In early October, Kitware had an active presence at the 15th International Conference on Medical Image Computing and Computer Assisted Intervention (MICCAI). As part of this event, Roland Kwitt, an R&D Engineer at Kitware, was awarded one of five Young Scientist Awards for his paper “Recognition in Ultrasound Videos: Where am I?,” on which he was first author, written in collaboration with Nuno Vasconcelos, Sharif Razzaque, and Stephen Aylward.
The Young Scientist awards are prizes awarded to first authors of high quality MICCAI papers. The authors had to be a full-time student at a recognized university or have been a student at most two years before the paper submission deadline. The awards are determined by a committee of five members who voted on their selections. Among the 30 eligible papers corresponding to oral presentations and papers with the highest-average reviewing score, 13 papers were selected and ranked by a second vote.
While many developed western countries have immediate access to expensive imaging modalities, such as MRI or CT, rural parts of the world or developing countries usually do not possess that kind of advanced imaging equipment. For that reason, ultrasound (US) imaging has become increasingly popular in these areas, especially with the emergence of portable probes that can be hooked-up to mobile phones or tablet PCs. Since experienced radiologists are usually unavailable as well, it has been proposed to have supporting staff perform ultrasound examinations. This can be problematic, though, for one particular reason: Locating an organ or an area of interest is hard for inexperienced personnel, mainly due to variation in human anatomy and high noise levels in ultrasound images. In this talk, we discuss an approach to guide the localization of interesting regions by modeling appearance changes in ultrasound video sequences by a generative model and comparing that model to an atlas of previously acquired key locations. We argue that the change of appearance of a particular anatomical structure as we move the ultrasound probe is more distinctive than a single, still image of the same area. Technically, we draw on recent advances in action recognition literature and model the appearance changes as a non- linear dynamical system. Similarity among US video sequences is then defined as similarity in the parameter space of that model. We present several experiments on US sequences acquired on a handmade noodle-phantom and a 3-D abdominal phantom. We further show preliminary results on the impact of anatomical variations, simulated by (non-linear) spatial distortion of the video material.