OBJECTIVE Birth weight (BW) prediction is crucial to assess the risk of fetal and neonatal morbidity and mortality related to growth restriction or macrosomia. Our objective here was to devise an accurate BW prediction method for small fetuses (BW ≤ 2500 g) and large fetuses (BW ≥ 4000 g), based on ultrasound data collected starting at 20 weeks' gestational age. These new models were tested on a new sample of validation consisted of 397 small fetuses and 135 large fetuses. METHODS We included 3486 fetuses whose age and weight at birth were known and for whom standard data were available from ultrasound scans. Extrapolation to full-term of the estimated fetal weight computed using Hadlock's equation identified 212 small and 291 large fetuses. We built a new regression model for predicting the BWs of small fetuses and large fetuses. The results obtained using these models were compared to those provided by Hadlock's equation. RESULTS Predicted BWs of small and large fetuses obtained using the regression models showed significantly less systematic bias (1.0 versus 4.9) and (-3.4 versus -7.0, respectively) and random error (10.1 versus 11.6 and 7.5 versus 8.9, respectively) than Hadlock's estimate. These results are confirmed on the sample of validation. CONCLUSIONS Our results support the accuracy and reliability of our BW prediction models for small and large fetuses. In addition, our models provide BW estimates as early as 5 ± 1 weeks before birth.