Phase unwrapping based on transformer-enhanced residual network for Doppler OCT flow measurement

Non-invasive and high-resolution blood flow velocity (BFV) measurement using Doppler optical coherence tomography (DOCT) is critical in diagnosing and monitoring clinical diseases. However, the inherent phase wrapping problem with inevitable noise in the acquired phase data severely limits the accuracy of flow quantification. To overcome this challenge, this paper takes phase unwrapping as a regression task and proposes a deep neural network called Transformer-enhanced residual network (TRNet) for automatic and noise-robust phase unwrapping in DOCT flow measurement. Considering the importance of long-range dependency for resolving phase discontinuities in noisy environments, TRNet integrates a Transformer block after each convolutional block, which is equipped with a residual strategy and row attention mechanism in the encoder path, forming a hybrid network that synergizes local feature extraction with global contextual awareness through self-attention mechanisms. To mitigate the lack of paired phase images used for network training, we construct a practically representative dataset with DOCT real phase images by leveraging Mamba-YOLO-based object detection and morphological image processing, avoiding reliance on simulated or synthetic data. Extensive evaluations on the rat middle cerebral artery (MCA) Doppler images demonstrate that TRNet outperforms traditional and other deep learning methods under varying signal-to-noise ratios (0-25 dB) in both visual inspection and quantitative evaluation. Notably, TRNet-derived flow velocities exhibit near-perfect correlation with the actual flow velocities in phantom milk flow experiments and show strong agreement with manual measurements in in vivo blood flow experiments, further validating the effectiveness of the proposed phase unwrapping method and its clinical feasibility for precision hemodynamic analysis.