Quantization and similarity measure selection for discrimination of lymphoma subtypes under k-nearest neighbor classification

Molecular classification of tumors holds great potential for cancer research, diagnosis, and treatment. In this study, we apply a novel classification technique to cDNA microarray data for discriminating between three subtypes of malignant lymphoma: CD5+ diffuse large B-cell lymphoma, CD5- diffuse large B-cell lymphoma, and mantle cell lymphoma. The proposed technique combines the k -Nearest Neighbor (k -NN) algorithm with optimized data quantization. The feature genes on which the classification is based are selected by ranking them according to their separability criteria computed by taking into account between-class and within-class scatter. The classification errors, estimated using cross-validation, are significantly lower than those produced by classical variants of the k -NN algorithm. Multidimensional scaling and hierarchical clustering dendrograms are used to visualize the separation of the three subtypes of lymphoma.