(622) Deep learning to assess twin gestation from blindly obtained ultrasound sweeps

We collected blind ultrasound sweeps (~10 second cines in cranio-caudal and lateral directions) from pregnant people in North Carolina. The cohort was randomly divided at the patient level into discreet training/tuning (80%) and testing (20%) datasets. We built deep learning AI models to diagnose twins and estimate their gestational age (GA). We also explored assessment of chorionicity.

Results:

From Feb 2020 thru March 2023, 16,557 pregnancies contributed 30,106 studies containing blind sweeps. Of these, 490 pregnancies (1,215 studies) were twins. To evaluate twin diagnosis, we restricted the test set to studies ≥14 weeks GA (3,163 pregnancies; 5,424 studies). Among the 87% of studies in which the model could make a prediction, the area under the receiver operating curve (AUC-ROC) was 99.4% (sens: 98.1%; spec: 98.2%). To evaluate twin GA estimation, we restricted the test set to twin pregnancies whose GA had previously been established by early CRL or IVF (92 pregnancies; 185 scans). The model made a prediction in all studies with a mean absolute error of 4.6 ±0.3 days vs 4.5 ±0.3 days for biometry (difference, 0.1 days; 95% CI: -0.6, 0.8). To evaluate chorionicity assessment, we limited the test set to twins with scans performed < 24 weeks (67 pregnancies, 93 scans). Among the 73% of studies in which the model could make a prediction, AUC-ROC was 87.1% (sens: 82.6%; spec: 82.2%).

Conclusion: AI-assisted ultrasound, which can be deployed on low-cost point-of-care devices, can diagnose twins and assess their gestational age with high performance. Preliminary results for chorionicity assessment are encouraging but more training data from earlier GA are needed. These results presage a future where all patients carrying twins – not just those in rich countries – can access the diagnostic benefits of sonography.