Purchase this article with an account.
Eric Tang, Mohamed T El-Haddad, Joseph D Malone, Yuankai Tao; Deep-learning based automated instrument-tracking and adaptive-sampling for 4D imaging of ophthalmic surgical maneuvers. Invest. Ophthalmol. Vis. Sci. 2020;61(7):2544.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
Intraoperative optical coherence tomography (iOCT) benefits clinical decision-making and allows for verification of surgical goals by providing high-resolution volumetric images of tissue microstructure. Current microscope-integrated OCT systems are limited by slow scan speeds and static OCT fields-of-view (FOVs), which preclude real-time 4D visualization of surgical maneuvers and require periodic adjustment of the OCT field, respectively. Here, we present deep-leaning based automated instrument-tracking and adaptive-sampling using spectrally encoded coherence tomography and reflectometry (SECTR) for 4D imaging of ophthalmic surgical maneuvers.
A custom 400 kHz swept-source OCT engine was used to perform SECTR imaging in ex vivo porcine eyes. En face spectrally encoded reflectometry (SER) and cross-sectional OCT images were simultaneously acquired at 1664x1000x200 pix. (spectral x lateral x lateral). A convolutional neural network (CNN) was trained to detect 25G ILM forceps using 2808 SER images. Bounding box outputs produced by the CNN were transferred to an FPGA to offset scan waveforms in real-time to localize and densely sample surgical instruments within the OCT FOV (Fig. 1).
Automated instrument-tracking was performed in intact ex vivo porcine eyes. Benchtop imaging was performed at 2 volumes per second with a tracking rate of 18 Hz. SER images show dense sampling of the instrument tip despite lateral movement. Corresponding OCT B-scans and OCT volumes show tracking of the instrument following out-of-plane motion (Fig. 2).
SECTR imaging allows for real-time localization of surgical instruments and dynamic adjustment of OCT scanning. Adaptive-sampling enhances visualization of instrument-tissue interactions by increasing sampling density without sacrificing speed. The proposed method overcomes limitations of conventional iOCT by providing 4D visualization of surgical maneuvers without the need for manual OCT aiming.
This is a 2020 ARVO Annual Meeting abstract.
(a) SER images (blue) are passed through the CNN. An FPGA is used to offset X (green) and Y (orange) galvanometer positions to densely sample the instrument tip and adjust OCT scanning position.
(a), (b) SER images (blue) show adaptive-sampling in an ex vivo porcine eye following out-of-plane motion. (c)-(f) OCT B-scans and volumes (red) demonstrate automated localization of the instrument (yellow asterisk) within the OCT FOV.
This PDF is available to Subscribers Only