OC01.03: Development and external validation of a deep learning model for gestational age estimation using blind ultrasound video sweeps

Gestational age (GA) estimation is essential for optimal prenatal care. The objective of this study is to develop deep learning models for GA prediction using blind ultrasound (US) videos without biometric information. Prospective observational study during which blindly acquired US videos (N = 748), using step 1 of the CALOPUS protocol, were collected from 636 participants enrolled in the Garbh-Ini cohort in North India from 8 to 37 weeks. The data set was divided into training (N = 560), validation (N = 50) and test (N = 138) sets, and an attention-based convolutional neural network model was developed to predict GA. The model was externally validated on videos (N = 377) collected from 377 participants in the UK. To make predictions interpretable, we calculated an importance score for each frame. The frames with high importance scores were assessed on a subset of labelled videos from the internal test set (N = 45) to determine which anatomical features most contribute to GA prediction. On the internal test set, our model had an overall MAE of 6.28 days (N = 138) and 7.46 (N = 38), 5.36 (N = 65), and 6.71 (N = 35) days in the first, second and third trimesters respectively, when compared to gold standard CRL-based GA estimation. On the labelled subset, our model, on average, focuses 94.3% on frames containing either head, abdomen, or femur while 2.8%, 2.4% and 0.4% on frames containing pelvis, spine or heart, respectively. On external validation (N = 377), the overall MAE was 10 days. We developed and externally validated a US video-based model for GA estimation. Our model predicts GA across trimesters without the need to capture standard planes and is interpretable in terms of the anatomy used in GA estimation.