Cervical Precancer Classification: A Support Vector Machine Study

• Activity Number: 89
• Type: Contributed
• Date/Time: Monday, August 4, 2003 : 8:30 AM to 10:20 AM
• Sponsor: Section on Statistical Computing
• Author(s): Shu Hui Chen*+ and Will Gersch
• Companies: University of Hawaii, Manoa and University of Hawaii
• Keywords: precancer ; classification ; support vector machine ; SVM

Abstract

The classification of cervical intraepithelial neoplasia (CIN) by use of support vector machine (SVM) is presented. Support vector machines and some details of our computations, including our choice of SVM kernel, estimate of the generalized error, margin and training errors, a gradient descent algorithm (Chapelle et al. 2000), and multi-lass SVMs are discussed. A previous study of CIN data using neural networks (NNs) achieved 2.53% classification error. For comparison purposes, nearest neighbor (1NN) classification was also studied. The CIN data set consist of 5 classes, and SVM analysis was done via one-versus-all (OVA) and error-correcting output codes. With parameter optimizations, our SVMs achieved an OVA classification error of 0.9 +/- 0.5%. The nearest neighbor classification error was 2.0 +/- 0.5%. The SVM performance is more than three standard deviations (SDs) better than achieved in the previous study using NNs and more than two SDs better than the nearest neighbor classifier.