Lung Nodules Detection System Based On Massive Training Artificial Neural Network

Computer-aided diagnosis (CAD) can improve the doctors’ efficiency as well ashelp doctors improve recognition rate of lung nodules, methods currently used in thescheme of lung nodules identification are based on the analysis of the characteristic,After segmentation of candidate area, characteristics are calculated, and then design aclassification algorithm of analysis the characteristic vector to determine whether thelung nodules, In this kind of scheme, the key process is segmentation andcharacteristics’ calculation. And, it’s likely to form false positive results. Therefore,the most important in CAD research and application is to segment and identify lungnodules in CT image more accurately, improve the accuracy of diagnosis results,make the best to avoid misdiagnosis.This paper’s main job is to introduce a new pattern recognition method based onneural network, design and develop a prototype system of identify lung nodules inCT image. This new technology called the massive training neural network(MTANN), and its feature is take the pixels value of candidates nodules in CT imageas input directly, and outputs an image. Output image to different distributioncharacteristics of different types of lung nodules said. By using the weighted sum getthe score of output image, and get the result by comparing with the threshold. Falsepositive nodules is divided into different types according to its shape, in order toreduce the false positive ratio, extended the single MTANN got the Multi-MTANN,each one of these neural network training to use different types of false positive lungnodules, and according to the type of false positive lung nodules use differentthreshold in judgment, aims to remove as much as false positive lung nodules.This paper used neural filter in image processing and pattern recognition, designand developed lung nodules CAD system based on massive training neural network.Adjust parameters of the neural network, the analysis parameters on the effect of theneural network training. Experiments show that the method can effectivelydistinguish real nodules and different types of false positive nodules, and it can keepthe sensitivity as same as greatly reduce false positive results ratio.