Machine Learning-based Anomaly Recognition System

Routine second trimester anomaly scan has become a routine part of antenatal care. Early detection of fetal anomalies permits patient counselling, consideration of termination if detected anomalies are considerable, and arrangement of delivery and immediate neonatal care if indicated. Furthermore, with the expanding role of fetal interventions, early detection of fetal anomalies may expand management options, some of which may lead superior outcomes compared to postnatal interventions. However, fetal anatomy scan necessitates a particular level of training and expertise, either by sonographers or obstetricians. Unfortunately, availability of experienced personals may be globally limited. Furthermore, first trimester anatomy scan has been evolving rapidly as ultrasound machine continues to develop and clinical research yields more information on first trimester normal standards and abnormal ranges. Accordingly, first trimester scan is anticipated to be a part of routine care in the near future. Although this tool should provide substantial benefits to obstetric patients, this would require more providers with specific training, which is unlikely to be readily available. Artificial intelligence has been incorporated in the medical field for more than 20 years. With the advancement of deep learning algorithms, deep learning has yielded exceptional accuracy in image recognition. In the last decade, deep learning exhibits high quality performance that may exceed human performance at times. One of the earliest and most prevalent applications of deep learning in medicine are radiology-related. In the current study, the investigators will create a series of deep learning models that appraise and identify common fetal anomalies in a series of frames including recorded videos or real time ultrasound. Deep learning algorithms will be fed by labelled images of known normal and abnormal findings representing common fetal anomalies for both training and validation. These images will be collected retrospectively through medical records of contributing centers. Their diagnostic performance will be tested on retrospectively collected videos including normal and abnormal findings. In the second stage of the study, These models will be applied to prospectively collected videos of fetal anatomy scan for further validation.