Implementation of Generative Adversarial Neural Networks for Lung Ultrasound Image Synthesis: Quality-Based Optimal Latent Space Dimension Selection Using FID Score

This paper presents the optimization of generative adversarial neural networks with the objective of improving the generation of synthetic lung ultrasound images, using the Frechet distance in Inception as a quality metric. The latent space or input noise is transformed into synthetic data, so its size is a fundamental hyperparameter for the network. An optimal latent space dimension was found for this network and the training data, improving the FID value by 19.26% in comparison with the worst case achieved. The quality of the images generated by the optimized model was quantified comparing their FID with those obtained by degrading the original images using various filters. The results showed that the quality of the generated images is similar to those that suffered low level of degradation. This study contributes to the field of lung ultrasound image synthesis since optimization based on a quality metric provides a quantitative approach to improve the quality of the generated images. These findings could be relevant to improve the generation of medical images, which could be used to perform data augmentation in order to improving the fitting of classifier models that use such images for training.