Target-Driven Autonomous Robotic Ultrasound Probe Positioning for Anatomies Occluded by Acoustic Shadows

Nowadays, Robotic Ultrasound Systems (RUSS) are increasingly being adopted to reduce operator dependence and enable consistent, repeatable acquisitions across patients, but Ultrasound (US) images quality can be strongly affected by acoustic artifacts. Among these, bone-induced acoustic shadows are particularly challenging: highly attenuating osseous structures block the propagation of the ultrasound beam, creating dark regions where relevant anatomy becomes partially or completely hidden. As a result, the effective field of view is reduced and target interpretability may degrade, often forcing the operator to perform repeated probe repositioning and fine angular sweeps to recover the acoustic window. This thesis develops a target-driven robotic ultrasound workflow that improves the visibility of clinically relevant target anatomy in the presence of acoustic shadowing. The main motivation of this work is to translate the practical expertise of the human sonographers into a robotic pipeline that supports more standardized acquisitions while preserving clinical control, patient comfort, and safety. The project is the result of a collaboration between ARSControl (University of Modena and Reggio Emilia) and the Technical University of Munich (TUM), carried out within the IFL Lab of the Chair of Computer Aided Medical Procedures directed by Prof. Nassir Navab. The work was developed during a six-month research internship in Munich. The proposed system follows a two-stage workflow. First, the robot performs a tracked survey acquisition with the probe held approximately perpendicular to the skin, collecting longitudinal and axial views to build an initial representation of the scanned region. These data are used to compute image-derived confidence maps and aggregate them into a 3D confidence volume in physical coordinates, providing an interpretable estimate of signal reliability and highlighting shadow-affected regions. Second, once a clinician selects a target of interest in the reconstructed volume, the system defines an ``apex'' anchor on the estimated skin surface and samples a set of feasible probe poses around it. Candidate poses mimic sonographer-like fine adjustments through bounded tilts and small shifts, and are evaluated through a scoring function that balances predicted visibility with clinically realistic constraints, including target inclusion within the field of view, tilt limits for comfort and safety, and motion regularization to avoid unnecessary repositioning. The robot then autonomously moves the probe to the best-ranked pose. To enable repeatable development and systematic testing, a simulation environment was implemented by coupling ROS/RViz-based robotic control with ImFusion Hybrid Ultrasound Simulation through a dedicated bridge, which generates the corresponding ultrasound frame from the current probe pose and supports closed-loop evaluation of acquisition, reconstruction, and view planning. Experiments on spinal ultrasound scenarios, with targets sampled within clinically relevant regions used for spinal injections show that the proposed approach improves target visibility compared to perpendicular-only scanning strategies while maintaining controlled, clinically plausible probe motions. Overall, this thesis demonstrates that target-driven, shadow-aware probe positioning with clinician-in-the-loop target selection can make robotic ultrasound acquisitions more robust in challenging settings and represents a step toward task-driven, deployable RUSS workflows.