Does Multispectral Texture Features Really Improve Cervical Cancer Detection?

For cervical cancer detection, the performance of multispectral texture (MST) features extracted from multispectral Pap smear images is evaluated. In this study we carried out pairwise comparisons between different image features, including MST versus average spectral texture features (AST, without spectral information), and MST versus multispectral intensity features (MSI, without texture information). We demonstrate, experimentally, that well-selected MST features combining both multispectral and texture information can achieve better classification results (ROC curves) for cervical cancer detection from multispectral Pap smear images. Furthermore, we investigate which type of wavelet texture features (orthogonal, bi-orthogonal or non-orthogonal) individually or in combination is most effective.