Fisher vector encoding for detecting objects of interest in ultrasound videos

One of the main factors limiting the wider adoption of ultrasound imaging for diagnosis and therapy is requiring highly skilled sonographers. In this paper we consider the challenge of making this technology easier to use for non-experts. Our approach follows some of the recently proposed frameworks that break the process into firstly data acquisition through a simple and task-specific scan protocol followed by using machine learning methodologies to assist non-experts in performing diagnostic tasks. We present an object classification pipeline to identify the fetal skull, heart and abdomen from all the other frames in an ultrasound video, using Fisher vector features. We describe the full proposed method and provide a comparison with a recently proposed approach based on Bag of Visual Words (BoVW) to demonstrate that the new approach is superior in terms of accuracy (98.9% versus 87.1%).