Accurate diagnosis of neuroblastoma, non-Hodgkin lymphoma, rhabdomyosarcoma, and Ewing sarcoma, is often difficult because these cancers appear similar in routine histology. Finding a few useful biomarkers (not all related genes) that can discriminate between the subgroups will help designing better diagnostic systems. In an earlier study we reported a set of seven genes having excellent discrimination power. In this investigation we extend that study and find other distinct sets of genes with strong class specific signatures. This is achieved analyzing the correlation between genes. This led us to find another set of seven genes with better discriminating power. Our original gene selection method used a neural network whose output may significantly depend on initialization of the network, network size as well as the training data set. To address these issues we propose a scheme based on re-sampling. This method can also reduce the effect wide variation in number of data points in the training set from different classes. This method led us to find a set of five genes with good discriminating power. The genes identified by the proposed methods have roles in cancer biology.