Automatic methods developed for detection of lung nodules in chest radiographs usually present an excessive number of false positive detections. In this work the authors show how the local curvature image of suspected nodule pixels provides a new description that permits to distinguish true nodules from those false positives. A multilayer perceptron network with supervised learning is able to recognize the images of nodule local curvature peaks. The results obtained with a set of 23 chest images each one with at least one nodule show a sensibility in the global detection process of 93% with a mean number of 2 false positives per image.
展开▼
