DIFFERENTIATION BETWEEN NODULES AND END-ON VESSELS USING A CONVOLUTION NEURAL-NETWORK ARCHITECTURE

In recent years, many computer-aided diagnosis schemes have been propo sed to assist radiologists in detecting lung nodules. The research eff orts have been aimed at increasing the sensitivity while decreasing th e false-positive detections on digital chest radiographs. Among the pr oblems of reducing the number of false positives, the differentiation between nodules and end-on vessels is one of the most challenging task s performed by computer, Most investigators have used a conventional t wo-stage pattern recognition approach, ie, feature extraction followed by feature classification. The performance of this approach depends t otally on good feature definition in the feature extraction stage. Unf ortunately, suitable feature definition and corresponding extraction i mplementation algorithms proved to be very difficult to define and spe cify. A convolution neural network (CNN) architecture, trained by dire ct connection to the raw image is proposed to tackle the problem. The CNN, which uses locally responsive activation function, is directly an d locally connected to the raw image. The performance of the CNN is ev aluated in comparison to an expert radiologist. We used the receiver o perating characteristics (ROC) method with area under the curve (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). The CNN approach can potentially he applied to other appli cations, such as the differentiation of film defects and microcalcific ations in mammography, in which the image features are difficult to de fine or not known a priori. (C) 1995 by W.B. Saunders Company