Enhancing reproducibility in autonomous robotic kidney ultrasound imaging via Bayesian optimization and multi-metric image similarity

Ultrasound (US) imaging is essential for renal assessment but remains limited by operator dependence, patient-specific variability, and difficulty in consistently acquiring diagnostic cross-sectional views. To address these challenges, we propose a robotic ultrasound framework that integrates Bayesian optimization (BO) with a composite image similarity index combining the Dice coefficient, normalized cross-correlation (NCC), and structural similarity (SSIM). The framework formulates renal US acquisition as an anatomical view identification problem and autonomously searches for probe poses that maximize similarity to a predefined target image. Phantom experiments were conducted to evaluate multiple weight combinations of the similarity metrics. Among the tested configurations, equal weighting of Dice, NCC, and SSIM achieved the highest average similarity score (0.93 ± 0.03) and demonstrated stable convergence. Subsequent feasibility studies in six healthy volunteers (12 trials per subject) achieved an average similarity score of 0.83 ± 0.14 with a mean search time of 53.40 ± 30.70 s, despite rib shadowing and respiratory motion. These results suggest that integrating complementary structural and intensity-based similarity measures within a BO framework improves robustness in renal US probe positioning. The proposed method provides a step toward more consistent and operator-independent kidney ultrasound acquisition.