The Study On The Identification Of Primary Lesion Unknown Cancer With Artificial Neural Network And Serum Tumor Markers

Objective:Tumor is a kind of complex disease with the participation of multiple genes, the development in multiple steps and the interaction of internal and external environment.The serum levels of the tumor markers are closely associated with the occurrence-recurrence and regression of malignancy. The positive rate and accuracy of using a-single tumor marker for early diagnosis of cancer is not high, and it cannot benefit for clinical reference. But the joint detection of multiple tumor markers can avoid these disadvantages, and increase the sensitivity and specificity of detection in different tumors. It has reached to a common view by most scholars that the use of bio-informational methods combined with multiple tumor markers can relatively show tumor characteristics so as to improve the diagnostic accuracy.This study further found that this method can better identify the suspected tumors. Clinical studies on malignant tumors suggest that the early diagnosis is more important than any existing therapeutic regimen. The morbidity and mortality of lung cancer, liver cancer, gastric cancer, colon cancer-breast cancer and other malignant tumors has been ranked to be the forefront of other tumors. How to improve the differential diagnostic accuracy of tumors, especially give an early prompt diagnosis has been recognized to be the key of improving the therapeutic effect and the prognosis. This study further analyzed the application value of the tumor markers combined with ANN as an auxiliary hint of the primary site of the unknown metastatic tumors.Artificial neural network (Artificial Neural Network, ANN) is a kind of calculating structure on the base of modern neurobiological research which simulate the biological processes to reflect some characteristics of human brain. It consists of a series of things called node (Node) similar to neurons consisting of processing units. Nodes through the network are connected to each other to determine the working principle of data model. It has strong fault tolerant ability-resistance ability and the category simulating ability for unknown data. The experimental data were put into the network through repeated iteration to create a data model reflecting experiment data relation. Using the feed-forward back-propagation algorithm of artificial neural network (ANN), we could establish several diagnosis models of serum tumor markers, then verify their application of the joint detection models in differential malignant tumors and tip for the unknown primary site of some metastatic malignant tumors. The input data transferred from the input layer to hidden layer, after calculated by ANN, then reach the output layer. If the data of output layer had not reached a predetermined desired value, the system would automatically calculate the error of the output layer, and the error would propagate along the original connection path. The models would not stop modifying the weights of various layers of neurons and repeatedly iterating to calculate the data and output them, until they reach to the desired target. Along with the computer technology and the development of artificial neural network, artificial intelligence technology assisting medical diagnosis will be more and more widely used. Materials and methods:1. Collect73cases of lung cancer,60cases of hepatocellular carcinoma,76cases of colon cancer,78cases of gastric carcinoma and51cases of patients with mammary carcinoma2Then use ELISA and the way of time-distinguishing to detect variant levels of8kinds of tumor markers [carcinoembryonic antigen(CEA),α-fetoprotein(AFP), cancer antigen(CA199), cancer antigen(CA242), cancer antigen(CA724), cancer antigen (CA211), neuron specific enolase(NSE) and tissue polypeptide antigen(TPA)] levels in all patients’serum3We analyzed data and erected models by the way of feed forward back propagation of software Matlab of ANN, used the blinded validation model to verify the accuracy of the models, then further studied this method in tumors of unknown primary sites. Each of these models were checked by cross validation, and we made the training samples as much as possible to produce various training sets and test sets with various combinations to reduce the maximum error.Results:1. The model of lung cancer and liver cancer (see Figure4):The computer randomly chosen89cases of lung cancer and liver cancer (51cases of lung cancer,38cases of liver cancer)from133patients as a training set, to set a neural network model. Lung cancer and liver cancer were set as land0in the network. The results got from the model between0and0.5corresponded to hepatocellular carcinoma, and between0.5and1corresponded to lung cancer. Using the built model for the remaining44patients (22cases of lung cancer,22cases of liver cancer) for blind diagnosis analysis, we got that the distinguishing diagnostic sensitivity was95.5%, the sensitivity was86.4%in hepatocellular carcinoma. In44patients, only4cases were mistaken, so the accurate identification rate was90.9%, the positive predictive value was95%, and the differential accuracy was up to90.9%for lung and liver cancer.2The model of colorectal cancer and hepatocellular carcinoma (see Figure5):The computer randomly chosen91cases of colorectal cancer and hepatocellular carcinoma (56cases of colorectal cancer,35cases of liver cancer) from136patients as a training set to make a neural network model. The network set colon cancer to be0and set liver cancer to be1; the results got from the model between0and0.5corresponded to colorectal cancer, and those between0.5and1corresponded for hepatocellular carcinoma. Using the built model for the remaining45patients (20cases of colorectal cancer,25cases of liver cancer) as a blind test analysis, we drew that the sensitivity for colorectal cancer was85%, and the diagnosis sensitivity for hepatocellular carcinoma was92%. In45patients, only5cases were mistaken, the accuracy rate was88.9%and the positive predictive value was89.5%, and differential accuracy rate was88.9%to distinguish colorectal cancer from hepatocellular carcinoma.3The model of gastric and hepatocellular carcinoma (see Figure6):The computer selected randomly92cases of gastric cancer and liver cancer (52cases of gastric cancer, liver cancer a report of40cases) from138patients to train as a neural network model. The network set gastric cancer as0, and set liver cancer as1; we got results between0and0.5corresponding to gastric cancer, and that between0.5to1corresponding for hepatocellular carcinoma. With the completed model for the remaining46patients (26cases of gastric cancer,20cases of hepatocellular carcinoma) for blind detection analysis, we drew that the sensitivity for gastric cancer was92.3%, and the diagnosis sensitivity for hepatocellular carcinoma was95%. In46patients, we found only3cases were misdiagnosed, and got93.5%of the accuracy rate,90.5%of the positive predictive value, To distinguish the hepatocellular carcinoma from gastric cancer, the identification accuracy rate was91.3%;4Breast cancer and lung cancer model (see Figure7):Use the same procedure to select randomly83cases of breast cancer and lung cancer (34cases of breast cancer,49cases of lung cancer) from total124cases to train as a neural network model. The network set breast cancer and lung cancer as0and1; results between0and0.5corresponded to breast cancer, between0.5and1corresponded to lung cancer. Using the model to test the remainder41cases (17cases of breast cancer,24cases of lung cancer) as blind tests, we got that the diagnosis sensitivity of breast cancer was70.6%, the sensitivity for lung cancer was87.5%. The model recognized wrong in8cases from41patients, so the accuracy rate was80.5%, the positive predictive value was80.8%, and the identification accuracy rate for breast cancer and lung cancer was80.5%;5Using colorectal cancer and hepatocellular carcinoma model to identify2tumor cases of unknown primary sites, or using hepatocellular carcinoma and gastric cancer model to identificate another2cases of unknown primary tumors, we both got consistent results with postoperative pathological diagnosis.Conclusions:Using the models set by tumor markers combined with ANN, we could effectively identify certain malignancies, and get a hint of the primary site for some unknown primary malignant tumor.Due to a limited number of existing cases, we only had done a preliminary study, the conclusion will be further confirmed with more and more cases, and will also benefit a lot for clinical diagnosis.