A machine learning strategy for automatic detection of high risk pregnancies in Uganda

Abstract

High risk pregnancies have a significant contribution to the increasing maternal and infant mortality rates in Uganda. One of the major parameters for assessing pregnancy risk is the fetal skull circumference of unborn babies. The circumference is determined using an obstetric ultrasound scan, followed by manual annotation of the image and subsequent estimation; a laborious and error-prone process, which could potentially lead to misdiag- nosis. To address this challenge, we set out to develop a machine learning strategy for automated estimation of the fetal-skull circumference. We developed a model, based on the U-Net architecture. The model was trained on the HC- 18 open-source dataset: 999 images were used for training while 150 images were used for model testing. Our model achieved an accuracy of 95% and an Intersection over Union (IoU) of 94%. Comparison of the model with alternative approaches reveals superior performance of the proposed approach. The model was integrated in a web-based decision support system. Future work will focus on evaluating our model’s performance on Ugandan data.