HEp-2 Cell Classification Using Multi-resolution Local Patterns and Ensemble SVMs

Abstract

We describe a pattern recognition system for classifying immunofluorescence images of HEp-2 cells into six classes: homogeneous, speckled, nucleolar, centromere, golgi, and nuclear membrane. We use locality-constrained linear coding to encode multiple local features and two-level cell pyramids to capture spatial structure of cells. An ensemble of linear support vector machines is used to classify each cell image. Leave-one-specimen-out experiments on the I3A Contest Task 1 training data set predicted a mean class accuracy of 80.25%.