Three-dimensional virtual colonoscopy for automatic polyps detection by artificial neural network approach: New tests on an enlarged cohort of polyps

摘要: Abstract Introduction and objective In computer aided diagnosis (CAD) tools searching for colonrectal polyps based on three dimensions virtual colonoscopy (3DVC) using computed tomography (CT) images, the reduction of occurrence false-positives (FPs) still represents a challenge because they are source unreliability. Following an encouraging previous supervised approach Bevilacqua et al., Three-dimensional Virtual Colonoscopy Polyps Detection by Supervised Artificial Neural Networks D.-S. Huang al. (Eds.): ICIC, LNBI 6840, Springer-Verlag, Berlin Heidelberg, (2011), pp. 596–603, aim this work is to discuss, in details, how adopted strategies, designed tested initial reduced data set, reveals good performance robustness terms FPs enlarged cohort new cases. Materials methods At beginning, materials consisted only 10 different polyps, diagnosed, expert radiologists, 6 patients, scanning 16 rows helical CT multi slices with resolution 1 mm. Moreover from those 7 were initially used analysis, excluding 2 tumors diameter bigger than 1 cm, one polyp hardly recognizable due fecal stool. paper, thanks accurate phase collecting data, grow impressively then consist total 43 all useful study. The whole set was merged former exams clinical operative unit called “Sezione di Diagnostica per Immagini” Di.M.I.M.P. Policlinico Bari ones coming two collaborations: Oncology department Faculty Medicine University Pisa participating, as former, IMPACT study (Italian Multicenter Accuracy CTC Study) Regge, Linear nonlinear feedforward neural network classifiers: comprehensive understanding, J. Intell. Syst., 9 (1) 1999, 1–38 and, more recently, radiology “Istituto Tumori Giovanni Paolo II” Bari. Starting colonography (CTC) several volumes scanned means artificial networks (ANNs) architectures error back propagation training algorithm Ma, 1–38. All sets built non-polyps sub-volume samples, whose correlated volume be detected. Results best ANN architecture, trained 27 sessile available dataset, evaluated false-negatives compared results shown 596–603 where cross validation strategy overcome small number old dataset Huang, bottleneck behaviour linear classifiers their breakthrough, Comput. Sci. Technol., 14 34–43. Good performances generalization work, fact that free-response operator characteristic analysis do not change significantly enlargement data. Conclusions This testing determined consistent performance; at same time it fairly intuitive necessary train method samples that, reason, overall could improved larger diagnosed radiologists.