Electronic nose lung cancer early diagnosis system based on neural network

Lung cancer is one of the most common malignant tumors in the world, and its morbidity and mortality rates are rising. The reason is that the cause of lung cancer is unknown, the onset time is short, the metastasis is fast, the malignancy is high, it is difficult to diagnose in the early stage, and the opportunity for surgery is lost in the middle and late stages. The five-year survival rate of patients is only about 15%. Early detection can increase the patient's survival rate to 70% to 80% within five years. Therefore, early detection, early diagnosis, and early treatment are the key to improving the survival rate and reducing the mortality rate of lung cancer. Lung cancer usually has no special symptoms in the early stage, and is almost not noticed by doctors and patients. In addition, it is difficult to achieve early detection and early qualitative diagnosis with commonly used diagnostic methods.

At present, the research on disease diagnosis based on electronic nose mainly focuses on the early diagnosis of kidney disease and diabetes and the identification of the types and growth stages of some bacteria. As one of the important directions of non-destructive medical diagnosis, disease diagnosis based on electronic nose has achieved many results, but there is no report on certified respiratory diagnostic instruments. How to further improve the diagnosis technology of lung cancer and improve the efficacy of various treatments has become a top priority in the field of tumor research around the world in recent years. my country has also listed lung cancer as a national key research topic. Finding a more advanced instrument and technology that can detect and diagnose when local tissues become cancerous is the work to be accomplished by this electronic nose system. This paper designs the key technologies of the electronic nose lung cancer early diagnosis system from the aspects of detection and collection of respiratory gases related to pathology, selection and optimization design of gas sensor arrays, and selection of pattern recognition technology, and has achieved good results.

1. Structure of the electronic nose lung cancer early diagnosis system

An electronic nose is an electronic system that uses the response pattern of a gas sensor array to identify gases. The electronic nose is mainly composed of three functional components: a gas sampling operator, a gas sensor array, and a signal processing system. The main mechanism by which the electronic nose identifies gases is that each sensor in the array has a different sensitivity to the gas being measured, and the entire sensor array has different response patterns to different gases. It is this difference that enables the system to identify odors based on the sensor's response pattern.

The typical workflow of an electronic nose is: first, a respiratory gas collection device (after respiratory gas purification and flow control) is used to draw respiratory gas into a small container chamber containing an electronic sensor array; then, the initialized sensor array is exposed to the gas to be measured. When volatile compounds (VOCs) come into contact with the surface of the sensor active material, an instantaneous response is generated. This response is recorded and transmitted to the signal processing unit for analysis, and compared and identified with a large number of VOC patterns stored in the database to determine the gas type; finally, the cleaning gas flushes the surface of the sensor active material to remove the measured gas mixture. Before entering the next round of new measurements, the sensor still needs to be initialized again (that is, before working, each sensor needs to be cleaned with dry gas or some other reference gas to reach a baseline state). The structure of the electronic nose lung cancer early diagnosis system is shown in Figure 1.

2 Design of the electronic nose lung cancer early diagnosis system

This paper designs key technologies from the aspects of detection of pathologically significant gases in the respiratory gases of lung cancer patients that are highly correlated with the disease, design of respiratory gas collection devices, selection and optimization of gas sensor arrays, and selection of pattern recognition technology.

2.1 Detection of respiratory gases

About 200 compounds have been detected in human exhaled breath, some of which are related to lung cancer, so using exhaled breath to detect diseases is a feasible method. Its advantages are non-invasive, simple and fast, so it has a very high potential for application development. In conjunction with hospitals, selecting appropriate gas sensors and detection methods, and detecting the concentrations of 22 organic volatile (VOCs) characteristic gases such as styrene, decane, and undecane in the exhaled breath of lung cancer patients is a promising non-invasive diagnosis and monitoring method for lung cancer.

2.2 Design of respiratory gas collection device

Since the concentration of lung cancer-related gases produced in exhaled breath is very low (usually at the ppb level), the traditional breath detection method is to use a gas chromatography-mass spectrometer to determine the type of compound according to the standard breath sampling procedure. The analysis process must concentrate a large amount of breath samples (about 3L of breath samples) before reaching the limit value that the instrument can detect. This method is not only expensive and time-consuming, but also requires a large number of sample specimens. The analysis cost required for the electronic nose is not high, and the required breath sample volume is only about 10ml. The operation is simple and the response is fast (a few minutes). The collection of respiratory gas occupies an extremely important position in the electronic nose's early diagnosis system for lung cancer. The structure of the gas collection device is shown in Figure 2.

The arrows in Figure 2 indicate the flow direction of the cleaning gas and the breathing gas. After the entire gas collection device is cleaned by gas, the test subject's breathing gas is exhaled through the air port. After a series of water and irrelevant gas removal, the gas flow rate is controlled by the flow meter and the microprocessor is collected at a fixed time, and then the inactive gas is removed by the heater.

2.3 Selection and optimization design of gas sensor array

In this electronic nose system, the gas sensor array is a key factor. The main factors affecting the performance of gas sensors include materials and molding technology, the application of sol-gel technology to prepare sensitive molds, working conditions and working environment. In addition, the effects of initial process response and oxygen partial pressure on the characteristics of gas sensors must also be considered.

The performance of the gas sensor array directly determines the system's recognition capability, recognition range, service life, etc. Therefore, how to construct the array to improve the performance of the electronic nose system has become an important research topic. The parameters of the sensor array are mainly: array size, sensor type and its selectivity, stability, noise level, and thermal sensitivity.

The sensor array in the electronic nose system can be a monolithic integrated array or composed of multiple discrete components. When more array units are used, the monolithic integrated array shows the advantages of small size and low power consumption; on the other hand, the performance of discrete devices is also constantly improving. Regardless of the type of array used, the scale and size of the array are very important. Properly increasing the number of array units will result in better system recognition capabilities, but sometimes the increase in array units does not improve the recognition effect of the system, and the larger array has higher power consumption and more serious thermal interference between units, which will increase the difficulty of system integration. When constructing the array, the selectivity of each unit in the array must also be considered. If each unit has good selectivity for a specific gas, the array will have a stronger ability to recognize these gases and their mixed gases, but the number of gases it can recognize will be reduced, and the recognition ability for complex mixed gases with more components will be weaker. When constructing the sensor array, sensors with low selectivity and a wide response range can be used to improve the selectivity and accuracy of the system through pattern recognition technology. At the same time, for different recognition objects, individual units with good selectivity are added to simplify the array. In the selection of array units, methods such as normal distribution characteristics of test results, relative standard variance analysis, correlation coefficient analysis, etc. In this system, cross-response characteristics and array stability are the main goals of sensor array unit selection.

2.4 Choice of Pattern Recognition Technology

By utilizing the cross-selectivity of the gas sensors in the array to form a high-dimensional response pattern for the measured medium, combined with pattern recognition technology, a single gas can be qualitatively analyzed or specific components in a mixed gas can be determined. The response of gas sensors is usually highly nonlinear, so conventional pattern recognition methods such as principal component analysis, partial least squares regression, and Euclidean cluster analysis are limited (most conventional classification methods are linear methods, assuming that the response vector is in Euclidean space and the concentration of the measured object is linearly related to the sensor's response. This is only the case when the concentration of gas and odor is very low). Artificial neural networks can process nonlinear data, tolerate sensor drift and noise, have good robustness, and have a higher prediction accuracy than conventional methods.

Since the relationship between the sensor response value and the measured gas composition is very complex and difficult to express with a clear mathematical relationship, neural network technology is used to establish the mapping relationship between the sensor array response signal and the measured gas. The radial basis function RBF (Radial Basis Function) neural network can overcome the problems of local minimum and low efficiency to a certain extent, and has obvious advantages over the BP neural network in function approximation. Based on the above analysis, this system adopts the RBF neural network pattern recognition method. Figure 3 is the topological structure of the RBF neural network.

The RBF neural network consists of an input layer, an intermediate layer (hidden layer) and an output layer. Here, the input layer only transmits data information without any transformation. The kernel function (or action function) of the hidden layer neurons is a Gaussian function, which performs spatial mapping transformation on the input information. The action function of the output layer neurons is a Sigmoid function, which linearly weights the information output by the hidden layer neurons and outputs it as the output result of the network. The supervised learning method is used to train the neural network to determine the center, width and adjustment weight of the network. From the test samples, 60 samples out of 80 are randomly selected as the training set, and the remaining 20 are the test set. Three experiments are performed under different temperature and humidity conditions.

The network training parameters are momentum factor α = 0.09, learning factor η = 10.12, the maximum number of training times is 20,000, the target error is 0.01, the training time is about 3 minutes, and the network meets the target error requirements. The trained network is tested on the samples, and the results are shown in Table 1. For the three experiments, the correct discrimination results reached more than 90%. Such results are satisfactory, indicating that this application can detect lung cancer patients early.

This paper established an electronic nose system that can quickly and accurately diagnose lung cancer. The electronic nose system consists of a sensor array. In data processing, the obtained sensor data was processed by RBF neural network for pattern recognition. Three experiments were conducted under different temperature and humidity conditions. In the entire test process, except for the headspace gas stabilization of about 2 minutes when the sample is placed and the acquisition of sensor and sample gas reaction data, which takes about 2 minutes, the other data processing takes less than half a minute, so the time for testing a sample does not exceed 5 minutes. However, since the developed electronic nose is still in the laboratory stage, there are still many issues that need further research, such as how to improve the existing device and optimize the sensor array in terms of device; in terms of data processing, the extraction of eigenvalues ​​and the improvement of pattern recognition algorithms.