Ultrasound Microvascular Imaging Using Deep Knowledge Distillation.

Area of Science:

• Medical Imaging
• Ultrasound Technology
• Artificial Intelligence in Medicine

Background:

• Ultrafast Doppler imaging is crucial for tissue perfusion assessment.
• Traditional methods like SVD filtering require long acquisitions and are computationally intensive.
• There is a need for faster, more efficient vascular imaging techniques.

Purpose of the Study:

• To develop a real-time deep learning framework (SONIC) for high-quality vascular imaging from limited Doppler frames.
• To improve the performance of Doppler imaging under sparse data conditions.
• To enable portable and low-data microvascular ultrasound imaging.

Main Methods:

• Proposed SONIC, a real-time deep learning framework using a teacher-student paradigm.
• Employed a dual-loss strategy: deep supervision for clutter suppression and knowledge distillation for spatiotemporal learning.
• Trained a fast student model guided by a pre-trained teacher network.

Main Results:

• SONIC demonstrated superior performance over SVD filtering with fewer frames on in vivo datasets.
• Ablation studies confirmed synergistic benefits of deep supervision and knowledge distillation, improving segmentation and signal fidelity.
• Achieved real-time inference speeds on GPU hardware, suitable for clinical workflows.
• Successfully applied SONIC to free-hand 3D microvascular imaging.

Conclusions:

• SONIC offers a significant advancement in real-time vascular imaging from limited Doppler data.
• The framework's efficiency and accuracy support integration into time-constrained clinical settings.
• SONIC has the potential to expand microvascular ultrasound imaging capabilities in portable and low-data environments.