Detection of spatiotemporal stochastic motion in echocardiography videos

A Patra, M Ali

Sep 30, 2018 NIPS 2018 Workshop Spatiotemporal Blind Submission readers: everyone
  • Keywords: fetal echocardiography., stochastic motion prediction, fetal motion
  • TL;DR: Unpredictable stochastic motion in fetal ultrasound, requires statistical decision process based methods for detection, necessary for corrections in automated fetal echocardiography estimations
  • Abstract: A challenge in fetal echocardiography is the unpredictable relative motions between the fetus being assessed and the probe which manifests as a change in the viewing state of a relevant anatomy being assessed on the ultrasound video monitor. While such fetal motion is natural, it is not deterministically predictable and is a function of several complex physiological factors in the in-utero environment. This unpredictable motion manifests in fetal ultrasound video sweeps and causes observational inaccuracies in the accurate physiological assessment of anatomical structures of the developing fetus. Additionally, it is a difficult transition case for automatic medical video analysis pipelines which are usually trained for extraction of deterministic patterns by representation learning on provided data and are therefore not inherently equipped to handle or correct anomalous or stochastic events in the input space. We treat the detection of such spatiotemporally stochastic transitions in fetal ultrasound videos as an anomaly detection case and use a novel statistical decision method to encode partwise distance measures across local regions from multiple frames in video segments as discrete probability distributions compared by a Matusita coefficient score.This helps compute divergences between discrete probability distributions so obtained and allows segregation between normal and anomalous spatiotemporal representations out of video segments. This is validated on a real clinical dataset and demonstrates significant performance gains over extant methods.
0 Replies

Loading