Abstract
Purpose :
Intraoperative optical coherence tomography (iOCT) enables real-time volumetric imaging of ophthalmic surgical maneuvers and benefits clinical decision-making. Previous studies have demonstrated the effectiveness of iOCT for guiding surgery, but its utility is limited by static field-of-views (FOVs). Alignment of the OCT imaging plane to surgical regions-of-interest is complex, requires real-time tracking of instrument motions during surgery, and is a critical barrier to broad clinical adoption of iOCT. Here, we demonstrate deep-learning-based adaptively-sampled spectrally encoded coherence tomography and reflectometry (SECTR) for real-time automated surgical instrument-tracking.
Methods :
SECTR imaging was performed using a custom 1060 nm swept-source OCT engine. Simultaneous en face spectrally encoded reflectometry (SER) and cross-sectional OCT images were acquired with 1664x1000 pix. (spectral x lateral). A GPU-accelerated convolutional neural network (CNN) was trained to detect 25G internal limiting membrane forceps using 2236 manually-labelled SER images. CNN outputs were used to track surgical instrument positions and densely sample instrument tips by dynamically modifying SECTR galvanometer waveforms for each B-scan using custom-designed hardware.
Results :
Automated instrument-tracking and adaptive-sampling was demonstrated in ex vivo bovine eyes. Simultaneously-acquired en face SER and corresponding cross-sectional OCT images show dense sampling and tracking of the instrument tip despite motion in the lateral plane (Fig. 1).
Conclusions :
Deep-learning-based adaptively-sampled SECTR enables real-time automated tracking of instrument motion. The proposed method overcomes limitations of conventional iOCT by taking advantage of the inherent co-registration between SER and OCT and using deep-learning to automatically identify the instrument tip. This allows for acquisition of densely-sampled OCT volumes of surgical maneuvers at high volumetric frame-rates to guide clinical decision-making.
This abstract was presented at the 2019 ARVO Annual Meeting, held in Vancouver, Canada, April 28 - May 2, 2019.