Abstract
The purpose of this paper is to develop a new algorithm to decrease the number of false-positive candidates which were obtained with a previous detection scheme, based on a genetic-algorithm (GA), for lung nodules in digital chest radiography. The algorithm consists of the following three steps. Firstly, the candidates with the marginal area just outside the lung were eliminated by the histogram analysis. Secondly, the Gaussian operator was used to enhance the bone edge of the candidates. After analyzing the kurtosis of orientation-amplitude histogram and entropy of interpolative difference signal of central region, the candidates including the shadows of bones were eliminated. Thirdly, the modified Laplacian-Gaussian operator which suppresses the high-frequency noise and the low-frequency components was used to enhance the nodules. After this filtering, the candidate area was divided into three rectangular regions (R1, R2 and R3). Three features, which consist of the entropy of interpolative difference signal in R1, the ratio of variances between R2 and R1, and that between R3 and R 1, were used to eliminate falsepositive candidates of the shadows of vessels and others. The 341 candidates (7 true-positive and 334false-positive candidates) from 20 clinical chest images (10 normal and 10 abnormal cases) were analyzed and 297 false-positive ones were eliminated by our method. In other words, we were able to decrease the number of false-positive candidates from approximately 17 to 1.9 per image without reducing any true nodules.
