Deep Learning for 3D Ultrasound Tomography Denoising

We have established the effectiveness of machine learning based denoising applied to 3D ultrasound tomography images (3D UT) from the QT Imaging Breast Acoustic CT Scanner. The images are subject to slight noise due to the large scale optimization employed in the reconstruction and imperfections in the collected projection data. While this does not affect diagnostic capabilities of the image, they can be potentially distracting to the interpreting readers. This paper extends the previous work on denoising in breast images to pediatric and orthopedic images. The algorithm is outlined as a variation of a Resnet deep learning architecture (20 layers) with residual connections. Orthopedic (Cadaver knee) and Pediatric (pig neonate as animal model) images are shown as proof of concept (POC). The original and denoised images are compared quantitatively to verify that the quantitative values of the images and hence diagnostic value are preserved.