Towards Multi-Sweep Ultrasound Video Understanding: Application in Detection of Breech Position Using Statistical Priors

We propose a novel way of creating graphs from multiple ultrasound (US) video sweeps. A three-node graph models three video sweeps which are obtained from a fetal US sweep protocol. The nodes are assigned binary sequences which represent the frame-level detection of the fetal head across all video frames in a sweep. We build 382 subject-level graphs and use them for automatic breech detection using a graph convolutional network. We experiment with weighting the edges of the graphs using three metrics: 1) discrete Fréchet distance, 2) dynamic time warping, and 3) Pearson correlation coefficient. These metrics are computed from the node signals, per subject. We find superior performance is achieved using Pearson correlation weighted graphs, over a baseline multi-layer perceptron (MLP), achieving an increase in classification accuracy of 28%. Finally, we experiment with creating statistical priors of graphs that correspond to the overall pattern seen for breech and for cephalic presentation, using a set of subjects independent to the 382 cohort. We compute the weights of the edges of these statistical graphs and assemble each set into two templates of edge weights that model the relationship between different video sweeps for breech and cephalic presentation respectively. We find that using the Pearson correlation template of edge weights during testing increases classification accuracy by 14%, over a baseline MLP.