Deep-Learning-based Velocity Prediction Model for Fast Optical Coherence Elastography Processing

Optical coherence elastography (OCE) enables quantitative assessment of tissue biomechanical properties by analyzing surface acoustic wave (SAW) velocities. However, current estimation methods require significant expertise and lengthy image processing times, limiting the clinical translation of this promising technology. This study presents a SAW velocity prediction network (SP-Net), a novel end-to-end deep learning model designed for rapid and accurate prediction of SAW velocities from raw OCE phase data of human skin in vivo, addressing the need for rapid and high-accuracy analysis in clinical settings. By using a convolutional-transformer which combines the advantages of local feature extraction and global information processing, SP-Net can extract wave propagation features efficiently, achieving robust performance and high accuracy across diverse tissue types, including tissue-mimicking agar phantoms, healthy human skin and facial acne. The SP-Net has the closest matched predicted velocity compared to the time-of-flight method (10.18 ± 1.27 vs. 10.13 ± 1.53 m/s) in the facial acne case. The experiment results show that SP-Net has the best balance among accuracy, efficiency, and model complexity, significantly improving prediction accuracy by 17% over VP-Net. This method is beneficial for future skin research to allow the real-time tissue biomechanical properties visualization.