Searching for Structures of Interest in an Ultrasound Video Sequence

Abstract

Ultrasound diagnosis and therapy is typically protocol driven but often criticized for requiring highly-skilled sonographers. However there is a shortage of highly trained sonographers worldwide, which is limiting the wider adoption of this cost-effective technology. The challenge therefore is to make the technology easier to use. We consider this problem in this paper. Our approach combines simple standardized clinical US scanning protocols (defined by our clinical partners) with machine learning driven image analysis solutions to enable a non-expert to perform ultrasound-based diagnostic tasks with minimal training. Motivated by recent work on dynamic texture analysis within the computer vision community, we have developed, and evaluated on clinical data, a framework that given a training set of Ultrasound Sweep Videos (USV), models the temporal evolution of objects of interest as a kernel dynamic texture which can form the basis of a metric for detecting structures of interest in new unseen videos. We describe the full original method, and demonstrate that it outperforms a simpler recently proposed approach on phantom data, and is significantly superior in performance on real clinical data.