Cross domain knowledge compression in realtime optical flow prediction on ultrasound sequences

Abstract

In this paper, we propose an approach for realtime optical flow estimation in ultrasound sequences of vein and arteries based on knowledge distillation. Knowledge distillation is a technique to train a faster, smaller model by learning from cues of other models. Mobile devices with limited resources could be key in providing effective point-of-care healthcare and motivate the search of more lightweight solutions in the deep learning based image analysis. For ultrasound video analysis motion correspondences of image contents (anatomies) have to be computed for temporal context and for real time application, fast solutions are required. We use a PWC-Net's optical flow estimation output as soft targets to train a lightweight optical flow estimator. We analyse how well it works on the challenging task of fast segmentation propagation of vein and arteries in ultrasound images. Experiments show that even though we did not fine-tune the teachers on this task, a model trained with soft targets outperformed a model trained directly with labels and without a teacher.

Tasks

Reproductions