Using Deep Learning to Estimate Frame-to-Frame Angle Displacements in 2D Ultrasound Image Sequences of an Infant Hip

Abstract

To assess developmental dysplasia of the hip in infants, evaluations are currently conducted based on 2D ultrasound (US) images. Using 3D US has been shown to markedly reduce inter-rater variability, but 3D scanners are not widely available in pediatric practices. Here, we propose using deep learning to estimate the spatial positions of 2D US image sequences; this can then be used to form 3D reconstructions. In this study, we extracted fan-shaped sets of slices from a database of 1403 3D US volumes and trained a previously proposed standard convolutional neural network (CNN) as well as two variations of a deeper CNN (one augmented with optical flow (OF) information) to estimate the angular distances between separated slices. The deeper CNN most accurately predicted the inter-slice angular displacements, with a mean absolute error of 0.02˚, for displacements of up to 3.0˚ (corresponding to a center-frame displacement of 5.3mm). OF did not appear to improve prediction accuracy in angle estimation. The deeper CNN also achieved a mean end-to-end sweep angle error of -0.8% ± 13.2%, compared with an error of 25.3% ± 14.7% for the previously proposed standard CNN. This relatively low error suggests that it may be feasible to accurately reconstruct a 3D representation of an infant hip using a 2D US video stream alone, without requiring additional probe-tracking devices.