Efficient Pose Estimation and Sensor feedback Optimization for Enhancement of Robotic Ultrasound


Autonomous Robotic Ultrasound presents significant challenges in ensuring high-quality image acquisition across various patients. The Autonomous robotic ultrasound systems face a significant challenge in acquiring high-quality images across different patients. The orientation of the robotic probe is crucial for ultrasound image quality. To address this challenge, we propose a sample-efficient method to automatically adjust the probe's orientation normal to the scanning surface, improving acoustic coupling and image quality. Our approach utilizes Bayesian Optimization (BO) to efficiently find the normalized probe orientation without using ultrasound image data. We formulate a novel objective function for BO that leverages contact force measurements and underlying mechanics. A regularization scheme handles noisy objective function data. We validate our strategy on urinary bladder phantoms with various surface shapes and 3D human mesh models. Results show a mean absolute angular error of 2.4 ± 0.7◦ for phantoms and 2.1 ± 1.3◦ for 3D models.

Forces at end effector

7 DOF Sawyer Arm

Ultrasound imaging is a valuable medical tool, offering non-invasiveness and safety. However, image quality depends heavily on operator skill, leading to variability in diagnoses. To address this, autonomous Robotic Ultrasound Systems (RUS) have been developed, using robotic arms to control the probe. Maintaining proper probe orientation is vital for good image quality. In this work, we focus on optimizing probe orientation for optimal acoustic coupling, highlighting the impact on image quality.

Related Work on Robotized Probe Normalization:

Previous studies have attempted to normalize robotized ultrasound probe orientations. Methods include 3D point cloud data analysis, visual servoing, and hardware solutions. However, these methods often rely on computational-intensive algorithms, extensive exploration, or are affected by sensor noise and deformations.


This project addresses the limitations of previous methods by proposing a sample-efficient Bayesian Optimization-based approach for identifying the necessary probe orientation adjustments. Key contributions include:

Sample-Efficient Bayesian Optimization: We employ Bayesian Optimization to efficiently search for the normal direction without extensive exploration.

Objective Function Formulation: We introduce an objective function that leverages force sensor data and mechanics to guide the normal identification process while handling noisy measurements.

Validation: The proposed method is validated on urinary bladder phantoms with different surface shapes and 3D human mesh models to simulate human physiology.

Images of phantoms used in simulation and the Sawyer experiments with the observations: