An application of convolution neural networks: reducing false-positives in lung nodule detection

Recently, various computer-aided diagnosis (CADx) schemes have been proposed to tackle the problem of detecting lung nodules on digital chest radiographs. The research efforts are aimed at increasing the “true-positive fraction” while decreasing the “false-positive fraction” of the CADx. Among the problems of decreasing the number of false-positives, the differentiation between nodules and end-on vessels is one of the most challenging tasks performed by computers. Most investigators have used a conventional two-step pattern recognition approach, i.e., feature extraction followed by feature classification, The principal difficulty in those methodologies is in specifying the kind of features which will differentiate nodules from end-on vessels. Unfortunately, suitable feature definition, and corresponding extraction implementation algorithms, proved to be very difficult to define and specify. A convolution neural network (CNN) architecture, trained by direct connection to the raw image is proposed to tackle the problem. The CNN, which used locally responsive activation function, was directly and locally connected to the raw image. The performance of the CNN is evaluated in comparison to an expert radiologist. We employed receiver operating characteristics (ROC) method with A_z as the performance index to evaluate all the simulation results. The CNN showed superior performance (A_z=0.99) to the radiologist´s (A_z =0.83)