The significant features corresponding to skull structures on cephalograms are clinically useful for cephalometric diagnosis and superimposition. Accordingly the specific anatomical landmarks need to be firstly located for cephalometric measurements. In this paper, a novel method combining the multilayer perceptron and genetic algorithm is proposed to extract the specific feature areas. Thus, the useful landmarks may then be easily found from these feature areas instead of the whole image. The multilayer perceptron is used to approximate a fitness function for the genetic algorithm. In each iteration, eighty randomly selected subimages are grouped as the population for a GA search. Based on the feature characteristics, the selected subimages with the best fitness will survive to the last. From the experimental results, it is shown that the proposed algorithm does work better than our previous method of correlation.