Ultrasound Speckle Filtering Using Deep Learning

In this work we present a technique to reduce speckle in ultrasound images that lead to sharper image when compared to other speckle reducing techniques. To achieve this goal, Cycle Generative Adversarial Networks (CycleGAN) are used, which map specific characteristics of a set of input images that will later be used to give origin to transformed images. The discriminator used in this article tries to differentiate between images with and without speckle by the network and the set of high-quality images. The images without speckle used in the network training were obtained from online databases and from a phantom using an Ultrasonix platform and later postprocessed to remove the speckle. The quality of the generated images was evaluated using the metrics Signal to Noise Ratio (SNR) and Peak Signal to Noise Ratio (PSNR) and compared to the Perona-Malik diffusion filter. The proposed technique had an improvement of 4.9 dB in the SNR and 4.7 dB in the PSNR when compared with Perona-Malik filtering.