Introduction
Peripheral Arterial Disease (PAD) affects approximately 20% of individuals over the age of 55. Despite its prevalence, it remains under-diagnosed due to the operator-dependent nature of manual ultrasound screenings. Robotic ultrasound systems (RUSS) offer a path toward standardization, but segmenting small, deep-seated vessels like the radial or ulnar arteries is challenging. Standard B-mode ultrasound is often obscured by speckle noise and artifacts. DopUS-Net addresses this by integrating Doppler flow signals into a quality-aware robotic control loop.
Methods
DopUS-Net utilizes a dual-encoder deep learning architecture to fuse anatomical (B-mode) and functional (Doppler) information.
- Dual-Encoder Architecture: One encoder processes duplex images (B-mode + Doppler) for spatial context, while a second encoder focuses solely on Doppler images to act as a region proposal module.
- Temporal Memory: Convolutional Gated Recurrent Units (convGRU) are employed at the bottleneck to maintain temporal consistency across frames, ensuring smooth segmentation of tubular structures.
- Quality-Aware Robotic Loop: The system monitors Doppler signal stability in real-time. If the signal degrades, the robot initiates an Artery Re-Identification procedure, automatically tilting the probe (±10°) to recover the optimal signal.
- Hardware: The framework is implemented on a KUKA LBR iiwa 7-DoF robot arm using impedance control for safe patient contact.
Key Insights
- Doppler as a Prior: Doppler signals provide high-confidence "hints" about vessel location, allowing the network to ignore B-mode artifacts.
- Active Sensing: Unlike passive models, DopUS-Net recognizes poor input quality and commands the robot to adjust the probe pose.
- Tubular Consistency: The use of RNNs (convGRUs) prevents segmentation "flickering" by leveraging the continuous nature of blood vessels.
Key Results
The integration of Doppler signals and robotic feedback significantly improved performance:
- Accuracy: The Dice score increased from 0.44 (standard U-Net) to 0.86 with the full DopUS-Net system.
- Robustness: Intersection over Union (IoU) improved from 0.47 to 0.78.
- Real-time Performance: Inference takes only 20ms, enabling real-time control.
- Robotic Success: The re-identification module successfully recovered lost signals in 86% to 92% of physical tests.
Limitations and Future Work
While promising, the system currently lacks compensation for large patient movements and is limited to one-dimensional probe "fanning" during re-identification. Future work will explore 3D search patterns and testing on clinical populations with PAD, where calcifications and irregular flow patterns present additional challenges.
Reference: Jiang, Z., Duelmer, F., & Navab, N. (2023). DopUS-Net: Quality-Aware Robotic Ultrasound Imaging based on Doppler Signal. arXiv:2305.08938v1.
Summary updated