Application Of Tumor Marker Group Based On Artificial Neural Network In Diagnosis Of Hepatic Carcinoma

BACKGROUNDAND OBJECTIVEHepatic carcinoma is one of the most common cancers in the world. There has been high incidence of hepatitis and liver cancer in China. Its mortality rate is lower than that of lung cancer, and it becomes second largest. In recently years, the incidence and mortality rate have been increasing. For the lack of typical symptoms of liver cancer in early stage, it is easy to neglect, and easy to confuse with other gastrointestinal diseases. Therefore it is difficult to diagnose the early stage of liver cancer. When there are typical symptoms of liver cancer, it has been often already advanced. This has resulted in high cost of therapy of Hepatic Carcinoma (HCC) and the situation of poor prognosis in China. Currently serum AFP combined with ultrasound imaging are applied to achieve the monitoring of high-risk groups in clinic. Some HCC can be diagnosed in sub-clinical stage. Long-term effect of early excision is particularly significant. So early detection, early diagnosis and early treatment can effectively reduce the mortality rate of liver cancer. But now people rely too much on liver cancer diagnostic imaging. The sensitivity of AFP for liver cancer is not high. There is also defective in ultrasound. Therefore, it is particularly important to study specific and sensitive diagnostic indicators in early stage. People try to look for new tumor markers or combination to help early diagnosis of HCC. Artificial neural network is rapidly developed. It is a new type of information processing systems in recent years. It is particularly suited to medical pattern classification and judging. In this paper, chemiluminescence immunoassay and spectrophotometry were separately used to detect the serum a-fetoprotein (AFP), Carbohydrate antigen 125 (CA125), Carcinoembryonic Antigen (CEA), Sialic Acid (SA) and Calcium (Ca) in three groups. Artificial Neural Network (ANN) technology can be applied to extract effective features of liver cancer and developed the diagnostic model to improve the accuracy of diagnosis of liver cancer. MATERIALS AND METHODS1. Serum samples were collected from 50 cases of liver cancer patients,40 cases of patients with benign liver disease,50 cases of normal human.2. AFP, CA125 and CEA were detected by chemiluminescence immunoassay. SA was determined by spectrophotometry. Ca was detected by calcium assay kit (Azo-end method of arsenicâ…¢).3. Samples of each group were randomly divided into training set and test set. The model was developed by training set, then the samples of test set were used to validate the quality of model. The five tumor markers were used as input data from 35 cases of liver cancer patients,30 cases of patients with benign liver disease, and 35 cases of normal human. Network parameter settings were 5 nodes as input layer, 15 nodes as hidden layer,1 node as output layer,0.001 as target error,0.7 as learning rate,0.95 as momentum factor.RESULTS1. Development of artificial neural network model and discriminant resultBack-Propagation neural network was applied to develop the diagnostic model by training set according to setting error. Then the model was validated by test set. The sensitivity of liver cancer was 96.0%, the specificity was 98.9%, the accuracy was 94.3%, the positive predictive value was 98.0%, the negative predictive value was 97.8%.2. Results of discriminant analysis model for liver cancer-liver benign disease-normal classificationThe sensitivity of liver cancer was 46.0%, the specificity was 98.9%, the accurate rate was 79.3%, the positive predictive value was 95.8%, the negative predictive value was 97.8% for three groups by discriminant model.3. Comparison between artificial neural networks and discriminant analysis ModelsThe sensitivity, accuracy and negative predictive value of artificial neural network model were greater than those of discriminant analysis. The area under the Receiver Operating Characteristic (ROC) curve (0.986) was more than discriminant analysis model (0.842). The difference was significant (P<0.05)CONLUSION1. The model based on artificial neural network combined with 5 tumor markers was superior to single tumor marker or combination of tumor markers group. And it can differentiate hepatocellular carcinoma and benign or normal.2. The model based on artificial neural network combined with 5 tumor markers was superior to traditional statistical methods. It is better than conventional statistical methods for clinical data analysis.