Deep Learning Based Quantitative Uncertainty Estimation for Ultrasound Shear Wave Elasticity Imaging

Ultrasound shear wave elasticity (SWE) imaging measures tissue stiffness by tracking the shear wave speed (SWS), which is directly related to elastic moduli. One application of SWE is to quantify cervical softening during pregnancy in an effort to detect risk of spontaneous preterm birth and predict the success of induction of labor. Accurate SWS estimation in clinical data is challenging; existing robust methods are slow and do not provide a calibrated uncertainty metric. Here, we use a deep learning approach to estimate SWS from tracked displacement data and simultaneously produce a well-calibrated, quantitative metric of uncertainty. Our dataset consists of in vivo point-SWE cervix measurements acquired with a Siemens scanner at multiple timepoints during pregnancy in 28 subjects. We train a deep neural network to estimate the two parameters
$m$
and σ of a lognormal distribution and derive the loss-function in a maximum likelihood framework. We use the median of 5 SWS estimators as ground truth and hold-out data from 3 subjects for testing the trained network. The average relative error in SWS was 13%, and the estimated uncertainty closely approximated the spread in true error, as measured by binned root-mean-square error. Our deep learning approach is the first SWS estimator in the toolbox that provides a well-calibrated and easily-interpretable uncertainty metric.