Bayesian-neural-network-based strain estimation approach for optical coherence elastography

Abstract

Strain estimation is critical for quantitative elastography in quasi-static phase-sensitive optical coherence elastography (PhS-OCE). Deep-learning methods have achieved exceptional performance in estimating high-quality strain distributions. However, they cannot often assess their predictive accuracy and reliability rigorously. To navigate these challenges, a Bayesian-neural-network (BNN)-based strain estimation is proposed. The method can provide the uncertainty distribution of the results beyond achieving high-quality strain estimation. Such an uncertainty distribution can assess the reliability of the strain results. Moreover, the uncertainty degree can function as an indicator for compensating for phase decorrelation and thus significantly enhancing the SNR and dynamic range of PhS-OCE. Thermal and three-point bending deformation experiments validated that the predicted uncertainty distribution can effectively address phase decorrelation and allow for a more comprehensive understanding of the estimated strain results.