| Background and ObjectiveLung cancer is one of the most common cancers. Every year more than one million people die of lung cancer in the world. The incidence of lung cancer has risen to the second place in China, its mortality in the top of the urban population, and its incidence rate tends to rise. It is due to lacking an effective early diagnosis means of lung cancer. Since the majority of diagnosed patients with lung cancer is already in the advanced stage, so the early diagnosis of lung cancer is to determine the prognosis of patients. Therefore, early detection, early diagnosis and early treatment are crucial to improve the survival rate of patients with lung cancer, the key to reducing mortality. Currently, there are three main ways in diagnosis of lung cancer: Imaging diagnosis, chemical diagnosis (serology and immunology) and cytology histology diagnosis.Artificial neural network (Artificial Neural Networks, ANN) is a nonlinear structural engineering information processing system which is on the basis of the simulation of the human brain mechanism of understanding and wisdom conducts. It is a simplification and simulation of information processing systems such as biological neural network structure function. Artificial neural networks has massive parallel processing, distributed information storage capacity, good adaptive ability, strong self-organization learning function and fault tolerance functions. At present, artificial neural networks have been widely used in finance, commerce, information, medical and other fields. In the field of medicine, Artificial neural networks has been used in clinical diagnosis of the disease (expert system), disease screening and diagnosis, disease-related factors studies, predict disease risk, survival analysis , gene identification, DNA and RNA sequence analysis and protein structure analysis.Tumor markers for the early detection and early diagnosis of lung cancer has been widespread attention in clinical , because the clinical stage of lung cancer can decide to the survival of patients, and the cure rate of early lung cancer is better than those advanced. Bronchoscopy has become a more conventional inspection method. The complications were fewer and the patients were more easily received compared with other invasive means. Bronchoscopy also can enhance disease diagnostic level combined with clinical manifestations. But images of bronchoscopy are combined the clinician unskilled in operation makes that the observation can not easy make the right judgement. If an artificial aided diagnosis system can established, it will prevent the subjective of physicians who lacked experience and knowledge and analyzed the characteristics of the image, it can greatly improve the effectiveness of endoscopic examination and the ratio of lung cancer diagnosis. The application of Artificial neural networks in the diagnosis of medical imaging are also based on artificial neural network learning, adaptive, fault-tolerant, non-linear processing functions.In this study , the artificial neural network combining with tumor markers was applied to fiberoptic bronchoscopy in the diagnosis of lung cancer, expected to enhance the accurate and objectivity of the diagnosis, offer clinician more accurate and more effective reference, then, compared with traditional method to discuss the development prospects.Materials and Methods1. There are 119 fiberoptic bronchoscopy (bronchoscopy) images, including 64 benign tumor images and 55 lung cancer images, serum samples.2. There are four tumor markers carcino-embryonic antigen (CEA), neuron-specific enolase (NSE), squamous cell carcinoma antigen (SCC-Ag), cytokeratin-fragment antigen (CYFRA21-1) .Three physicians give a score for every image. Input the 11 score results and 6 clinical parameters, training the network, building models and blinding the simulation.3.Matlab6.5,SPSS13.04. Use trained artificial neural networks model to predict the forecast set. We compare the performance of serum model, image data model and the combine data model, compared artificial neural networks with logistic regression by means of ROC.Results1. The results of artificial neural network trainingUsing back-propagation algorithm, artificial neural network achieved the desired objectives, stopped training.2. Building models to predict the lung cancer by use of serum tumor markers, bronchoscopy images and the combine data.The BP neural network model builded by use of serum tumor markers to predict lung cancer, with the sensitivity, specificity, accuracy, positive predictive value, negative predictive value and area under the curve were 64.0 percent, 82.8 percent, 74.1 percent, 76.2 percent, 72.7 percent and 0.782; the model builded by use of images to predict the lung cancer with the sensitivity, specificity, accuracy, positive predictive value, negative predictive value and area under the curve were 68.0 percent, 82.8 percent, 75.9 percent, 77.3 percent, 75.0 and 0.890; the model builded by use of combine data to predict lung cancer with the sensitivity, specificity, accuracy, positive predictive value, negative predictive value and area under the curve were 88.0 percent, 93.1 percent, 90.7 percent, 91.7 percent, 90.0 percent and 0.897.3. The forecast results of the training set and the forecast set using artificial neural network model and Logistic regression model.Artificial neural networks and logistic regression forecast all samples of the training group and test group. The prediction sensitivity of BP neural networks was 94.5 percent, specificity was 96.9 percent, the accuracy was 95.8 percent; the prediction sensitivity of Logistic regression was 74.5 percent, specificity was 85.9 percent, the accuracy was 80.7 percent, the area under ROC curve of BP neuralnetworks and Logistic regression are 0.950 (95%CI: 0.894-0.982) and 0.848 (95%CI:0.778-0.917).4. Comparation between the model builded in this study and early research.The prediction sensitivity of BP neural networks in this study was 94.5%, specificity, accuracy, positive predictive value, negative predictive value were 96.9 percent, 95.8 percent, 96.3 percent and 95.4 percent, the sensitivity specificity, accuracy, positive predictive value, negative predictive value of early research were 100 percent, 98.5 percent, 96.9 percent, 97.9 percent and 100 percent.Conclusion1. The artificial neural network model builded by use of bronchoscopic data and serum tumor markers combine data is better than the model builded only by use of images or serum tumor markers in the forecast of lung cancer.2. The discrimination of the BP neural network model was superior to thelogistic model.3. Artificial neural networks can be used in auxiliary diagnosis of lung cancer as a potential useful tool. |
PDF Full Text Request