Multinetwork classification scheme for detection of colonic polyps in CT colonography data sets

摘要: A multi-network decision classification scheme for colonic polyp detection is presented. The approach based on the results of voting over several neural networks using different variable sets size N which are selected randomly or by an expert from a general set M. Detection polyps complicated large variety polypoid looking shapes (haustral folds, leftover stool) colon surface. Using various shape and curvature characteristics, intensity, measurements texture features to distinguish real false positives leads intricate problem. We used 17 including region density, Gaussian average sphericity, lesion size, wall thickness, their means standard deviations in vicinity prospective polyp. Selection most important parameters reduce feature acceptable generally unsolved method suggested this paper uses collection subsets variables. These variables weighted effectiveness. effectiveness cost function calculated basis training test sample mis-classification rates obtained net with given set. final majority vote across generated subsets, takes into account votes all nets. This reduces flst positive rate factor 1.7 compared single decisions. overall sensitivity specificity reached 100% 95% correspondingly. Best were back propagation nets one hidden layer trained Levenberg-Marquardt algorithm. Ten-fold cross-validation better estimate true error rates.