Deep Learning-Based Inclusion Boundary Identification Using Wave Propagation in Optical Coherence Elastography

This study developed a deep learning framework called Boundary Detection Network, or BDNet, to identify internal inclusion boundaries directly from optical coherence elastography wave propagation data. Conventional elastography approaches often infer boundaries indirectly from velocity or stiffness maps, which can be distorted, shifted, or difficult to interpret when stiffness gradients are subtle. The authors instead used depth-resolved OCE phase images as model inputs and trained BDNet to localize boundaries without relying on intermediate velocity estimation. The network combines a UNet-based encoder-decoder for local spatial feature extraction with Mamba state-space sequence modelling to capture long-range boundary continuity, and it uses a custom Margin-Aware Soft Asymmetric loss function to preserve weak boundary signals while reducing false detections. The method was evaluated using single-boundary and dual-boundary agar phantoms with known stiffness contrasts, as well as in vivo human facial acne data. For single-boundary agar phantoms, all models performed well, with MAE values near 2-3 pixels. In more challenging dual-boundary phantoms, BDNet showed stronger robustness at low stiffness contrasts where elastography maps became distorted or misaligned. On the acne dataset, BDNet achieved the best base-model performance, while the larger BDNet_L variant achieved the lowest MAE. Overall, the study demonstrates that deep learning applied directly to OCE wave propagation data can improve boundary localization in heterogeneous tissue-mimicking phantoms and may support future clinical applications in dermatology, fibrosis or scar assessment, and tumour boundary delineation.