Measurement-informed model-mediated human teleoperation for remote ultrasound

Access to diagnostic ultrasound imaging remains limited in remote communities, motivating the development of tele-ultrasound to bridge this gap. This thesis presents advances to human teleoperation, a mixed reality and haptics-based tele-ultrasound approach in which a novice on-site operator acts as a flexible, cognitive robot guided remotely by an expert sonographer. Three main contributions are described. We demonstrated the first clinical use of human teleoperation in a feasibility study conducted across 754 km between Skidegate, Haida Gwaii and Vancouver, Canada. Across 11 scans with 10 novices and two sonographers, 92% of acquired epigastric images were rated as diagnostically sufficient by two radiologists, with the remaining 8% attributable to body habitus. Novices reported below-reference task load scores and unanimously positive usability. We additionally introduced an ellipsoid-based haptic feedback model, calibrated using the system's position and force sensors. Our second contribution established an ongoing comparative study of human teleoperation, tele-mentored ultrasound, and direct ultrasound for abdominal aortic aneurysm screening. Preliminary results highlighted the importance of sonographer training for human teleoperation and the need for more accurate patient modelling for force feedback. Our final and most significant contribution addresses the challenge of accurate force feedback under communication delay by introducing a dynamic model-mediated teleoperation framework. We developed a voxel-based implementation of pressure field contact and incorporated measured positions, forces, and torques to improve teleoperation transparency. In our framework, a patient point cloud is captured and the volume contained by the point cloud is voxelized in cylindrical coordinates. Each voxel is assigned a potential, and force and torque are computed based on interaction with a point shell virtual tool. We initialize the potential field with the spatial Laplace's equation solution, then update it with real-time measurements using a least squares system. Evaluation with pseudo-ultrasound scans on human subjects (n=4) and a phantom showed the addition of measurements to the model reduced the force magnitude error by an average of 7.38 N, the force vector angle error by an average of 3.63°, and the torque vector angle error by an average of 74.3°. Additionally, the model's stiffness error was reduced to an average of 4.6%.