Point-of-care ultrasound in the global south: A case of fetal heart anomaly assessment with mobile devices

A major challenge in pre-natal healthcare delivery is the lack of devices and clinicians in several areas of the developing world. While the advent of portable ultrasound machines and more recently, handheld probes, have brought down the capital costs, the shortage of trained manpower is a serious impediment towards ensuring the mitigation of maternal and infant mortality. Diagnosis of pre-natal ultrasound towards several key pre-natal health indicators can be modelled as an image analysis problem amenable to present day state-of-the art deep learning based image and video understanding pipelines. However, deep learning based analysis typically involves memory intensive models and the requirement of significant computational resources, which is a challenging prospect in point-of-care healthcare applications in the developing world. With the advent of portable ultra-sound systems, it is increasingly possible to expand the reach of prenatal health diagnosis. To accomplish that, there is a need for lightweight architectures that can perform image analysis tasks without a large memory or computational footprint. We propose a lightweight convolutional architecture for assessment of ultrasound videos, suitable for those acquired using mobile probes or converted from a DI-COM standard from portable machines. As exemplar of approach, we validated our pipeline for fetal heart assessment (a first step towards identification of congenital heart defects) inclusive of viewing plane identification and visibility prediction in fetal echocardiography. This was attempted by models using optimised kernel windows and the construction of image representations using salient features from multiple scales with relative feature importance gauged at each of these scales using weighted attention maps for different stages of the convolutional operations. Such a representation is found to improve model performances at significant economization of model size, and has been validated on real-world clinical videos.