Blind Approach Using Convolutional Neural Networks to a New Ultrasound Image Denoising Task

Speckle noise reduction from ultrasound images still represents a challenge. One limiting factor is the difficulty in noise characterization and modeling. Specifically, the removal of this image-correlated noise is often obtained at the expense of signal degradation. Convolutional neural networks have been shown to reach good performance along with generalizability properties. In this work, we tested an unsupervised neural network, Noise2Noise, to reduce speckle noise from ultrasound images. The network training used only noisy images obtained adding synthetic noise to real ultrasound images. The trained networks have been tested both on synthetic noisy images and on real ultrasound acquisitions, showing promising results that speak in favor of this novel denoising approach.