The ultrasonic estimation of fetal weight at delivery is of important prognostic significance in obstetrical practice. The convertional regression formulas used for estimating fetal weight have the disadvantage of less reliability. In this study, we used the back propagation neural network (BP) to estimate Fetal Weight. Some input variables were adopted in constructing the BP model: biparietal diameter (BPD), cerebellum transverse diameter (TCD), abdominal circumference (AC), liver length (LL), femur length (FL), fetal thigh soft tissue thickness (FSTT), and gestational age (GA). The fetal weights of 109 singleton fetuses were estimated. In the training group and validation group, coincidence rates were 89.77% and 76.19% respectively. The results show that the estimation based on neural network is more accurate than that by regression method. GA, its unit is not week but day in our formulas, is very valuable in combination with other ultrasonic parameters on estimation.