Background
Robotic ultrasound can improve scan repeatability and standardization, but image quality still depends strongly on probe contact force, probe pose, and the local acoustic window. Fixed-path scanning with fixed force control may not handle insufficient contact, over-compression, probe misalignment, or local acoustic-window loss during scanning.
Therefore, this project aims to develop a real-time image-quality-guided closed-loop acquisition method for robotic musculoskeletal ultrasound. During scanning, the robot will use B-mode image quality feedback to adapt the contact force and probe pose, and perform local re-scanning when low-quality regions are detected.
Project Objectives
The objective of this project is to develop a real-time quality-aware robotic ultrasound acquisition system. The system should be able to detect image-quality degradation during scanning and perform corresponding recovery actions. Specifically, the project will focus on:
- Real-time image quality assessment: develop a method to evaluate B-mode ultrasound image quality during robotic scanning and detect low-quality frames.
- Quality-guided force and pose adaptation: design a recovery strategy that adjusts the probe contact force and performs small pitch/roll corrections when image quality decreases.
- Local re-scanning and validation: implement a quality-triggered local re-scanning mechanism and evaluate the system using phantom experiments, healthy-volunteer scans, and an existing segmentation network.
Methodology
The robot will follow a predefined musculoskeletal ultrasound scanning path while continuously acquiring B-mode ultrasound images, robot pose, and contact force information. For each frame, the system will compute an image quality score. A sliding window or exponential moving average will be used to smooth the quality score over time, reducing unnecessary triggers caused by single-frame speckle noise.
When image quality is normal, the robot continues along the original scan path. When continuous low-quality frames are detected, the system estimates the likely failure cause. If the degradation is caused by insufficient contact or poor acoustic coupling, the robot slightly increases the contact force. If over-compression is detected, the robot reduces the contact force. If the contact force is reasonable but the bone cortex or muscle boundary remains unclear, the system performs small pitch/roll adjustments or local lateral shifts. If image quality remains poor over consecutive frames, the robot backtracks by approximately 5–15 mm and re-scans the local region. Once image quality recovers, the robot resumes the original scan path.
Experimental Design
The experiments will include phantom validation and healthy-volunteer validation. Phantom experiments will be used to create controlled failure cases, such as weak contact, over-compression, probe tilt, and local acoustic-window loss. These experiments will evaluate whether the system can detect and recover from image-quality degradation in real time.
Healthy-volunteer experiments will be conducted with at least six healthy participants, mainly focusing on the forearm radius/ulna region. Each participant will undergo comparative scans under different acquisition strategies, including fixed-path scanning, force-only adaptation, force-and-pose adaptation, and the complete quality-aware local re-scanning method.
Main evaluation metrics will include mean image quality score, low-quality frame ratio, usable-frame ratio, recovery success rate, number of re-scanning triggers, additional scanning time, contact-force safety, bone cortex visibility, and muscle-boundary clarity.
Segmentation-Based Validation and Expected Outcomes
This project will use an existing musculoskeletal ultrasound segmentation network developed by PhD researchers as a downstream validation tool. The master’s student is not expected to develop a new segmentation model. Instead, the existing network will be used to evaluate whether images acquired under different scanning strategies are more suitable for automatic segmentation of the bone cortex, muscle boundary, or related anatomical structures.
The expected outcomes include a real-time closed-loop robotic musculoskeletal ultrasound acquisition prototype, phantom and healthy-volunteer experimental results, image-quality analysis, and downstream validation results based on the existing segmentation network.
The daily supervisor of this project will be Dezhi Sun.