Synthetic Ultrasound Signal-Pairs Generation using Gans

Abstract

Learning-based approaches have shown promising results in several medical imaging applications. However, the performance of the models is affected by the scarcity of the available data. This problem is partially circumvented by using simulation tools to create a large dataset. However, the generation of realistic 3D-simulated input-output pairs, specifically in ultrasound imaging, is usually computationally complex which makes the creation of an adequate training dataset a burdening and time-consuming task. In this work, we investigate the capability of Generative Adversarial Networks (GANs) to generate two interrelated RF signals from random noise while learning the relationship between them. This allows us to incorporate the generated input-output signal-pairs in the training dataset to enhance the performance of a deep learning-based ultrasound application, specifically harmonic imaging. Convolutional based generator and discriminator networks are implemented with three different loss functions to evaluate the performance of the model across different settings. The generated signals are then used to augment the training dataset for harmonic imaging. The dataset augmented with the generated synthetic data enhanced the desired signals estimation by 6.7% compared to the non-augmented dataset. These results demonstrate the GANs ability to generate high quality and diverse ultrasound signal-pairs with a close distribution to the simulated data, yet at a considerably higher speed. This approach could be further extended to enhance the prediction accuracy in other ultrasound applications.