Abstract
Purpose :
To develop a platform that uses deep-learning neural networks to distinguish the level of experience of vitreoretinal surgeons when analyzing their instrument maneuvers and areas of visual attention-via gaze tracking-, when performing standardized tasks with a surgical simulator.
Methods :
Four attending surgeons, three fellows, and two resident surgeons were invited to perform a series of ophthalmic surgical tasks using a surgical simulator. These tasks included membrane peeling, hyaloid manipulation, endolaser photocoagulation, general vitrector use, and retinal detachment repair. An instance segmentation neural network was trained to track instrument maneuvers and to extract the gaze position provided by an eye-tracking bar. A second spatio-temporal neural network (CNN + LSTM) was trained to classify the level of experience of each subject by analyzing the acquired instrument maneuvers and areas of visual attention.
Results :
Combining instrument maneuvers and gaze behavior proves most effective in discerning surgeons' experience levels, especially within core vitrectomy and membrane peeling tasks (M = 0.983, SD = 0.017). Notably, endolaser tasks exhibit lower efficacy (M = 0.32, SD = 0.159). Cross-task validation models successfully identify surgeons' experience (M = 0.733, SD = 0.216). Exclusive reliance on instrument maneuvers for training and evaluation outperforms gaze behavior assessment in predicting surgical experience (M = 0.456, SD = 0.319 vs. M = 0.254, SD = 0.241). Membrane peeling task models consistently demonstrate superior performance across all scenarios: combined maneuvers with gaze (M = 0.938, SD = 0.051), maneuvers alone (M = 0.707, SD = 0.284), and gaze alone (M = 0.242, SD = 0.277).
Conclusions :
Vitreoretinal surgeons' experience levels can be distinguished by analyzing their surgical maneuvers and gaze behavior using deep-learning neural networks. Combining assessment of instrument maneuvers with gaze behavior was the most effective approach.
This abstract was presented at the 2024 ARVO Annual Meeting, held in Seattle, WA, May 5-9, 2024.