In this paper, we present a neural network ensemble method called DS-NNE which is based on dataset splitting. In order to get better classification results, the DS-NNE method performs the following tasks: (1) performs gene selection using t-test and f-test to remove the redundant genes. (2) divides the original training dataset into k disjoint subsets; (3) performs random re-sampling k-1 out of k subsets to get a training dataset and trains a neural network classifier on the generated dataset, then repeats the training procedure n times to obtain n neural networks. (4) predicts the class label from the unknown data using the ensemble classifiers through majority vote method. The DS-NNE method has two salient advantages: it is more stable than bagging and boosting methods; it can obtain higher prediction accuracy compared with other methods. Experiments on five benchmark tumor datasets confirm the validity of the proposed method.