Multiple Pregnancy Detection from Ultrasound Blind Sweeps

Timely detection of high-risk pregnancies is of paramount interest to maternal-fetal health. With advances in AI and the advent of low-cost handheld ultrasound, automated detection of high-risk pregnancies, without the need for a trained sonographer in primary care setting, is a highly impactful yet feasible technology. The major challenge of building detection models for high risk pregnancies, such as multiple pregnancy (MP), fetal demise, placenta previa, is the scarcity of available data due to low-occurring nature of high-risk situations. Often model selection requires data hungry nested cross-validation. To mitigate the problem, we introduce a novel model selection criterion using non-nested cross-validation.We showcase the pipeline we built for developing a MP detection model with both deep learning and a less expressive machine learning (ML) model as components. The pipeline (A) efficiently uses limited multiple pregnancy data collected by blindly sweeping over maternal abdomen (blind to images), (B) can be deployed in a low-compute setting, and (C) is fully interpretable and hence inspectable for any alarming learned pattern.<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">1</sup>