SR-Ultra is a self-supervised deep residual super-resolution neural network designed to upscale and enhance 3D and 4D volumetric ultrasound data through a novel method known as salient signal recovery. It uses a self-supervised self-to-same quasi-unpaired training paradigm wherein training examples are generated via various image processing techniques and input images serve as their own ground truth labels. Using Tensorflow and supported by a bespoke parallel asynchronous data processing pipeline, SR-Ultra was trained on over 18,363 unique volumes.

SR-Ultra was conceived to support the SonoSim® Ultrasound Training Solution, a medical education platform designed to empower students and clinicians from diverse fields to increase their ultrasonography expertise by providing affordable and highly-portable ultrasound simulation technology, thereby reducing historically formidable barriers to access.

SR-Ultra improves interpretation time by an average of 47%, as measured by feedback from ultrasound technicians and clinicians. Findings presented at the 2018 CRESST Conference.

Teofilo E. Zosa, Matthew Wang, and Eric Savitsky. "Deep Learning for Ultrasound Image Enhancement." CRESST Conference 2018, Oct 1-2, 2018. [POSTER]