Ultrasound Microvascular Imaging Using Deep Knowledge Distillation

Ultrafast Doppler imaging provides critical insights into tissue perfusion but traditionally requires long acquisitions and computationally expensive filtering, such as singular value decomposition (SVD). In this work, we propose SONIC, a real-time, deep learning framework that reconstructs high-quality vascular images from limited Doppler frames using a teacher-student paradigm. The student model, designed for fast inference, is trained under the guidance of a teacher network pretrained on full-length sequences. To further enhance performance under sparse data conditions, we introduce a dual-loss strategy combining deep supervision and knowledge distillation. The deep supervision loss aids in learning clutter-suppressed intermediate features, while the knowledge distillation loss transfers high-level spatiotemporal knowledge from the teacher to the student model. Experiments on in vivo datasets demonstrate that SONIC achieves superior performance compared to SVD filtering when operating on fewer frames. Our ablation study confirms that the combination of deep supervision and knowledge distillation provides synergistic benefits, significantly improving segmentation accuracy and signal fidelity. Furthermore, SONIC achieves real-time inference speeds on GPU hardware, supporting its integration into time-constrained clinical workflows. Finally, we demonstrate the applicability of SONIC to free-hand 3D microvascular imaging by stacking high-quality 2D Doppler slices acquired during handheld scanning. This capability highlights the framework's potential for extending microvascular ultrasound imaging into portable and low-data clinical environments.