Trajectory-Optimized Sensing for Active Search of Tissue Abnormalities in Robotic Surgery

Hadi Salman, Elif Ayvali, Rangaprasad Arun Srivatsan, Yifei Ma, Nicolas Zevallos, Rashid Yasin, Long Wang, Nabil Simaan, Howie Choset

TL;DR

This paper presents a framework for surgical robots to autonomously locate and map tumors by "feeling" tissue stiffness. By combining Gaussian Process regression with trajectory-optimized active learning, the system efficiently identifies the shapes and locations of multiple abnormalities while respecting robot kinematics and avoiding obstacles.

What Problem Does This Paper Address?

In minimally invasive surgery (MIS), surgeons lose the direct "haptic" feedback—the sense of touch—they would normally have in open surgery. Palpation (pressing on tissue) is critical for identifying hidden tumors, blood vessels, or tendons that are stiffer than surrounding healthy tissue.

While robotic tools can restore some of this through force sensors, manual robotic palpation is slow and mentally taxing. Previous autonomous attempts often relied on "blind" raster scans (scanning everything, which is slow) or focused solely on finding the single stiffest point (missing the overall shape or multiple tumors). Furthermore, these methods rarely accounted for the physical constraints of the robot or the need to avoid obstacles like bones or other surgical instruments.

Methods

The authors propose a "Trajectory-Optimized Sensing" approach that treats tumor search as an active learning problem:

  1. Gaussian Process (GP) Regression: The robot uses a GP to build a probabilistic map of tissue stiffness. This allows the system to predict stiffness in unsampled areas and quantify its own uncertainty. Crucially, they incorporate "input uncertainty" to account for the fact that the robot's position sensing isn't perfect.
  2. Active Learning (AL) Strategies: The robot decides where to move next by maximizing an "acquisition function." The paper compares four strategies:
  3. Active Area Search (AAS): Focuses on regions where the average stiffness is likely above a threshold.
  4. Active Level Set Estimation (LSE): Aims to find the boundaries (the "contour lines") of stiff regions.
  5. Uncertainty Sampling (UNC): Simply goes where the map is least certain.
  6. Bayesian Optimization (BOA): Searches for the global maximum (the stiffest point).
  7. Stochastic Trajectory Optimization: Instead of jumping between discrete points, the robot plans continuous paths. They use a Dubins car model to ensure the paths are smooth and physically achievable.
  8. Cross-Entropy (CE) Method: This is the "engine" that optimizes the trajectory. It iteratively samples potential paths and hones in on the one that provides the most information (highest acquisition value) while avoiding obstacles.

Datasets

The researchers evaluated their approach using:

  • Simulated Environments: 100 simulations with randomly generated "ground truth" stiffness maps to statistically compare the different active learning algorithms.
  • Physical Phantoms: Silicone "organs" with embedded stiff inclusions of various shapes.
  • Robotic Platforms: The system was validated on three distinct setups:
  • A 6-DoF industrial arm (for high-precision ground truth gathering).
  • The da Vinci Research Kit (dVRK) (a standard research platform for robotic surgery).
  • The Insertable Robotic Effector Platform (IREP) (a specialized continuum robot for single-port surgery).

Key Insights and Ideas

  • AAS is the "Goldilocks" Strategy: While Bayesian Optimization is popular, it tends to get "stuck" at the single stiffest peak. AAS provides a better balance, successfully finding multiple tumors and mapping their boundaries more effectively than LSE or UNC.
  • Continuous Palpation > Discrete Probing: Moving the sensor along an optimized path provides a continuous stream of data, which builds a more accurate map faster than stopping to "poke" the tissue at discrete locations.
  • Obstacles as Constraints: By integrating obstacle avoidance directly into the trajectory optimization (using a proximity penalty), the robot can safely search for tumors even in cluttered surgical scenes.
  • Handling Robot "Wobble": By incorporating robot position uncertainty into the GP covariance, the system becomes more robust to the mechanical inaccuracies inherent in flexible surgical robots like the IREP.

Key Results

  • Higher Recall: In simulations, the Active Area Search (AAS) algorithm consistently achieved higher recall (finding more of the tumor area) with fewer samples compared to the other three methods.
  • Efficiency: The system successfully mapped multiple stiff inclusions in silicone phantoms using the dVRK and IREP, revealing the shape and location of abnormalities without wasting time exploring "soft" healthy regions.
  • Real-time Performance: The framework is computationally efficient enough to run in real-time, allowing the robot to update its path as new force measurements arrive.

Limitations and Future Work

  • Simplified Tissue Models: The current work assumes a linear stiffness model. Real human tissue is visco-elastic (its stiffness changes based on how fast you press) and non-linear.
  • Deformation: The paper assumes the organ is relatively static. In a real body, organs shift and deform when pressed, which would require more complex "dynamic" mapping.
  • Future Directions: The authors suggest incorporating more sophisticated tissue mechanics models and exploring how this autonomous search can better assist surgeons in a "shared-control" setting.

Summary updated

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, include prior information about location of stiff inclusions while respecting the robot-kinematics. To the best of our knowledge this is the first work in literature that considers all the above conditions while localizing tumors. The proposed framework is evaluated via simulation and experimentation on three different robot platforms: 6-DoF industrial arm, 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 the stiff inclusions while reducing exploration time.
BibTeX
@inproceedings{salman2018trajectory,
  author = {Hadi Salman and Elif Ayvali and Rangaprasad Arun Srivatsan and Yifei Ma and Nicolas Zevallos and Rashid Yasin and Long Wang and Nabil Simaan and Howie Choset},
  title = {{Trajectory-Optimized Sensing for Active Search of Tissue Abnormalities in Robotic Surgery}},
  booktitle = {IEEE International Conference on Robotics and Automation},
  year = {2018},
  url = {https://openreview.net/forum?id=616PpwNdPWn},
}

Loading related content...

Figure