Abstract
Purpose :
Point-scan optical coherence tomography (OCT) is limited by a tradeoff between scan resolution and frame rate, which may introduce unacceptable latency for surgical scene imaging. To overcome this issue, we exploit an adaptive scanning strategy to scan regions of higher interest with greater frequency, enabling faster performance without reducing quality.
Methods :
We extend our probabilistic adaptive scanning paradigm with machine learning and path planning to intelligently determine when and where to collect OCT data. To enable anticipation of object motion, we leverage the convolutional long short-term memory (ConvLSTM) neural network architecture to extract spatio-temporal characteristics of frame sequences that generalize well across different scenes (Fig. 1). This network determines the lateral positions for data acquisition, which we combine with random exploration to discover unknown objects. We then derive the corresponding scan path by formulating and solving a clustered traveling salesperson problem (CTSP) that seeks the shortest duration path subject to the known dynamic limits of our scanner hardware. The full scene is reconstructed from only the visited positions. To evaluate our system, we collected real-time OCT data on surgically-relevant imaging phantoms, including one designed to simulate ophthalmic surgical forceps.
Results :
When imaging our forceps-shaped phantom, adaptive scanning with ConvLSTM and CTSP operated 2.97x faster than dense raster scanning and created high-fidelity images while imaging only 9.7% of the scene on average. Despite limited scene coverage, qualitative inspection of the resulting OCT data (Fig. 2) demonstrates an improvement in object boundaries compared to our prior approach.
Conclusions :
Trajectory generation using CTSPs for scan planning and the predictive capabilities of ConvLSTM not only enhance the scan efficiency but also reduce the image latency. This approach maintains a high level of scene fidelity throughout the scanning process.
This abstract was presented at the 2024 ARVO Annual Meeting, held in Seattle, WA, May 5-9, 2024.