volatile Organic Compounds); for a core-recognition algorithm of the electronic nose, a time sequence enhanced PRM-CNN algorithm is provided, and the data classification accuracy of the recognition algorithm is improved.

Description

Technical Field

[0001] The present invention relates to the field of gas detection, and in particular to a design of an electronic nose for early screening of lung cancer based on a MEMS sensor array and a hybrid algorithm. Background Art

[0002] Since ancient times, smell has been an important way for humans to observe the world. Gas detection also holds significant value in fields such as disease detection, environmental monitoring, industrial safety, and food safety. For example, early disease screening in clinical research, monitoring for benzene leaks in industrial sites, monitoring for PM2.5 precursors in the atmosphere, and detecting amine compounds produced by food spoilage all require gas composition analysis for safety warnings and quality control.

[0003] Currently, gas detection primarily relies on laboratory-grade equipment such as gas chromatography-mass spectrometry (GC-MS) and Fourier transform infrared spectroscopy (FTIR). While these methods offer high-precision detection capabilities, they suffer from inherent drawbacks such as bulky instruments (requiring a carrier gas system and vacuum equipment), high single-test costs (consumables and maintenance fees can reach thousands of yuan), and long analysis cycles (typically exceeding 30 minutes).

[0004] Electronic nose (E-Nose) technology, using a biomimetic olfactory perception mechanism, employs a multimodal sensor array and pattern recognition algorithms, providing an innovative solution for on-site detection of complex gas mixtures. In the field of food safety, researchers have developed an E-Nose system based on a quartz microbalance (QCM) sensor that can distinguish the degree of meat spoilage. In environmental monitoring, an E-Nose that combines nano-gas-sensitive materials with deep learning can detect VOCs at the ppb level. However, existing E-Nose systems for lung cancer detection, for example, still face two common technical bottlenecks: First, sensor selection lacks a systematic optimization method and is highly subjective. Most systems utilize a "broad-spectrum coverage" strategy, integrating 8-12 sensors, resulting in device power consumption exceeding 5W and a volume greater than 200 cm³. Second, currently used pattern recognition algorithms are unable to effectively extract spatiotemporal features from data, resulting in low recognition accuracy. Summary of the Invention

[0005] In view of the above shortcomings of the existing technologies, in order to solve the problems that the sensor array design of the existing electronic nose system for early screening of lung cancer is highly subjective or the algorithm is complex, the recognition algorithm cannot effectively extract the spatiotemporal characteristics of the data, and the recognition accuracy is low. The purpose of the present invention is to propose a design method and device for an electronic nose for early screening of lung cancer based on a MEMS sensor array and a hybrid algorithm to improve the lung cancer detection capability of the electronic nose. The present invention adopts the method of correlation coefficient + differentiation performance value to improve the target detection capability of the sensor array according to the relationship between the target gas and the sensor response and the type of MEMS gas sensor; and proposes a time-series enhanced PRM-CNN algorithm to improve the classification accuracy of the recognition algorithm for data samples.

[0006] In order to achieve the above objectives, the present invention adopts and implements the following technical solutions:

[0007] An electronic nose for early screening of lung cancer based on a MEMS sensor array and a hybrid algorithm is designed. The system includes: a MEMS gas sensor array, an enrichment device, a solenoid valve, a gas chamber, a single-chip microcomputer main control board, an air pump, a host computer, a recognition algorithm, etc.

[0008] The method for selecting the sensor type of the gas sensor array is:

[0009] (1) The experiment was repeated many times to test the dynamic response characteristic curves of 18 MEMS gas sensors to hexanal, toluene, and isoprene at a concentration of 1 ppm.

[0010] (2) Calculate the correlation coefficient between each two gas sensors (such as sensor x and sensor y) for the same gas based on the dynamic response characteristic curve .

[0011]

[0012] in, and They represent the response values ​​of the i-th sampling point in the timing response curves of sensors X and Y for a specific gas sample; and is the mean response value of the corresponding gas sensor over the entire response cycle; N represents the total number of sampling points in a single gas detection. The correlation coefficient ranges from [-1 to 1]. The larger the absolute value, |Rxy|, the more similar the response patterns of the two sensors are. A similarity above 0.9 is considered highly similar and interchangeable. A similarity below this threshold is considered low. (This evaluation mechanism emphasizes global consistency; low similarity in the response of any one gas negates overall interchangeability.)

[0013] (3) Calculate the discrimination performance value of each gas sensor for different target gases based on the dynamic response characteristic curve .

[0014] i=1,2,…,nk=1,2,…,m

[0015] in is the inter-class dispersion of samples, Intra-class dispersion; Represents a standard gas sample detected by the kth gas sensor; It represents the average value of the total samples of a certain gas corresponding to the gas sensor; u represents the average value of the total samples of all gases corresponding to the gas sensor; n represents the number of test gas types; m represents the number of test samples for each gas. The larger the value of , the greater the inter-class dispersion of the gas samples of the sensor, the smaller the intra-class dispersion, and the better the sensor's ability to distinguish gas types.

[0016] (4) First, the gas sensor SMD1001 with the largest discrimination performance value is selected, recorded as S1, and then the sensors with correlation coefficients greater than 0.9 are eliminated.

[0017] (5) Repeat step (4) for the remaining gas sensors until the correlation coefficients are all less than 0.9. Select SMD1001 (S1), GM-202B (S2), GM-512B (S3), SYS1000 (S4), and SY-CH2-4 (S5) to form the sensor array.

[0018] The CNN pattern recognition algorithm with PRM coding timing enhancement in the host computer specifically includes the following steps:

[0019] (1) Calculate the response value of the sensor response curve ,in is the load resistance value in the voltage divider detection circuit, is the voltage across the load resistor, is the gas sensitive resistance of the gas sensor, It is the stable gas sensitive resistance value of the gas sensor under clean air;

[0020]

[0021]

[0022] (2) Perform PRM encoding on the response data and convert the one-dimensional time data of each sensor into a two-dimensional tensor as the input of CNN, where and are all data points in the sensor's one-dimensional response data. is the total number of data points;

[0023] in

[0024] (3) The convolutional neural network adopts a layered and progressive feature extraction architecture, specifically including:

[0025] Convolutional layer 1: 32 3×3 convolution kernels, ReLU activation function, L2 regularization coefficient 0.001; batch normalization layer; Dropout layer with dropout rate 0.2; 2×2 max pooling layer;

[0026] Second convolutional layer: 64 3×3 convolution kernels, ReLU activation function, L2 regularization coefficient 0.0005; batch normalization layer; 2×2 maximum pooling layer;

[0027] The third convolutional layer: 128 3×3 convolution kernels, ReLU activation function; batch normalization layer; 2×2 maximum pooling layer;

[0028] Fully connected layer: 64 neurons, ReLU activation function, L2 regularization coefficient 0.001; Dropout layer, dropout rate 0.6;

[0029] Output layer: 2 neurons, activation function is Softmax; Adam optimizer is used.

[0030] Compared with existing inventions, the advantages of the present invention are:

[0031] The lung cancer electronic nose system of this invention focuses on sensor array design and recognition algorithm design. This electronic nose system builds upon the traditional electronic nose system by focusing on the construction of a lung cancer sensor array and algorithm design. For the sensor array, the hardware core of the electronic nose, a novel dual screening method using "correlation coefficient + discrimination performance value" is employed to streamline the number of sensors while ensuring target gas selectivity, providing technical support for the efficient detection of VOCs, a hallmark of lung cancer. For the recognition algorithm, the software core of the electronic nose, a time-series enhanced PRM-CNN algorithm is proposed, improving the recognition algorithm's classification accuracy.

[0032] The present invention adopts a correlation coefficient-discrimination performance value joint screening strategy when selecting sensor types. It has scientific data support and can better improve the sensor array's ability to distinguish target gases. Compared with the traditional subjective sensor selection, it is more scientific and simpler to calculate and more interpretable than the machine learning screening algorithm. In response to the problem that conventional algorithms such as LDA, SVM, and 1D-CNN are insufficient in extracting the temporal features of exhaled breath samples, a temporally enhanced PRM-CNN algorithm is proposed. It fully utilizes the enhancement of data temporal features by PRM coding and the characteristics of autonomous iteration and strong spatial attention of the CNN algorithm. Compared with algorithms such as LDA, SVM, and 1D-CNN, the classification accuracy of data samples by the recognition algorithm is effectively improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Figure 1 It is the sensor correlation matrix diagram.

[0034] Figure 2 Fi value diagram of each sensor.

[0035] Figure 3This is the confusion matrix diagram of the PRM-CNN model training set.

[0036] Figure 4 This is a comparison chart of multi-model performance.

[0037] Figure 5 This is a structural block diagram of the electronic nose of the present invention.

[0038] Figure 6 This is the upper PCB diagram of the sensor array in the present invention.

[0039] Figure 7 This is the lower PCB diagram of the sensor array in the present invention.

[0040] Figure 8 This is a schematic diagram of the upper layer of the single-chip microcomputer main control board in the present invention.

[0041] Figure 9 This is a schematic diagram of the lower layer of the single-chip microcomputer main control board in the present invention.

[0042] In the figure: 1. Positive pressure air pump, 2. First solenoid valve, 3. Enrichment device, 4. Blank gas line, 5. Second solenoid valve, 6. Gas chamber, 7. Temperature control device, 8. Single chip microcomputer main control board, 9. Host computer. DETAILED DESCRIPTION

[0043] In order to make the construction and use of the present invention clearer and easier for researchers in this field to understand and use, the technical solution of the present invention will be described in detail below with reference to the accompanying drawings. The specific embodiments described here are only used to illustrate the present invention.

[0044] Example 1 Sensor Array Construction

[0045] The specific construction method is:

[0046] (1) The experiment was repeated many times to test the dynamic response characteristic curves of 18 MEMS gas sensors to hexanal, toluene, and isoprene at a concentration of 1 ppm.

[0047] (2) Calculate the correlation coefficient between each two gas sensors (such as sensor x and sensor y) for the same gas based on the dynamic response characteristic curve .

[0048]

[0049] in, and They represent the response values ​​of the i-th sampling point in the timing response curves of sensors X and Y for a specific gas sample; and is the mean response value of the corresponding gas sensor over the entire response cycle; N represents the total number of sampling points in a single gas detection. The correlation coefficient ranges from [-1 to 1]. The larger the absolute value, |Rxy|, the more similar the response patterns of the two sensors are. A similarity above 0.9 is considered highly similar and interchangeable. A similarity below this threshold is considered low. (This evaluation mechanism emphasizes global consistency; low similarity in the response of any one gas negates overall interchangeability.)

[0050] (3) Calculate the discrimination performance value of each gas sensor for different target gases based on the dynamic response characteristic curve .

[0051] i=1,2,…,nk=1,2,…,m

[0052] in is the inter-class dispersion of samples, Intra-class dispersion; Represents a standard gas sample detected by the kth gas sensor; It represents the average value of the total samples of a certain gas corresponding to the gas sensor; u represents the average value of the total samples of all gases corresponding to the gas sensor; n represents the number of test gas types; m represents the number of test samples for each gas. The larger the value of , the greater the inter-class dispersion of the gas samples of the sensor, the smaller the intra-class dispersion, and the better the sensor's ability to distinguish gas types.

[0053] (4) Select the gas sensor with the largest Fi value (SMD1001) and combine it with the correlation coefficient matrix diagram, as shown in the figure below: Figure 1 , eliminating gas sensors similar to (SMD1001). Then, from the remaining gas sensors, select the gas sensor with the next highest Fi value (GM-202B) and repeat the above process. After optimization and screening, a MEMS gas sensor array consisting of five types, SMD1001 (S1), GM-202B (S2), GM-512B (S3), SYS1000 (S4), and SY-CH2-4 (S5), was constructed. Figure 5 shown.

[0054] Example 2

[0055] The CNN pattern recognition algorithm with PRM coding timing enhancement in the host computer specifically includes the following steps:

[0056] (1) Calculate the response value of the sensor response curve ,in is the load resistance value in the voltage divider detection circuit, is the voltage across the load resistor, is the gas sensitive resistance of the gas sensor, It is the stable gas sensitive resistance value of the gas sensor under clean air;

[0057]

[0058]

[0059] (2) Perform PRM encoding on the response data and convert the one-dimensional time data of each sensor into a two-dimensional tensor as the input of CNN, where and are all data points in the sensor's one-dimensional response data. is the total number of data points;

[0060] in

[0061] (3) The convolutional neural network adopts a layered and progressive feature extraction architecture, specifically including:

[0062] Convolutional layer 1: 32 3×3 convolution kernels, ReLU activation function, L2 regularization coefficient 0.001; batch normalization layer; Dropout layer with dropout rate 0.2; 2×2 max pooling layer;

[0063] Second convolutional layer: 64 3×3 convolution kernels, ReLU activation function, L2 regularization coefficient 0.0005; batch normalization layer; 2×2 maximum pooling layer;

[0064] The third convolutional layer: 128 3×3 convolution kernels, ReLU activation function; batch normalization layer; 2×2 maximum pooling layer;

[0065] Fully connected layer: 64 neurons, ReLU activation function, L2 regularization coefficient 0.001; Dropout layer, dropout rate 0.6;

[0066] The output layer has two neurons, a Softmax activation function, and an Adam optimizer. The pattern recognition algorithm is the PRM-CNN algorithm. The microcontroller sampling frequency is 0.2 Hz, the acquisition time is 15 minutes, and the sensor array contains five sensors, resulting in a raw data size of 5 x 900. After cropping the first 350 seconds of data, the sample size is 5 x 110. The one-dimensional time data from each sensor is PRM-encoded, generating a 110 x 110 two-dimensional tensor, which ultimately serves as the CNN input data. PRM encoding enhances the temporal characteristics of the one-dimensional time signal. Since CNN algorithms are sensitive to local spatial features, the PRM-CNN algorithm can fully extract the spatiotemporal features in the sample data, improving recognition accuracy. The CNN architecture consists of a three-layer convolutional cascade. Each convolution layer is followed by a ReLU activation function, L2 weight decay, and batch normalization. Spatial dropout regularization is additionally introduced in the first two layers to enhance model robustness. The output layer generates a binary probability distribution through the Softmax function, which enables accurate distinction between healthy volunteers (HE) and lung cancer patients (CA) samples.

[0067] Example 3 Electronic Nose Hardware System Construction

[0068] Based on the sensor array, an electronic nose detection system is built, which mainly includes a MEMS gas sensor array, an enrichment device, a solenoid valve, a gas chamber, a single-chip main control board, an air pump, etc. Figure 5 As shown, the sensor array and the closed gas chamber combination constitute the core of the detection device; the Tenax-GR enrichment tube and the heating wire constitute the enrichment device 3; the positive pressure air pump 1-first solenoid valve 2-blank gas path 4-second solenoid valve 5-gas chamber 6 constitute the first gas path; the first solenoid valve 2-enrichment device 3-second solenoid valve 5-gas chamber 6 constitute the second gas path; the single-chip microcomputer 8 serves as the hardware main control unit, which is responsible for collecting the sensor array signal and uploading it to the host computer 9 through UART communication, and receiving commands from the host computer to control the gas path switching and heating and desorption of the enrichment device.

[0069] The enrichment device is a Tenax-TA enrichment tube (2,6-diphenyl-p-phenyl oxide). When powered on, it is heated to 200°C using a heating wire for desorption, and the desorbed gas reacts with the gas sensor.

[0070] The top of the chamber is a semi-ellipsoid with a height of 10mm and a base radius of 30mm, while the bottom is a cylindrical structure with a height of 10mm and a base radius of 30mm. Air enters through a circular inlet with a diameter of 6mm at the top, diffuses evenly through the cavity, and exits through circular outlets with a diameter of 3mm on all sides. The chamber is made entirely of high-temperature-resistant and non-adsorbing polytetrafluoroethylene, with a total volume of 47mL. This significantly reduces the volume of the electronic nose's chamber and improves the efficiency of sample gas utilization.

[0071] Figure 6 and Figure 7 They are the upper and lower PCB diagrams of the sensor array. Figure 8 and Figure 9 They are schematic diagrams of the upper and lower layers of the microcontroller main control board respectively.

[0072] Example 4 Electronic Nose Detection Sample Process and PRM-CNN Algorithm Recognition of Sample Set

[0073] The overall scheme of the electronic nose detection system includes the following steps.

[0074] Step 1: Sample collection and pretreatment

[0075] Active sampling method was used to collect exhaled gas samples from volunteers. The samples were controlled at a constant flow rate of 150 sccm (standard cubic centimeters per minute), and samples containing volatile organic compounds (VOCs) were continuously introduced into Tenax-GR adsorption tubes for target enrichment.

[0076] Step 2: Thermal desorption and detection timing control

[0077] After the enrichment tube is loaded into the electronic nose detection cabin, the detection process control program is started, and the system performs precise timing control:

[0078] Baseline calibration phase (0-5 minutes): Clean air is introduced into the first gas path to purge the gas path to ensure that the baseline of the sensor array is stable;

[0079] Thermal desorption stage (5-10 minutes): At the 5th minute, the heating module of the second gas path is activated, causing the enrichment tube to reach 200°C at a heating rate of 20°C / s and maintain a constant temperature to promote the thermal desorption of VOCs. At the 6th minute, the three-way solenoid valve is energized, and the gas path is switched to the second gas path. The carrier gas carrying the desorption products enters the gas sensing chamber with a volume of 50 cm³ at a flow rate of 150 mL / min, undergoing surface adsorption reaction with the gas sensor array.

[0080] System reset stage (10-15 minutes): In the 10th minute, the thermal desorption module is turned off and the gas path is reset. Purified air is input through the first gas path to reset the sensor response and clean the gas path.

[0081] Step 3: Signal acquisition and data storage

[0082] The sensor array's resistance change is converted into a voltage signal via a voltage divider circuit. After high-frequency noise is eliminated by a low-pass filter, a voltage follower composed of rail-to-rail operational amplifiers provides impedance matching. An STM32F103 microcontroller performs 12-bit A / D conversion at a 10Hz sampling frequency, transmitting the sensor array signal curve in real time to a host computer via a serial port protocol. The host computer interface displays the dynamic curve simultaneously, and the raw data is stored in a local database in CSV format.

[0083] Step 4: Data Analysis

[0084] After the gas detection is completed, the sensor's response data has been transmitted to the host computer through the serial port for storage. Click "Data Analysis" on the host computer to call the host computer's analysis algorithm, obtain the diagnosis results and display them on the host computer interface.

[0085] Among the 187 subjects enrolled during 2024-2025, after strict exclusion criteria screening (excluding 25 cases with concurrent benign lung tumors or malignant tumors of other organs), 162 eligible subjects were randomly divided into independent training sets (129 cases) and test sets (33 cases) in an 8:2 ratio. Based on this dataset, the present invention systematically evaluated the performance of four classification models, namely PCA, SVM, 1D-CNN, and PRM-CNN, in lung cancer diagnosis through a strictly designed experimental framework. Figure 4 As shown in the figure, the PRM-CNN algorithm in this application is higher than other algorithms in sensitivity, specificity, F1 score and accuracy, which proves the overall superiority of this algorithm.

[0086] The above embodiments are only used to illustrate the present invention. Any equivalent transformations and improvements based on the technical solution of the present invention should not be excluded from the protection scope of the present invention.

Claims

1. An electronic nose for early lung cancer screening based on a MEMS sensor array and a hybrid algorithm, comprising a MEMS gas sensor array, an enrichment device, a gas chamber, a single-chip microcomputer control board, an air pump, a host computer, and a recognition algorithm; characterized by: The electronic nose also includes a MEMS gas sensor array obtained by a dual screening method of "correlation coefficient + discrimination performance value", and the host computer adopts a CNN pattern recognition algorithm with PRM coding timing enhancement; The sensor array is: SMD1001, GM-202B, GM-512B, SYS1000, SY-CH2-4.

2. The electronic nose according to claim 1, wherein: The process of screening and obtaining a MEMS gas sensor array includes the following steps: (1) Test the dynamic response characteristic curves of various MEMS gas sensors to lung cancer marker gases at a concentration of 1 ppm; (2) Calculate the correlation coefficient between each two gas sensors for the same gas based on the dynamic response characteristic curve : ； in, and They represent the response values ​​of the i-th sampling point in the timing response curves of sensors X and Y for a specific gas sample; and is the average response value of the corresponding gas sensor in the entire response cycle; N represents the total number of sampling points in a single gas detection; The range of the correlation coefficient is [-1,1], and its absolute value |R xy The larger the |, the higher the similarity of the response patterns of the two sensors. A similarity above 0.9 is considered to be highly similar to the gas sensors. If it is below this threshold, it is directly judged as a low similarity relationship. (3) Calculate the discrimination performance value of each gas sensor for different target gases based on the dynamic response characteristic curve : i=1,2,…,n k=1,2,…,m; in, is the inter-class dispersion of samples, Intra-class dispersion; Represents a standard gas sample detected by the kth gas sensor; Indicates the average value of the total samples of a certain gas corresponding to the gas sensor; u represents the average value of the total samples of all gases corresponding to the gas sensor; n represents the number of test gas types; m represents the number of test samples for each gas; (4) Taking the sensor's specific recognition ability for gas as the core, the gas sensor SMD1001 with the largest gas discrimination performance value is first selected, recorded as S1, and then the sensors with correlation coefficients greater than 0.9 are eliminated; (5) Repeat step (4) for the remaining gas sensors until the correlation coefficients between the sensors are all less than 0.9; SMD1001 (S1), GM-202B (S2), GM-512B (S3), SYS1000 (S4), and SY-CH2-4 (S5) are selected to form the sensor array.

3. The electronic nose according to claim 1, wherein: The process of the recognition algorithm is: (1) Calculate the response value of the sensor response curve : ； ； in is the load resistance value in the voltage divider detection circuit, is the voltage across the load resistor, is the gas sensitive resistance of the gas sensor, It is the stable gas sensitive resistance value of the gas sensor under clean air; (2) Perform PRM encoding on the preprocessed data and convert the one-dimensional time data of each sensor into a two-dimensional tensor as the input of CNN: in ; in, and are all data points in the sensor's one-dimensional response data. is the total number of data points; (3) Convolutional neural networks use a layered and progressive feature extraction architecture, specifically including: Convolutional layer 1: 32 3×3 convolution kernels, ReLU activation function, L2 regularization coefficient 0.001; batch normalization layer; Dropout layer with dropout rate 0.2; 2×2 max pooling layer; Second convolutional layer: 64 3×3 convolution kernels, ReLU activation function, L2 regularization coefficient 0.0005; batch normalization layer; 2×2 maximum pooling layer; The third convolutional layer: 128 3×3 convolution kernels, ReLU activation function; batch normalization layer; 2×2 maximum pooling layer; Fully connected layer: 64 neurons, ReLU activation function, L2 regularization coefficient 0.001; Dropout layer, dropout rate 0.6; Output layer: 2 neurons, activation function is Softmax; Adam optimizer is used.

Images