Traditional classifiers, such as support vector machine and Bayesian classifier, require data normalization for removing experimental batch effects, which limit their applications at the individual level. In this paper, we aim to build a classifier to distinguish lung cancer and non-cancer lung tissues (pneumonia and normal lung tissues). We identified gene pairs as signatures to build a classifier based on the within-sample relative expression orderings of gene pairs in a particular type of tissues (cancer or non-cancer). Using multiple independent datasets as the training data, including a total of 197 lung cancer cases and 189 non-cancer cases, we identified three gene pairs. Classifying a sample by the majority voting rule, the average accuracy reached 95.34% in the training data. Using multiple independent validation datasets, including a total of 251 lung cancer samples and 141 non-cancer samples without data normalization, the average accuracy was as high as 96.78%. The rank-based signature is robust against experimental batch effects and can be used to diagnose lung cancer using samples measured by different laboratories at the individual level.