Development of automated detection system for lung nodules in chest radiograms

We developed a computerized scheme to detect nodular shadows in chest radiography by a genetic algorithm (GA) and artificial neural networks (ANN). The scheme is based on template matching. The GA technique was used in the searching part of the matching. The automated detection method consists of the following five steps: (1) preprocessing, (2) detection by GA template matching, (3) detection by ANN, (4) false positive elimination by ANN, and (5) false positive elimination by feature analysis. At the first step, the lung areas were extracted automatically by thresholding and region expansion. At the detection step (2), GA was employed to identify the suitable positions and the shape of reference images by calculating the similarities between an extracted image from whole chest radiograms and 448 artificial nodular shadows with various 2D Gaussian distributions indicated by three parameters. At detection step (3) and false positive elimination step (4), ANNs that were trained by the regulated pixel values of nodular shadows on other images were scanned in the extracted lung area. The positions of nodules were indicated by thresholding the ANN output values. At false positive elimination step (5), we used some statistical parameters for histogram analysis. We employed 10 abnormal and 10 normal chest radiograms in which nodular sizes and diagnostic difficulty for the detection had been predetermined by a radiologist. The true positive fraction at this initial detection stage was approximately 80% with 2.4 false positive findings per image. These results show that the GA based pattern matching method is useful for detecting nodular shadows on chest images.