MRI-based fetal gestational age estimation using a structure-aware self-supervised network.

OBJECTIVES: To develop a structure-aware self-supervised deep learning model that leverages global edge information to reduce gestational age (GA) estimation error from fetal brain MRI and improve clinical reliability. MATERIALS AND METHODS: A retrospective collection of 1630 fetal brain coronal T2-weighted MR images from 207 singleton pregnancies (mean GA, 30 ± 8 weeks; median, 34 weeks; range, 22-38 weeks) acquired between January 2019 and July 2023. GA, determined from the last menstrual period and the first-trimester ultrasound, served as the reference standard. Subjects were randomly split into a training set (80%) and an independent test set (20%) on a patient level. Model performance on the test set was evaluated using mean absolute error (MAE) and the coefficient of determination (R²), with 95% confidence intervals obtained by bootstrap resampling. RESULTS: The proposed model achieved an MAE of 0.793 weeks (95% CI, 0.574-0.860) and an R² of 0.934 (95% CI, 0.918-0.967) for GA estimation from routine clinical MRI. Predicted GAs showed a strong linear association with the reference GAs (p < 0.001). CONCLUSION: The proposed structure-aware self-supervised model enables accurate MRI-derived estimation of clinical GA from fetal brain MRI. This approach may complement conventional dating in selected cases; prospective and multi-center validation is required before clinical deployment. KEY POINTS: Question Can fetal brain MRI provide an accurate estimation of clinical gestational age when conventional clinical dating methods are unreliable in mid-to-late pregnancy? Findings In 207 pregnancies, SSFN-Net achieved lower MAE (0.79 weeks) and higher R² than existing deep-learning baselines on fetal brain MRI. Clinical relevance MRI-derived gestational age estimation may complement ultrasound or LMP-based clinical dating in selected cases, providing an additional quantitative reference for prenatal assessment.