Dynamics-aware deep predictive adaptive scanning optical coherence tomography

Conventional scanned optical coherence tomography (OCT) suffers from the frame rate/resolution tradeoff, whereby increasing image resolution leads to decreases in the maximum achievable frame rate. To overcome this limitation, we propose two variants of machine learning (ML)-based adaptive scanning approaches: one using a ConvLSTM-based sequential prediction model and another leveraging a temporal attention unit (TAU)-based parallel prediction model for scene dynamics prediction. These models are integrated with a kinodynamic path planner based on the clustered traveling salesperson problem to create two versions of ML-based adaptive scanning pipelines. Through experimental validation with novel deterministic phantoms based on a digital light processing board, our techniques achieved mean frame rate speed-ups of up to 40% compared to conventional raster scanning and the probabilistic adaptive scanning method without compromising image quality. Furthermore, these techniques reduced scene-dependent manual tuning of system parameters to demonstrate better generalizability across scenes of varying types, including those of intrasurgical relevance. In a real-time surgical tool tracking experiment, our technique achieved an average speed-up factor of over 3.2× compared to conventional scanning methods, without compromising image quality.