UltraNet: Deep Learning Tools for Modeling Acoustic Wall Clutter

Ultrasonic image quality is often challenged by speckle noise and acoustic clutter which reduces the anatomical conspicuity of medically relevant features. Deep learning approaches show promise in correcting these sources of noise, but there currently lacks a large dataset that i) contains enough feature variation to prevent model over-fitting and ii) captures higher-order effects such as reverberation, aberration, and harmonic generation. To address these needs, we simulated a 180-channel linear transducer acquiring 850,000 transmit-receive events to beamform 10,000 ultrasound images on unique scatter field maps. We fit fully convolutional neural networks (CNNs) to this dataset as a proof-of-concept for noise reduction tasks. We show that CNNs are able to translate from in silico to in vivo images for speckle reduction. However, naive fully convolutional networks are challenged to correct for clutter and aberration simultaneously, even within an in silico training dataset. We hope that this dataset, termed UltraNet, will provide analogous benefits to ultrasound image reconstruction as the ImageNet dataset did for image recognition. These data and tooling will be made available at: https://github.com/ouwen/ultranet.