Abstract: Genetic algorithm is a random search algorithm that simulates natural selection and evolution. It searches through the total solution space and can find the optimal solution globally over a domain. Recently, the popular encoding scheme is to encode the structure and weights, etc. into a string, which is not easy for the reservation of sub-structure during the process of genetic evolution. Generally, BP training scheme used in feed-forward neural network is to train all the offspring equally, which obviously wastes resources. A new method named SEFNN is proposed, which uses compact matrix encoding scheme, a new crossover operator, a properly modified mutate operator and rules of training elites. The efficiency of evolutionary feed-forward neural network is improved by properly considering the relationship between genotype and phenotype, thus improving the mutation speed and adopting a scheme of selective training. Experiments show that the proposed method can get good performance in accuracy. It has also found good application in a lung cancer diagnosis system.