Study of Generative Adversarial Networks for Generating Synthetic Data and its Application on Optoacoustic Tomography

Abstract

This work proposes the use of a Generative Adversarial Network (GAN) to perform data augmentation with the goal of improving image reconstruction in Optoacustic Tomography (OAT) applications. We employ the FastGAN model, a compact net capable of generating high resolution images from small datasets. The quality of the generated data was assessed by two methods. First, the Fréchet distance (FID) was measured, observing a decreasing trend throughout the entire GAN training. Then, a U-Net neural network designed for a OAT system with and without augmented data was trained. In this case, the model trained with the extra data generated by the GAN achieved an appreciable improvement in the figures of merit associated with the reconstruction.

Keywords

• tomografía optoacústica
• aprendizaje profundo
• redes generativas de confrontación
• datos sintéticos