In order to obtain reference curves for data sets when the covariate is multidimensional, we propose a new methodology based on dimension-reduction and nonparametric estimation of conditional quantiles. This semiparametric approach combines sliced inverse regression (SIR) and a kernel estimation of conditional quantiles. The convergence of the derived estimator is shown. By a simulation study, we compare this procedure to the classical kernel nonparametric one for different dimensions of the covariate. The semiparametric estimator shows the best performance. The usefulness of this estimation procedure is illustrated on a real data set collected in order to establish reference curves for biophysical properties of the skin of healthy French women.