A Multimodal AI Copilot for Obstetric Ultrasound Interpretation

Ultrasound interpretation in Rwanda is largely confined to tertiary facilities: approximately 115 obstetricians serve 3.7 million women of reproductive age, and the median first rural scan occurs at 32 weeks. Artificial intelligence could extend interpretation to frontline health workers, but algorithms trained on non-African data transfer poorly to regional populations, dropping 7 to 12 percent in performance. This poster reports a bench evaluation of an offline obstetric ultrasound decision-support system designed for low-resource settings. Six TensorFlow Lite models run fully on-device, routing standard planes and segmenting skull, abdomen, femur, and the intrapartum pubic symphysis to fetal head. Fetal biometry is computed deterministically from geometry, using ellipse fitting and tangent construction rather than learned regressors. Plain-language advisories are constrained by a numeric ground check, and a rules engine abstains when input is insufficient. Results: plane classification 0.862; skull Dice 0.98; brain sub-plane classification 0.605. Biometry mean absolute error was 2.37 mm for head circumference, 4.4 mm for abdomen and 1.0 mm for femur, with angle of progression MAE of 5.6 degrees (bias -0.6 degrees). Fetal biometry accuracy falls within reported inter-observer variability, indicating that specialist-free antenatal ultrasound support is feasible in low-resource settings. Pairing deterministic geometry with an abstaining rules engine adds transparency and safety over black-box regressors. Brain sub-plane classification remains experimental. These are bench results only; clinical diagnostic accuracy is not yet established. A two-phase prospective diagnostic-accuracy study of 20,000 images across 2,000 participants in Rwanda is proposed to establish clinical performance against local conditions. African-context validation and demographic sub-group analysis should precede deployment of obstetric AI in Rwanda.