Automated Assessment of Uterine Coverage in Blind Sweep Obstetric Imaging Protocol

Blind sweep ultrasound imaging protocols have increased the accessibility of obstetric ultrasound imaging in austere and low-resource environments due to its ease of adoption for midwives and novice users. Image processing algorithms can be used to extract key indicators of fetal and maternal health from blind sweep datasets, without expert interpretation. However, the accuracy of these indicators depends on markers of integrity in the dataset: for example, coverage of the uterine interior. Thus, automated assessment of blind sweep completeness would be beneficial to increase algorithm accuracy and consistency, and user confidence. In this paper, we describe the use of a support vector machine (SVM) model to classify blind sweep acquisitions as complete or incomplete, based on ground truth annotations from expert sonographers. Our results show that this model can be used to classify sweep completeness with an overall accuracy of 0.83, and a positive predictive value (PPV) of 0.80. These findings suggest that our model can adequately flag incomplete blind sweep acquisition. Next steps in this work include investigating the performance of downstream algorithms on the complete and incomplete data cohorts.