Artificial Neural Network Assisted Cancer Risk Prediction of Oral Precancerous Lesions.

The incidence of oral cancer is still increasing. It has become very common in patients with malignant tumors, which has forced medical personnel to continuously explore its treatment methods. What kind of method can effectively and correctly diagnose the disease in the early stage and improve the survival rate has become one of the research topics that have attracted much attention. Aiming at this problem, it has great research significance for the field of oral precancerous lesions diagnosis. With the in-depth research on oral precancerous diagnosis, the research on artificial neural network (ANN) in medical diagnosis is gradually carried out. Its performance advantage is of great significance to solve the problem of early and correct disease diagnosis. This paper aimed to investigate the application of ANN-assisted cancer risk prediction method in risk prediction of oral precancerous lesions. Through the analysis and research of ANN and oral cancer, the construction of oral cancer risk prediction model was applied to solve the problem of improving the survival rate of oral cancer patients. In this paper, ANN and oral precancerous lesions were analyzed, the performance of the algorithm was experimentally analyzed, and the relevant theoretical formulas were used to explain. The results showed that the method had higher accuracy than traditional forecasting methods. When N = 2, the output accuracy was above 90%. It can be seen that the algorithm can meet the needs of the diagnosis of high-risk groups of oral cancer lesions, and the diagnosis efficiency and patient survival rate has been greatly improved.

1. Introduction

In the current era of continuous advancement of medical technology, traditional diagnostic technology is undergoing technological transformation. Common disease diagnosis methods have become incapable of treating increasingly complex disease conditions. Both in terms of speed and accuracy, it cannot meet the ever-increasing demands of patient disease diagnosis. Cancer risk prediction is an important diagnosis and treatment method for cancer treatment at present, which can solve the problem of low survival rate caused by patients seeking treatment too late. Due to its advantages in improving the survival rate of patients, it has been applied to the diagnosis and treatment of various diseases to successfully solve various cancer blocking problems. The secret to finding oral cancer is to correctly identify precancerous lesions. If correct diagnosis, continuous monitoring, and timely intervention can be achieved, it is important to stop or reverse changes in cancer. For the diagnosis of oral precancerous lesions, its characteristics exist in the form of data changes. How to process a large number of index data quickly and efficiently has far-reaching significance for the diagnosis of oral precancerous lesions. ANNs are nonlinear, adaptive information processing systems composed of a large number of interconnected processing units. ANN has better effect and less restrictions on the risk prediction problem of oral cancer lesions to be solved. Therefore, its application range is very wide. In recent years, scholars have used ANN for medical risk prediction, but there are relatively few applications and researches of ANN in this area. Therefore, it is of great significance to apply ANN in this paper to solve the cancer risk prediction of oral precancerous lesions.

At present, with the continuous updating of medical methods in the era of big data, more and more scholars have explored the risk prediction of oral cancer. Among them, Said B. I. developed and validated a predictive score for locoregional failure and distant metastasis in oral squamous cell carcinoma (OSCC) for risk assessment of patients [1]. To assess postoperative risk and prevention, Tanda N. performed intensive oral care for each patient and measured swallowing and coughing reflexes in 12 of 13 patients [2]. Dhami S. believed that the incidence and mortality of cancer were strongly associated with the formation of venous thromboembolism. He proposed that understanding the biological mechanisms underlying venous thromboembolism (VTE) risk may provide insights into novel targeted prevention in cancer patients [3]. Using risk prediction, Leader A determined the incidence of intracranial hemorrhage in patients with brain metastases receiving direct oral anticoagulant (DOAC) or low molecular weight heparin (LMWH) for venous thromboembolism or atrial fibrillation [4]. Liu F. assessed the potential of different oral species and subject characteristics in predicting sodium diacetate (SDA) risk by constructing a conditional logistic regression model [5]. However, the accuracy of the method used for the risk prediction of oral cancer is not high.

ANN can be used in oral cancer risk prediction, and has good performance in data processing and prediction accuracy. Among them, Leili T. used the ANN method to identify prognostic factors associated with renal transplant rejection and compared the results with those obtained by logistic regression [6]. In order to achieve accurate prediction results, Huang L. J. tested the ANN model with 15 influencing factors and 7 influencing factors obtained by principal component analysis (PCA), respectively [7]. Yang B. proposed an intelligent RUL prediction method based on dual CNN model architecture to predict cancer risk [8]. Daoud M. reviewed recently proposed models to highlight the role of neural networks in predicting cancer from gene expression data [9]. These methods improve the accuracy of oral cancer risk prediction from data to a certain extent, but the methods themselves are too complicated.

In order to solve the above-mentioned problem of low accuracy of cancer risk prediction of oral precancerous lesions, this paper used ANN to analyze the risk of oral cancer. Through the simulation of oral cancer risk prediction, the effect of improving the prediction accuracy can be achieved. The innovation of this paper is that it introduces the use of ANN and oral precancerous lesions to analyze how ANN, oral precancerous lesions, and risk prediction models play a role in ANN-assisted oral precancerous lesions in cancer risk prediction research. The proposed risk prediction method is described. Through experiments, it is found that the method has a good effect on the prediction of oral precancerous lesions.

2. Methods of Cancer Risk Prediction of Oral Precancerous Lesions

The prevalence of oral cancer is on the rise, but it can be cured if treated properly. The best stage of treatment is in the early stages of cancer. If the risk prediction of precancerous lesion characteristics of oral cavity can be carried out as soon as possible, the survival rate of oral cancer patients can be effectively improved. Oral precancerous lesions include leukoplakia, oral lichen planus, etc. Therefore, risk prediction for patients with oral precancerous lesions is very important for the diagnosis and treatment of oral cancer [10]. Oral precancerous lesions are pathological changes in body tissues that may evolve into cancers. According to investigations, many cancers have different forms of precancerous states before they occur. Common oral precancerous lesions are shown in Figure 1.

Figure 1

Common oral precancerous lesions.

Among them, Figure 1(a) shows that leukoplakia appears in the oral cavity; Figure 1(b) shows that lichen planus appears in the oral cavity. Through the investigation, it is found that the current research on ANN-assisted cancer risk prediction of oral precancerous lesions is not complete enough. Most of the research has focused on case reviews and theoretical summative analyses of past treated cases. Therefore, this paper proposes a research on the cancer risk prediction of ANN-assisted oral precancerous lesions [11, 12]. This paper analyzes the related methods of oral precancerous lesions, ANN and risk prediction model construction. A canceration risk prediction model for oral precancerous lesions is also proposed. The experimental analysis shows that the model prediction results are more accurate.

This paper mainly introduces the research background based on the application of ANN in oral precancerous risk prediction, and elicits the problems to be solved to illustrate the purpose and significance of this paper. Then, a general analysis of the research status in the field of oral cancer risk prediction and the application of ANN is carried out, and the content and innovation of this study are explained. The organization structure and methods of the full text of this paper are described, and the related methods of oral precancerous lesions, ANN and risk prediction models are analyzed and described. Next, the data source of this paper is explained in detail, and the conclusion is drawn after analyzing the result data after arranging the data. Finally, the full text is summarized.

2.1. ANN Method

A complex network structure consisting of many easily connected basic processing units constitutes an ANN model. The development of the BP algorithm of the multilayer network makes the most popular network BP network possible [13, 14]. Due to the development of the BP algorithm for multilayer networks, BP network is currently the most popular network [15]. The field of using ANN is shown in Figure 2.

Figure 2

ANN application areas.

As shown in Figure 2, smart healthcare is now a hot area in the healthcare industry. ANNs can be used for intelligent guidance, building electronic medical records, medical image analysis, early cancer detection, patch design, and optimizing transplant time [16-18]. Now, many hospitals are promoting ANN for doctors' diagnosis support and providing diagnosis and treatment advice. This technology is very important for remote areas with low medical standards. In the field of communication, ANN has been applied to image and data compression, automatic information service, real-time spoken language translation, customer payment processing system, etc. In the aviation field, ANNs are suitable for automatic maneuvering, circuit simulation, flight control systems, driving enhancement, component fault diagnosis, etc. From the perspective of transportation system, ANN is used for urban traffic guidance, driving route planning, and vehicle scheduling. Military applications in ANN are weapons operations, target tracking and identification, and sonar and radar signal processing.

Regarding the application scheme of ANN, it is necessary to mention the two most concerned areas of computer vision and natural language recognition. These are simulations of human vision and hearing, respectively. The most popular application scenarios in these two fields are autonomous driving and various speech recognition applications such as smart speakers, voice input, and instant translation [6]. The function of ANN is very powerful, and its functional characteristics are shown in Figure 3.

Figure 3

Features of ANN.

As shown in Figure 3, the representation of the pattern classification problem in ANN is to map n-dimensional feature vectors to scalar or vector-represented classification labels. The powerful nonlinear function of ANN can properly describe the nonlinear classification surface, which leads to better pattern recognition function. Clustering does not need to provide classified samples to ANN, but only needs to cluster them according to the provided samples themselves. Regression and fitting are relatively basic statistical methods of data. The optimization calculation finds a combination of a set of parameters under a set of constraints that minimizes the objective function determined by the combination. The weight matrix is adjusted to minimize output error. This process has a specific randomness and does not require an export. Data compression is that ANN stores specific knowledge in weights. That is, the original sample is described with less data and its properties for simplicity.

The construction of artificial neurons is learned from the structure and working principles of biological neurons [19]. The artificial neuron output is shown in

Among them, h is the relation function; Q is the link strength; Z is the output vector; α is the threshold.

BP neural network is a feedforward network, which usually consists of three layers: input layer, hidden layer, and output layer. Its development is based on the development of biological neural networks. Through continuous learning and error correction, BP neural network (BPNN) can replicate brain stimuli, respond to external inputs, and adapt to changes in the experimental environment. It has powerful adaptive, self-learning and generalization capabilities. The network can be trained on known data samples and then process and model complex data by looking at the outputs and the internal relationships between the inputs and outputs.

BP neurons are comparable to other neurons. However, the difference is that it has a nonlinear transfer function and the output layer uses a linear function purelin, the most popular being the logsig and tansig functions [20]. Its output expression is shown in

Among them, the expressions of the logsig and tansig functions are shown in

BPNN lowers the input learning samples and the two processes of modifying the full-time and threshold until convergence, but there are some problems with this algorithm, such as problem complexity, slow convergence, network easily limited to minima, learning process vibration, etc.

During network training, it is easy to have small errors on samples from the training set, but large errors on new sample data outside the training set. It is often necessary to modify the size of the network to make it “fit” to increase the generality of the network. Normalization is often used to increase the generalization ability of the network. The mean squared error function is a typical error performance function of a forward feedback network. The basic idea is to modify the network error performance function, as shown in

The network enhances generalization by adding the mean square of the weights and thresholds of the carrying network to the error performance function. The process is shown in

Among them, β is the error performance adjustment rate.

The perceptron is the basic unit of a neural network. It consists of weights, offsets, and transfer functions [21]. z is the value after linear processing, and its expression is shown in

Among them, Q is the weight; p is paranoia.

When processed by the transfer function, the final value is obtained, as shown in

Among them, p is the input signal of the previous perceptron; h(z) is the final value.

If no transfer function is added, learning multiple hidden layers has the same effect as using only one hidden function, as shown in

Therefore, it is necessary to process the transfer function.

The application of the node design of the hidden layer in this paper belongs to the classification of pattern recognition, and its design can be referred to as shown in

Among them, x is the number of nodes in the hidden layer; x is the number of input nodes; x0 is the number of output nodes; δ is the constant term.

In the design of the transfer function, the sigmoid function is usually used, and its expression is shown in

Among them, if the last layer is a signaling function, the output of the entire network is limited to the range of 0-1. If the last layer is a pure function, the output of the entire network can take arbitrary values [22].

If one of the (last layer) functions is a signaling function, the output of the entire network is constrained to a range of 0–1. If the last layer is a pure function, the output of the entire network can take any value.

2.2. Risk Prediction Model Based on ANN

In this paper, ANN is used to establish a cancer risk prediction model of oral precancerous lesions. The weighted sum of all the values of the previous layer is used to create the value output in the network, which is then processed by the transfer function. After processing each layer, the output value is obtained. The weights can then be modified by using backpropagation. When the training set is completed after a few epochs, the algorithm updates the learning times. If the number of learning times is within the threshold, the training process would end. Currently, the input signal comes from the input layer, and the final prediction result can be obtained by layer-by-layer computation [23].

The collected data needs to be standardized due to different dimensions. Usually, the process of data standardization should be carried out first, and different data attribute types can use different standardization methods [24]. Attribute types can be divided into benefit type, cost type, interval type, etc. Taking the benefit type as an example, it is shown in

If the attribute type is interval, it is as shown in

As soon as the data packets are substituted into the formula, the data normalization process is complete.

The components for BPNN for risk prediction are relatively simple. The BPNN algorithm flow is shown in Figure 4.

Figure 4

BPNN algorithm flow.

As shown in Figure 4, the algorithm includes preoutput, back-propagation, update weights, and other parameters [25]. First, the signal parameters of the input layer are set, and weights are randomly assigned. By calculating according to the above process, a set of calculation results are obtained at the output layer. The error between the predicted value and the actual value is then calculated. Vice versa, the relationship between the final error and the initial weight is calculated. Next, a probabilistic gradient descent algorithm is used to find the wrong minimum. Finally, by repeatedly updating the weights, the prediction results can be obtained after training is complete.

The input value reaches the first unit value of the hidden layer through weight calculation as shown in

Among them, q is the weight; u is the order of the sensing unit; l is the weight of the previous layer.

Then the signal value processed by the transfer function is shown in

Among them, h(x) is the transfer function.

The calculation to obtain the total error after calculating the errors individually is shown in

Among them, W is the total error value; p is paranoia; k is the number of layers; d is the label.

If the error value of the output result is the smallest, according to the above Formula, the value of εW/εQ needs to be calculated, as shown in

Among them, s is the output value.

Finally, the weights are updated to obtain the updated weights as shown in

In the following, after the establishment of the cancer risk prediction model of oral precancerous lesions based on ANN, the effect of its use is tested through experiments.

3. Data Sources for Cancer Risk Prediction of Oral Precancerous Lesions

The data used for the algorithm test in this paper are the case data parameters of patients with oral precancerous lesions that are screened out from the web page information through data mining. A total of 230 case samples were included in the experiment in this paper, and they were sorted to form the training and testing of the cancer risk prediction model for oral precancerous lesions. Some data contents are shown in Table 1.

Table 1

Partial raw data of the case dataset.

Project	Case sample1	Case sample2	Case sample3
Year of birth	1976	1988	1964
Age	46	34	58
Gender	Male	Male	Female
Smokes	1	0	1
Oral disease	1	1	1
True value of cancer	0.17	0.04	0.34

The information included includes the case's year of birth, age, gender, smoking, oral disease, and true cancer risk. In the Table, 1 represents yes, 0 represents no.

Due to the different dimensions, normalized data is required to be used. For example, the younger one is, the less likely one is to get cancer. After the attributes are normalized, some data are obtained as shown in Table 2.

Table 2

Partial data after normalization.

Project	Case sample1	Case sample2	Case sample3
Year of birth	0.635	0.514	0.684
Age	0.614	0.498	0.674
Gender	0	0	0
Smokes	0.514	0	0.547
Oral disease	0.741	0.671	0.657
True value of cancer	0.17	0.04	0.34

The data is plugged into the algorithm to obtain the final prediction. Part is used for training and part is used for prediction. Prediction results can be obtained after training on a partial dataset.

4. Results and Discussion of Cancer Risk Prediction in Oral Precancerous Lesions

4.1. Case Grouping Prediction Model Test and Results

This paper used 230 case samples and divided them into training sample set and test sample set according to the ratio of 7 : 3. The samples were grouped, in which group A was patients with oral cancer, and group B was patients with nonoral cancer. The specific results of training and testing are shown in Figure 5.

Figure 5

Model prediction results.

Figure 5 shows the training and testing results of the prediction model. Among the 161 training samples shown in Figure 5(a), 86 of the 91 nonoral cancer patients were correctly assigned to group B. Of the 70 patients with oral cancer, 62 were correctly assigned to group A, and the overall percentage of correct predictions was 92.5%. Among the 69 test samples shown in Figure 5(b), 30 of the 34 nonoral cancer patient samples were correctly assigned to group B. Twenty-eight out of 35 oral cancer samples were correctly assigned to group A, with an overall correct percentage of predictions of 94.2%. It can be seen that the correct rate of the model in both training and testing reaches more than 90%, which can meet the grouping prediction requirements of this paper.

The results of the tested models are summarized as shown in Table 3.

Table 3

Summary of results.

Train	Sum of squares error	15.347
	Error rate	7.5%
	Training time	46 s
Text	Sum of squares error	9.324
	Error rate	5.7%

In the summary of results in Table 3, the training sample error rate of 7.5% corresponds to the test sample prediction error rate of 5.7%. The predicted quasi-probabilities for the combined training and test samples can be displayed for the observation prediction plot. The specific results are shown in Figure 6.

Figure 6

Training and test samples predict pseudo-probabilities.

As shown in Figure 6, Figure 6(a) showed the predicted pseudoprobabilities of class A for the observed samples of class A. The part with correct prediction value greater than 0.5 and the part less than 0.5 were marked as wrong prediction value. Figure 6(b) showed the predicted pseudoprobabilities of class B for samples of observation class B. Likewise, predictions greater than 0.5 fractions were correct values, while wrong predictions were fractions less than 0.5.

4.2. Abnormal Identification Test of Oral Mucosal Cell Smear

This paper used an ANN to take the gender, age, smoking, and drinking status of 682 patients as input values. Oral cancer and precancerous lesions of this population were screened, and pathological diagnosis was used as the evaluation standard, which was compared with the results of manual identification. Among them, the data of 447 patients was fed into a three-layer feedforward network as training data. The BP algorithm was used for training, and the network was tested on the remaining 205 patients. The specific results are shown in Figure 7.

Figure 7

Abnormal discrimination test results on oral mucosal smear.

As shown in Figure 7, on the whole, the neural network and human discrimination results were comparable. Figure 7(a) showed that the neural network was 81.6% sensitive to samples and the artificial was 74.2% sensitive to samples. Figure 7(b) showed that the specificity of the neural network to the sample was 77.3%, and the specificity of the artificial to the sample was 98%. The use of neural networks in screening can help screen high-risk groups. ANN does not require screening personnel training, and also saves time for professional doctors to check.

4.3. Tissue Autofluorescence Spectroscopy Experiments

In this paper, the autofluorescence intensities of the lesion specimens of 34 patients with oral leukoplakia and 22 normal oral mucosa specimens of volunteers were normalized and input to the BP network. And when the number of neurons in the output layer N is 2, 3, 4, and 5, the set accuracy of the output result is shown in Figure 8.

Figure 8

Detection results of autoluminescent fluorescent substances in oral cells and tissues.

As shown in Figure 8, using a neural network, the recognition of autofluorescence in oral mucosa specimens from patients with oral leukoplakia was overall excellent. Figure 8(a) showed that when N = 2, the output accuracy was above 90%. When N = 3, the output accuracy decreased slightly, the highest was 89.4%, and the lowest was 80.3%. Figure 8(b) showed that when N = 4, the output accuracy reached the highest 89.6% and the lowest 80.1%. When N = 5, the output accuracy rate reached 87.6% at the highest and 82.3% at the lowest. It can be seen that the output results were best when N = 2 was set, where the sensitivity and specificity for distinguishing abnormal and normal tissue were 100% and 83.7%, respectively. When N = 3, 4, and 5 were set, the overall recognition accuracy had decreased, but it can be maintained above 80%. It can be seen that the neural network can meet the needs of identifying abnormal cells and tissues in patients with oral leukoplakia.

5. Conclusions

The rising incidence of oral cancer deserves attention. People's requirements for the diagnosis and treatment of oral cancer are also getting higher and higher. The development of oral cancer risk prediction is inseparable from the contribution of ANN. ANN has been widely used in many fields because of its data processing advantages. Through comprehensive experimental tests, it can be seen that ANN-assisted oral cancer risk prediction can improve the prediction accuracy, which is conducive to early detection of treatment. Through the training and testing of the case grouping model, the accuracy rate of the model in both training and testing is more than 90%, which can meet the needs of the case grouping prediction in this paper. Through the analysis of the abnormal identification of oral mucosal smear, the results are comparable to those of manual screening, with sensitivities of 81.6% and 74.2, and specificities of 77.3% and 98%, respectively. It can be seen that the use of neural networks in screening helps to screen high-risk groups and saves the time for professional doctors to check. Through the experimental analysis of tissue autofluorescence spectrum, it is found that the output results are the best when N = 2, which can reach more than 90%. It can be seen that the neural network can meet the needs of identifying abnormal cells and tissues in patients with oral leukoplakia. At the end of the experiment, the experimental process of this paper has reviewed and the problems have reflected from it. If more nonimportant factors are eliminated in the experimental test, the final result should be more accurate. It is hoped that this paper can provide reference for future research and can be improved in future research.