Classification of White Blood Cells Using Bispectral Invariant Features of Nuclei Shape

Classification of white blood cells from microscope images is a challenging task, especially in the choice of feature representation, considering intra-class variations arising from non-uniform illumination, stage of maturity, scale, rotation and shifting. In this paper, we propose a new feature extraction scheme relying on bispectral invariant features which are robust to these challenges. Bispectral invariant features are extracted from the shape of segmented white blood cell nuclei. Segmentation of white blood cell nuclei is achieved using a level set algorithm via geometric active contours. Binary support vector machines and a classification tree are used for classifying multiple classes of the cells. Performance of the proposed method is evaluated on a combined dataset of 10 classes with 460 white blood cell images collected from 3 datasets and using 5-fold cross validation. It achieves an average classification accuracy of 96.13% and outperforms other popular representations including local binary pattern, histogram of oriented gradients, local directional pattern and speeded up robust features with the same classifier over the same data. The classification accuracy of the proposed method is also compared and benchmarked with the other existing techniques for classification white blood cells into 10 classes over the same datasets and the results show that the proposed method is superior over other approaches.