Abstract
Purpose :
Various tools are available to analyze the workflow in ophthalmic surgery, facilitating the detection of deviations from normal procedures. However, there is currently no approach for predicting technical errors during cataract surgery and providing surgeons with alerts and recommendations preemptively, anticipating errors before they occur.
Given the limited availability of surgery videos showcasing errors, defining and modeling them remains challenging.
On the other hand, surgery simulators enable the generation of a large number of scenarios, providing access to a variety of errors.
In this study, we make the most of data from a realistic surgical simulator to automatically predict technical errors during the capsulorhexis step in cataract surgery.
Methods :
We leverage video streams and tool-related kinematic data from the capsulorhexis module of the EyeSi Surgical virtual reality simulator for cataract surgery. Our dataset comprises 421 exercises obtained from 10 operators of varying skill levels, totaling 8 hours and 50 minutes, with a cumulative technical error time of 42 minutes.
In this context, we developed a conditional AI algorithm that enables spatial and temporal encoding. It takes an observation time sequence (variable duration) and prediction time as inputs to predict various types of errors up to T seconds in advance.
The errors include radial extension of the capsule, capsule tear, irregular opening (smoothness), and being out of red reflex (eye error).
The mean calibrated accuracy, cAP (%), was employed as the evaluation metric.
Results :
Overall, we achieve cAP of 66.4% for T = 5 s using kinematic data. Performances increase to 74% with video, demonstrating a notable improvement with visual information.
Notably, not all errors in cataract surgery are equal in their predictability. For instance, the sudden nature of tear events seems to pose a challenge for advanced detection while the geometric features of the eye appear to be well captured by the model from images for predicting eye error.
Conclusions :
Despite the difficulty of this task, our efforts in predicting errors during cataract surgery showcased the feasibility of anticipating error detection, especially by leveraging visual data. Extending this research into real-world scenarios to ascertain the potential of learning from simulation data should deserve further investigation.
This abstract was presented at the 2024 ARVO Annual Meeting, held in Seattle, WA, May 5-9, 2024.