Purchase this article with an account.
H. Bagherinia, X. Chen, C. Flachenecker, R. Angeles, D. Burger, P. Caroline, J. Dishler, D. Tanzer, D. Schanzlin, K. Reeder; Support Vector Machine (SVM)-Based Classification of Corneal Topography. Invest. Ophthalmol. Vis. Sci. 2008;49(13):1023.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
To demonstrate the feasibility of a SVM-based learning algorithm to discriminate normal corneas from other corneal conditions using topography exams collected by the ATLAS® Model 9000 Corneal Topographer (Carl Zeiss Meditec, Dublin, CA).
A Support Vector Machine (SVM) based classifier was developed to discriminate normal corneas from other corneal conditions (such as suspect keratoconus, keratoconus, orthokeratology, pellucid marginal degeneration, myopic and hyperopic laser vision correction). The feature vector consists of 12 parameters: CIM, Shape Factor, TKM, Convexity, Centroid X, Centroid Y, Max Mean Power, Max Mean Power X, Max Mean Power Y, Mean I-S, Mean IN-ST, and Mean IT-SN. The first 3 parameters were used in the original PathFinderTM Corneal Analysis program, while the remaining 9 parameters are derived from the Mean Curvature Map. Data from the eyes of 85 normal subjects, and 239 other corneal condition subjects were used to train and evaluate the current version of the algorithm. The SVM classifier was trained using 80% of the dataset determined at random, and evaluated using the remaining 20%. The current algorithm version was used to maximize the separation between these two groups in this 12 dimensional feature space.
Early verification results based on the training data set using the current version of the algorithm show that the SVM classifier was able to discriminate normal corneas from the other corneal conditions with ≥ 90% sensitivity, specificity, and accuracy. A performance validation study is currently ongoing using a different data set. The results from this study will be presented.
The use of a SVM classifier with parameters derived from the mean curvature map may become a useful tool to discriminate normal corneas from other corneal conditions based on early verification results using a randomly selected subset of the training data set.
Clinical Trial: :
This PDF is available to Subscribers Only