Developing GANs for Synthetic Medical ImagingData: Enhancing Training and Research

Abstract

Medical imaging has become integral to modern healthcare, enabling non-invasivevisualization and assessment of anatomical structures. However, medical imagingdatasets are often limited in size and diversity, constraining development of robust analysis algorithms. Meanwhile, generative adversarial networks (GANs) haveachieved remarkable synthetic image generation capabilities. This paper comprehensively reviews contemporary GAN techniques and evaluates their effectiveness producing synthetic medical images to augment scarce training data. Six prevalent GAN architectures were trained on diverse medical imaging datasets. A systematic hyperparameter optimization strategy coupled with quantitative imageanalysis reveal substantial variability in output fidelity and diversity. Downstreamsegmentation task performance provides further domain-specific assessments on theutility of the generated datasets. The study reveals that while select advanced GANscan produce seemingly realistic medical images, the synthetic data consistentlyunderperforms real datasets on specialized tasks. The results caution against indiscriminate use of GAN-produced medical images but highlight paths for developing tailored GAN solutions for enhanced training. Keywords deep learning; generative adversarial networks; medical imaging; synthetic data