Visuotactile and Explicitly Force-Controlled Robotic Ultrasound for Abdominal Volumetric Reconstruction

Technical Summary: Visuotactile and Explicitly Force-Controlled Robotic Ultrasound for Abdominal Volumetric Reconstruction

Problem Statement
Ultrasound is the most widely used diagnostic imaging modality globally due to its safety, portability, and low cost. However, its reliance on highly trained sonographers creates a significant bottleneck in healthcare delivery, exacerbated by a global shortage of professionals and the ergonomic risks associated with repetitive manual scanning. While robotic systems offer a potential solution, existing approaches often lack the ability to autonomously adapt to patient-specific anatomical variations, particularly the complex topography of the abdomen where the rib cage obstructs soft tissue imaging. Furthermore, manual scanning typically yields only two-dimensional (2D) data, whereas three-dimensional (3D) volumetric data is increasingly valuable for advanced diagnostics but difficult to acquire consistently by hand.

Methodology
The authors present a robotic ultrasound acquisition system that integrates stereo vision, touch-based feedback, and expert-informed strategies to perform autonomous and adaptive abdominal scans. The system utilizes a compliant, torque-controlled seven degree-of-freedom (7-DOF) robotic manipulator equipped with an ultrasound transducer.

• Unified Motion and Force Control: The system employs an operational space controller that decouples motion and force control. In the force space, the robot maintains consistent contact pressure against the deformable phantom surface. In the motion space, it tracks desired trajectories. Redundancy is managed via joint configuration-holding control.
• Expert Strategy Recording and Replay: To replicate human expertise, the system records freehand motion and force data from expert radiologists. A mechanism captures transducer position, orientation, and contact forces. The authors use Proper Orthogonal Decomposition (POD) on contact force snapshots to identify the primary direction of force control. This data is used to parameterize a "human-inspired" force strategy where the robot modulates applied force based on position error; specifically, force is reduced when the robot detects resistance (surface deformation) to facilitate tangential sliding, mimicking the expert's technique of applying less force while moving and more when stationary.
• Vision-Based Path Planning: A stereo camera system generates a 3D topography map of the patient's abdomen (represented here by a phantom). The point cloud is processed into a watertight mesh using Poisson surface reconstruction. Initial serpentine scan paths are computed to sweep the transducer face over the mesh surface.
• Touch-Based Rib Boundary Delineation: To address acoustic shadowing caused by ribs, the system refines the vision-based paths using tactile feedback. The robot probes the surface at key points to measure linear stiffness (force/displacement). Regions with stiffness above a specific threshold are classified as rib cage, while softer regions are classified as soft tissue.
• Autonomous Scan Execution: Using the delineated rib boundary, the system generates two distinct scanning paths:
  1. Under-Rib Sweep: A geodesic path along the rib boundary where the transducer is angled (rotated -30° from the surface normal) to visualize structures beneath the ribs.
  2. Soft Tissue Sweep: A perpendicular sweep across soft tissue regions where the transducer remains normal to the surface.
• 3D Volume Reconstruction: The system tracks the transducer pose (position and orientation) synchronized with ultrasound images at 7 Hz. These pose-tracked images are compounded to reconstruct 3D volumes, enabling lesion segmentation.

Key Contributions

1. Expert Strategy Integration: A framework for recording and replaying expert freehand scanning strategies, including motion, force, and anatomical observations, to create a foundation for autonomous capabilities.
2. Visuotactile Autonomy: A novel method combining stereo vision for initial path planning and tactile stiffness measurement for refining those paths to distinguish between rib cages and soft tissue.
3. Adaptive Force Control: A compliant control strategy that dynamically adjusts applied force based on surface deformation, allowing the robot to navigate deformable surfaces and replicate expert "sliding" techniques.
4. 3D Volumetric Acquisition: The demonstration of a robotic system capable of autonomously acquiring 3D ultrasound volumes, surpassing the 2D limitations of manual scanning.

Experimental Results
Experiments were conducted on an abdominal phantom containing a liver, gallbladder, pancreas, kidneys, and simulated lesions (cysts, tumors).

• Replay Performance: The robot successfully replayed a reduced expert scan trajectory. The average motion space position error was approximately 0.5–0.7 cm, and orientation error was 1.6–2.4°. Force tracking errors were generally low, with a notable deviation in the z-direction (2.20 N) attributed to the complex interaction with the deformable surface.
• Image Quality: Ultrasound images acquired by the robot during replay showed functional equivalence to expert scans. Lesions and anatomical structures (e.g., inferior vena cava) were clearly visible in both modalities. Image similarity analysis using distance-based hashing and Hamming distance confirmed that robotically acquired images were significantly more similar to their expert counterparts than to unrelated scans.
• Autonomous Scanning: The system successfully executed the under-rib and soft tissue sweeps. The resulting 3D volumes allowed for the segmentation of a 20 mm simulated hepatic lesion. The segmented volumes measured 7.11 cm³ and 6.43 cm³ for the two sweeps, respectively, compared to a nominal volume of 8 cm³ (a discrepancy the authors attribute to the segmentation process).

Significance and Claims
The paper claims that this work represents a significant step toward generalizing robotic imaging for diverse patients and clinical scenarios. By integrating expert knowledge with physical reasoning (stiffness-based boundary detection), the system achieves high-quality imaging comparable to expert scans while dynamically adapting to patient-specific topographies.

The authors emphasize that the robotic system surpasses expert capabilities in the specific domain of 3D volume acquisition. While experts are limited to 2D planes during manual scanning, the robot's ability to execute consistent, pose-tracked sweeps enables the reconstruction of volumetric data. This capability enhances diagnostic potential by allowing for post-acquisition plane reconstruction and quantitative measurements (e.g., organ volume estimation), addressing the operator dependence and variability inherent in manual ultrasound. The work underscores the potential of combining perception-based autonomy with physical reasoning to improve diagnostic performance and accessibility.