Early detection of severe fetal growth restriction using multimodal deep learning based on ultrasound and prenatal biomarkers.

A deep learning model combining ultrasound and serum biomarkers can predict severe fetal growth restriction with high accuracy, but further validation in broader populations is needed.

Summary and findings

The study aimed to develop a multimodal deep learning model for early prediction of severe fetal growth restriction (FGR) using ultrasound images and maternal serum biomarkers. A total of 598 singleton pregnancies were analyzed, with 299 cases of severe FGR and 299 controls. The model achieved high predictive accuracy with an AUC of approximately 0.96.

Abstract

Objective To develop and evaluate a multimodal deep learning model that integrates first- and second-trimester ultrasound images with first-trimester maternal serum biomarkers for early prediction of severe fetal growth restriction (FGR). Materials and methods In this prospective study, 13,991 pregnant women were initially recruited, and after applying inclusion and exclusion criteria, 598 singleton pregnancies (299 with severe FGR and 299 controls) were analyzed. Severe FGR was diagnosed at or after 24 weeks' gestation and defined as an estimated fetal weight (EFW) or abdominal circumference (AC) below the 3rd percentile. Seven ultrasound views were collected (first-trimester crown-rump length, nuchal translucency; second-trimester head and abdominal circumference, femur length, umbilical artery S/D ratio, and amniotic fluid depth), along with maternal age and first-trimester serum markers (PAPP-A and free β-hCG, expressed as MoM). A convolutional neural network (CNN) with multi-level attention and dynamic convolution modules was trained end-to-end using stochastic gradient descent with cross-entropy loss. Model performance was evaluated with five-fold cross-validation. Results The integrated model achieved high predictive accuracy for severe FGR, with an AUC of approximately 0.96, accuracy of 91%, and F1-score of 0.912. Matthews correlation coefficient (MCC) was 0.798. The model outperformed single-modality baselines, such as a ResNet-50 CNN trained on images alone. Ablation experiments showed that adding maternal serum biomarkers and advanced CNN modules significantly improved performance. The model demonstrated excellent discrimination between severe FGR and control cases, with high sensitivity and precision in classification. Conclusions A deep learning model combining routine ultrasound and first-trimester serum biomarkers can accurately predict severe FGR before clinical signs appear. This model significantly enhances early severe FGR risk stratification and suggests potential for AI-based prediction tools in prenatal care to guide earlier surveillance and intervention.

Key findings

• Not reported in abstract.
• Not reported in abstract.
• Not reported in abstract.

Limitations

• Not reported in abstract.
• Not reported in abstract.
• Not reported in abstract.

Elsewhere in the HCG (Human Chorionic Gonadotropin) corpus