Performance Analysis of Neighborhood Component Feature Selection for Oral Histopathology Images

Precise feature selection leads to improved accuracy in automatic oral cancer detection using digital Hematoxylin and Eosin (H&E)-stained histopathology images of oral squamous cell carcinoma. In this work, we have used neighborhood component analysis (NCA) feature selection technique with stochastic gradient descent based feature weight estimator. To verify the efficiency of the feature selection technique and independency of choice of classifier, three well known classifiers are used. The histopathology images of oral mucosa are classified using selected features. It has been observed that misclassification rate is reduced considerably after feature selection is employed.