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

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