Quality-Aware Robotic Knee Ultrasound Images Acquisition Using Multiscale Feature Extraction Model

Ultrasound (US) examination is critical for diagnosing knee joint disorders, but its quality and reproducibility are heavily reliant on sonographers’ experience, which causes fatigue and imaging inconsistency. This article presents an autonomous knee US scanning robot system integrated with a multiscale feature extraction model to address these issues. The robotic platform integrates a seven-degrees of freedom (DoF) manipulator, dual RGB-Depth cameras, and a force sensor to ensure safe human–robot interaction and accurate knee joint localization. Meanwhile, a novel multiscale image quality assessment network is proposed, which combines global features, target-specific regions, and local details to evaluate the visibility and clarity of knee cartilage. Gaussian process regression (GPR) is employed to model the relationship between probe angles and image quality. This enables the robot to autonomously search for the optimal probe contact angle. Experimental validation involving 20 trials on five volunteers demonstrated a 95% procedure completion rate and an angular error of 2.45^◦ relative to expert adjustments. The system enables standardized and autonomous acquisition of diagnostic-grade knee US images.