We propose two new filter-based feature selection criteria for identifying differentially expressed genes, namely the average difference score (ADS) and the mean difference score (MDS). These criteria replace the serial noise estimator used in existing criteria by a parallel noise estimator. The result is better detection of changes in the variance of expression levels, which t-statistic type criteria tend to under-emphasize. We compare the performance of the new criteria to that of several commonly used feature selection criteria, including the Welch t-statistic, the Fisher correlation score, the Wilcoxon rank sum, and the Independently Consistent Expression discriminator, on synthetic data and real biological data obtained from acute lymphoblastic leukemia and acute myeloid leukemia patients. We find that ADS and MDS outperform the other criteria by exhibiting higher sensitivity and comparable specificity. ADS is also able to flag several biologically important genes that are missed by the Welch t-statistic.