Title: Comparison of scleroderma detection techniques in CT images using supervised and unsupervised methods
Authors: M. Prasad, M.S. Brown, F. Abtin, S. Vasunilashorn, J.G. Goldin
Addresses: Department of Radiological Sciences, David Geffen School of Medicine, University of California, Los Angeles, USA. ' Department of Radiological Sciences, David Geffen School of Medicine, University of California, Los Angeles, USA. ' Department of Radiological Sciences, David Geffen School of Medicine, University of California, Los Angeles, USA. ' Department of Radiological Sciences, David Geffen School of Medicine, University of California, Los Angeles, USA. ' Department of Radiological Sciences, David Geffen School of Medicine, University of California, Los Angeles, USA
Abstract: We compare supervised and unsupervised computer-aided techniques trained to detect Scleroderma Lung Disease (SLD) abnormalities using textural features in Computer Tomography (CT) images. Unsupervised learning can be very valuable since obtaining valid regions of interest are subject to intra- and inter-reader variation. An unsupervised technique, K-Means and a semisupervised technique, Seeded K-Means were trained with minimal user input and few training examples. The techniques| performance was compared with supervised techniques, naive Bayes and C4.5. Results showed that texture-based semisupervised and supervised methods achieve comparable performance (i.e., an overall accuracy of 85.9% using seeded K-Means, 88.3% using naive Bayes and 89% using C4.5). The preliminary results suggest that SLD abnormalities can be characterised with minimal expert input.
Keywords: unsupervised learning; supervised learning; computer-aided techniques; classification; computer tomography; CT images; texture; scleroderma detection; scleroderma lung disease; textural features.
DOI: 10.1504/IJBET.2008.020073
International Journal of Biomedical Engineering and Technology, 2008 Vol.1 No.4, pp.453 - 464
Published online: 25 Aug 2008 *
Full-text access for editors Full-text access for subscribers Purchase this article Comment on this article