Label Efficient Localization of Fetal Brain Biometry Planes in Ultrasound Through Metric Learning.

For many emerging medical image analysis problems, there is limited data and associated annotations. Traditional deep learning is not well-designed for this scenario. In addition, for deploying deep models on a consumer-grade tablet, it requires models to be efficient computationally. In this paper, we describe a framework for automatic quality assessment of freehand fetal ultrasound video that has been designed and built subject to constraints such as those encountered in low-income settings: ultrasound data acquired by minimally trained users, using a low-cost ultrasound probe and android tablet. Here the goal is to ensure that each video contains good neurosonography biometry planes for estimating the head circumference (HC) and transcerebellar diameter (TCD). We propose a label efficient learning framework for this purpose that it turns out generalises well to unseen data. The framework is semi-supervised consisting of two major components: 1) a prototypical learning module that learns categorical embeddings implicitly to prevent the model from overfitting; and, 2) a semantic transfer module (to unlabelled data) that performs “temperature modulated” entropy minimization to encourage a low-density separation of clusters along categorical boundaries. The trained model is deployed on an Andriod tablet via TensorFlow Lite and we report on real-time inference with the deployed models in terms of model complexity and performance.