Development and validation of a deep learning model for predicting small-for-gestational-age birth using routine mid-trimester ultrasound

Background Small-for-gestational-age (SGA) birth is associated with increased perinatal morbidity and mortality. Earlier identification of pregnancies at risk could inform antenatal surveillance. We aimed to develop and externally test a deep learning model for predicting SGA birth using routine fetal ultrasound images acquired at 20–23 weeks of gestation. Methods This two-center retrospective study used case-enriched samples comprising 594 pregnancies (306 SGA; 2,970 images) at Center A and 407 pregnancies (164 SGA; 2,035 images) at Center B. SGA was defined as birth weight below the 10th percentile for gestational age and sex. Model development and five-fold cross-validation used Center A images, followed by evaluation on a held-out internal test set. Center B provided independent external testing without model updating. The model incorporated split-group convolution, channel attention, and multi-scale feature fusion. Performance was evaluated at the image level and compared with established deep learning architectures. Ablation analyses and gradient-weighted class activation mapping were also performed. Results Comparative and ablation experiments supported the contribution of the proposed modules. In the held-out internal image test set, the proposed model achieved an AUC of 0.8667, an F1-score of 0.8006, an MCC of 0.5820, and an accuracy of 0.7912. In the independent external image test set, it achieved an AUC of 0.8046, an F1-score of 0.7132, an MCC of 0.4845, and an accuracy of 0.7273. Conclusions The model showed image-level discrimination of subsequent SGA birth using routine mid-trimester ultrasound images in internal and independent external testing. Patient-level validation and assessment of calibration and clinical utility are needed before clinical implementation.