3D-BNN-accelerated phase tracking with optimal sub-region selection for high-dynamic-range OCE

Phase decorrelation can limit the dynamic range of phase-sensitive optical coherence elastography (PhS-OCE). Here, we reported a three-dimensional (3D) Bayesian neural network with an optimized sub-region size determination approach to overcome phase decorrelation rapidly in 3D-PhS-OCE. Our method can automatically provide a reasonable sub-region size as small as possible without requiring any prior knowledge. As a result, 3D pixel-level offsets induced by large deformations can be tracked rapidly, benefiting from both the smaller matching subregion and the network's parallel computing capability. The results demonstrate that the proposed 3D-BNN with optimal sub-region size determination improves computational efficiency by about 24% over the empirical 3D-BNN and by 26 times over the PVC method.