Speaker
Description
Prenatal neurodevelopment research relies on 4D ultrasound imaging, yet manual frame-by-frame annotation of fetal body parts remains labor-intensive and inconsistent across human coders. This work adapts DeepLabCut for automated markerless fetal pose estimation across 12 anatomical bodyparts in 301 videos from 42 fetuses. We introduce a robust pipeline featuring fetus-level data splitting to prevent leakage, five-fold cross-validation, gestational-age-balanced training, and ensemble predictions via mean and majority voting. Our enhanced models substantially reduce localization error and recover recall lost in the naïve approach, with the age-balanced ensemble approaching human-level F1 performance on internal data and generalizing more reliably to previously unseen external videos.