Machine Learning for image denoising in 3D ultrasound tomography
Tuesday, April 9, 2024
11:34am – 11:41am
Location: 412
Authors: James Wiskin, QT Imaging John Klock, QT Imaging Inc
3D ultrasound tomography (3DUT) is gaining acceptance in clinical environments, producing two correlated images, speed of sound (SOS) and refraction corrected speckle free 360 compounded reflection images.
We show for the first time a ML/AI denoising technique applied to 3D ultrasound tomography (3DUT) images of breast, knee, pediatric surrogate, eye, kidney and liver and show quality improvement as measured by a 3D UT expert MD who has viewed over 10,000 images from this FDA cleared device. We have established the sub-mm resolution and quantitative accuracy of 3D UT elsewhere. The denoising method here is prototyped on a Natron app but will be optimized for speed in PyTorch. This AI technology provides artifact reduced images in approximately 20 minutes without compromising image quality or relevance for diagnostic purposes. 3D UT is presently deployed in several national and international sites for breast imaging. "Clutter" in images is a problem for the readers. This method shows the viability of denoising with AI in breast, orthopedic and pediatric scenarios.