Ultrasound speckle reduction using generative adversial networks

Abstract

Generative adversial networks (GANs) have shown its ability to create realistic and accurate image-to-image transformation. The goal of this work was to investigate whether deep convolutional GANs can learn to perform advanced ultrasound speckle reduction in real-time. The GAN was trained using a dataset of cardiac images from 200 patients and tested on a separate dataset from 55 patients. A U-net type of generator was used together with a patch-wise discriminator. Three different generator sizes were tested in order to see the tradeoff between speckle reduction accuracy and runtime. The results show that GANs can learn ultrasound speckle reduction. Even though the training set consisted only of cardiac ultrasound images, results from other parts of the body and scanners indicate that the method learns speckle reduction in general, and not just for cardiac images. By reducing the number of filters in the generator, real-time performance was achieved with an average of 11 ms per frame.