P22.03: Searching for structures of interest in an ultrasound video sequence with an application for detection of breech

Abstract

The shortage of highly trained sonographers worldwide limits the wider adoption of ultrasound (US) as a diagnostic tool in obstetrics. The aim of this study was to develop a learning software solution to analyse US video-clips acquired along the major uterine longitudinal axis. In particular, the automatic computer analysis was tested to diagnose non-cephalic presentation. Transabdominal 2D fetal US video clips were acquired from 30 subjects participating in the INTERGROWTH-21st project (www.intergrowth21st.org) at the University of Oxford, following a simple and repeatable protocol: a single sweep starting at the symphysis pubis of the mother moving toward the fundus of the uterus over 6-8 seconds was acquired. Data acquisition was carried out using a mid-range US machine (Philips HD9 with a V7-3 transducer). Each video sequence was broken down into smaller sequences, and machine learning techniques were used to look for the fetal skull. When the fetal skull was detected in the last third of the video sequence it was assumed that the fetus is in a non-cephalic presentation. Video clips were acquired and analysed in 100% of the cases. Automatic head detection and thus detection of non-cephalic presentations was possible with an average accuracy of 90% in all the videos obtained through our proposed protocol. We have proposed a general framework for US video analysis. Cost effective tools for automatic US video clips analysis can assist in the diagnosis of fetal presentation; in this case only a longitudinal uterine axis sweep is required by the sonographer. The described tool should be developed to assess other structures of interest and be tested for use in areas where basic US training is lacking.