One-Shot Autofocus Via User-Adaptive Gaze Control for Robot-Assisted Microsurgery
Yao Guo, Guang-Zhong Yang
AI summary
Problem
Existing gaze-based autofocus methods in robot-assisted microsurgery rely on complex triggering rules and inefficient multi-step optimization, making them difficult to adapt to diverse surgeons and dynamic surgical environments.
Approach
The framework integrates real-time gaze tracking with a bi-directional defocus estimation network to predict trigger likelihood and absolute defocus distance from a single image, allowing surgeons to adjust sensitivity via a slider for rapid one-shot focusing.
Key results
- Reduced defocus estimation error to 0.671 mm with 75.4% accuracy
- Enabled real-time trigger prediction at 20 Hz with adjustable sensitivity thresholds
- Demonstrated superior user-friendliness and operational efficiency in user studies
- Provided a practical data collection pipeline for training surgical defocus models
Why it matters
This approach streamlines surgical workflows by delivering fast, intuitive, and personalized autofocus control, directly benefiting surgeons and developers of automated microsurgical systems.
Abstract
Robot-assisted microsurgery (RAMS) is rapidly advancing with increasing levels of automation. Given the inherently shallow depth-of-field characteristics of surgical microscopes, integrating autofocus capabilities into RAMS has emerged as an urgent trend. Among existing solutions, gaze- induced autofocus has gained prominence due to its natural alignment with the surgeon’s visual attention. However, gaze autofocus often relies on complex and non-intuitive triggering mechanisms, making it difficult to adapt for diverse users. Additionally, although the hill-climbing strategy is commonly employed to find the optimal focus plane, this process is inefficient for RMAS due to its slow convergence and inability to accommodate dynamic surgical scenarios. To address these limitations, we propose a novel gaze-controlled autofocus system featuring user-adaptive triggering and one-shot focusing. When a region is defocused and under the surgeon’s gaze, our system rapidly achieves optimal focus with a single-step lens movement. Surgeons can easily adjust trigger sensitivity using a slider. Ex- periments validate the accuracy of our defocus estimation and triggering prediction algorithms. A user study demonstrates that the proposed system offers superior user-friendliness and operational efficiency compared to conventional systems.