Abstract
In this work, we develop an approach for guiding robots to automatically localize and find the shapes of tumors and other stiff inclusions present in the anatomy. Our approach uses Gaussian processes to model the stiffness distribution and active learning to direct the palpation path of the robot. The palpation paths are chosen such that they maximize an acquisition function provided by an active learning algorithm. Our approach provides the flexibility to avoid obstacles in the robot's path, incorporate uncertainties in robot position and sensor measurements, and include prior information about the location of stiff inclusions while respecting robot kinematics. The proposed framework is evaluated via simulation and experimentation on three different robot platforms: a 6-DoF industrial arm, the da Vinci Research Kit (dVRK), and the Insertable Robotic Effector Platform (IREP). Results show that our approach can accurately estimate the locations and boundaries of stiff inclusions while reducing exploration time.
Introduction
In minimally invasive surgery (MIS), surgeons often lose the direct "haptic" feedback—the sense of touch—available in open surgery. Palpation, or pressing on tissue to identify hidden tumors, blood vessels, or tendons, is critical because these structures are typically stiffer than surrounding healthy tissue. While robotic tools can restore some of this through force sensors, manual robotic palpation is slow and mentally taxing.
Existing autonomous attempts often rely on inefficient "blind" raster scans or focus solely on finding the single stiffest point, thereby missing the overall shape or multiple tumors. Furthermore, these methods rarely account for the physical constraints of the robot or the need to avoid obstacles like bones or other surgical instruments. This paper addresses these gaps by presenting a framework for autonomous, trajectory-optimized palpation.
Methodology
The proposed "Trajectory-Optimized Sensing" approach treats tumor search as an active learning problem composed of four key components:
- Gaussian Process (GP) Regression: The robot builds a probabilistic map of tissue stiffness, predicting values in unsampled areas and quantifying uncertainty. The model incorporates "input uncertainty" to account for mechanical inaccuracies in the robot's position sensing.
- Active Learning (AL) Strategies: The robot decides its next move by maximizing an "acquisition function." The study compares four strategies: Active Area Search (AAS), Active Level Set Estimation (LSE), Uncertainty Sampling (UNC), and Bayesian Optimization (BOA).
- Stochastic Trajectory Optimization: Instead of jumping between discrete points, the robot plans continuous, smooth paths using a Dubins car model to ensure they are physically achievable.
- Cross-Entropy (CE) Method: This optimization engine iteratively samples potential paths to find the one providing the most information while respecting obstacle constraints.
Experimental Evaluation
The framework was validated across simulated environments and physical phantoms. Simulations involved 100 randomly generated stiffness maps to statistically compare active learning algorithms. Physical experiments utilized silicone "organs" with embedded stiff inclusions of various shapes, tested on three distinct robotic platforms: a high-precision 6-DoF industrial arm, the standard dVRK research platform, and the specialized IREP continuum robot for single-port surgery.
Key Results and Insights
Strategy Performance
Active Area Search (AAS) emerged as the most effective strategy. While Bayesian Optimization (BOA) often gets "stuck" at a single stiffest peak, AAS provides a superior balance, successfully identifying multiple tumors and mapping their boundaries more effectively than LSE or UNC.
Efficiency and Robustness
- Continuous Palpation: Moving the sensor along an optimized path provides a continuous stream of data, building an accurate map faster than discrete "poking."
- Obstacle Awareness: By integrating obstacle avoidance directly into the trajectory optimization, the robot can safely search for tumors in cluttered surgical scenes.
- Mechanical Robustness: Incorporating robot position uncertainty into the GP covariance makes the system robust to the mechanical "wobble" inherent in flexible surgical robots like the IREP.
Conclusion
This framework demonstrates that surgical robots can autonomously and efficiently map tissue abnormalities. By combining probabilistic modeling with trajectory optimization, the system achieves higher recall and real-time performance. Future work will focus on addressing the visco-elastic nature of real human tissue and the dynamic deformations that occur during surgery, rather than just before, the palpation process.
Summary updated